Brian L. Strom, Stephen E. Kimmel, Sean Hennessy - Textbook of Pharmacoepidemiology-Wiley-Blackwell (2021)

Textbook of Pharmacoepidemiology
Textbook of Pharmacoepidemiology
Third Edition
Edited by
Brian L. Strom, MD, MPH
Chancellor, Rutgers Biomedical & Health Sciences
Executive Vice President for Health Affairs
University Professor
Rutgers, The State University of New Jersey
Newark, NJ, USA
Stephen E. Kimmel, MD, MSCE

Dean’s Professor and Chair of Epidemiology
College of Public Health and Health Professions and College of Medicine
University of Florida
Gainesville, FL, USA
Sean Hennessy, PharmD, PhD

Professor of Epidemiology
Director, Center for Pharmacoepidemiology Research and Training
University of Pennsylvania Perelman School of Medicine
Philadelphia, PA, USA
This edition first published 2022
© 2022 John Wiley & Sons Ltd
Edition History
John Wiley & Sons Ltd (1e, 2006);
John Wiley & Sons Ltd (2e, 2013)
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by
any means, electronic, mechanical, photocopying, recording or otherwise, except as permitted by law. Advice on how to obtain
permission to reuse material from this title is available at http://www.wiley.com/go/permissions.
The right of Brian L. Strom, Stephen E. Kimmel, Sean Hennessy to be identified as the authors of the editorial material in this work
has been asserted in accordance with law.
Registered Offices
John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA
John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex, PO19 8SQ, UK
Editorial Office
9600 Garsington Road, Oxford, OX4 2DQ, UK
For details of our global editorial offices, customer services, and more information about Wiley products visit us at www.wiley.com.
Wiley also publishes its books in a variety of electronic formats and by print-on-demand. Some content that appears in standard
print versions of this book may not be available in other formats.
Limit of Liability/Disclaimer of Warranty
The contents of this work are intended to further general scientific research, understanding, and discussion only and are not
intended and should not be relied upon as recommending or promoting scientific method, diagnosis, or treatment by physicians
for any particular patient. In view of ongoing research, equipment modifications, changes in governmental regulations, and the
constant flow of information relating to the use of medicines, equipment, and devices, the reader is urged to review and evaluate
the information provided in the package insert or instructions for each medicine, equipment, or device for, among other things,
any changes in the instructions or indication of usage and for added warnings and precautions. While the publisher and authors
have used their best efforts in preparing this work, they make no representations or warranties with respect to the accuracy or
completeness of the contents of this work and specifically disclaim all warranties, including without limitation any implied
warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives,
written sales materials or promotional statements for this work. The fact that an organization, website, or product is referred to in
this work as a citation and/or potential source of further information does not mean that the publisher and authors endorse the
information or services the organization, website, or product may provide or recommendations it may make. This work is sold with
the understanding that the publisher is not engaged in rendering professional services. The advice and strategies contained herein
may not be suitable for your situation. You should consult with a specialist where appropriate. Further, readers should be aware that
websites listed in this work may have changed or disappeared between when this work was written and when it is read. Neither the
publisher nor authors shall be liable for any loss of profit or any other commercial damages, including but not limited to special,
incidental, consequential, or other damages.
Library of Congress Cataloging-in-Publication Data
Names: Strom, Brian L., editor. | Kimmel, Stephen | Hennessy, Sean, editor.
E., editor.
Title: Textbook of pharmacoepidemiology / edited by Brian L. Strom, Sean
Hennessy, Stephen E. Kimmel.
Description: Third edition. | Hoboken, NJ : Wiley-Blackwell, 2021. |
Includes bibliographical references and index.
Identifiers: LCCN 2021029285 (print) | LCCN 2021029286 (ebook) | ISBN
9781119701071 (paperback) | ISBN 9781119701088 (adobe pdf) | ISBN
9781119701118 (epub)
Subjects: MESH: Pharmacoepidemiology--methods
Classification: LCC RM302.5 (print) | LCC RM302.5 (ebook) | NLM QV 771 |
DDC 615.7/042–dc23
LC record available at https://lccn.loc.gov/2021029285
LC ebook record available at https://lccn.loc.gov/2021029286
Cover Design: Wiley

Cover Image: © ShutterWorx/Getty Images r.classen/Shutterstock
Set in 9.5/12.5pt STIXTwoText by Straive, Pondicherry, India
10 9 8 7 6 5 4 3 2 1
v
Contents
Contributors xvii
Preface xxi
Acknowledgements xxv
Part I Introduction to Pharmacoepidemiology 1
1 What is Pharmacoepidemiology? 3
Brian L. Strom
Introduction 3
Definition of Pharmacoepidemiology 3
Pharmacoepidemiology Versus Clinical Pharmacology 4
Pharmacoepidemiology Versus Epidemiology 5
Historical Background 5
Early Legislation 5
Drug Crises and Resulting Regulatory Actions 7
Legislative Actions Resulting from Drug Crises 9
Intellectual Development of Pharmacoepidemiology Emerging from Drug Crises 10
The Current Drug Approval Process 13
Drug Approval in the US 13
Drug Approval in Other Countries 14
Potential Contributions of Pharmacoepidemiology 15
Supplementary Information 16
New Types of Information Not Available from Premarketing Studies 17
General Contributions of Pharmacoepidemiology 17
Key Points 18
Further Reading 18
2 Study Designs Available for Pharmacoepidemiologic Studies 20

Brian L. Strom
Introduction 20
Overview of the Scientific Method 20
Types of Errors that one Can Make in Performing a Study 22
Criteria for the Causal Nature of an Association 23
Epidemiologic Study Designs 26
Case Reports 26
Case Series 26
vi Contents
Analyses of Secular Trends 27

Case–Control Studies 28
Cohort Studies 28
Analysis of Case–Control and Cohort Studies 29
Randomized Clinical Trials 30
Discussion 31
Conclusion 32
Key Points 32
Further Reading 33
3 Sample Size Considerations for Pharmacoepidemiologic Studies 35

Brian L. Strom
Introduction 35
Sample Size Calculations for Cohort Studies 35
Sample Size Calculations for Case–Control Studies 40
Sample Size Calculations for Case Series 41
Discussion 43
Key Points 45
Further Reading 45
4 Basic Principles of Clinical Pharmacology Relevant to Pharmacoepidemiologic Studies 47

Jeffrey S. Barrett
Introduction 47
Clinical Pharmacology and Pharmacoepidemiology 48
Basics of Clinical Pharmacology 48
Pharmacokinetics 49
Absorption 49
Volume of Distribution 49
Metabolism 49
Elimination 51
Special Populations 52
Elderly 52
Pediatrics 53
Pregnancy 55
Organ Impairment 55
Drug Interactions 56
Pharmacodynamics 56
Overview 57
Pharmacogenomics 59
Model-Informed Drug Development 59
Conclusion 60
Key Points 60
Further Reading 61
5 When Should One Perform Pharmacoepidemiologic Studies? 62

Brian L. Strom
Introduction 62
Reasons to Perform Pharmacoepidemiologic Studies 62
Contents vii
Regulatory 63
Marketing 64
Legal 65
Clinical 66
Safety Versus Risk 67
Risk Tolerance 67
Features of the Adverse Outcome 68
Characteristics of the Exposure 68
Perceptions of the Evaluator 69
Conclusion 70
Key Points 70
Further Reading 71
6 Views from Academia, Industry, Regulatory Agencies, and the Legal System 73
Joshua J. Gagne, Jerry Avorn, Nicolle M. Gatto, Jingping Mo, Gerald J. Dal Pan, June Raine,
Shinobu Uzu, Aaron S. Kesselheim, and Kerstin N. Vokinger
The View from Academia 73
Introduction 73
The Drug Approval Process 74
Prescribing Practices 75
Evaluation of Patients’ Use of Drugs in the Health Care System 76
Assessment of the Quality and Outcomes of Medication Use in Populations 76
Policy Analysis 77
Interventional Pharmacoepidemiology 77
Economic Assessment of Medication-Related Issues 78
The Academic Medical Center 78
Consortia of Academic Medical Center Programs for Pharmacoepidemiologic
Research 78
The Future 79
Summary Points for the View from Academia 79
The View from Industry 81
Introduction 81
Regulatory and Industry Focus on Risk Management and Epidemiology 81
Epidemiology in Drug Safety Evaluation 83
Epidemiology in Evaluation of Risk Mitigation Interventions 86
Conclusion 87
Summary Points for the View from Industry 88
The View from Regulatory Agencies 90
Introduction 90
Assessing the Need for Medicines 91
Orphan Drugs 91
Planning Drug Development Programs 92
Pre-approval Review of Clinical Safety Data 93
Planning for Post-approval Studies 94
Monitoring Post-approval Safety 94
Assessing Actual Use Patterns of a Medicine 95
Assessing Impact of Regulatory Actions 95
Advancing the Science of Pharmacoepidemiology 96
viii Contents
Conclusion 97
Summary Points for the View from Regulatory Agencies 97
The View from the Legal System 98
Introduction 98
Tort Law and Product Liability Lawsuits 98
Pharmacoepidemiology and Contract Law 102
Pharmacoepidemiology and Intellectual Property Law 103
Conclusion 105
Summary Points for the View from the Legal System 105
Further Reading 107
The View from the Legal System 110
Contract-Related Issues in Pharmacoepidemiology 110
Patent Law and Pharmacoepidemiology 110
Part II Sources of Pharmacoepidemiology Data 113
7 Postmarketing Spontaneous Pharmacovigilance Reporting Systems 115

Gerald J. Dal Pan, Marie Lindquist, and Kate Gelperin
Introduction 115
Description 116
Adverse Events and Adverse Drug Reactions 116
Overview of Pharmacovigilance Reporting Systems 117
The Concept of Spontaneous AE/ADR Reporting 118
Report Characteristics 119
Social Media 121
National Pharmacovigilance Systems 121
National and International Postmarketing Adverse Event Databases 123
Detecting Signals from a Postmarketing Adverse Event Database 124
Review of Individual Case Safety Reports 126
Reporting Ratios 127
Strengths 128
Signal Detection 128
Opportunity for the Public to Report AEs/ADRs 129
Scope 129
Limitations 129
Quality of Reports 129
Underreporting 130
Non-uniform Temporal Trends in Reporting 130
Particular Applications 131
Case Series and Reporting Rates 131
Data Mining Signals 131
Signals from Developing Countries 131
The Future 132
Key Points 132
Further Reading 134
Contents ix
8 Overview of Electronic Databases in Pharmacoepidemiology 136

Brian L. Strom
Introduction 136
Description 137
Claims and Other Administrative Databases 137
Electronic Health Record Databases 138
Strengths 138
Weaknesses 139
The Future 140
Key Points 141
Further Reading 141
9 Encounter Databases 142

Tobias Gerhard, Yola Moride, Anton Pottegård, and Nicole Pratt
Introduction 142
Description 142
Attributes of Encounter Databases 145
Selected Encounter Databases 149
Strengths 161
Limitations 162
Typical Activities Involved in Studies Using Encounter Databases 163
Deciding Between Individual Encounter Databases 164
The Future 166
Key Points 167
Further Reading 171
US Databases 172
European Databases 172
Canadian Databases 173
Asian Databases 173
10 Electronic Health Record Databases 174

Daniel B. Horton, Harshvinder Bhullar, Francesca Cunningham,
Janet Sultana, and Gialuca Trifirò
Introduction 174
Description 174
Overview of Health Care Systems and Populations 174
Overview of Databases 181
Data Collection and Structure 182
Data Quality: Accuracy and Completeness 183
Data Access for Researchers 184
Strengths 184
Population-Based Data, Sample Size, and Length of Follow-up 184
Validity of Clinical Information 184
Accuracy of Drug Information 184
Ability to Access Original Health Records 184
Linkage to External Patient-Level Data 185
x Contents
Limitations 185
Incompleteness of Clinical Data 185
Incompleteness of Drug Data 185
The Future 186
Summary Points for Electronic Health Record Databases 187
Acknowledgment 187
Further Readings 189
11 Primary Data Collection for Pharmacoepidemiology 192

Priscilla Velentgas
Introduction 192
Research Questions that Require Primary Data 192
Hybrid or Enriched Designs 195
Methods of Primary Data Collection 195
Site-Based Data Collection 195
Clinician or Site-Reported Outcomes (ClinROs) 195
Patient-Generated Data 196
Registries as Means of Data Collection 196
Biobanks/Specimen Banks 197
Guidelines on the Quality of Data Collection 197
Strengths 197
Limitations 197
Conclusions 199
Key Points 199
Further Reading 201
12 How Should One Perform Pharmacoepidemiologic Studies?

Choosing Among the Available Alternatives 203
Brian L. Strom
Introduction 203
Choosing Among the Available Approaches to Pharmacoepidemiologic Studies 203
Comparative Characteristics of Pharmacoepidemiologic Data Resources 208
Characteristics of Research Questions and their Impact on the Choice
of Pharmacoepidemiologic Data Resources 211
Examples 215
Conclusion 216
Key Points 216
Further Reading 216
Part III Special Issues in Pharmacoepidemiology Methodology 219
13 Validity of Drug and Diagnosis Data in Pharmacoepidemiology 221

Mary Elizabeth Ritchey, Suzanne L. West, and George Maldonado
Introduction 221
Clinical Problems to be Addressed by Pharmacoepidemiologic Research 221
Methodological Problems to be Solved by Pharmacoepidemiologic Research 222
Indices of Measurement Error Relevant to Pharmacoepidemiologic Research 222
Contents xi
Quantitative Measurement of Validity 222

Quantitative Measurement of Reliability 224
Measurement Error in Pharmacoepidemiologic Research 225
Adjusting Measures of Association for Measurement Error 227
Self-Reported Drug Data from Ad hoc Survey Studies: Recall Accuracy 228
The Influence of Medication Class 228
The Influence of Questionnaire Design 228
The Influence of Patient Population 229
Self-Reported Diagnosis and Hospitalization Data from Ad hoc Studies: Recall Accuracy 230
The Influences of Medical Condition Type 230
The Influences of Timing of Diagnosis and Its Emotional Effects on the Patient 230
The Influence of Patient Population 232
The Influence of Questionnaire Design 232
Currently Available Solutions 233
Following Best Practices for Questionnaire Design 233
Developing a De novo Questionnaire 233
Conducting Validation Studies to Assess Self-Reported Data 235
Considering the Influence of Comparator Selection on Validation Studies 235
Validation of Pharmacoepidemiologic Drug and Diagnosis Data from Electronic Encounter
Databases 236
Special Considerations with Drug Data 237
Special Considerations with Diagnosis and Hospitalization Data 237
Special Considerations with Distributed Data Systems 239
Best Practices 239
The Future 242
Key Points 242
Further Reading 243
14 Assessing Causality from Case Reports 246

Bernard Bégaud and Judith K. Jones
Introduction 246
The Two Paradigms of Causality Assessment 246
When is Assessing Causation from Cases Reports Useful? 247
Spontaneous Reporting 247
Clinical Trials and Pharmacoepidemiology 248
Clinical Practice and Prescription 248
Reports of Adverse Drug Reactions to Medical Journals 248
Hypothesis Generation and Research 248
Methodological Problems to be Addressed by Pharmacoepidemiologic Research 248
Approaches for Assessing Causation from Individual Cases 249
Expert Judgment/Global Introspection 249
Structured Guidelines and Algorithms 250
Probabilistic Approaches 251
Calibration 253
Choosing the Appropriate Approach 253
The Future 254
Key Points 255
Further Reading 255
xii Contents
15 Molecular Pharmacoepidemiology 257

Christine Y. Lu and Stephen E. Kimmel
Introduction 257
Definitions and Concepts 258
Genetic Variability 258
Pharmacogenetics and Pharmacogenomics 259
The Interface of Pharmacogenetics and Pharmacogenomics with Molecular
Pharmacoepidemiology 259
Three Ways That Genes Can Affect Drug Response 260
The Interplay of Various Mechanisms 263
The Progression and Clinical Application of Molecular Pharmacoepidemiology 264
Interactions 266
Type I Error 267
Type II Error 267
Confounding by Population Admixture 268
Identifying Additional Genetic Contributions to Drug Response 269
Interactions 270
Type I Error and Replication 270
Type II Error 271
Confounding by Population Admixture 271
The Future 271
Key Points 273
Further Reading 274
16 Bioethical Issues in Pharmacoepidemiologic Research 276

Laura E. Bothwell, Annika Richterich, and Jeremy Greene
Introduction 276
The Emergence, Changing Methods, and Moral Stakes of Pharmacoepidemiology in
Twentieth Century North America 276
European Pharmacoepidemiologic Trends and Ethics 280
East Asian Pharmacoepidemiologic Trends and Ethics 282
Methodologic Problems to be Solved by Pharmacoepidemiologic Research 283
Informed Consent 284
Ethics of Surveillance 285
Ethical Benefits of Pharmacoepidemiologic Research for Data Integrity 285
Problems of Conflicts of Interest for Drug Industry Research 286
Good Pharmacoepidemiology and Pharmacovigilance Practices 286
Protections against Conflicts of Interest for Drug Industry-Sponsored Research 288
The Future 289
Acknowledgement 291
Key Points 291
Further Reading 293
Contents xiii
17 The Use of Randomized Controlled Trials for Pharmacoepidemiology 294

Samuel M. Lesko, Allen A. Mitchell, and Robert F. Reynolds
Introduction 294
Overview of Classic RCTs 296
Limitations of RCTs 298
Large Simple Trials 298
When is an LST Feasible? 301
Logistics of Conducting an LST 302
Analysis 303
Primary Analysis 303
Subgroup Analyses 303
Data Monitoring/Interim Analyses 304
The Future 304
Key Points 305
Further Reading 305
18 Pharmacoeconomics: Economic Evaluation of Pharmaceuticals 307

Kevin A. Schulman
Introduction 307
Clinical Problems to be Addressed by Pharmacoeconomic Research 307
The Economics of Drug Development 307
Health Economics and Health Care Financing 308
Methodological Problems to be Addressed by Pharmacoeconomic Research 312
Types of Analysis 312
Types of Costs 315
Perspective of Analysis 316
Using Economic Data 317
The Future 320
Acknowledgements 320
Key Points 320
Further Reading 320
19 Patient Engagement and Patient Reported Outcomes 322

Esi M. Morgan and Adam C. Carle
Introduction 322
Patient Reported Outcomes in Clinical Trials 323
Patient Reported Outcomes in Routine Care 323
Patient Reported Outcomes as Motivation to Develop New Therapeutic Strategies 325
Ensuring PRO Completion and Results Review 326
PRO Selection, Score Interpretation and Interventions 326
Patient Engagement and Individualized Assessment and Treatment Plans 327
Barriers to Measuring PROs in Clinical Practice and Using PROs to Guide
Interventions 327
0005160718.INDD 13 09-13-2021 15:12:51

xiv Contents
ethodologic Problems to be Solved by Pharmacoepidemiologic Research 328

M
Discordance in Perspectives between Patients, Clinicians and Researchers 328
Measuring within Person Change 329
Selection of Patient Reported Outcomes and Implementation into Practice 330
The Future 330
Key Points 331
Further Reading 331
20 The Use of Meta-analysis in Pharmacoepidemiology 334

Brenda J. Crowe, Stephen J.W. Evans, H. Amy Xia, and Jesse A. Berlin
Introduction 334
Susceptibility of the Original Studies to Bias 336
Combinability of Studies 336
Publication Bias 337
Bias in the Abstraction of Data 338
Steps Involved in Performing a Meta-analysis 338
Publication Bias 344
Indirect Comparison and Simultaneous Comparison of Treatments Available for Specific
Conditions 346
Case Studies of Applications of Meta-analysis 346
The Future 350
Key Points 351
Further Reading 352
21 Studies of Medication Adherence 355

Julie Lauffenburger, Trisha Acri, and Robert Gross
Introduction 355
Analysis Issues in Adherence 362
Use of Adherence Data in Clinical Trials and Comparative Effectiveness Studies 362
Selecting Adherence Intervals 362
Statistical Analysis 363
Time-Varying Nature of Adherence and Trajectory Modeling 364
Prediction of Adherence for Interventions 365
The Future 365
Key Points 365
Further Reading 366
22 Advanced Approaches to Controlling Confounding in Pharmacoepidemiologic Studies 368

Sebastian Schneeweiss and Samy Suissa
Introduction 368
Contents xv
linical Problems to be Addressed by Pharmacoepidemiologic Research 368

C
Efficient Sampling Designs within a Cohort Study 370
Analytic Approaches for Improved Confounding Control 377
Conclusion 382
Key Points 382
Further Reading 384
Part IV Special Applications and the Future of Pharmacoepidemiology 387
23 Special Applications of Pharmacoepidemiology 389

David Lee, Björn Wettermark, Christine Y. Lu, Stephen B. Soumerai, Robert T. Chen, Sharon-Lise T.
Normand, Art Sedrakyan, Danica Marinac-Dabic, Daniel B. Horton, Sonia Hernandez-Diaz, Tamar
Lasky, Krista F. Huybrechts, Claudia Manzo, Emil Cochino, Hanna M. Seidling, David W. Bates,
Bennett Levitan, Rachael L. DiSantostefano, and Scott Evans
Studies of Drug Utilization 389
Introduction 389
Examples of Currently Available Solutions 393
The Future 396
Key Points for Studies of Drug Utilization 396
Evaluating and Improving Prescribing 398
The Future 402
Key Points for Evaluating and Improving Prescribing 402
Special Methodological Issues in Pharmacoepidemiologic Studies of Vaccine Safety 403
The Future 408
Key Points for Special Methodological Issues in Pharmacoepidemiologic Studies of Vaccine
Safety 408
Epidemiologic Studies of Implantable Medical Devices 408
What Is a Medical Device and how Is it Different from a Drug? 409
Clinical Problems to be Addressed by Medical Device Epidemiologic Research 409
Methodological Problems to be Solved by Medical Device Epidemiologic Research 410
The Future 415
Key Points for Epidemiologic Studies of Devices 418
Research on the Effects of Medications in Pregnancy and in Children 418
xvi Contents

The Future 424
Key Points for Research on the Effects of Medications in Pregnancy and in Children 426
Risk Management 427
The Future 434
Key Points for Risk Management 436
The Pharmacoepidemiology of Medication Errors 436
Safety Theory 436
Patient Safety Concepts as Applied to Pharmacoepidemiology 437
The Future 442
Key Points for the Pharmacoepidemiology of Medication Errors 442
Benefit–Risk Assessments of Medical Treatments 442
The Future 451
Key Points for Benefit-Risk Assessments of Medical Treatments 454
Further Reading 454
24 The Future of Pharmacoepidemiology 464

Brian L. Strom, Stephen E. Kimmel, and Sean Hennessy
Introduction 464
Scientific Developments 465
Funding 469
Personnel 470
The View from the Law 473
Conclusion 473
Key Points 473
Further Reading 474
Appendix A — Sample Size Tables 475

Appendix B — Glossary 493
Index 505
xvii
List of Contributors
Trisha Acri Laura E. Bothwell

Department of Family and Community Yale School of Public Health
Medicine Temple University School of New Haven, Connecticut, USA
Medicine Philadelphia, PA, USA
Currently, Director of Community Research Adam C. Carle
Health Services, AIDS Care Group, AIDS Care Cincinanti Children’s Hospital Medical Center
Group, Sharon Hill, PA, USA Cincinnati, OH, USA
University of Cincinnati, College of Medicine
Jerry Avorn Cincinnati, OH, USA
Harvard Medical School and Brigham and University of Cincinnati, College of
Women’s Hospital Arts and Sciences
Boston, MA, USA Cincinnati, OH, USA
Jeffrey S. Barrett Robert T. Chen

Critical Path Institute Brighton Collaboration, Task Force for
Tucson, AZ, USA Global Health
Decatur, GA, USA
David W. Bates
Division of General Internal Medicine and
Emil Cochino
Primary Care
European Medicines Agency
Brigham and Women’s Hospital and Harvard
Amsterdam, The Netherlands
Medical School
Boston, MA, USA
Brenda J. Crowe
Bernard Bégaud Eli Lilly and Company
Clinical Pharmacology and Indianapolis, IN, USA
Pharmacoepidemiology, Medical School
University of Bordeaux Francesca Cunningham
Bordeaux, France US Department of Veterans Affairs
Hines, IL, USA
Jesse A. Berlin
Johnson & Johnson Gerald J. Dal Pan
Titusville, NJ, USA Office of Surveillance and Epidemiology
Center for Drug Evaluation and Research
Harshvinder Bhullar US Food and Drug Administration
Independent Consultant London, UK Silver Spring, MD, USA
xviii List of Contributors
Rachael L. DiSantostefano Center for Pharmacoepidemiology Research

Department of Epidemiology and Training, Perelman School of Medicine
Janssen Research & Development University of Pennsylvania, Philadelphia
Titusville, NJ, USA PA, USA
Scott Evans Sean Hennessy

Biostatistics Center University of Pennsylvania Perelman School
The George Washington University of Medicine
Rockville, MD, USA Philadelphia, PA, USA
Stephen J.W. Evans
Sonia Hernandez Diaz
The London School of Hygiene and Tropical
Harvard T.H. Chan School of Public Health
Medicine, London, UK
Boston, MA, USA
Joshua J. Gagne
Daniel B. Horton
Harvard Medical School and Brigham and
Rutgers Robert Wood Johnson
Women’s Hospital, Boston, MA, USA
Medical School, Rutgers Center for
and
Pharmacoepidemiology and Treatment
Johnson and Johnson
Science, Rutgers School of Public Health
New Brunswick, NJ, USA
New Brunswick, NJ, USA
Nicolle M. Gatto
Aetion Inc., New York, NY, USA Krista F. Huybrechts
Brigham and Women’s Hospital
Kate Gelperin Harvard Medical School
Division of Epidemiology Office of Boston, MA, USA
Surveillance and Epidemiology
Center for Drug Evaluation and Research Judith K. Jones
US Food and Drug Administration †
Formerly Principal Consultant PharmaLex,
Silver Spring, MD, USA Inc., Fairfax, VA and Adjunct Faculty
The University of Michigan School of Public
Tobias Gerhard Health Summer Program
Ernest Mario School of Pharmacy Ann Arbor, USA
Rutgers Biomedical and Health Sciences
Piscataway, NJ, USA Aaron S. Kesselheim
and Harvard Medical School and Brigham and
Center for Pharmacoepidemiology and Women’s Hospital
Treatment Science Boston, MA, USA
Rutgers Biomedical and Health Sciences
New Brunswick, NJ, USA Stephen E. Kimmel
University of Florida College of Public
Jeremy Greene Health and Health Professions & College of
Johns Hopkins University Medicine
Baltimore, Maryland, MD, USA Gainesville, FL, USA
Robert Gross Tamar Lasky

Center for Clinical Epidemiology and US Food and Drug Administration
Biostatistics Silver Spring, MD, USA
List of Contributors xix
Julie Lauffenburger Allen A. Mitchell

Brigham and Women’s Hospital and Harvard Slone Epidemiology Center at Boston
Medical School University, Boston, MA, USA
Boston, MA, USA
Jingping Mo
David Lee Epidemiology, Worldwide Research &
Center for Pharmaceutical Management Development, Pfizer Inc.
Management Sciences for Health New York, NY, USA
Arlington, VA, USA
Esi M. Morgan
Samuel M. Lesko Seattle Children’s Hospital
Northeast Regional Cancer Institute and Seattle, WA, USA
Geisinger Commonwealth School of Medicine University of Washington
Scranton, PA, USA Seattle, WA, USA
Bennett Levitan
Yola Moride
Department of Epidemiology
Center for Pharmacoepidemiology and
Janssen Research & Development
Treatment Science, Rutgers Biomedical and
Titusville, NJ, USA
Health Sciences
Marie Lindquist New Brunswick, NJ, USA
Uppsala Monitoring Centre
WHO Collaborating Centre for International Sharon-Lise T. Normand
Drug Monitoring Harvard Medical School and
Uppsala, Sweden Harvard School of Public Health
Boston, MA, USA
Christine Y. Lu
Harvard Medical School and Harvard Pilgrim Anton Pottegård
Health Care Institute, Boston, MA, USA Clinical Pharmacology and Pharmacy
Department of Public Health
George Maldonado University of Southern Denmark
Division of Environmental Health Sciences Odense, Denmark
School of Public Health, University of
Minnesota, Minneapolis, MN, USA Nicole Pratt
Quality Use of Medicines and Pharmacy
Claudia Manzo Research Centre, Clinical and Health Sciences
Division of Risk Management Office of University of South Australia, Adelaide,
Surveillance and Epidemiology South Australia, Australia
Center for Drug Evaluation and Research
US Food and Drug Administration
June Raine
Silver Spring, MD, USA
Vigilance and Risk Management of Medicine
Medicines and Healthcare Products
Danica Marinac-Dabic
Regulatory Agency, London, UK
Division of Epidemiology
Office of Surveillance and Biometrics Center
for Devices and Radiological Health Robert F. Reynolds
US Food and Drug Administration Epidemiology, Research and Development,
Silver Spring, MD, USA GlaxoSmithKline, New York, NY, USA
xx List of Contributors
Annika Richterich Brian L. Strom

Maastricht University Rutgers Biomedical and Health Sciences
Maastricht, The Netherlands Newark, NJ, USA
Mary Elizabeth Ritchey Samy Suissa

Med Tech Epi, LLC, Philadelphia, PA, USA McGill University and Jewish General Hospital
Center for Pharmacoepidemiology and Montreal, Quebec, Canada
Treatment Science, Rutgers University
New Brunswick, NJ, USA Janet Sultana
Mater Dei Hospital, Msida, Malta
Sebastian Schneeweiss and Exeter College of Medicine and Health
Departments of Medicine and Epidemiology University of Exeter, Exeter, UK
Harvard Medical School and
Gianluca Trifirò
Division of Pharmacoepidemiology
Department of Diagnostics and Public Health
Department of Medicine
University of Verona, Verona, Italy
Brigham & Women’s Hospital
Boston, MA, USA
Shinobu Uzu
Pharmaceuticals and Medical Devices Agency
Kevin A. Schulman Tokyo, Japan
Clinical Excellence Research Center
Stanford University Priscilla Velentgas
Stanford, CA, USA Real World Solutions, IQVIA, Inc.
Cambridge, MA, USA
Art Sedrakyan
Department of Public Health Kerstin N. Vokinger
New York Presbyterian Hospital and Harvard Medical school and Brigham and
Weill Cornell Medical College Women’s Hospital
New York, NY, USA Boston, MA, USA
and
Institute for Law University of Zurich
Hanna M. Seidling
Zurich, Switzerland
Head of Cooperation Unit Clinical Pharmacy
Department of Clinical Pharmacology and
Suzanne L. West
Pharmacoepidemiology Cooperation Unit
Gillings School of Global Public Health
Clinical Pharmacy
University of North Carolina
University of Heidelberg
Chapel Hill, NC, USA
Heidelberg, Germany
Björn Wettermark
Stephen B. Soumerai Disciplinary Domain of Medicine and Pharmacy
Department of Population Medicine Director Uppsala University, Uppsala, Sweden
Drug Policy Research Group
Harvard Medical School and Harvard Pilgrim H. Amy Xia
Health Care Institute Amgen, Thousand Oaks
Boston, MA, USA CA, USA
xxi
Preface
It was a remarkable 33 years ago that the first grown into a major international scientific
edition of Strom’s Pharmacoepidemiology was force, with over 1476 members from 63 coun-
published. The preface to that book stated that tries, an extremely successful annual meeting
pharmacoepidemiology was a new field with a attracting more than 1800 attendees, a large
new generation of pharmacoepidemiologists number of very active committees and scien-
arising to join the field’s few pioneers. Over the tific interest groups, and its own journal. In
ensuing 32 years, the field indeed has grown addition, a number of established journals
and no longer deserves to be called “new.” have targeted pharmacoepidemiology manu-
Many of those “new generation” scientists scripts as desirable. As new scientific develop-
(including two of the editors of this book) are ments occur within mainstream epidemiology,
now “middle-aged” pharmacoepidemiologists. they are rapidly adopted, applied, and advanced
Despite its relatively brief academic life, a within our field as well. We have also become
short history of pharmacoepidemiology and institutionalized as a subfield within the field
review of its current state will set the stage for of clinical pharmacology, with scientific sec-
the purpose of this textbook. tions of the American Society for Clinical
Pharmacoepidemiology originally arose from Pharmacology and Therapeutics and with
the union of the fields of clinical pharmacology pharmacoepidemiology a required part of the
and epidemiology. Pharmacoepidemiology clinical pharmacology board examination.
studies the use of and the effects of medical Most of the major international pharmaceuti-
products in large numbers of people and cal companies have founded dedicated units to
applies the methods of epidemiology to the organize and lead their efforts in pharmacoepi-
content area of clinical pharmacology. This demiology, pharmacoeconomics, and quality-of-
field represents the science underlying post- life studies. The continuing parade of drug safety
marketing medical product surveillance, stud- crises emphasizes the need for the field, and
ies of the effects of medical products (i.e. drugs, some foresighted manufacturers have begun to
biologicals, devices) performed after a product perform “prophylactic” pharmacoepidemiology
has been approved for use. In recent years, studies, to have data in hand and available when
pharmacoepidemiology has expanded to questions arise, rather than waiting to begin to
include many other types of studies, as well. collect data after a crisis has developed.
The field of pharmacoepidemiology has Pharmacoepidemiologic data are now routinely
grown enormously since the first publication used for regulatory decisions, and many govern-
of Strom. The International Society of mental agencies have been developing and
Pharmacoepidemiology, an early idea when expanding their own pharmacoepidemiology
the first edition of this book was written, has programs. Risk evaluation and mitigation
xxii Preface
strategies are now required by regulatory bodies recipients, introduced to pharmacoepidemiol-

with the marketing of new drugs, as a means of ogy a new database with a stable population
improving drugs’ benefit/risk balance, and approaching 50 million in what may be the
manufacturers are identifying ways to respond. largest healthcare system in the world. The US
Requirements that a drug be proven to be cost Congress has recognized the importance of the
effective have been added to many national, field, with the founding of the Sentinel
local, and insurance health care systems, either Program, and new requirements that FDA
to justify reimbursement or even to justify drug focus on “real world evidence.” A new move-
availability. A number of schools of medicine, ment has arisen in the US of “comparative
pharmacy, and public health have established effectiveness research,” which in many ways
research programs in pharmacoepidemiology, learns from much longer experience in Europe,
and a few of them have also established pharma- as well as decades of experience in pharma-
coepidemiology training programs in response coepidemiology. These developments portend
to a desperate need for more pharmacoepide- major changes for our field.
miology personnel. Pharmacoepidemiologic In summary, there has been tremendous
research funding is now more plentiful, and growth in the field of pharmacoepidemiology
even limited support for training is available. and a fair amount of maturation. With the
In the United States, drug utilization review growth and maturation of the field, Strom’s
programs are required, by law, of each of the 50 Pharmacoepidemiology has grown and
state Medicaid programs, and have been imple- matured right along. Pharmacoepidemiology
mented as well in many managed care organi- thus represents a comprehensive source of
zations. Now, years later, the utility of drug information about the field. As a reflection of
utilization review programs is being ques- the growth of the field, the fourth Edition
tioned. In addition, the Joint Commission of Strom was over twice as long as the first!
requires that every hospital in the US have an Now, seven years after the fifth edition, the
adverse drug reaction monitoring program and field continues to change and garner wide-
a drug use evaluation program, turning every spread interest, leading to the recent publica-
hospital into a mini-pharmacoepidemiology tion of the sixth edition in 2020.
laboratory. Stimulated in part by the interests So, why, one may ask, do we need a Textbook
of the World Health Organization and the of Pharmacoepidemiology? The need arose
Rockefeller Foundation, there is even substan- precisely because of the growth of the field.
tial interest in pharmacoepidemiology in the With that, and the corresponding growth in
developing world. Yet, throughout the world, the parent book, Strom’s Pharmacoepidemiology
the increased concern by the public about pri- has really become more of a reference book
vacy has made pharmacoepidemiologic than a book usable as a textbook. Yet, there is
research much more difficult to conduct. increasing need for people to be trained in the
In recent years, major new changes have field and an increasing number of training
been made in drug regulation and organiza- programs. With the maturity of the field comes
tion, largely in response to a series of accusa- therefore the necessity for both comprehensive
tions about myocardial infarction caused by approaches (such as Strom’s Pharmacoep
analgesics, which was detected in long-term idemiology) and more focused approaches.
prevention trials rather than in normal use of Therefore, Textbook of Pharmacoepidemiology
the drugs. For example, FDA was given new was intended as a modified and shortened
regulatory authority after drug marketing. version of its parent, designed to meet the need
Further, the development, since 1 January of students. We believe that students can
2006, of Medicare Part D, a US federal program benefit from an approach that focuses on the
to subsidize prescription drugs for Medicare core of the discipline, along with learning aids.
Preface xxiii
Textbook of Pharmacoepidemiology attempts Data” and includes important chapters about

to fill this need, providing a focused educa- spontaneous pharmacovigilance reporting
tional resource for students. It is our hope that systems, electronic databases, and other
this book will serve as a useful textbook for stu- approaches to pharmacoepidemiology stud-
dents at all levels: upper-level undergraduates, ies. Part III summarizes “Special Issues in
graduate students, post-doctoral fellows, and Pharmacoepidemiology Methodology” that we
others who are learning the field. To achieve feel are important to more advanced pharma-
our goals, we have substantially edited down coepidemiology students. Although no student
from Strom’s Pharmacoepidemiology, with a is likely to become an expert in all of these
focus on what is needed by students, eliminat- methods, they form a core set of knowledge
ing some chapters and shortening others. We that we believe all pharmacoepidemiologists
also have provided case examples for most should have. In addition, one never knows
chapters and key points for all chapters. Each what one will do later in one’s own career, nor
chapter is followed by a list of further reading. when one may be called upon to help others
So why update it? In looking at the sixth with the use of these methods. Part IV con-
Edition of Strom, most chapters in the new cludes the textbook with a collection of
edition were thoroughly revised to provide “Special Applications” of the field, and specu-
updated content. New chapters were added, lation about its future, always an important
along with many new authors. The second edi- consideration for new investigators in charting
tion of the textbook was simply getting out of a career path.
date in comparison to the recently published Pharmacoepidemiology may be maturing,
sixth edition of the parent book. but many exciting opportunities and chal-
Specifically, we have tried to emphasize the lenges lie ahead as the field continues to grow
methods of pharmacoepidemiology and the and respond to unforeseeable future events. It
strengths and limitations of the field, while is our hope that this book can continue to serve
minimizing some of the technical specifica- as a useful introduction and resource for stu-
tions that are important for a reference book dents of pharmacoepidemiology, both those
but not for students. Therefore, the first five enrolled in formal classes and those learning
chapters of Part I, “Introduction to Pharmaco in “the real world,” who will respond to the
epidemiology,” lay out the cores of the disci- challenges that they encounter. Of course, we
pline, and remain essentially unchanged from are always students of our own discipline, and
Strom’s Pharmacoepidemiology, with the excep- the process of developing this textbook has
tion of the inclusion of key points and lists of been educational for us. We hope that this
further reading. We have also included a chap- book will also be stimulating and educational
ter on different perspectives of the field (from for you.
academia, industry, regulatory agencies, and Brian L. Strom, MD, MPH.
the legal system), as a shortened form of several Stephen E. Kimmel, MD, MSCE.
chapters from the reference book. Part II Sean Hennessy, PharmD, PhD.
focuses on “Sources of Pharmacoepidemiology January 2022
xxv
Acknowledgments
There are many individuals and institutions to their own careers, and see them blossom into
whom we owe thanks for their contributions to star senior pharmacoepidemiologists. It is
our efforts in preparing this book. Mostly, we wonderful to be able to share with them this
would like to thank all of the contributors for book, which has been an important part of
the work that they did in revising their book BLS’s life and career.
chapters and sections for this textbook and BLS would also like to thank his parents,
providing case examples, key points, and sug- now deceased, for the support and education
gested readings. Over the years, our pharma- that were critical to him being successful in his
coepidemiology work has been supported career. BLS would also like to thank the late
mostly by numerous grants from government, Paul D. Stolley, M.D., M.P.H. and the late
foundations, and industry. While none of this Kenneth L. Melmon, M.D., for their direction,
support was specifically intended to support guidance, and inspiration in the formative
the development of this book, without this years of his career. He would also like to thank
assistance, we would not have been able to his trainees, from whom he learns at least as
support our careers in pharmacoepidemiology. much as he teaches. Last, but certainly not
We would like to thank our publisher, John least, BLS would like to thank his family –
Wiley & Sons, Ltd., for their assistance and Lani, Shayna, and Jordi – for accepting the
insights, both in support of this book, and in time demands of the book, for tolerating his
support of the field’s journal, Pharma endless hours working at home (on its earlier
coepidemiology and Drug Safety. editions, for the kids), and for their ever pre-
John Hemphill’s contributions to this book sent love and support.
were instrumental, encompassing the role of SEK expresses his sincere gratitude to BLS
project manager where he coordinated the for his almost 30 years as a mentor and col-
entire process of contacting the authors and league and for the chance to work on this book,
pulling the book together, while additionally to SH for all of his years as an amazing col-
providing editorial assistance. league, to his parents for providing the founda-
BLS would like to thank Steve Kimmel and tion for all of his work, and to his family – Alison,
Sean Hennessy for joining him as co-editors. David, Benjamin, and Jonathan – for all their
Steve did the bulk of the work on the first edi- support and patience during the many late eve-
tion of this textbook, and Steve and Sean joined nings that SEK worked on the book.
BLS as co-editors for the fifth and sixth edition SH also thanks BLS, his longtime friend and
of the parent book, Pharmacoepidemiology. career mentor, and all of his students, mentees,
These are two very special and talented men. It and collaborators. Finally, he thanks his par-
has been BLS’s pleasure to help train them – ents, and his family – Kristin, Landis, and
now, too many years ago – help them cultivate Bridget – for their love and support.
1
Part I
Introduction to Pharmacoepidemiology
3
What is Pharmacoepidemiology?
Brian L. Strom
Rutgers Biomedical and Health Sciences, Newark, NJ, USA
“A desire to take medicine is, perhaps, the the field has expanded its focus to include
great feature which distinguishes man many issues other than adverse reactions, as
from other animals.” well.
Sir William Osler, 1891 To clarify what is, and what is not, included
within the discipline of pharmacoepidemiol-
ogy, this chapter will begin by defining phar-
Introduction macoepidemiology, differentiating it from
other related fields. The history of drug regula-
In recent decades, modern medicine has been tion will then be briefly and selectively
blessed with a pharmaceutical armamentar- reviewed, focusing on the US experience as an
ium that is much more powerful now than it example, demonstrating how it has led to the
had had before. Although this has given health development of this new field. Next, the cur-
care providers the ability to provide better rent regulatory process for the approval of new
medical care for their patients, it has also drugs will be reviewed, in order to place the
resulted in the ability to do much greater harm. use of pharmacoepidemiology and postmar-
It has also generated an enormous number of keting drug surveillance into proper perspec-
product liability suits against pharmaceutical tive. Finally, the potential scientific and clinical
manufacturers, some appropriate and others contributions of pharmacoepidemiology will
inappropriate. In fact, the history of drug regu- be discussed.
lation parallels the history of major adverse
drug reaction (ADR) “disasters.” Each change
in pharmaceutical law was a political reaction Definition
to an epidemic of ADRs. A 1998 study esti- of Pharmacoepidemiology
mated that 100 000 Americans die each year
from ADRs, and 1.5 million US hospitaliza- Pharmacoepidemiology is the study of the use
tions each year result from ADRs; yet, 20–70% of and the effects of drugs in large numbers of
of ADRs may be preventable. The harm that people. The term pharmacoepidemiology obvi-
drugs can cause has also led to the develop- ously contains two components: “pharmaco”
ment of the field of pharmacoepidemiology, and “epidemiology.” In order to better appreci-
which is the focus of this book. More recently, ate and understand what is and what is not
Textbook of Pharmacoepidemiology, Third Edition. Edited by Brian L. Strom, Stephen E. Kimmel, and Sean Hennessy.
© 2022 John Wiley & Sons Ltd. Published 2022 by John Wiley & Sons Ltd.
4 1 What is Pharmacoepidemiology?
included in this new field, it is useful to com- level achieved. It deals with drug absorption,
pare its scope to that of other related fields. distribution, metabolism, and excretion.
The scope of pharmacoepidemiology will first Pharmacodynamics is the study of the relation-
be compared to that of clinical pharmacology, ship between drug level and drug effect.
and then to that of epidemiology. Together, these two fields allow one to predict
the effect one might observe in a patient from
administering a certain drug regimen.
Pharmacoepidemiology Versus
Pharmacoepidemiology encompasses ele-
Clinical Pharmacology
ments of both of these fields, exploring the
Pharmacology is the study of the effects of effects achieved by administering a drug regi-
drugs. Clinical pharmacology is the study of the men. It does not normally involve or require
effects of drugs in humans (see also Chapter 4). the measurement of drug levels. However,
Pharmacoepidemiology obviously can be con- pharmacoepidemiology can be used to shed
sidered, therefore, to fall within clinical phar- light on the pharmacokinetics of a drug when
macology. In attempting to optimize the use of used in clinical practice, such as exploring
drugs, one central principle of clinical pharma- whether aminophylline is more likely to cause
cology is that therapy should be individualized, nausea when administered to a patient simul-
or tailored, to the needs of the specific patient taneously taking cimetidine. However, to date
at hand. This individualization of therapy this is a relatively novel application of the field.
requires the determination of a risk/benefit Specifically, the field of pharmacoepidemiol-
ratio specific to the patient at hand. Doing so ogy has primarily concerned itself with the
requires a prescriber to be aware of the poten- study of adverse drug effects. Adverse reac-
tial beneficial and harmful effects of the drug in tions have traditionally been separated into
question and to know how elements of the those which are the result of an exaggerated
patient’s clinical status might modify the prob- but otherwise usual pharmacologic effect of
ability of a good therapeutic outcome. For the drug, sometimes called Type A reactions,
example, consider a patient with a serious versus those which are aberrant effects,
infection, serious liver impairment, and mild called Type B reactions. Type A reactions tend
impairment of his or her renal function. In con- to be common, dose-related, predictable, and
sidering whether to use gentamicin to treat his less serious. They can usually be treated by
infection, it is not sufficient to know that gen- simply reducing the dose of the drug. They
tamicin has a small probability of causing renal tend to occur in individuals who have one of
disease. A good clinician should realize that a three characteristics. First, the individuals may
patient who has impaired liver function is at a have received more of a drug than is customar-
greater risk of suffering from this adverse effect ily required. Second, they may have received a
than one with normal liver function. conventional amount of the drug, but they
Pharmacoepidemiology can be useful in pro- may metabolize or excrete the drug unusually
viding information about the beneficial and slowly, leading to drug levels that are too high
harmful effects of any drug, thus permitting a (see also Chapter 4). Third, they may have nor-
better assessment of the risk/benefit balance mal drug levels, but for some reason are overly
for the use of any particular drug in any par- sensitive to them.
ticular patient. In contrast, Type B reactions tend to be
Clinical pharmacology is traditionally uncommon, not related to dose, unpredictable,
divided into two basic areas: pharmacokinetics and potentially more serious. They usually
and pharmacodynamics. Pharmacokinetics is require cessation of the drug. They may be due
the study of the relationship between the dose to what are known as hypersensitivity reactions
administered of a drug and the serum or blood or immunologic reactions. Alternatively, Type
Historical Backgroun 5
B reactions may be some other idiosyncratic epidemiology, as well. Epidemiology is also

reaction to the drug, either due to some inher- traditionally subdivided into two basic areas.
ited susceptibility (e.g. glucose-6-phosphate The field began as the study of infectious dis-
dehydrogenase deficiency) or due to some eases in large populations, i.e. epidemics. It
other mechanism. Regardless, Type B reactions has since been expanded to encompass the
are the most difficult to predict or even detect, study of chronic diseases. The field of pharma-
and represent the major focus of many phar- coepidemiology uses the techniques of chronic
macoepidemiologic studies of ADRs. disease epidemiology to study the use of and
One typical approach to studying ADRs has the effects of drugs. Although application of
been the collection of spontaneous reports of the methods of pharmacoepidemiology can be
drug-related morbidity or mortality (see useful in performing the clinical trials of drugs
Chapter 7), sometimes called pharmacovigi- that are performed before marketing, the major
lance (although other times that term is used to application of these principles is after drug
refer to all of pharmacoepidemiology). marketing. This has primarily been in the con-
However, determining causation in case reports text of postmarketing drug surveillance,
of adverse reactions can be problematic (see although in recent years the interests of phar-
Chapter 14), as can attempts to compare the macoepidemiologists have broadened consid-
effects of drugs in the same class. Further, erably. Now, as will be made clearer in future
drug–drug interactions, predicted based on chapters, pharmacoepidemiology is consid-
pharmacokinetic data (see Chapter 4), require ered of importance in the whole life cycle of a
massive sample sizes to confirm in people. This drug, from the time when it is first discovered
has led academic investigators, industry, regu- or synthesized through when it is no longer
latory bodies, and the legal community to turn sold as a drug.
to the field of epidemiology. Specifically, studies Thus, pharmacoepidemiology is a relatively
of adverse effects have been supplemented with new applied field, bridging between clinical
studies of adverse events. In the former, investi- pharmacology and epidemiology. From clini-
gators examine case reports of purported ADRs cal pharmacology, pharmacoepidemiology
and attempt to make a subjective clinical judg- borrows its focus of inquiry. From epidemiol-
ment on an individual basis about whether the ogy, pharmacoepidemiology borrows its meth-
adverse outcome was actually caused by the ods of inquiry. In other words, it applies the
antecedent drug exposure. In the latter, con- methods of epidemiology to the content area
trolled studies are performed examining of clinical pharmacology. In the process, multi-
whether the adverse outcome under study ple special logistical approaches have been
occurs more often in an exposed population developed and multiple special methodologic
than in an unexposed population. This mar- issues have arisen. These are the primary foci
riage of the fields of clinical pharmacology and of this book.
epidemiology has resulted in the development
of a field: pharmacoepidemiology.
Historical Background
Pharmacoepidemiology Versus
Epidemiology Early Legislation
Epidemiology is the study of the distribution The history of drug regulation in the US is sim-
and determinants of diseases in populations. ilar to that in most developed countries, and
Since pharmacoepidemiology is the study of reflects the growing involvement of govern-
the use of and effects of drugs in large ments in attempting to assure that only safe
numbers of people, it obviously falls within and effective drug products were available and
that appropriate manufacturing and market- dramatic increase was seen in the frequency of
ing practices were used. The initial US law, the a previously rare birth defect, phocomelia--
Pure Food and Drug Act, was passed in 1906, the absence of limbs or parts of limbs, some-
in response to excessive adulteration and mis- times with the presence instead of flippers.
branding of the food and drugs available at Epidemiologic studies established its cause to
that time. There were no restrictions on sales be in utero exposure to thalidomide. In the
or requirements for proof of the efficacy or United Kingdom, this resulted in the estab-
safety of marketed drugs. Rather, the law sim- lishment in 1968 of the Committee on Safety
ply gave the federal government the power to of Medicines. Later, the World Health
remove from the market any product that was Organization established a bureau to collect
adulterated or misbranded. The burden of and collate information from this and other
proof was on the federal government. similar national drug monitoring organiza-
In 1937, over 100 people died from renal fail- tions (see Chapter 7).
ure as a result of the marketing by the The US had never permitted the marketing
Massengill Company of elixir of sulfanilimide of thalidomide and, so, was fortunately spared
dissolved in diethylene glycol. In response, this epidemic. However, the “thalidomide dis-
Congress passed the 1938 Food, Drug, and aster” was so dramatic that it resulted in regu-
Cosmetic Act. Preclinical toxicity testing was latory change in the US as well. Specifically, in
required for the first time. In addition, manu- 1962 the Kefauver–Harris Amendments were
facturers were required to gather clinical data passed. These amendments strengthened the
about drug safety and to submit these data to requirements for proof of drug safety, requir-
FDA before drug marketing. The FDA had ing extensive preclinical pharmacologic and
60 days to object to marketing or else it would toxicologic testing before a drug could be tested
proceed. No proof of efficacy was required. in man. The data from these studies were
Little attention was paid to ADRs until the required to be submitted to the US Food and
early 1950s, when it was discovered that chlo- Drug Administration (FDA) in an
ramphenicol could cause aplastic anemia. In Investigational New Drug (IND) application
1952, the first textbook of ADRs was published. before clinical studies could begin. Three
In the same year, the AMA Council on explicit phases of clinical testing were defined,
Pharmacy and Chemistry established the first which are described in more detail below. In
official registry of adverse drug effects, to col- addition, a new requirement was added to the
lect cases of drug-induced blood dyscrasias. In clinical testing, for “substantial evidence that
1960, the FDA began to collect reports of ADRs the drug will have the effect it purports or is
and sponsored new hospital-based drug moni- represented to have.” “Substantial evidence”
toring programs. The Johns Hopkins Hospital was defined as “adequate and well-controlled
and the Boston Collaborative Drug Surveillance investigations, including clinical investiga-
Program developed the use of in-hospital mon- tions.” Functionally, this has generally been
itors to perform cohort studies to explore the interpreted as requiring randomized clinical
short-term effects of drugs used in hospitals. trials to document drug efficacy before market-
This approach was later to be transported to ing. This new procedure also delayed drug
the University of Florida-Shands Teaching marketing until the FDA explicitly gave
Hospital, as well. approval. With some modifications, these are
In the winter of 1961, the world experi- the requirements still in place in the US today.
enced the infamous “thalidomide disaster.” In addition, the amendments required the
Thalidomide was marketed as a mild hypnotic, review of all drugs approved between 1938 and
and had no obvious advantage over other 1962, to determine if they too were efficacious.
drugs in its class. Shortly after its marketing, a The resulting DESI (Drug Efficacy Study
Implementation) process, conducted by the children” in Chapter 23). Acute flank pain
National Academy of Sciences’ National and reversible acute renal failure were noted
Research Council with support from a contract to be caused by suprofen. Isotretinoin was
from FDA, was not completed until years later, almost removed from the US market because
and resulted in the removal from the US mar- of the birth defects it causes. The Eosinophilia-
ket of many ineffective drugs and drug combi- Myalgia syndrome was linked to a particular
nations. The result of all these changes was a brand of L-tryptophan. Triazolam, thought by
great prolongation of the approval process, the Netherlands in 1979 to be subject to a dis-
with attendant increases in the cost of drug proportionate number of central nervous sys-
development, the so-called “drug lag.” tem side effects, was discovered by the rest of
However, the drugs that are marketed are pre- the world to be problematic in the early 1990s.
sumably much safer and more effective. Silicone breast implants, inserted by the mil-
lions in the US for cosmetic purposes, were
accused of causing cancer, rheumatologic dis-
Drug Crises and Resulting
ease, and many other problems, and restricted
Regulatory Actions
from use except for breast reconstruction
Despite the more stringent process for drug after mastectomy. Human insulin was mar-
regulation, subsequent years have seen a keted as one of the first of the new biotech-
series of major ADRs. Subacute myelo-optic- nology drugs, but soon thereafter was accused
neuropathy (SMON) was found in Japan to be of causing a disproportionate amount of
caused by clioquinol, a drug marketed in the hypoglycemia. Fluoxetine was marketed as a
early 1930s but not discovered to cause this major new important and commercially suc-
severe neurological reaction until 1970. In the cessful psychiatric product, but then lost a
1970s, clear cell adenocarcinoma of the cervix large part of its market due to accusations
and vagina and other genital malformations about its association with suicidal ideation.
were found to be due to in utero exposure to An epidemic of deaths from asthma in New
diethylstilbestrol two decades earlier. The Zealand was traced to fenoterol, and later
mid-1970s saw the UK discovery of the ocu- data suggested that similar, although smaller,
lomucocutaneous syndrome caused by pract- risks might be present with other beta-agonist
olol, five years after drug marketing. In 1980, inhalers. The possibility was raised of cancer
the drug ticrynafen was noted to cause deaths from depot-medroxyprogesterone, resulting
from liver disease. In 1982, benoxaprofen was in initial refusal to allow its marketing for this
noted to do the same. Subsequently the use of purpose in the US, then multiple studies, and
zomepirac, another nonsteroidal anti- ultimate approval. Arrhythmias were linked
inflammatory drug, was noted to be associ- to the use of the antihistamines terfenadine
ated with an increased risk of anaphylactoid and astemizole. Hypertension, seizures, and
reactions. Serious blood dyscrasias were strokes were noted from postpartum use of
linked to phenylbutazone. Small intestinal bromocriptine. Multiple different adverse
perforations were noted to be caused by a par- reactions were linked to temafloxacin. Other
ticular slow release formulation of indometh- examples include liver toxicity from
acin. BendectinR, a combination product amoxicillin-clavulanic acid; liver toxicity
indicated to treat nausea and vomiting in from bromfenac; cancer, myocardial infarc-
pregnancy, was removed from the market tion, and gastrointestinal bleeding from cal-
because of litigation claiming it was a terato- cium channel blockers; arrhythmias with
gen, despite the absence of valid scientific evi- cisapride interactions; primary pulmonary
dence to justify this claim (see “Research on hypertension and cardiac valvular disease
the effects of medications in pregnancy and in from dexfenfluramine and fenfluramine;
g astrointestinal bleeding, postoperative astemizole, cisapride, droperidol,

bleeding, deaths, and many other adverse grepafloxacin, halofantrine, pimozide, pro-
reactions associated with ketorolac; multiple poxyphene, rofecoxib, sertindole, sibutramine
drug interactions with mibefradil; thrombosis terfenadine, terodiline, thioridazine, veva-
from newer oral contraceptives; myocardial cetylmethadol, and ziprasidone, were subject
infarction from sildenafil; seizures with tram- to significant regulatory actions because of car-
adol; anaphylactic reactions from vitamin K; diac concerns.
liver toxicity from troglitazone; and intussus- Since 1993, trying to deal with drug safety
ception from rotavirus vaccine. problems, FDA morphed its extant spontane-
Later drug crises have occurred due to alle- ous reporting system into the MedWatch pro-
gations of ischemic colitis from alosetron; gram of collecting spontaneous reports of
rhabdomyolysis from cerivastatin; bronchos- adverse reactions (see Chapter 7), as part of
pasm from rapacuronium; torsades de pointes that issuing monthly notifications of label
from ziprasidone; hemorrhagic stroke from changes. Compared to the 20–25 safety-related
phenylpropanolamine; arthralgia, myalgia, label changes that were being made every
and neurologic conditions from Lyme vaccine; month by mid-1999, between 19 and 57 safety-
multiple joint and other symptoms from related label changes (boxed warnings, warn-
anthrax vaccine; myocarditis and myocardial ings, contraindications, precautions, adverse
infarction from smallpox vaccine; and heart events) were made every month in 2009. From
attack and stroke from rofecoxib. January of 2010 to July of 2016, there were
Major ADRs continue to plague new drugs, 3324 safety-related label changes. The range of
and in fact are as common if not more com- safety-related label changes per month was
mon in the last several decades. In total, 36 19–87 (median of 41). Among all safety-related
different oral prescription drug products label changes (January of 2010 to July of 2016),
have been removed from the US market, 8, 13, 56, and 65% were boxed warnings, con-
since 1980 alone (alosetron-2000, aprotinin- traindications, warnings, and precautions,
2007, astemizole-1999, benoxaprofen-1982, respectively.
bromfenac-1998, cerivastatin-2001, cisap According to a study from a number of years
ride-2000, dexfenfluramine-1997, efali ago by the US Government Accountability
zumab-2009, encainide-1991, etretinate-1998, Office, 51% of approved drugs have serious
fenfluramine-1998, flosequinan-1993, grepa- adverse effects not detected before approval.
floxin-1999, levomethadyl-2003, lumiracoxib- Further, there is recognition that the initial
2007, mibefradil-1998, natalizumab-2005, dose recommended for a newly marketed drug
nomifensine-1986, palladone-2005, pamo is often incorrect, and needs monitoring and
line-2005, pergolide-2010, phenylpropanola- modification after marketing.
mine-2000, propoxyphene-2010, rapacuronium- In some of the examples above, the drug was
2001, rimonabant-2010, rofecoxib-2004, never convincingly linked to the adverse reac-
sibutramine-2010, suprofen-1987, tegaserod- tion, yet many of these accusations led to the
2007, terfenadine-1998, temafloxacin-1992, removal of the drug involved from the market.
ticrynafen-1980, troglitazone-2000, val Interestingly, however, this withdrawal was
decoxib-2007, zomepirac 1983). The licensed not necessarily performed in all of the different
vaccines against rotavirus and Lyme were countries in which each drug was marketed.
also withdrawn because of safety concerns Most of these discoveries have led to litigation,
(see “Special methodological issues in phar- as well, and a few have even led to criminal
macoepidemiologic studies of vaccine safety” charges against the pharmaceutical manufac-
in Chapter 23). Further, between 1990 and turer and/or some of its employees (see
2004, at least 15 non-cardiac drugs including Chapter 6).
Legislative Actions Resulting the FDA was approving drugs too fast. There
from Drug Crises were also calls for the development of an inde-
pendent drug safety board, analogous to the
Through the 1980s, there was concern that an
National Transportation Safety Board, with a
underfunded FDA was approving drugs too
mission much wider than FDA’s regulatory
slowly, and that the US suffered, compared to
mission, to complement the latter. For exam-
Europe, from a “drug lag.” To provide addi-
ple, such a board could investigate drug safety
tional resources to the FDA to help expedite
crises such as those cited above, looking for
the drug review and approval process, in 1992
ways to prevent them, and could deal with
Congress passed the Prescription Drug User
issues such as improper physician use of
Fee Act (PDUFA), allowing the FDA to charge
drugs, the need for training, and the develop-
manufacturers a fee for reviewing New Drug
ment of new approaches to the field of
Applications. This legislation was reauthor-
pharmacoepidemiology.
ized by Congress three more times: PDUFA II,
Recurrent concerns about the FDA’s man-
also called the Food and Drug Modernization
agement of postmarketing drug safety issues
Act of 1997; PDUFA III, also called the Public
led to a systematic review of the entire drug
Health Security and Bioterrorism Preparedness
risk assessment process. In 2006, the US
and Response Act of 2002; and PDUFA IV, also
General Accountability Office issued its report
called the Food and Drug Administration
of a review of the organizational structure and
Amendments (FDAAA-PL 110-85) of 2007.
effectiveness of FDA’s postmarketing drug
The goals for PDUFA I, II, III, and IV was to
safety decision-making, followed in 2007 by
enable the FDA to complete the review of over
the Institute of Medicine’s independent assess-
90% of priority drug applications in 6 months,
ment. Important weaknesses were noted in the
and complete the review of over 90% of stand-
current system, including the failure of the
ard drug applications in 12 months (under
FDA’s Office of New Drugs and Office of Drug
PDUFA I) or 10 months (under PDUFA II, III,
Safety to communicate with each other on
and IV). In addition to reauthorizing the col-
safety issues, the failure of the FDA to track
lection of user fees from the pharmaceutical
ongoing postmarketing studies, the ambiguous
industry, PDUFA II allowed the FDA to accept
role of the FDA’s Office of Drug Safety in scien-
a single well-controlled clinical study under
tific advisory committees, the limited authority
certain conditions, to reduce drug develop-
by the FDA to require the pharmaceutical
ment time. The result was a system where
industry to perform studies to obtain needed
more than 550 new drugs were approved by
data, concerns about culture problems at the
the FDA in the 1990s.
FDA where recommendations by members of
However, whereas 1400 FDA employees in the FDA’s drug safety staff were not followed,
1998 worked with the drug approval process, and concerns about conflict of interest involv-
only 52 monitored safety; FDA spent only ing advisory committee members. This
$2.4 million in extramural safety research. This Institute of Medicine report was influential in
state of affairs has coincided with the growing shaping PDUFA IV.
numbers of drug crises cited above. With suc- Indeed, with the passage of PDUFA IV, the
cessive reauthorizations of PDUFA, this mark- authority of the FDA was substantially
edly changed. PDUFA III allowed the FDA for increased, with the ability, for example, to
the first time to use a small portion of the user require postmarketing studies and levy heavy
fees for post-marketing drug safety monitor- fines if these requirements were not met.
ing, to address safety concerns. Further, its resources were substantially
However, there was now growing concern, increased, with a specific charge to (i) fund epi-
in Congress and the US public, that perhaps demiology best practices and data acquisition
($7 million in fiscal 2008, increasing to drug companies looking for FDA approval on
$9.5 million in fiscal 2012), (ii) fund new drug new products or new indications on existing
trade name review ($5.3 million in fiscal 2008, drugs. It calls for the use of “data summaries”
rising to $6.5 million in fiscal 2012), and (iii) to support the approval of certain drugs for
fund risk management and communication new indications, rather than full clinical trial
($4 million in fiscal 2008, rising to $5 million data. It will also allow drug companies to pro-
in fiscal 2012) (see also “Risk management” in mote off-label uses to insurance companies,
Chapter 23). In addition, in another use of the allowing them to expand their markets. Of par-
new PDUFA funds, the FDA plans to develop ticular relevance to pharmacoepidemiology, it
and implement agency-wide and special- permitted the use of “real world evidence”
purpose postmarket IT systems, including the rather than full clinical trial results. Depending
MedWatch Plus Portal, the FDA Adverse Event on how these new rules are interpreted, this
Reporting System, the Sentinel System (a vir- could massively change drug development in
tual national medical product safety system), the US, and in particular the role of pharma-
and the Phonetic and Orthographic Computer coepidemiology in that drug development.
Analysis System to find similarities in spelling
or sound between proposed proprietary drug
Intellectual Development
names that might increase the risk of confu-
of Pharmacoepidemiology
sion and medication errors.
Emerging from Drug Crises
FDASIA, the fifth authorization of the
PDUFA, expanded the authority of the US Several developments of the 1960s can be
FDA with the ability to safeguard and advance thought to have marked the beginning of the
public health by: (i) “giving the authority to field of pharmacoepidemiology. The Kefauver-
collect user fees from industry to fund reviews Harris Amendments that were introduced in
of innovator drugs, medical devices, generic 1962 required formal safety studies for new
drugs and biosimilar biological products”; (ii) drug applications. The DESI program that was
“promoting innovation to speed patient access undertaken by the FDA as part of the Kefauver-
to safe and effective products”; (iii) “increasing Harris Amendments required formal efficacy
stakeholder involvement in FDA processes”, studies for old drugs that were approved ear-
and (iv) “enhancing the safety of the drug sup- lier. These requirements created the demand
ply chain.” Also enacted in 2012, GDUFA perfor new expertise and new methods. In addi-
mitted the FDA to assess industry user fees tion, the mid-1960s saw the publication of a
with the intention of increasing the predicta- series of drug utilization studies. These studies
bility and timeliness of generic drug applica- provided the first descriptive information on
tions reviews. The Biosimilar User Fee Act how physicians use drugs, and began a series
(BsUFA), also enacted in 2012, authorized the of investigations of the frequency and determi-
FDA to collect fees directly from biosimilar nants of poor prescribing (see also “Evaluating
drug product applicants to expedite the review and improving prescribing” in Chapter 23).
of biosimilar applications. The FDA In part in response to concerns about adverse
Reauthorization Act of 2017 (FDARA), reau- drug effects, the early 1970s saw the develop-
thorized PDUFA, GDUFA, and BsUFA through ment of the Drug Epidemiology Unit, now the
fiscal year 2022. Slone Epidemiology Center, which extended
Among other things, the twenty-first cen- the hospital-based approach of the Boston
tury Cures Act (enacted in December 2016 in Collaborative Drug Surveillance Program by
the United States) was intended to expedite the collecting lifetime drug exposure histories
process by which new drugs and devices are from hospitalized patients and using these to
approved by easing the requirements put on perform hospital-based case–control studies.
The year 1976 saw the formation of the Joint new program of Centers for Education and
Commission on Prescription Drug Use, an Research on Therapeutics (CERTs) was author-
interdisciplinary committee of experts charged ized under the FDA Modernization Act of 1997
with reviewing the state of the art of pharma- (as part of the same legislation that reauthor-
coepidemiology at that time, as well as provid- ized PDUFA II described earlier). Starting in
ing recommendations for the future. The 1999 and incrementally adding more centers in
Computerized Online Medicaid Analysis and 2002, 2006, and 2007, the Agency for
Surveillance System (COMPASS®) was first Healthcare Research and Quality (AHRQ) that
developed in 1977, using Medicaid billing data was selected to administer this program had
to perform pharmacoepidemiologic studies funded up to 14 Centers for Education and
(see Chapter 9). The Drug Surveillance Research on Therapeutics (CERTs), although
Research Unit, now called the Drug Safety this program ended in 2016 (see also
Research Trust, was developed in the United Chapter 6).
Kingdom in 1980, with its innovative system of The research and education activities spon-
Prescription Event Monitoring. Each of these sored by AHRQ through the CERTs program
represented major contributions to the field of since the late 1990s take place in academic
pharmacoepidemiology. These and newer centers. These CERTs centers conduct research
approaches are reviewed in Part II of this book. on therapeutics, exploring new uses of drugs,
In the examples of drug crises mentioned ways to improve the effective uses of drugs,
above, these were serious but uncommon drug and the risks associated with new uses or com-
effects, and these experiences have led to an binations of drugs. They also develop educa-
accelerated search for new methods to study tional modules and materials for disseminating
drug effects in large numbers of patients. This the research findings about medical products.
led to a shift from adverse effect studies to With the development of direct-to-consumer
adverse event studies, with concomitant advertising of drugs since the mid 1980s in the
increasing use of new data resources and new US, the CERTs’ role in educating the public
methods to study adverse reactions. The and health care professionals by providing
American Society for Clinical Pharmacology evidence-based information has become espe-
and Therapeutics issued, in 1990, a position cially important.
paper on the use of purported postmarketing Another impetus for research on drugs
drug surveillance studies for promotional pur- resulted from one of the mandates (in
poses, and the International Society for Section 1013) of the Medicare Prescription
Pharmacoepidemiology (ISPE) issued, in 1996, Drug, Improvement, and Modernization Act
Guidelines for Good Epidemiology Practices of 2003 to provide beneficiaries with scientific
for Drug, Device, and Vaccine Research in the information on the outcomes, comparative
United States, which were updated in 2007. clinical effectiveness, and appropriateness of
Since the late 1990s, pharmacoepidemiologic health care items and services. In response, the
research has also been increasingly burdened AHRQ created in 2005 the DEcIDE (Developing
by concerns about patient confidentiality (see Evidence to Inform Decisions about
also Chapter 16). Effectiveness) Network to support in academic
There is also increasing recognition that settings the conduct of studies on effective-
most of the risk from most drugs to most ness, safety, and usefulness of drugs and other
patients occurs from known reactions to old treatments and services. This, too ended in
drugs. As an attempt to address concerns about 2012.
underuse, overuse, and adverse events of med- Another major new initiative of close rele-
ical products and medical errors that may vance to pharmacoepidemiology is risk man-
cause serious impairment to patient health, a agement. There is increasing recognition that
the risk/benefit balance of some drugs can at its peak with over $50 million/year of fund-
only be considered acceptable with active ing. While only a portion of this is dedicated to
management of their use, to maximize their medication errors, they are clearly a focus of
efficacy and/or minimize their risk. In interest and relevance to many (see “The
response, in the late 1990s, there were new ini- pharmacoepidemiology of medication errors”
tiatives underway, ranging from FDA require- in Chapter 23).
ments for risk management plans, to a FDA The 1990s and especially the 2000s have seen
Drug Safety and Risk Management Advisory another shift in the field, away from its exclu-
Committee, and issuing risk minimization and sive emphasis on drug utilization and adverse
management guidances in 2005 (see Chapters 6 reactions, to the inclusion of other interests as
and 23). well, such as the use of pharmacoepidemiol-
Another initiative closely related to phar- ogy to study beneficial drug effects, the appli-
macoepidemiology is the Patient Safety move- cation of health economics to the study of drug
ment. In the Institute of Medicine’s report, effects, studies of patient engagement and
“To Err is Human: Building a Safer Health patient reported outcomes, meta-analysis, etc.
System,” the authors note that: (i) “even These new foci are discussed in more detail in
apparently single events or errors are due Part III of this book.
most often to the convergence of multiple con- Also, with the publication of the results from
tributing factors,” (ii) “preventing errors and the Women’s Health Initiative indicating that
improving safety for patients requires a sys- combination hormone replacement therapy
tems approach in order to modify the condi- causes an increased risk of myocardial infarc-
tions that contribute to errors,” and (iii) “the tion rather than a decreased risk, there has
problem is not bad people; the problem is that been increased concern about reliance solely
the system needs to be made safer.” In this on nonexperimental methods to study drug
framework, the concern is not about substand- safety after marketing. This has led to increased
ard or negligent care, but rather, is about use of massive randomized clinical trials as
errors made by even the best trained, bright- part of postmarketing surveillance (see
est, and most competent professional health Chapter 17). This is especially important
caregivers and/or patients. From this perspec- because often the surrogate markers used for
tive, the important research questions ask drug development cannot necessarily be relied
about the conditions under which people upon to map completely to true clinical
make errors, the types of errors being made, outcomes.
and the types of systems that can be put into Finally, with the advent of the Obama admin-
place to prevent errors altogether when possi- istration in the US, there was enormous interest
ble. Errors that are not prevented must be in comparative effectiveness research (CER).
identified and corrected efficiently and CER was defined in 2009 by the Federal
quickly, before they inflict harm. Turning spe- Coordinating Council for Comparative
cifically to medications, from 2.4 to 6.5% of Effectiveness Research as “the conduct and
hospitalized patients suffer ADEs, prolonging synthesis of research comparing the benefits
hospital stays by two days, and increase costs and harms of different interventions and strate-
by $2000–2600 per patient. Over 7000 US gies to prevent, diagnose, treat, and monitor
deaths were attributed to medication errors in health conditions in “real world” settings. The
1993. Although these estimates have been dis- purpose of this research is to improve health
puted, the overall importance of reducing outcomes by developing and disseminating
these errors has not been questioned. In recog- evidence-based information to patients, clini-
nition of this problem, AHRQ launched a cians, and other decision-makers, responding to
major new grant program of over 100 projects, their expressed needs, about which interventions
The Current Drug Approval Proces 13
are most effective for which patients under and develop a drug that successfully reached
specific circumstances.” By this definition, CER the market rose from over $800 million in 2004
includes three key elements: (i) evidence to an estimated $1.3 billion to 1.7 billion cur-
synthesis, (ii) evidence generation, and (iii) rently. In addition to the sizeable costs of
evidence dissemination. Typically, CER is con- research and development, a substantial part
ducted through observational studies of either of this total cost is determined also by the regu-
large administrative or medical record data- latory requirement to test new drugs during
bases (see Part II), or large naturalistic clinical several pre-marketing and post-marketing
trials (see Chapter 17). In many ways, the UK phases, as will be reviewed next.
has been focusing on CER for years, with its The current drug approval process in the US
National Institute for Health and Clinical and most other developed countries includes
Excellence (NICE), an independent organiza- preclinical animal testing followed by three
tion responsible for providing national guid- phases of clinical testing. Phase I testing is usu-
ance on promoting good health and preventing ally conducted in just a few normal volunteers,
and treating ill health. However, the Obama and represents the initial trials of the drug in
administration included $1.1 billion for CER in humans. Phase I trials are generally conducted
its federal stimulus package, and has plans by clinical pharmacologists, to determine the
for hundreds of millions of dollars of support metabolism of the drug in humans, a safe dos-
per year thereafter. While CER does not overage range in humans, and to exclude any
lap completely with pharmacoepidemiology, extremely common toxic reactions which are
the scientific approaches are very close. unique to humans.
Pharmacoepidemiologists evaluate the use and Phase II testing is also generally conducted
effects of medications. CER investigators com- by clinical pharmacologists, on a small number
pare, in the real world, the safety and benefits of patients who have the target disease. Phase
of one treatment compared to another. CER II testing is usually the first time patients are
extends beyond pharmacoepidemiology in that exposed to the drug. Exceptions are drugs that
CER can include more than just drugs; phar- are so toxic that it would not normally be con-
macoepidemiology extends beyond CER in that sidered ethical to expose healthy individuals to
it includes studies comparing exposed to unex- them, like cytotoxic drugs. For these, patients
posed patients, not just alternative exposures. are used for Phase I testing as well. The goals of
However, to date, most work done in CER has Phase II testing are to obtain more information
been done in pharmacoepidemiology. on the pharmacokinetics of the drug and on
any relatively common adverse reactions, and
to obtain initial information on the possible
he Current Drug Approval
T efficacy of the drug. Specifically, Phase II is
used to determine the daily dosage and regi-
Process
men to be tested more rigorously in Phase III.
Phase III testing is performed by clinician-
Drug Approval in the US
investigators in a much larger number of
Since the mid-1990s, there has been a decline patients, in order to rigorously evaluate a
in the number of novel drugs approved per drug’s efficacy and to provide more informa-
year, while the cost of bringing a drug to mar- tion on its toxicity. At least one of the Phase III
ket has risen sharply. The total cost of drug studies needs to be a randomized clinical trial
development to the pharmaceutical industry (see Chapter 17). To meet FDA standards, at
increased from $24 billion in 1999, to $32 bil- least one of the randomized clinical trials usu-
lion in 2002, and to $65.2 billion on research ally needs to be conducted in the US. Generally
and development in 2008. The cost to discover between 500 and 3000 patients are exposed to a
drug during Phase III, even if drug efficacy can review, approval for clinical trials, postmar-
be demonstrated with much smaller numbers, keting surveillance, and inspection of manu-
in order to be able to detect less common facturing practice. Examples for this are
adverse reactions. For example, a study includ- Health Canada, the China Food and Drug
ing 3000 patients would allow one to be 95% Administration (CFDA), the Medicines
certain of detecting any adverse reactions that Agency in Denmark, the Medicines Agency
occur in at least one exposed patient out of in Norway, the Center for Drug Administration
1000. At the other extreme, a total of 500 in Singapore and the Medicines & Medical
patients would allow one to be 95% certain of Devices Safety Authority in New Zealand. In
detecting any adverse reactions which occur in other countries, regulatory functions are dis-
six or more patients out of every 1000 exposed. tributed among different agencies. An exam-
Adverse reactions which occur less commonly ple of the latter is The Netherlands, where
than these are less likely to be detected in these the Ministry of Health, Welfare & Sports per-
premarketing studies. The sample sizes needed forms the functions of licensing; the
to detect drug effects are discussed in more Healthcare Inspectorate checks on general
detail in Chapter 4. Nowadays, with the manufacturing practice; and the Medicines
increased focus on drug safety, premarketing Evaluation Board performs the functions of
dossiers are sometimes being extended well product assessment and registration and
beyond 3000 patients. However, as one can tell ADR monitoring. As another example, in
from the sample size calculations in Chapter 3 Singapore, two independent agencies (the
and Appendix A, by itself these larger numbers Center for Pharmaceutical Administration
gain little additional information about ADRs, and the Center for Drug Evaluation) were
unless one were to increase to perhaps 30 000 previously responsible for medicinal regula-
patients, well beyond the scope of most pre- tion and evaluation, but are currently merged
marketing studies. into a single agency (the Center for Drug
Finally, Phase IV testing is the evaluation of Administration). Another dimension on
the effects of drugs after general marketing. which countries may vary is the degree of
The bulk of this book, is devoted to such autonomy of regulatory decisions from politi-
efforts. cal influence. Drug regulation in most coun-
tries is performed by a department within the
executive branch (Australia, Cuba, Cyprus,
Drug Approval in Other Countries
Tunisia, and Venezuela are examples cited by
Outside the US, national systems for the reg- the WHO report, and Denmark, India, and
ulation and approval of new drugs vary New Zealand are other examples). In other
greatly, even among developed countries and countries, this function is performed by an
especially between developed and developing independent commission or board. An exam-
countries. While in most developed coun- ple of the latter arrangement is The
tries, at least, the general process of drug Netherlands, where members of the
development is very analogous to that in the Medicines Evaluation Board are appointed
US, the implementation varies widely. A directly by the Crown, thereby enabling
WHO comparative analysis of drug regula- actions that are independent of interference
tion in 10 countries found that not all coun- by other government authorities, such as the
tries even have a written national drug policy Minister of Health. All 10 countries exam-
document. Regulation of medicines in some ined by the WHO require registration of
countries is centralized in a single agency pharmaceutical products, but they differ on
that performs the gamut of functions involv- the documentation requirements for evi-
ing product registration, licensing, product dence of safety and efficacy. Some countries
Potential Contributions of Pharmacoepidemiolog 15
carry out independent assessments while otential Contributions

P
others, especially many developing coun- of Pharmacoepidemiology
tries, rely on WHO assessments or other
sources. With the exception of Cyprus, the The potential contributions of pharmacoepi-
remaining nine countries surveyed by the demiology are now well recognized, even
WHO were found to regulate the conduct of though the field is still relatively new. However,
clinical trials, but with varying rates of par- some contributions are already apparent (see
ticipation of health care professionals in Table 1.1). In fact, in the 1970s the FDA
reporting ADRs. Another source noted that requested postmarketing research at the time
countries also differ on the extent of empha- of approval for about one third of drugs, com-
sis on quantitative or qualitative analysis for pared to over 70% in the 1990s. Now, since the
assessing pre-and post-marketing data. passage of the Food, and Drug Administration
Further, within Europe, each country has Amendments Act of 2007 (FDAAA-PL 110-85)
its own regulatory agency, e.g. the United noted above, the FDA has the right to require
Kingdom’s Medicines and Healthcare such studies be completed. In this section of
Products Regulatory Agency (MHRA), formed this chapter, we will first review the potential
in 2003 as a merger of the Medicines Control for pharmacoepidemiologic studies to supple-
Agency (MCA) and the Medical Devices ment the information available prior to mar-
Agency (MDA). In addition, since January keting, and then review the new types of
1998, some drug registration and approval information obtainable from postmarketing
within the European Union has shifted away
from the national licensing authorities of the Table 1.1 Potential contributions
EU members to that of the centralized author- of pharmacoepidemiology.
ity of the European Medicines Evaluation
Agency (EMEA), which was established in A) I nformation which supplements the
information available from premarketing
1993. To facilitate this centralized approval studies – better quantitation of the incidence
process, the EMEA pushed for harmonization of known adverse and beneficial effects
of drug approvals. While the goals of harmo- 1) Higher precision
nization are to create a single pharmaceutical 2) In patients not studied prior to marketing,
market in Europe and to shorten approval e.g. the elderly, children, pregnant women
times, concerns were voiced that harmonized 3) As modified by other drugs and other illnesses
safety standards would lower the stricter 4) Relative to other drugs used for the same
standards that were favored by some coun- indications
tries such as Sweden, for example, and would B) New types of information not available from
compromise patient safety. Now called the premarketing studies
European Medicines Agency (EMA), the 1) Discovery of previously undetected adverse
EMA is a decentralized body of the European and beneficial effects
Union, responsible for the scientific evalua- i) Uncommon effects
tion and supervision of medicines. These ii) Delayed effects
functions are performed by the EMA’s 2) Patterns of drug utilization
Committee for Medicinal Products for Human
3) The effects of drug overdoses
Use (CHMP). EMA authorization to market a
4) The economic implications of drug use
drug is valid in all European Union countries,
but individual national medicines agencies C) General contributions of
pharmacoepidemiology
are responsible for monitoring the safety of
1) Reassurances about drug safety
approved drugs and sharing this information
with the EMA. 2) Fulfillment of ethical and legal obligations
pharmacoepidemiologic studies but not not typically included in studies conducted

obtainable prior to drug marketing. Finally, we before drug marketing, usually for ethical rea-
will review the general, and probably most sons. Examples include the elderly, children,
important, potential contributions such stud- and pregnant women. Studies of the effects of
ies can make. In each case, the relevant infor- drugs in these populations generally must
mation available from premarketing studies await studies conducted after drug marketing
will be briefly examined first, to clarify how (see also “Research on the effects of medica-
postmarketing studies can supplement this tions in pregnancy and in children” in
information. Chapter 23).
Additionally, for reasons of statistical effi-
ciency, premarketing clinical trials generally
seek subjects who are as homogeneous as pos-
Supplementary Information
sible, in order to reduce unexplained variabil-
Premarketing studies of drug effects are neces- ity in the outcome variables measured and
sarily limited in size. After marketing, nonex- increase the probability of detecting a differ-
perimental epidemiologic studies can be ence between the study groups, if one truly
performed, evaluating the effects of drugs exists. For these reasons, certain patients are
administered as part of ongoing medical care. often excluded, including those with other ill-
These allow the cost-effective accumulation of nesses or those who are receiving other drugs.
much larger numbers of patients than those Postmarketing studies can explore how factors
studied prior to marketing, resulting in a more such as other illnesses and other drugs might
precise measurement of the incidence of modify the effects of the drugs, as well as look-
adverse and beneficial drug effects (see ing at the effects of differences in drug regi-
Chapter 3). For example, at the time of drug men, adherence, etc. For example, after
marketing, prazosin was known to cause a marketing, the ophthalmic preparation of tim-
dose-dependent first dose syncope, but the olol was noted to cause many serious episodes
FDA requested the manufacturer to conduct a of heart block and asthma, resulting in over 10
postmarketing surveillance study of the drug deaths. These effects were not detected prior to
in the US to quantitate its incidence more premarketing, as patients with underlying cardio-
cisely. In recent years, there has even been an vascular or respiratory disease were excluded
attempt, in selected special cases, to release from the premarketing studies.
selected critically important drugs more Finally, to obtain approval to market a drug,
quickly, by taking advantage of the work that a manufacturer needs to evaluate its overall
can be performed after marketing. Probably safety and efficacy, but does not need to evalu-
the best-known early example was zidovudine. ate its safety and efficacy relative to any other
More recently, this has been the case with a drugs available for the same indication. To the
number of cancer drugs, including at least one contrary, with the exception of illnesses that
where initial expectations of efficacy were not could not ethically be treated with placebos,
confirmed in definitive trials after marketing, such as serious infections and malignancies, it
and then proven again later in a subgroup, is generally considered preferable, or even
leading to the product being removed from the mandatory, to have studies with placebo con-
market and then marketed again. As noted trols. There are a number of reasons for this
above, the increased sample size available after preference. First, it is easier to show that a new
marketing also permits a more precise drug is more effective than a placebo than to
determination of the correct dose to be used. show it is more effective than another effective
Premarketing studies also tend to be very drug. Second, one cannot actually prove that a
artificial. Important subgroups of patients are new drug is as effective as a standard drug.
Potential Contributions of Pharmacoepidemiolog 17
A study showing a new drug is no worse than high doses is rarely possible before drug mar-
another effective drug does not provide assur- keting. Again, this must await postmarketing
ance that it is better than a placebo; one simply pharmacoepidemiologic studies.
could have failed to detect that it was in fact Finally, it is only in the past decade or two
worse than the standard drug. One could that pharmacoepidemiologists have become
require a demonstration that a new drug is more sensitive to the costs of medical care, and
more effective than another effective drug, but the techniques of health economics been
this is a standard that does not and should not applied to evaluate the cost implications of
have to be met. Yet, optimal medical care drug use. It is clear that the exploration of the
requires information on the effects of a drug costs of drug use requires consideration of
relative to the alternatives available for the more than just the costs of the drugs them-
same indication. This information must often selves. The costs of a drug’s adverse effects may
await studies conducted after drug marketing. be substantially higher than the cost of the
drug itself, if these adverse effects result in
additional medical care and possibly even hos-
New Types of Information Not
pitalizations. Conversely, a drug’s beneficial
Available from Premarketing
effects could reduce the need for medical care,
Studies
resulting in savings that can be much larger
As mentioned above, premarketing studies are than the cost of the drug itself. As with studies
necessarily limited in size (see also Chapter 3). of drug utilization, the economic implications
The additional sample size available in post- of drug use can be predicted prior to market-
marketing studies permits the study of drug ing, but can only be rigorously studied after
effects that may be uncommon, but important, marketing (see Chapter 18).
such as drug-induced agranulocytosis.
Premarketing studies are also necessarily
General Contributions
limited in time; they must come to an end, or
the drug could never be marketed. In contrast,
postmarketing studies permit the study of Lastly, it is important to review the general
delayed drug effects, such as the unusual clear contributions that can be made by pharma-
cell adenocarcinoma of the vagina and cervix, coepidemiology. As an academic or a clinician,
which occurred two decades later in women one is most interested in the new information
exposed in utero to diethylstilbestrol. about drug effects and drug costs that can be
The patterns of physician prescribing and gained from pharmacoepidemiology. Certainly,
patient drug utilization often cannot be pre- these are the findings that receive the greatest
dicted prior to marketing, despite pharmaceu- public and political attention. However, often
tical manufacturers’ best attempts to predict no new information is obtained, particularly
when planning for drug marketing. Studies of about new adverse drug effects. This is not a
how a drug is actually being used, and determi- disappointing outcome, but in fact, a very reas-
nants of changes in these usage patterns, can suring one, and this reassurance about drug
only be performed after drug marketing (see safety is one of the most important contribu-
“Studies of drug utilization” and “Evaluating tions that can be made by pharmacoepidemio-
and improving prescribing” in Chapter 23). logic studies. Related to this is the reassurance
In most cases, premarketing studies are per- that the sponsor of the study, whether manu-
formed using selected patients who are closely facturer or regulator, is fulfilling its organiza-
observed. Rarely are there any significant over- tional duty ethically and responsibly by looking
doses in this population. Thus, the study of the for any undiscovered problems which may be
effects of a drug when ingested in extremely there. In an era of product liability litigation,
this is an important assurance. One cannot ●● The history of pharmacoepidemiology is a

change whether a drug causes an adverse reac- history of increasingly frequent accusations
tion, and the fact that it does will hopefully about ADRs, often arising out of the sponta-
eventually become evident. What can be neous reporting system, followed by formal
changed is the perception about whether a studies proving or disproving those
manufacturer did everything possible to detect associations.
it and was not negligent in its behavior. ●● The drug approval process is inherently
limited, so it cannot detect before market-
ing adverse effects that are uncommon,
Key Points delayed, unique to high risk populations,
due to misuse of the drugs by prescribers or
●● Pharmacoepidemiology is the study of the patients, etc.
use of and the effects of drugs and other ●● Pharmacoepidemiology can contribute
medical devices in large numbers of people. information about drug safety and effective-
It uses the methods of epidemiology to study ness that is not available from premarketing
the content area of clinical pharmacology. studies.
Further Reading
Califf, R.M. (2002). The need for a national ISPE (2008). Guidelines for good
infrastructure to improve the rational use of pharmacoepidemiology practices (GPP).
therapeutics. Pharmacoepidemiol. Drug Saf. Pharmacoepidemiol. Drug Saf. 17: 200–208.
11: 319–327. Joint Commission on Prescription Drug Use
Caranasos, G.J., Stewart, R.B., and Cluff, L.E. (1980) Final Report. Washington, DC.
(1974). Drug-induced illness leading to Kimmel, S.E., Keane, M.G., Crary, J.L. et al.
hospitalization. JAMA 228: 713–717. (1999). Detailed examination of fenfluramine-
Cluff, L.E., Thornton, G.F., and Seidl, L.G. phentermine users with valve abnormalities
(1964). Studies on the epidemiology of adverse identified in Fargo, North Dakota. Am. J.
drug reactions I. Methods of surveillance. Cardiol. 84: 304–308.
JAMA 188: 976–983. Kono, R. (1980). Trends and lessons of SMON
Crane, J., Pearce, N., Flatt, A. et al. (1989). research. In: Drug-Induced Sufferings (ed. T.
Prescribed fenoterol and death from asthma in Soda), 11. Princeton, NJ: Excerpta Medica.
New Zealand, 1981–1983: case–control study. Lazarou, J., Pomeranz, B.H., and Corey, P.N.
Lancet 1: 917–922. (1998). Incidence of adverse drug reactions in
Erslev, A.J. and Wintrobe, M.M. (1962). hospitalized patients: a meta-analysis of
Detection and prevention of drug induced prospective studies. JAMA 279: 1200–1205.
blood dyscrasias. JAMA 181: 114–119. Lenz, W. (1966). Malformations caused by
Geiling, E.M.K. and Cannon, P.R. (1938). drugs in pregnancy. Am. J. Dis. Child. 112:
Pathogenic effects of elixir of sulfanilimide 99–106.
(diethylene glycol) poisoning. JAMA 111: Meyler, L. (1952). Side Effects of Drugs.
919–926. Amsterdam: Elsevier.
Herbst, A.L., Ulfelder, H., and Poskanzer, D.C. Miller, R.R. and Greenblatt, D.J. (1976). Drug
(1971). Adenocarcinoma of the vagina: Effects in Hospitalized Patients. New York:
association of maternal stilbestrol therapy John Wiley & Sons, Inc.
with tumor appearance in young women. N. Rawlins, M.D. and Thompson, J.W. (1977).
Engl. J. Med. 284: 878–881. Pathogenesis of adverse drug reactions.
Further Readin 19
In: Textbook of Adverse Drug Reactions (ed. D.M. postmarketing surveillance study. JAMA 275:
Davies), 44. Oxford: Oxford University Press. 376–382.
Strom, B.L. (1990). Members of the ASCPT Wallerstein, R.O., Condit, P.K., Kasper, C.K.
pharmacoepidemiology section. Position et al. (1969). Statewide study of
paper on the use of purported postmarketing chloramphenicol therapy and fatal aplastic
drug surveillance studies for promotional anemia. JAMA 208: 2045–2050.
purposes. Clin. Pharmacol. Ther. 48: 598. Wright, P. (1975). Untoward effects associated
Strom, B.L., Berlin, J.A., Kinman, J.L. et al. with practolol administration.
(1996). Parenteral ketorolac and risk of Oculomucocutaneous syndrome. BMJ 1:
gastrointestinal and operative site bleeding: a 595–598.
20
Study Designs Available for Pharmacoepidemiologic Studies

Brian L. Strom
Introduction principles of clinical pharmacology, the con-

tent area of pharmacoepidemiology.
Pharmacoepidemiology applies the methods
of epidemiology to the content area of clinical
pharmacology. Therefore, in order to under- verview of the
O
stand the approaches and methodological Scientific Method
issues specific to the field of pharmacoepide-
miology, the basic principles of the field of epi- The scientific method to investigate a research
demiology must be understood as well. To this question involves a three-stage process (see
end, this chapter will begin with an overview Figure 2.1). In the first stage, one selects a
of the scientific method in general. This will be group of subjects for study. These subjects may
followed by a discussion of the different types be patients or animals or biologic cells and are
of errors one can make in designing a study. the sources for data sought by the study to
Next, the chapter will review the “Criteria for answer a question of interest. Second, one uses
the causal nature of an association,” which is the information obtained in this sample of
how one can decide whether an association study subjects to generalize and draw a conclu-
demonstrated in a particular study is, in fact, a sion about a population in general. This con-
causal association. Finally, the specific study clusion is referred to as an association. Third,
designs available for epidemiologic studies, or one generalizes again, drawing a conclusion
in fact for any clinical studies, will be reviewed. about scientific theory or causation. Each will
The next chapter discusses a specific methodo- be discussed in turn.
logical issue which needs to be addressed in Any given study is performed on a selection
any study, but which is of particular impor- of individuals, who represent the study sub-
tance for pharmacoepidemiologic studies: the jects. These study subjects should theoretically
issue of sample size. These two chapters are represent a random sample of some defined
intended to be an introduction to the field of population. For example, one might perform a
epidemiology for the neophyte. More informa- randomized clinical trial of the efficacy of
tion on these principles can be obtained from enalapril in lowering blood pressure, randomly
any textbook of epidemiology or clinical epide- allocating a total of 40 middle aged hyperten-
miology. Finally, Chapter 4 will review basic sive men to receive either enalapril or a placebo
Overview of the Scientific Metho 21
Study Sample quantitate the probability that the observed

outcome in this study (i.e. the difference seen
Statistical Inference
between the two study groups) could have hap-
pened simply by chance. There are explicit
rules and procedures for how one should prop-
Conclusion about a Population
erly make this determination: the science of
(Association)
statistics. If the results of any study under con-
sideration demonstrate a “statistically signifi-
Biological Inference
cant difference” (i.e. ruling out the probability
of a chance occurrence), then one is said to
Conclusion about Scientific Theory have an association. The process of assessing
(Causation)
whether random variation could have led to a
study’s findings is referred to as statistical
Figure 2.1 Overview of the scientific method.
inference, and represents the major role for sta-
and observing their blood pressure six weeks tistical testing in the scientific method.
later. One might expect to see the blood pres- If there is no statistically significant differ-
sure of the 20 men treated with the active drug ence, then the process in Figure 2.1 stops. If
decrease more than the blood pressure of the there is an association, then one is tempted to
20 men treated with a placebo. In this example, generalize the results of the study even further,
the 40 study subjects would represent the study to state that enalapril is an antihypertensive
sample, theoretically a random sample of drug, in general. This is referred to as scientific
middle-aged hypertensive men. In reality, the or biological inference, and the result is a con-
study sample is almost never a true random clusion about causation, that the drug really
sample of the underlying target population, does lower blood pressure in a population of
because it is logistically impossible to identify treated patients. To draw this type of conclu-
every individual who belongs in the target pop- sion, however, requires one to generalize to
ulation and then randomly choose from among populations other than that included in the
them. However, the study sample is usually study, including types of people who were not
treated as if it were a random sample of the tar- represented in the study sample, such as
get population. women, children, and the elderly. Although it
At this point, one would be tempted to make may be apparent in this example that this is in
a generalization that enalapril lowers blood fact appropriate, that may well not always be
pressure in middle-aged hypertensive men. the case. Unlike statistical inference, there are
However, one must explore whether this obser- no precise quantitative rules for biological
vation could have occurred simply by chance, inference. Rather, one needs to examine the
i.e. due to random variation. If the observed data at hand in light of all other relevant data
outcome in the study was simply a chance in the rest of the scientific literature and make
occurrence, then the same observation might a subjective judgment. To assist in making that
not have been seen if one had chosen a differ- judgment, however, one can use the “Criteria
ent sample of 40 study subjects. Perhaps more for the Causal Nature of an Association,”
importantly, it might not exist if one were able described below. First, however, we will place
to study the entire theoretical population of all causal associations into a proper perspective,
middle-aged hypertensive men. In order to by describing the different types of errors that
evaluate this possibility, one can perform a sta- can be made in performing a study and the dif-
tistical test, which allows an investigator to ferent types of associations that each results in.
22 2 Study Designs Available for Pharmacoepidemiologic Studies
Table 2.1 Types of associations between factors interviewer probes could create an apparent
under study. but false association, which is referred to as
interviewer bias. Another example would be a
1) None (independent)
study of drug-induced birth defects that com-
2) Artifactual (spurious or false) pares children with birth defects to children
a) Chance (unsystematic variation) without birth defects. A mother of a child with
b) Bias (systematic variation) a birth defect, when interviewed about any
3) Indirect (confounded) drugs she took during her pregnancy, may be
4) Causal (direct or true) likely to remember drug ingestion during preg-
nancy with greater accuracy than a mother of a
healthy child, because of the unfortunate expe-
ypes of Errors that one Can
T rience she has undergone. The improved recall
Make in Performing a Study in the mothers of the children with birth
defects may result in false apparent associa-
There are four basic types of associations that tions between drug exposure and birth defects.
can be observed in a study (Table 2.1). The This systematic difference in recall is referred
basic purpose of research is to differentiate to as recall bias.
among them. Note that biases, once present, cannot be
First, of course, one could have no associa- corrected. They represent errors in the study
tion. Second, one could have an artifactual design that can result in incorrect results in the
association, i.e. a spurious or false association. study. It is important to note that a statistically
This can occur by either of two mechanisms: significant result is no protection against a bias;
chance or bias. Chance is unsystematic, or ran- one can have a very precise measurement of an
dom, variation. The purpose of statistical test- incorrect answer! The only protection against
ing in science is to evaluate this, estimating the biases is proper study design (See Chapter 22
probability that the result observed in a study for more discussion about biases in pharma-
could have happened purely by chance. coepidemiologic studies.)
The other possible mechanism for creating Third, one can have an indirect, or con-
an artifactual association is bias. founded, association. A confounding variable,
Epidemiologists’ use of the term bias is differ- or confounder, is a variable, other than the risk
ent from that of the lay public. To an epidemi- factor and other than the outcome under study,
ologist, bias is systematic variation, a consistent which is related independently to both the risk
manner in which two study groups are treated factor and the outcome and which may create
or evaluated differently. This consistent differ- an apparent association or mask a real one. For
ence can create an apparent association where example, a study of risk factors for lung cancer
one actually does not exist. Of course, it also could find a very strong association between
can mask a true association. having yellow fingertips and developing lung
There are many different types of potential cancer. This is obviously not a causal associa-
biases. For example, consider an interview tion, but an indirect association, confounded
study in which the research assistant is aware by cigarette smoking. Specifically, cigarette
of the investigator’s hypothesis. Attempting to smoking causes both yellow fingertips and
please the boss, the research assistant might lung cancer. Although this example is trans-
probe more carefully during interviews with parent, most examples of confounding are not.
one study group than during interviews with In designing a study, one must consider every
the other. This difference in how carefully the variable that can be associated with the risk
Criteria for the Causal Nature of an Associatio 23
Table 2.2 Approaches to controlling confounding. Table 2.3 Criteria for the causal nature of an
association.
1) Random allocation
2) Subject selection 1) Coherence with existing information
(biological plausibility)
a) Exclusion
2) Consistency of the association
b) Matching
3) Time sequence
3) Data analysis
4) Specificity of the association
a) Stratification
5) Strength of the association
b) Mathematical modeling
a) Quantitative strength
b) Dose–response relationship
factor under study or the outcome variable c) Study design
under study, in order to plan to deal with it as a
potential confounding variable. Preferably, one
will be able to specifically control for the vari- causal association. Essentially, the more crite-
able, using one of the techniques listed in ria that are present, the more likely it is that an
Table 2.2. (See Chapter 22 for more discussion association is a causal association. The fewer
about confounding in pharmacoepidemiologic criteria that are met, the less likely it is that an
studies.) association is a causal association. Each will be
Fourth, and finally, there are true, causal discussed in turn.
associations. The first criterion listed in Table 2.3 is coher-
Thus, there are three possible types of errors ence with existing information or biological
that can be produced in a study: random error, plausibility. This refers to whether the associa-
bias, and confounding. The probability of ran- tion makes sense, in light of other types of
dom error can be quantitated using statistics. information available in the literature. These
Bias needs to be prevented by designing the other types of information could include data
study properly. Confounding can be controlled from other human studies, data from studies of
either in the design of the study or in its analy- other related questions, data from animal stud-
sis. If all three types of errors can be excluded, ies, or data from in vitro studies, as well as sci-
then one is left with a true, causal association. entific or pathophysiologic theory. To use the
example provided above, it clearly was not bio-
logically plausible that yellow fingertips could
riteria for the Causal
C cause lung cancer, and this provided the clue
Nature of an Association that confounding was present. Using the exam-
ple of the association between cigarettes and
The “Criteria for the causal nature of an asso- lung cancer, cigarette smoke is a known car-
ciation” were first put forth by Sir Austin cinogen, based on animal data. In humans, it is
Bradford Hill, but have been described in vari- known to cause cancers of the head and neck,
ous forms since, each with some modification. the pancreas, and the bladder. Cigarette smoke
Probably the best known description of them also goes down into the lungs, directly expos-
was in the first Surgeon General’s Report on ing the tissues in question. Thus, it certainly is
Smoking and Health, published in 1964. These biologically plausible that cigarettes could
criteria are presented in Table 2.3, in no par- cause lung cancer. It is much more reassuring
ticular order. No one of them is absolutely nec- if an association found in a particular study
essary for an association to be a causal makes sense, based on previously available
association. Analogously, no one of them is information, and this makes one more com-
sufficient for an association to be considered a fortable that it might be a causal association.
Clearly, however, one could not require that virus, but even in this example, not everyone
this criterion always be met, or one would who becomes infected with the measles virus
never have a major breakthrough in science. develops clinical measles. Certainly, not every-
The second criterion listed in Table 2.3 is the one who smokes develops lung cancer, and not
consistency of the association. A hallmark of everyone who develops lung cancer was a
science is reproducibility: if a finding is real, smoker. This is one of the major points the
one should be able to reproduce it in a different tobacco industry stresses when it attempts to
setting. This could include different geographic make the claim that cigarette smoking has not
settings, different study designs, different pop- been proven to cause lung cancer. Some
ulations, etc. For example, in the case of ciga- authors even omit this as a criterion, as it is so
rettes and lung cancer, the association has now rarely met. When it is met, however, it provides
been reproduced in many different studies, in extremely strong support for a conclusion that
different geographic locations, using different an association is causal.
study designs. The need for reproducibility is The fifth criterion listed in Table 2.3 is the
such that one should never believe a finding strength of the association. This includes three
reported only once: there may have been an concepts: its quantitative strength, dose–
error committed in the study, which is not response, and the study design. Each will be
apparent to either the investigator or the discussed in turn.
reader. The quantitative strength of an association
The third criterion listed is that of time refers to the effect size. To evaluate this, one
sequence – a cause must precede an effect. asks whether the magnitude of the observed
Although this may seem obvious, there are difference between the two study groups is
study designs from which this cannot be deter- large. A quantitatively large association can
mined. For example, if one were to perform a only be created by a causal association or a
survey in a classroom of 200 medical students, large error, which should be apparent in evalu-
asking each if he or she were currently taking ating the methods of a study. A quantitatively
diazepam and also whether he or she were small association may still be causal, but it
anxious, one would find a strong association could be created by a subtle error, which would
between the use of diazepam and anxiety, but not be apparent in evaluating the study.
this does not mean that diazepam causes anxi- Conventionally, epidemiologists consider an
ety! Although this is obvious, as it is not a bio- association with a relative risk of less than 2.0
logically plausible interpretation, one cannot a weak association. Certainly, the association
differentiate from this type of cross-sectional between cigarette smoking and lung cancer is a
study which variable came first and which strong association: studies show relative risks
came second. In the example of cigarettes and ranging between 10.0 and 30.0.
lung cancer, obviously the cigarette smoking A dose–response relationship is an extremely
usually precedes the lung cancer, as a patient important and commonly used concept in clin-
would not survive long enough to smoke much ical pharmacology and is used similarly in epi-
if the opposite were the case. demiology. A dose–response relationship exists
The fourth criterion listed in Table 2.3 is when an increase in the intensity of an expo-
specificity. This refers to the question of sure results in an increased risk of the disease
whether the cause ever occurs without the pre- under study. Equivalent to this is a duration–
sumed effect and whether the effect ever response relationship, which exists when a
occurs without the presumed cause. This crite- longer exposure causes an increased risk of the
rion is almost never met in biology, with the disease. The presence of either a dose–
occasional exception of infectious diseases. response relationship or a duration–response
Measles never occurs without the measles relationship strongly implies that an
Criteria for the Causal Nature of an Associatio 25
a ssociation is, in fact, a causal association. harder to perform, but are progressively
Certainly in the example of cigarette smoking more convincing. In other words, associa-
and lung cancer, it has been shown repeatedly tions shown by studies using designs at the
that an increase in either the number of ciga- top of the list are more likely to be causal
rettes smoked each day or in the number of associations than associations shown by
years of smoking increases the risk of develop- studies using designs at the bottom of the
ing lung cancer. list. The association between cigarette smok-
Finally, study design refers to two concepts: ing and lung cancer has been reproduced in
whether the study was well designed, and multiple well-designed studies, using analy-
which study design was used in the studies ses of secular trends, case–control studies,
in question. The former refers to whether the and cohort studies. However, it has not been
study was subject to one of the three errors shown using a randomized clinical trial,
described earlier in this chapter, namely ran- which is the “cadillac” of study designs, as
dom error, bias, and confounding. Table 2.4 will be discussed below. This is the other
presents the study designs typically used for major defense used by the tobacco industry.
epidemiologic studies, or in fact for any clin- Of course, it would not be ethical or logisti-
ical studies. They are organized in a hierar- cally feasible to randomly allocate individu-
chical fashion. As one advances from the als to smoke or not to smoke and expect this
designs at the bottom of the table to those at to be followed for 20 years to observe the out-
the top of the table, studies get progressively come in each group.
Table 2.4 Advantages and disadvantages of epidemiologic study designs.
Study Design Advantages Disadvantages
Randomized clinical trial Most convincing design Most expensive

(Experimental study)
Only design which controls for unknown Artificial
or unmeasurable confounders Logistically most difficult
Ethical objections
Cohort study Can study multiple outcomes Possibly biased outcome data
Can study uncommon exposures More expensive
Selection bias less likely If done prospectively, may
take years to complete
Unbiased exposure data
Incidence data available
Case–control study Can study multiple exposures Control selection problematic
Can study uncommon diseases Possibly biased exposure data
Logistically easier and faster
Less expensive
Analyses of secular trends Can provide rapid answers No control of confounding
Case series Easy quantitation of incidence No control group, so cannot
be used for hypothesis testing
Case reports Cheap and easy method for generating Cannot be used for
hypotheses hypothesis testing
The issue of causation is discussed more in An example would be the ability of penicillin
Chapter 7 as it relates to the process of sponta- to cure streptococcal endocarditis, a disease
neous reporting of adverse drug reactions, and that is nearly uniformly fatal in the absence of
in Chapter 14 as it relates to assessing causa- treatment. Case reports can be particularly
tion from case reports. useful to document causation when the treat-
ment causes a change in disease course which
is reversible, such that the patient returns to
Epidemiologic Study his or her untreated state when the exposure is
Designs withdrawn, can be treated again, and when the
change returns upon repeat treatment.
In order to clarify the concept of study design Consider a patient who is suffering from an
further, each of the designs in Table 2.4 will be overdose of methadone, a long-acting narcotic,
discussed in turn, starting at the bottom of the and is comatose. If this patient is then treated
list and working upwards. with naloxone, a narcotic antagonist, and
immediately awakens, this would be very sug-
gestive that the drug indeed is efficacious as a
Case Reports
narcotic antagonist. As the naloxone wears off
Case reports are simply reports of events the patient would become comatose again, and
observed in single patients. As used in phar- then if he or she were given another dose of
macoepidemiology, a case report describes a naloxone the patient would awaken again.
single patient who was exposed to a drug and This, especially if repeated a few times, would
experiences a particular, usually adverse, out- represent strong evidence that the drug is
come. For example, one might see a published indeed effective as a narcotic antagonist. This
case report about a young woman who was type of challenge–rechallenge situation is rela-
taking oral contraceptives and who suffered a tively uncommon, however, as physicians gen-
pulmonary embolism. erally will avoid exposing a patient to a drug if
Case reports are useful for raising hypothe- the patient experienced an adverse reaction to
ses about drug effects, to be tested with more it in the past. This issue is discussed in more
rigorous study designs. However, in a case detail in Chapters 7 and 14.
report one cannot know if the patient reported
is either typical of those with the exposure or
Case Series
typical of those with the disease. Certainly, one
cannot usually determine whether the adverse Case series are collections of patients, all of
outcome was due to the drug exposure or whom have a single exposure, whose clinical
would have happened anyway. As such, it is outcomes are then evaluated and described.
very rare that a case report can be used to make Often they are from a single hospital or medi-
a statement about causation. One exception to cal practice. Alternatively, case series can be
this would be when the outcome is so rare and collections of patients with a single outcome,
so characteristic of the exposure that one looking at their antecedent exposures. For
knows that it was likely to be due to the expo- example, one might observe 100 consecutive
sure, even if the history of exposure were women under the age of 50 who suffer from a
unclear. An example of this is clear cell vaginal pulmonary embolism, and note that 30 of
adenocarcinoma occurring in young women them had been taking oral contraceptives.
exposed in utero to diethylstilbestrol. Another After drug marketing, case series are most
exception would be when the disease course is useful for two related purposes. First, they can
very predictable and the treatment causes a be useful for quantifying the incidence of an
clearly apparent change in this disease course. adverse reaction. Second, they can be useful for
Epidemiologic Study Design 27
being certain that any particular adverse effect Analyses of Secular Trends
of concern does not occur in a population which
Analyses of secular trends, also called “ecologi-
is larger than that studied prior to drug market-
cal studies,” examine trends in an exposure that
ing. The so-called “Phase IV” postmarketing
is a presumed cause and trends in a disease that
surveillance study of prazosin was conducted
is a presumed effect and test whether the trends
for the former reason, to quantitate the inci-
coincide. These trends can be examined over
dence of first dose syncope from prazosin. The
time or across geographic boundaries. In other
“Phase IV” postmarketing surveillance study of
words, one could analyze data from a single
cimetidine was conducted for the latter reason.
region and examine how the trend changes
Metiamide was an H-2 blocker, which was with-
over time, or one could analyze data from a sin-
drawn after marketing outside the US because it
gle time period and compare how the data dif-
caused agranulocytosis. Since cimetidine is
fer from region to region or country to country.
chemically related to metiamide there was a
Vital statistics are often used for these studies.
concern that cimetidine might also cause agran-
As an example, one might look at sales data for
ulocytosis. In both examples, the manufacturer
oral contraceptives and compare them to death
asked its sales representatives to recruit physi-
rates from venous thromboembolism, using
cians to participate in the study. Each participat-
recorded vital statistics. When such a study was
ing physician then enrolled the next series of
actually performed, mortality rates from venous
patients for whom the drug was prescribed.
thromboembolism were seen to increase in par-
In this type of study, one can be more certain
allel with increasing oral contraceptive sales,
that the patients are probably typical of those
but only in women of reproductive age, not in
with the exposure or with the disease, depend-
older women or in men of any age.
ing on the focus of the study. However, in the
absence of a control group, one cannot be cer- Analyses of secular trends are useful for rap-
tain which features in the description of the idly providing evidence for or against a hypoth-
patients are unique to the exposure, or out- esis. However, these studies lack data on
come. As an example, one might have a case individuals; they utilize only aggregated group
series from a particular hospital of 100 indi- data (e.g. annual sales data in a given geo-
viduals with a certain disease, and note that all graphic region in relation to annual cause-
were men over the age of 60. This might lead specific mortality in the same region). As such,
one to conclude that this disease seems to be they are unable to control for confounding
associated with being a man over the age of 60. variables. Thus, among exposures whose
However, it would be clear that this would be trends coincide with that of the disease, analy-
an incorrect conclusion once one noted that ses of secular trends are unable to differentiate
the hospital this case series was drawn from which factor is likely to be the true cause. For
was a Veterans Administration hospital, where example, lung cancer mortality rates in the US
most patients are men over the age of 60. In the have been increasing in women, such that lung
previous example of pulmonary embolism and cancer is now the leading cause of cancer mor-
oral contraceptives, 30% of the women with tality in women. This is certainly consistent
pulmonary embolism had been using oral con- with the increasing rates of cigarette smoking
traceptives. However, this information is not observed in women until the mid-1960s, and
sufficient to determine whether this is higher, so appears to be supportive of the association
the same as, or even lower than would have between cigarette smoking and lung cancer.
been expected. For this reason, case series are However, it would also be consistent with an
also not very useful in determining causation, association between certain occupational
but provide clinical descriptions of a disease or exposures and lung cancer, as more women in
of patients who receive an exposure. the US are now working outside the home.
Case–Control Studies generally confirm their results. As such, the

case–control design is a very useful approach
Case–control studies are studies that compare
for pharmacoepidemiologic studies.
cases with a disease to controls without the dis-
ease, looking for differences in antecedent
exposures. As an example, one could select Cohort Studies
cases of young women with venous thrombo-
Cohort studies are studies that identify subsets
embolism and compare them to controls with-
of a defined population and follow them over
out venous thromboembolism, looking for
time, looking for differences in their outcome.
differences in antecedent oral contraceptive
Cohort studies are generally used to compare
use. Several such studies have been performed,
exposed patients to unexposed patients,
generally demonstrating a strong association
although they can also be used to compare one
between the use of oral contraceptives and
exposure to another. For example, one could
venous thromboembolism.
compare women of reproductive age who use
Case–control studies can be particularly use- oral contraceptives to users of other contracep-
ful when one wants to study multiple possible tive methods, looking for the differences in the
causes of a single disease, as one can use the frequency of venous thromboembolism. When
same cases and controls to examine any num- such studies were performed, they in fact con-
ber of exposures as potential risk factors. This firmed the relationship between oral contra-
design is also particularly useful when one is ceptives and thromboembolism, which had
studying a relatively rare disease, as it guaran- been noted using analyses of secular trends
tees a sufficient number of cases with the dis- and case–control studies. Cohort studies can
ease. Using case–control studies, one can study be performed either prospectively, that is
rare diseases with markedly smaller sample simultaneous with the events under study, or
sizes than those needed for cohort studies (see retrospectively, that is after the outcomes
Chapter 3). For example, the classic study of under study had already occurred, by recreat-
diethylstilbestrol and clear cell vaginal adeno- ing those past events using medical records,
carcinoma required only 8 cases and 40 con- questionnaires, or interviews.
trols, rather than the many thousands of The major difference between cohort and
exposed subjects that would have been required case–control studies is the basis upon which
for a cohort study of this question. patients are recruited into the study (see
Case–control studies generally obtain their Figure 2.2). Patients are recruited into
information on exposures retrospectively, i.e.
by recreating events that happened in the past.
Case-Control Studies
Information on past exposure to potential risk
factors is generally obtained by abstracting
medical records or by administering question- Disease
naires or interviews. As such, case–control Present Absent
studies are subject to limitations in the validity (cases) (controls)
Cohort Studies
of retrospectively collected exposure informa- Present

Risk Factor
A B
(exposed)
tion. In addition, the proper selection of con-
trols can be a challenging task, and appropriate Absent
C D
control selection can lead to a selection bias, (not exposed)
which may lead to incorrect conclusions.
Nevertheless, when case–control studies are Figure 2.2 Cohort and case–control studies
provide similar information, but approach data
done well, subsequent well-done cohort stud- collection from opposite directions. Source:
ies or randomized clinical trials, if any, will Reprinted with permission from Strom (1986).
Epidemiologic Study Design 29
case–control studies based on the presence or exposure seems to protect against the disease.
absence of a disease, and their antecedent A relative risk of 1.0 means that exposed sub-
exposures are then studied. Patients are jects and unexposed subjects have the same
recruited into cohort studies based on the pres- risk of developing the disease, or that the expo-
ence or absence of an exposure, and their sub- sure and the disease appear unrelated.
sequent disease course is then studied. One can calculate a relative risk directly
Cohort studies have the major advantage of from the results of a cohort study. However, in
being free of the major problem that plagues a case–control study one cannot determine the
case–control studies: the difficult process of size of either the exposed population or the
selecting an undiseased control group. In addi- unexposed population that the diseased cases
tion, prospective cohort studies are free of the and undiseased controls were drawn from. The
problem of the questionable validity of retro- results of a case–control study do not provide
spectively collected data. For these reasons, an information on the incidence rates of the dis-
association demonstrated by a cohort study is ease in exposed and unexposed individuals.
more likely to be a causal association than Therefore, relative risks cannot be calculated
one demonstrated by a case–control study. directly from a case–control study. Instead, in
Furthermore, cohort studies are particularly reporting the results of a case–control study
useful when one is studying multiple possible one generally reports the odds ratio, which is a
outcomes from a single exposure, especially a close estimate of the relative risk when the dis-
relatively uncommon exposure. Thus, they are ease under study is relatively rare. Since case–
particularly useful in postmarketing drug sur- control studies are generally used to study rare
veillance studies, which are looking at any pos- diseases, there is generally very close agree-
sible effect of a newly marketed drug. However, ment between the odds ratio and the relative
cohort studies can require extremely large risk, and the results from case–control studies
sample sizes to study relatively uncommon are often loosely referred to as relative risks,
outcomes (see Chapter 3). In addition, pro- although they are in fact odds ratios.
spective cohort studies can require a prolonged Both relative risks and odds ratios can be
time period to study delayed drug effects. reported with p-values. These p-values allow
one to determine if the relative risk is statisti-
cally significantly different from 1.0, that is
Analysis of Case–Control
whether the differences between the two study
and Cohort Studies
groups are likely to be due to random variation
As can be seen in Figure 2.2, both case–control or are likely to represent real associations.
and cohort studies are intended to provide the Alternatively, and probably preferably, rela-
same basic information; the difference is how tive risks and odds ratios can be reported with
this information is collected. The key statistic confidence intervals, which are an indication of
reported from these studies is the relative risk. the range of relative risks within which the
The relative risk is the ratio of the incidence true relative risk for the entire theoretical pop-
rate of an outcome in the exposed group to the ulation is most likely to lie. As an approxima-
incidence rate of the outcome in the unex- tion, a 95% confidence interval around a
posed group. A relative risk of greater than relative risk means that we can be 95% confi-
1.0 means that exposed subjects have a greater dent that the true relative risk lies in the range
risk of the disease under study than unexposed between the lower and upper limits of this
subjects, or that the exposure appears to cause interval. If a 95% confidence interval around a
the disease. A relative risk less than 1.0 means relative risk excludes 1.0, then the finding is
that exposed subjects have a lower risk of the statistically significant with a p-value of less
disease than unexposed subjects, or that the than 0.05. A confidence interval provides
much more information than a p-value, how- As with relative risks, excess risks cannot be
ever. As an example, a study that yields a rela- calculated from case–control studies, as inci-
tive risk (95% confidence interval) of 1.0 dence rates are not available. As with the other
(0.9–1.1) is clearly showing that an association statistics, p-values can be calculated to deter-
is very unlikely. A study that yields a relative mine whether the differences between the two
risk (95% confidence interval) of 1.0 (0.1–100) study groups could have occurred just by
provides little evidence for or against an asso- chance. Confidence intervals can be calculated
ciation. Yet, both could be reported as a relative around excess risks as well, and would be
risk of 1.0 and a p-value greater than 0.05. As interpreted analogously.
another example, a study that yields a relative
risk (95% confidence interval) of 10.0 (9.8–
Randomized Clinical Trials
10.2) precisely quantifies a 10-fold increase in
risk that is also statistically significant. A study Finally, experimental studies are studies in
that yields a relative risk (95% confidence which the investigator controls the therapy
interval) of 10.0 (1.1–100) says little, other than that is to be received by each participant.
an increased risk is likely. Yet, both could be Generally, an investigator uses that control to
reported as a relative risk of 10.0 (p < 0.05). As randomly allocate patients between or among
a final example, a study yielding a relative risk the study groups, performing a randomized
(95% confidence interval) of 3.0 (0.98–5.0) is clinical trial. For example, one could theoreti-
strongly suggestive of an association, whereas cally randomly allocate sexually active women
a study reporting a relative risk (95% confi- to use either oral contraceptives or no contra-
dence interval) of 3.0 (0.1–30) would not be. ceptive, examining whether they differ in their
Yet, both could be reported as a relative risk of incidence of subsequent venous thromboem-
3.0 (p > 0.05). bolism. The major strength of this approach is
Finally, another statistic that one can calcu- random assignment, which is the only way to
late from a cohort study is the excess risk, also make it likely that the study groups are compa-
called the risk difference or, sometimes, the rable in potential confounding variables that
attributable risk. Whereas the relative risk is are either unknown or unmeasurable. For this
the ratio of the incidence rates in the exposed reason, associations demonstrated in rand-
group versus the unexposed groups, the excess omized clinical trials are more likely to be
risk is the arithmetic difference between the causal associations than those demonstrated
incidence rates. The relative risk is more using one of the other study designs reviewed
important in considering questions of causa- above.
tion. The excess risk is more important in con- However, even randomized clinical trials
sidering the public health impact of an are not without their problems. The rand-
association, as it represents the increased rate omized clinical trial outlined above, allocating
of disease due to the exposure. For example, women to receive contraceptives or no contra-
oral contraceptives are strongly associated ceptives, demonstrates the major potential
with the development of myocardial infarction problems inherent in the use of this study
in young women. However, the risk of myocar- design. It would obviously be impossible to
dial infarction in nonsmoking women in their perform, ethically and logistically. In addition,
20s is so low, that even a fivefold increase in randomized clinical trials are expensive and
that risk would still not be of public health artificial. Inasmuch as they have already been
importance. In contrast, women in their 40s performed prior to marketing to demonstrate
are at higher risk, especially if they are ciga- each drug’s efficacy, they tend to be unneces-
rette smokers as well. Thus, oral contraceptives sary after marketing. They are likely to be used
should not be as readily used in these women. in pharmacoepidemiologic studies mainly for
Discussio 31
supplementary studies of drug efficacy. Table 2.5 Epidemiologic study designs.

However, they remain the “gold standard” by
which the other designs must be judged. A) Classified by how subjects are recruited into
the study
Indeed, with the publication of the results
from the Women’s Health Initiative indicating 1) Case–control (case-history, case-referent,
retrospective, trohoc) studies
that combination hormone replacement ther-
2) Cohort (follow-up, prospective) studies
apy causes an increased risk of myocardial
infarction rather than a decreased risk, there a) Experimental studies (clinical trials,
intervention studies)
has been increased concern about reliance
solely on nonexperimental methods to study B) Classified by how data are collected for the
study
drug safety after marketing, and we are begin-
1) Retrospective (historical, nonconcurrent,
ning to see the use of massive randomized
retrolective) studies
clinical trials as part of postmarketing
2) Prospective (prolective) studies
surveillance.
3) Cross-sectional studies
Discussion viewed as a subset of cohort studies, a type of

cohort study in which the investigator controls
Thus, a series of different study designs are the allocation of treatment, rather than simply
available (Table 2.4), each with respective observing ongoing medical care. From the per-
advantages and disadvantages. Case reports, spective of timing, data can be collected pro-
case series, analyses of secular trends, case– spectively, that is simultaneously with the
control studies, and cohort studies have been events under study, or retrospectively, that is
referred to collectively as observational study after the events under study had already devel-
designs or nonexperimental study designs, in oped. In the latter situation, one re-creates
order to differentiate them from experimental events that happened in the past using medical
studies. In nonexperimental study designs the records, questionnaires, or interviews. Data
investigator does not control the therapy, but can also be collected using cross-sectional stud-
simply observes and evaluates the results of ies, studies that have no time sense, as they
ongoing medical care. Case reports, case series, examine only one point in time. In principle,
and analyses of secular trends have also been either cohort or case–control studies can be
referred to as descriptive studies. Case–control performed using any of these time frames,
studies, cohort studies, and randomized clini- although prospective case–control studies are
cal trials all have control groups, and have unusual. Randomized clinical trials must be
been referred to as analytic studies. The ana- prospective, as this is the only way an investi-
lytic study designs can be classified in two gator can control the therapy received.
major ways, by how subjects are selected into The terms presented in this chapter, which
the study and by how data are collected for the are those that will be used throughout the
study (see Table 2.5). From the perspective of book, are probably the terms used by a major-
how subjects are recruited into the study, case– ity of epidemiologists. Unfortunately, however,
control studies can be contrasted with cohort other terms have been used for most of these
studies. Specifically, case–control studies select study designs, as well. Table 2.5 also presents
subjects into the study based on the presence several of the synonyms that have been used in
or absence of a disease, while cohort studies the medical literature. The same term is some-
select subjects into the study based on the pres- times used by different authors to describe dif-
ence or absence of an exposure. From this ferent concepts. For example, in this book we
perspective, randomized clinical trials can be are reserving the use of the terms “retrospective
study” and “prospective study” to refer to a two long-term, large-scale cohort studies. This
time sense. As is apparent from Table 2.5, how- question might even be worth the investment
ever, in the past some authors used the term of a randomized clinical trial, except it would
“retrospective study” to refer to a case–control not be feasible or ethical. In contrast, when
study and used the term “prospective study” to thalidomide was marketed, it was not a major
refer to a cohort study, confusing the two con- breakthrough; other hypnotics were already
cepts inherent in the classification schemes available. Case reports of phocomelia in
presented in the table. Other authors use the exposed patients were followed by case–con-
term “retrospective study” to refer to any non- trol studies and analyses of secular trends.
experimental study, while others appear to use Inasmuch as the adverse effect was so terrible
the term to refer to any study they do not like, and the drug was not of unique importance,
as a term of derision! Unfortunately, when the drug was then withdrawn, without the
reading a scientific paper, there is no way of delay that would have been necessary if cohort
determining which usage the author intended. studies and/or randomized clinical trials had
What is more important than the terminology, been awaited. Ultimately, a retrospective
however, are the concepts underlying the cohort study was performed, comparing those
terms. Understanding these concepts, the exposed during the critical time period to those
reader can choose to use whatever terminology exposed at other times.
he or she is comfortable with. In general, however, clinical, regulatory,
commercial, and legal decisions need to be
made based on the best evidence available at
Conclusion the time of the decision. To quote Sir Austin
Bradford Hill (1965):
From the material presented in this chapter, it
is hopefully now apparent that each study All scientific work is incomplete–
design has an appropriate role in scientific pro- whether it be observational or experi-
gress. In general, science proceeds from the mental. All scientific work is liable to be
bottom of Table 2.4 upward, from case reports upset or modified by advancing knowl-
and case series that are useful for suggesting an edge. That does not confer upon us a
association, to analyses of trends and case– freedom to ignore the knowledge we
control studies that are useful for exploring already have, or to postpone the action
these associations. Finally, if a study question that it appears to demand at a given
warrants the investment and can tolerate the time.
delay until results become available, then Who knows, asked Robert Browning,
cohort studies and randomized clinical trials but the world may end tonight? True,
can be undertaken to assess these associations but on available evidence most of us
more definitively. make ready to commute on the 8:30 next
For example, regarding the question of day.
whether oral contraceptives cause venous
thromboembolism, an association was first
suggested by case reports and case series, then
was explored in more detail by analyses of Key Points
trends and a series of case–control studies.
Later, because of the importance of oral con- ●● Many different types of potential biases can
traceptives, the number of women using them, create artifactual associations in a scientific
and the fact that users were predominantly study. Among them are: interviewer bias,
healthy women, the investment was made in recall bias, and confounding.
Further Reading 33
●● Four basic types of association can be ●● Study design options, in hierarchical order
observed in studies that examine whether of progressively harder to perform but more
there is an association between an exposure convincing, are: case reports, case series,
and an outcome: no association, artifactual analyses of secular trends, case–control
association (from chance or bias), indirect studies, retrospective cohort studies, pro-
association (from confounding), or true spective cohort studies, and randomized
association. clinical trials.
●● A series of criteria can be used to assess the ●● Associations between an exposure and an
causal nature of an association, to assist in outcome are reported with relative risk ratios
making a subjective judgment about whether (in cohort studies), odds ratios (in case–con-
a given association is likely to be causal. trol studies), confidence intervals, and
These are: biological plausibility, consist- p-values. Sometimes also as attributable
ency, time sequence, specificity, and quanti- (excess) risk.
tative strength.
Further Reading
Bassetti, W.H.C. and Woodward, M. (2004). Kelsey, J.L., Whittemore, A.S., and Evans, A.S.
Epidemiology: Study Design and Data Analysis, (1996). Methods in Observational
2e. Boca Raton, Florida: Chapman & Hall/ Epidemiology, 2e. New York: Oxford
CRC. University Press.
Bhopal, R.S. (2008). Concepts of Epidemiology: Lilienfeld, D.E. and Stolley, P. (1994).
Integrating the Ideas, Theories, Principles and Foundations of Epidemiology, 3e. New York:
Methods of Epidemiology, 2e. New York: Oxford University Press.
Oxford University Press. MacMahon, B. (1997). Epidemiology: Principles
Fletcher, R.H. and Fletcher, S.W. (2005). Clinical and Methods, 2e. Hagerstown MD:
Epidemiology: The Essentials, 4e. Lippincott Lippincott-Raven.
Williams & Wilkins. Mausner, J.S. and Kramer, S. (1985).
Friedman, G. (2003). Primer of Epidemiology, 5e. Epidemiology: An Introductory Text, 2e.
New York: McGraw Hill. Philadelphia, PA: Saunders.
Gordis, L. (2009). Epidemiology, 4e. Philadelphia, Rothman, K.J. (2002). Epidemiology: An
PA: Saunders. Introduction. New York: Oxford University
Hennekens, C.H. and Buring, J.E. (1987). Press.
Epidemiology in Medicine. Boston, MA: Little Rothman, K.J., Greenland, S., and Lash, T.L.
Brown. (2008). Modern Epidemiology, 3e.
Hill, A.B. (1965). The environment and disease: Philadelphia, PA: Lippincott Williams &
association or causation? Proc. Royal Soc. Med. Wilkins.
58: 295–300. Sackett, D.L. (1979). Bias in analytic research.
Hulley, S.B., Cummings, S.R., Browner, W.S. J. Chronic Dis. 32: 51–63.
et al. (2006). Designing Clinical Research: An Sackett, D.L., Haynes, R.B., and Tugwell, P.
Epidemiologic Approach, 3e. Baltimore, MD: (1991). Clinical Epidemiology: A Basic Science
Lippincott Williams & Wilkins. for Clinical Medicine, 2e. Boston, MA: Little
Katz, D.L. (2001). Clinical Epidemiology and Brown.
Evidence-based Medicine: Fundamental Strom, B.L. (1986). Medical databases in
Principles of Clinical Reasoning and Research. post-marketing drug surveillance. Trends
Thousand Oaks, CA: Sage Publications. Pharmacol. Sci. 7: 377–380.
Szklo, M. and Nieto, F.J. (2006). Epidemiology: Weiss, N.S. (1996). Clinical Epidemiology: The
Beyond the Basics. Sudbury, MA: Jones & Study of the Outcome of Illness, 2e. New York:
Bartlett. Oxford University Press.
US Public Health Service (1964). Smoking and Weiss, N.S., Koepsall, T., and Koepsell, T.D.
Health. Report of the Advisory Committee to the (2004). Epidemiologic Methods: Studying the
Surgeon General of the Public Health Service, 20. Occurrence of Illness. New York: Oxford
Washington DC: Government Printing Office. University Press.
35
Sample Size Considerations for Pharmacoepidemiologic Studies

Brian L. Strom
Introduction can examine the resulting confidence intervals

in order to determine the smallest differences
Chapter 1 pointed out that between 500 and between the two study groups that the study
3000 subjects are usually exposed to a drug had sufficient sample size to exclude.
prior to marketing, in order to be 95% certain Alternatively, one can approach the question
of detecting adverse effects that occur in in a manner similar to the way one would
between one and six in a thousand exposed approach it if one were planning the study de
individuals. While this seems like a reasonable novo. Nomograms can be used to assist a
goal, it poses some important problems that reader in interpreting negative clinical trials in
must be taken into account when planning this way.
pharmacoepidemiologic studies. Specifically, In contrast, in this chapter we will discuss in
such studies must generally include a suffi- more detail how to determine a proper study
cient number of subjects to add significantly to sample size, from the perspective of one who is
the premarketing experience, and this require- designing a study de novo. Specifically, we will
ment for large sample sizes raises logistical begin by discussing how one calculates the
obstacles to cost-effective studies. This central minimum sample size necessary for a pharma-
special need for large sample sizes is what has coepidemiologic study, to avoid the problem of
led to the innovative approaches to collecting a study with a sample size that is too small. We
pharmacoepidemiologic data that are will first present the approach for cohort stud-
described in Section II of this book Sources of ies, then for case–control studies, and then for
Pharmacoepidemiology Data. case series. For each design, one or more tables
The approach to considering the implica- will be presented to assist the reader in carry-
tions of a study’s sample size is somewhat dif- ing out these calculations.
ferent depending on whether a study is already
completed or is being planned. After a study is
completed, if a real finding was statistically Sample Size Calculations
significant, then the study had a sufficient for Cohort Studies
sample size to detect it, by definition. If a find-
ing was not statistically significant, then one The sample size required for a cohort study
can use either of two approaches. First, one depends on what you are expecting from the
36 3 Sample Size Considerations for Pharmacoepidemiologic Studies
study. To calculate sample sizes for a cohort decreasing it, as expected? If the investigator’s
study, one needs to specify five variables (see response to this would be: “Boy, what a sur-
Table 3.1). prise, but I believe it,” then a two-tailed test
The first variable to specify is the alpha (α) or should be performed. If the investigator’s
type I error that one is willing to tolerate in the response would be: “I don’t believe it, and I will
study. Type I error is the probability of conclud- interpret this simply as a study that does not
ing there is a difference between the groups show the expected decrease in coronary artery
being compared when in fact a difference does disease in the group treated with the study
not exist. Using diagnostic tests as an analogy, a drug,” then a one-tailed test should be per-
type I error is a false positive study finding. The formed. The more conservative option is the
more tolerant one is willing to be of type I error, two-tailed test, assuming that the results could
the smaller the sample size required. The less turn out in either direction. This is the option
tolerant one is willing to be of type I error, the usually, although not always, used.
smaller one would set alpha, and the larger the The second variable that needs to be speci-
sample size that would be required. fied to calculate a sample size for a cohort
Conventionally the alpha is set at 0.05, although study is the beta (β) or type II error that one is
this certainly does not have to be the case. Note willing to tolerate in the study. A type II error is
that alpha needs to be specified as either one- the probability of concluding there is no differ-
tailed or two-tailed. If only one of the study ence between the groups being compared
groups could conceivably be more likely to when in fact a difference does exist. In other
develop the disease and one is interested in words, a type II error is the probability of miss-
detecting this result only, then one would spec- ing a real difference. Using diagnostic tests as
ify alpha to be one-tailed. If either of the study an analogy, a type II error is a false negative
groups may be likely to develop the disease, and study finding. The complement of beta is the
either result would be of interest, then one power of a study, i.e. the probability of detect-
would specify alpha to be two-tailed. To decide ing a difference if a difference really exists.
whether alpha should be one-tailed or two- Power is calculated as (1-β). Again, the more
tailed, an investigator should consider what his tolerant one is willing to be of Type II errors,
or her reaction would be to a result that is sta- i.e. the higher the beta, the smaller the sample
tistically significant in a direction opposite to size required. The beta is conventionally set at
the one expected. For example, what if one 0.1 (i.e. 90% power) or 0.2 (i.e. 80% power),
observed that a drug increased the frequency of although again this need not be the case. Beta
dying from coronary artery disease instead of is always one-tailed.
Table 3.1 Information needed to calculate a study’s sample size.
For cohort studies For case–control studies
1) Alpha, or type I error, considered tolerable, 1) Alpha, or type I error, considered tolerable, and
and whether it is one-tailed or two-tailed whether it is one-tailed or two tailed
2) Beta, or type II error, considered tolerable 2) Beta, or type II, error considered tolerable
3) Minimum relative risk to be detected 3) Minimum relative risk to be detected
4) Incidence of the disease in the unexposed 4) Prevalence of the exposure in the undiseased
control group control group
5) Ratio of unexposed controls to exposed study 5) Ratio of undiseased controls to diseased study
subjects subjects
Sample Size Calculations for Cohort Studie 37
The third variable one needs to specify in three controls for each exposed subject
order to calculate sample sizes for a cohort increases the power further. However, the
study is the minimum effect size one wants to increment in power achieved by increasing the
be able to detect. For a cohort study, this is ratio of control subjects to exposed subjects
expressed as a relative risk. The smaller the from 2 : 1 to 3 : 1 is smaller than the increment
relative risk that one wants to detect, the larger in power achieved by increasing the ratio from
the sample size required. Note that the relative 1 : 1 to 2 : 1. Each additional increase in the size
risk often used by investigators in this calcula- of the control group increases the power of the
tion is the relative risk the investigator is study further, but with progressively smaller
expecting from the study. This is not correct, as gains in statistical power. Thus, there is rarely a
it will lead to inadequate power to detect rela- reason to include greater than three or four
tive risks which are smaller than expected, but controls per study subject. For example, one
still clinically important to the investigator. In could design a study with an alpha of 0.05 to
other words, if one chooses a sample size that detect a relative risk of 2.0 for an outcome vari-
is designed to detect a relative risk of 2.5, one able that occurs in the control group with an
should be comfortable with the thought that, if incidence rate of 0.01. A study with 2319
the actual relative risk turns out to be 2.2, one exposed individuals and 2319 controls would
may not be able to detect it as a statistically sig- yield a power of 0.80, or an 80% chance of
nificant finding. detecting a difference of that magnitude. With
In a cohort study one selects subjects based the same 2319 exposed subjects, ratios of con-
on the presence or absence of an exposure of trol subjects to exposed subjects of 1 : 1, 2 : 1,
interest and then investigates the incidence of 3 : 1, 4 : 1, 5 : 1, 10 : 1, and 50 : 1 would result in
the disease of interest in each of the study statistical powers of 0.80, 0.887, 0.913, 0.926,
groups. Therefore, the fourth variable one 0.933, 0.947, and 0.956, respectively.
needs to specify is the expected incidence of It is important to differentiate between the
the study outcome in the unexposed control number of controls (as was discussed and illus-
group. Again, the more you ask of a study trated above) and the number of control
(e.g. power to detect very small differences), groups. It is not uncommon, especially in
the larger the sample size need. Specifically, case–control studies, where the selection of a
the rarer the outcome of interest, the larger the proper control group can be difficult, to choose
sample size needed. more than one control group (for example, a
The fifth variable one needs to specify is the group of hospital controls and a group of com-
number of unexposed control subjects to be munity controls). This is done for reasons of
included in the study for each exposed study validity, not for statistical power, and it is
subject. A study has the most statistical power important that these multiple control groups
for a given number of study subjects if it has not be aggregated in the analysis. In this situa-
the same number of exposed and unexposed tion, the goal is to assure that the comparison
subjects (controls). However, sometimes the of the exposed subjects to each of the different
number of exposed subjects is limited and, control groups yields the same answer, not to
therefore, inadequate to provide sufficient increase the available sample size. As such, the
power to detect a relative risk of interest. In comparison of each control group to the
that case, additional power can be gained by exposed subjects should be treated as a sepa-
increasing the number of controls alone. rate study. The comparison of the exposed
Doubling the number of controls, that is group to each control group requires a separate
including two controls for each exposed subsample size calculation.
ject, results in a modest increase in the statisti- Once the five variables above have been
cal power, but it does not double it. Including specified, the sample size needed for a given
study can be calculated. Several different studied using a cohort study design and,
formulas have been used for this calcula- depending upon the values chosen for alpha,
tion, each of which gives slightly different beta, the incidence of the disease in the unex-
results. The formula that is probably the posed population, the relative risk one wants
most often used is modified from to be able to detect, and the ratio of control to
Schlesselman: exposed subjects, the sample sizes needed
could differ markedly (see Table 3.2). For
1 1 example, what if your goal was to study hepa-
N 2
Z1 1 U 1 U
p1 R K titis that occurs, say, in 0.1% of all unexposed
2 individuals? If one wanted to design a study
p1 p with one control per exposed subject to detect
Z1 pR 1 Rp
K a relative risk of 2.0 for this outcome variable,
assuming an alpha (two-tailed) of 0.05 and a
where p is the incidence of the disease in the beta of 0.1, one could look in Table A1 and see
unexposed, R is the minimum relative risk to that it would require 31 483 exposed subjects,
be detected, α is the Type I error rate which is as well as an equal number of unexposed con-
acceptable, β is the Type II error rate which is trols. If one were less concerned with missing
acceptable, Z1-α and Z1-β refer to the unit nor- a real finding, even if it was there, one could
mal deviates corresponding to α and β, Κ is the change beta to 0.2, and the required sample
ratio of number of unexposed control subjects size would drop to 23 518 (see Table 3.2 and
to the number of exposed subjects, and Table A5). If one wanted to minimize the
Kp pR number of exposed subjects needed for the
U . study, one could include up to four controls
K 1
for each exposed subject (Table 3.2 and Table
Z1-α is replaced by Z1-α /2 if one is planning to A8). This would result in a sample size of
analyze the study using a two-tailed alpha. 13 402, with four times as many controls, a
Note that K does not need to be an integer. total of 67 010 subjects. Finally, if one consid-
A series of tables are presented in the ers it inconceivable that this new drug could
Appendix, which were calculated using this protect against liver disease and one is not
formula. In Tables A1 through A4 we have interested in that outcome, then one might
assumed an alpha (two-tailed) of 0.05, a beta of use a one-tailed alpha, resulting in a some-
0.1 (90% power), and control to exposed ratios what lower sample size of 10 728, again with
of 1 : 1, 2 : 1, 3 : 1, and 4 : 1, respectively. Tables four times as many controls. Much smaller
A5 through A8 are similar, except they assume sample sizes are needed to detect relative
a beta of 0.2 (80% power). Each table presents risks of 4.0 or greater; these are also presented
the number of exposed subjects needed to in Table 3.2.
detect any of several specified relative risks, for In contrast, what if one’s goal was to study
outcome variables that occur at any of several elevated liver function tests, which, say, occur
specified incidence rates. The total study size in 1% of an unexposed population? If one
will be the sum of exposed subjects (as listed in wants to detect a relative risk of 2 for this
the table) plus the controls. more common outcome variable, only 3104
For example, what if one wanted to investi- subjects would be needed in each group,
gate a new nonsteroidal anti-inflammatory assuming a two-tailed alpha of 0.05, a beta of
drug that is about to be marketed, but premar- 0.1, and one control per exposed subject.
keting data raised questions about possible Alternatively, if one wanted to detect the
hepatotoxicity? This would presumably be same relative risk for an outcome variable
Table 3.2 Examples of sample sizes needed for a cohort study.
Incidence rate assumed Relative risk to be Control: Sample size needed Sample size needed
Disease in unexposed α β detected exposed ratio in exposed group in control group
Abnormal liver 0.01 0.05 (2-tailed) 0.1 2 1 3104 3104

function tests 0.01 0.05 (2-tailed) 0.2 2 1 2319 2319
0.01 0.05 (2-tailed) 0.2 2 4 1323 5292
0.01 0.05 (1-tailed) 0.2 2 4 1059 4236
0.01 0.05 (2-tailed) 0.1 4 1 568 568
0.01 0.05 (2-tailed) 0.2 4 1 425 425
0.01 0.05 (2-tailed) 0.2 4 4 221 884
0.01 0.05 (1-tailed) 0.2 4 4 179 716
Hepatitis 0.001 0.05 (2-tailed) 0.1 2 1 31 483 31 483
0.001 0.05 (2-tailed) 0.2 2 1 23 518 23 518
0.001 0.05 (2-tailed) 0.2 2 4 13 402 53 608
0.001 0.05 (1-tailed) 0.2 2 4 10 728 42 912
0.001 0.05 (2-tailed) 0.1 4 1 5823 5823
0.001 0.05 (2-tailed) 0.2 4 1 4350 4350
0.001 0.05 (2-tailed) 0.2 4 4 2253 9012
0.001 0.05 (1-tailed) 0.2 4 4 1829 7316
Cholestatic jaundice 0.0001 0.05 (2-tailed) 0.1 2 1 315 268 315 268
0.0001 0.05 (2-tailed) 0.2 2 1 235 500 235 500
0.0001 0.05 (2-tailed) 0.2 2 4 134 194 536 776
0.0001 0.05 (1-tailed) 0.2 2 4 107 418 429 672
0.0001 0.05 (2-tailed) 0.1 4 1 58 376 58 376
0.0001 0.05 (2-tailed) 0.2 4 1 43 606 43 606
0.0001 0.05 (2-tailed) 0.2 4 4 22 572 90 288
0.0001 0.05 (1-tailed) 0.2 4 4 18 331 73 324
that occurred as infrequently as 0.0001, per- e xposure, and then studies whether or not the
haps cholestatic jaundice, one would need disease of interest develops in each group.
315 268 subjects in each study group. Therefore, the fourth variable to be specified
Obviously, cohort studies can require very for a case–control study is the expected preva-
large sample sizes to study uncommon dis- lence of the exposure in the undiseased control
eases. A study of uncommon diseases is often group, rather than the incidence of the disease
better performed using a case–control study of interest in the unexposed control group of a
design, as described in the previous chapter. cohort study.
Finally, analogous to the consideration in
cohort studies of the ratio of the number of
Sample Size Calculations unexposed control subjects to the number of
for Case–Control Studies exposed study subjects, one needs to consider
in a case–control study the ratio of the number
The approach to calculating sample sizes for of undiseased control subjects to the number
case–control studies is similar to the approach of diseased study subjects. The principles in
for cohort studies. Again, there are five varia- deciding upon the appropriate ratio to use are
bles that need to be specified, the values of similar in both study designs. Again, there is
which depend on what the investigator expects rarely a reason to include a ratio greater than
from the study (see Table 3.1). Three of these 3 : 1 or 4 : 1. For example, if one were to design
are alpha, or the type I error one is willing to a study with a two-tailed alpha of 0.05 to detect
tolerate; beta, or the type II error one is willing a relative risk of 2.0 for an exposure which
to tolerate; and the minimum odds ratio (an occurs in 5% of the undiseased control group, a
approximation of the relative risk) one wants study with 516 diseased individuals and 516
to be able to detect. These are defined and controls would yield a power of 0.80, or an 80%
described in the section on cohort studies, chance of detecting a difference of that size.
above. Studies with the same 516 diseased subjects
In addition, in a case–control study one and ratios of controls to cases of 1 : 1, 2 : 1, 3 : 1,
selects subjects based on the presence or 4 : 1, 5 : 1, 10 : 1, and 50 : 1 would result in statis-
absence of the disease of interest, and then tical powers of 0.80, 0.889, 0.916, 0.929, 0.936,
investigates the prevalence of the exposure of 0.949, and 0.959, respectively.
interest in each study group. This is in contrast The formula for calculating sample sizes for
to a cohort study, in which one selects subjects a case–control study is similar to that for
based on the presence or absence of an cohort studies:
2
1 1
N 2
Z1 1 U 1 U Z1 p 1 p /K V 1 V
p V K
where R, α, β, Z1-α, and Z1-β are as above, p is pR

the prevalence of the exposure in the control V .
1 p R 1
group, and K is the ratio of undiseased control
subjects to diseased cases,
Again, a series of tables that provide sample
p R sizes for case–control studies is presented in
U K
K 1 1 p R 1 the Appendix. In Tables A9 through A12, we
have assumed an alpha (two-tailed) of 0.05, a
and beta of 0.1 (90% power), and control to case
Sample Size Calculations for Case Serie 41
ratios of 1 : 1, 2 : 1, 3 : 1, and 4 : 1, respectively. In contrast, what if one’s estimates of the

Tables A13 through A16 are similar, except new drug’s sales were more conservative? If
they assume a beta of 0.2 (80% power). Each one wanted to detect a relative risk of 2.0
table presents the number of diseased subjects assuming sales to 0.1% of the population, per-
needed to detect any of a number of specified haps similar to tolmetin, then 31 588 subjects
relative risks, for a number of specified expo- would be needed in each group, assuming a
sure rates. two-tailed alpha of 0.05, a beta of 0.1, and one
For example, what if again one wanted to control per diseased subject. In contrast, if one
investigate a new nonsteroidal anti- estimated the drug would be used in only
inflammatory drug that is about to be mar- 0.01% of the population (i.e. in controls with-
keted but premarketing data raised questions out the study disease of interest), perhaps like
about possible hepatotoxicity? This time, phenylbutazone, one would need 315 373 sub-
however, one is attempting to use a case–con- jects in each study group.
trol study design. Again, depending upon the Obviously, case–control studies can require
values chosen of alpha, beta, and so on, the very large sample sizes to study relatively
sample sizes needed could differ markedly uncommonly used drugs. In addition, each
(see Table 3.3). For example, what if one disease of interest requires a separate case
wanted to design a study with one control group and, thereby, a separate study. As such,
(undiseased subject) per diseased subject, as described in the prior chapter, studies of
assuming an alpha (two-tailed) of 0.05 and a uncommonly used drugs and newly marketed
beta of 0.1? The sample size needed to detect drugs are usually better done using cohort
a relative risk of 2.0 for any disease would study designs, whereas studies of rare dis-
vary, depending on the prevalence of use of eases are better done using case–control
the drug being studied. If one optimistically designs.
assumed the drug will be used nearly as com-
monly as ibuprofen, by perhaps 1% of the
population, then one could look in Table A9 Sample Size Calculations
and see that it would require 3210 diseased for Case Series
subjects and an equal number of undiseased
controls. If one were less concerned with As described in Chapter 2, the utility of case
missing a real association, even if it existed, series in pharmacoepidemiology is limited, as
one could opt for a beta of 0.2, and the the absence of a control group makes causal
required sample size would drop to 2398 (see inference difficult. Despite this, however, this
Table 3.3 and Table A13). If one wanted to is a design that has been used repeatedly. There
minimize the number of diseased subjects are scientific questions that can be addressed
needed for the study, one could include up to using this design, and the collection of a con-
four controls for each diseased subject trol group equivalent in size to the case series
(Table 3.3 and Table A16). This would result would add considerable cost to the study. Case
in a sample size of 1370, with four times as series are usually used in pharmacoepidemiol-
many controls. Finally, if one considers it ogy to quantitate better the incidence of a par-
inconceivable that this new drug could protect ticular disease in patients exposed to a newly
against liver disease, then one might use a marketed drug. For example, in the “Phase 4”
one-tailed alpha, resulting in a somewhat postmarketing drug surveillance study con-
lower sample size of 1096, again with four ducted for prazosin, the investigators collected
times as many controls. Much smaller sample a case series of 10 000 newly exposed subjects
sizes are needed to detect relative risks of 4.0 recruited through the manufacturer’s sales
or greater and are also presented in Table 3.3. force, to quantitate better the incidence of first
Table 3.3 Examples of sample sizes needed for a case–control study.
Prevalence rate Odds ratio to Control: Sample size needed Sample size needed
Hypothetical drug assumed in undiseased α β be detected case ratio in case group in control group
Ibuprofen 0.01 0.05 (2-tailed) 0.1 2 1 3210 3210

0.01 0.05 (2-tailed) 0.2 2 1 2398 2398
0.01 0.05 (2-tailed) 0.2 2 4 1370 5480
0.01 0.05 (1-tailed) 0.2 2 4 1096 4384
0.01 0.05 (2-tailed) 0.1 4 1 601 601
0.01 0.05 (2-tailed) 0.2 4 1 449 449
0.01 0.05 (2-tailed) 0.2 4 4 234 936
0.01 0.05 (1-tailed) 0.2 4 4 190 760
Tolmetin 0.001 0.05 (2-tailed) 0.1 2 1 31 588 31 588
0.001 0.05 (2-tailed) 0.2 2 1 23 596 23 596
0.001 0.05 (2-tailed) 0.2 2 4 13 449 53 796
0.001 0.05 (1-tailed) 0.2 2 4 10 765 43 060
0.001 0.05 (2-tailed) 0.1 4 1 5856 5856
0.001 0.05 (2-tailed) 0.2 4 1 4375 4375
0.001 0.05 (2-tailed) 0.2 4 4 2266 9064
0.001 0.05 (1-tailed) 0.2 4 4 1840 7360
Phenylbutazone 0.0001 0.05 (2-tailed) 0.1 2 1 315 373 315 373
0.0001 0.05 (2-tailed) 0.2 2 1 235 579 235 579
0.0001 0.05 (2-tailed) 0.2 2 4 134 240 536 960
0.0001 0.05 (1-tailed) 0.2 2 4 107 455 429 820
0.0001 0.05 (2-tailed) 0.1 4 1 58 409 58 409
0.0001 0.05 (2-tailed) 0.2 4 1 43 631 43 631
0.0001 0.05 (2-tailed) 0.2 4 4 22 585 90 340
0.0001 0.05 (1-tailed) 0.2 4 4 18 342 73 368
Discussio 43
dose syncope, which was a well-recognized 0.003 = 0.000618. Analogously, the upper
adverse effect of this drug. Case series are usu- boundary would be the product of the corre-
ally used to determine whether a disease sponding “Upper Limit Factor (U)” multiplied
occurs more frequently than some predeter- by the observed incidence rate. In the above
mined incidence in exposed patients. Most example, this would be 2.92 × 0.003 = 0.00876.
often, the predetermined incidence of interest In other words, the incidence rate (95% confi-
is zero, and one is looking for any occurrences dence interval) would be 0.003 (0.000618–
of an extremely rare illness. As another exam- 0.00876). Thus, the best estimate of the
ple, when cimetidine was first marketed, there incidence rate would be 30 per 10 000, but
was a concern over whether it could cause there is a 95% chance that it lies between 6.18
agranulocytosis, since it was closely related per 10 000 and 87.6 per 10 000.
chemically to metiamide, another H-2 blocker, In addition, a helpful simple guide is the so-
which had been removed from the market in called “rule of threes,” useful in the common
Europe because it caused agranulocytosis. This situation where no events of a particular kind
study also collected 10 000 subjects. It found are observed. Specifically, if no events of a par-
only two cases of neutropenia, one in a patient ticular type (i.e. the events of interest to the
also receiving chemotherapy. There were no study) are observed in a study of X individuals,
cases of agranulocytosis. then one can be 95% certain that the event
To establish drug safety, a study must include occurs no more often than 3/X. For example, if
a sufficient number of subjects to detect an 500 patients are studied prior to marketing a
elevated incidence of a disease, if it exists. drug, then one can be 95% certain that any
Generally, this is calculated by assuming the event which does not occur in any of those
frequency of the event in question is vanish- patients may occur with a frequency of 3 or
ingly small, so that the occurrence of the event less in 500 exposed subjects, or that it has an
follows a Poisson distribution, and then one incidence rate of less than 0.006. If 3000 sub-
generally calculates 95% confidence intervals jects are exposed prior to drug marketing, then
around the observed results. one can be 95% certain that any event which
Table A17 in the Appendix presents a table does not occur in this population may occur
useful for making this calculation. In order to no more than three in 3000 subjects, or the
apply this table, one first calculates the inci- event has an incidence rate of less than 0.001.
dence rate observed from the study’s results, Finally, if 10 000 subjects are studied in a post-
that is the number of subjects who develop the marketing drug surveillance study, then one
disease of interest during the specified time can be 95% certain that any events which are
interval, divided by the total number of indi- not observed may occur no more than three in
viduals in the population at risk. For example, 10 000 exposed individuals, or that they have
if three cases of liver disease were observed in an incidence rate of less than 0.0003. In other
a population of 1000 patients exposed to a new words, events not detected in the study may
nonsteroidal anti-inflammatory drug during a occur less often than one in 3333 subjects in
specified period of time, the incidence would the general population.
be 0.003. The number of subjects who develop
the disease is the “Observed Number on Which
Estimate is Based (n)” in Table A17. In this Discussion
example, it is three. The lower boundary of the
95% confidence interval for the incidence rate The above discussions about sample size deter-
is then the corresponding “Lower Limit Factor minations in cohort and case–control studies
(L)” multiplied by the observed incidence rate. assume one is able to obtain information on
In the example above, it would be 0.206 X each of the five variables that factor into these
sample size calculations. Is this in fact realis- requirement, and note that the study will have
tic? Four of the variables are, in fact, totally in even more power for the other outcome (or
the control of the investigator, subject to his or exposure) variables. Regardless, it is usually
her specification: alpha, beta, the ratio of con- better to have a somewhat larger than expected
trol subjects to study subjects, and the mini- sample size than the minimum, to allow some
mum relative risk to be detected. Only one of leeway if any of the underlying assumptions
the variables requires data derived from other were wrong. This also will permit subgroup
sources. For cohort studies, this is the expected analyses with adequate power. In fact, if there
incidence of the disease in the unexposed con- are important subgroup analyses that repre-
trol group. For case–control studies, this is the sent a priori hypotheses that one wants to be
expected prevalence of the exposure in the able to evaluate, one should perform separate
undiseased control group. In considering this sample size calculations for those subgroups.
needed information, it is important to realize In this situation, one should use the incidence
that the entire process of sample size calcula- of disease or prevalence of exposure that
tion is approximate, despite its mathematical occurs in the subgroups, not that which occurs
sophistication. There is certainly no compel- in the general population.
ling reason why an alpha should be 0.05, as Note that sample size calculation is often an
opposed to 0.06 or 0.04. The other variables iterative process. There is nothing wrong with
specified by the investigator are similarly arbi- performing an initial calculation, realizing that
trary. As such, only an approximate estimate is it generates an unrealistic sample size, and
needed for this missing variable. Often the then modifying the underlying assumptions
needed information is readily available from accordingly. What is important is that the inves-
some existing data source, for example vital tigator examines his or her final assumptions
statistics or commercial drug utilization data closely, asking whether, given the compromises
sources. If not, one can search the medical lit- made, the study is still worth undertaking.
erature for one or more studies that have col- Note that the discussion above was restricted
lected these data for a defined population, to sample size calculations for dichotomous
either deliberately or as a by-product of their variables, i.e. variables with only two options: a
data collecting effort, and assume that the pop- study subject either has a disease or does not
ulation you will study will be similar. If this is have a disease. Information was not presented
not an appropriate assumption, or if no such on sample size calculations for continuous out-
data exist in the medical literature, one is left come variables, i.e. variables that have some
with two alternatives. The first, and better, measurement, such as height, weight, blood
alternative is to conduct a small pilot study pressure, or serum cholesterol. Overall, the use
within your population, in order to measure of a continuous variable as an outcome varia-
the information you need. The second is sim- ble, unless the measurement is extremely
ply to guess. In the second case, one should imprecise, will result in a marked increase in
consider what a reasonable higher guess and a the power of a study. Details about this are
reasonable lower guess might be, as well, to see omitted because epidemiologic studies unfor-
if your sample size should be increased to take tunately do not usually have the luxury of
into account the imprecision of your estimate. using such variables. Readers who are inter-
Finally, what if one is studying multiple out- ested in more information on this can consult
come variables (in a cohort study) or multiple a textbook of sample size calculations.
exposure variables (in a case–control study), All of the previous discussions have focused
each of which differs in the frequency you on calculating a minimum necessary sample
expect in the control group? In that situation, size. This is the usual concern. However, two
an investigator might base the study’s sample other issues specific to pharmacoepidemiology
size on the variable that leads to the largest are important to consider as well. First, one of
Further Readin 45
the main advantages of postmarketing pharma- addition, it must be kept in mind that subtle
coepidemiologic studies is the increased sensi- findings, even if statistically and clinically
tivity to rare adverse reactions that can be important, could easily have been created by
achieved, by including a sample size larger than biases or confounders (see Chapter 2). Subtle
that used prior to marketing. Since between 500 findings should not be ignored, but should be
and 3000 patients are usually studied before interpreted with caution.
marketing, most pharmacoepidemiologic
cohort studies are designed to include at least
10 000 exposed subjects. The total population Key Points
from which these 10 000 exposed subjects would
be recruited would need to be very much larger, ●● Premarketing studies of drugs are inherently
of course. Case–control studies can be much limited in size, meaning larger studies are
smaller, but generally need to recruit cases and needed after marketing in order to detect
controls from a source population of equivalent less common drug effects.
size as for cohort studies. These are not com- ●● For a cohort study, the sample size needed is
pletely arbitrary figures, but are based on the determined by specifying the Type I error
principles described above, applied to the ques- one is willing to tolerate, the Type II error
tions which remain of great importance to one is willing to tolerate, the smallest rela-
address in a postmarketing setting. Nevertheless, tive risk which one wants to be able to detect,
these figures should not be rigidly accepted but the expected incidence of the outcome of
should be reconsidered for each specific study. interest in the unexposed control group, and
Some studies will require fewer subjects, many the ratio of the number of unexposed control
will require more. To accumulate these sample subjects to be included in the study to the
sizes while performing cost-effective studies, number of exposed study subjects.
several special techniques have been developed, ●● For a case–control study, the needed sample
which are described in Section II of this book. size is determined by specifying the Type I
Second, because of the development of these error one is willing to tolerate, the Type II
new techniques and the development of large error one is willing to tolerate, the smallest
electronic data systems (see Chapters 8, 9, odds ratio which one wants to be able to
and 10), pharmacoepidemiologic studies have detect, the expected prevalence of the expo-
the potential for the relatively unusual prob- sure of interest in the undiseased control
lem of too large a sample size. It is even more group, and the ratio of the number of undis-
important than usual, therefore, when inter- eased control subjects to be included in the
preting the results of studies that use these study to the number of exposed study
data systems to examine their findings, differ- subjects.
entiating clearly between statistical signifi- ●● As a rule of thumb, if no events of a particu-
cance and clinical significance. With a very lar type are observed in a study of X individ-
large sample size, one can find statistically sig- uals, then one can be 95% certain that the
nificant differences that are clinically trivial. In event occurs no more often than 3/X.
Further Reading
Cohen, J. (1977). Statistical Power Analysis for postmarket outpatient surveillance program.
the Social Sciences. New York: Academic JAMA 243: 1532–1535.
Press. Graham, R.M., Thornell, I.R., Gain, J.M. et al.
Gifford, L.M., Aeugle, M.E., Myerson, R.M., and (1976). Prazosin: the first dose phenomenon.
Tannenbaum, P.J. (1980). Cimetidine BMJ 2: 1293–1294.
Haenszel, W., Loveland, D.B., and Sirken, M.G. Schlesselman, J.J. (1974). Sample size
(1962). Lung cancer mortality as related to requirements in cohort and case–control studies
residence and smoking history. I. White of disease. Am. J. Epidemiol. 99: 381–384.
males. J. Natl. Cancer Inst. 28: 947–1001. Stolley, P.D. and Strom, B.L. (1986). Sample size
Joint Commission on Prescription Drug Use calculations for clinical pharmacology studies.
(1980) Final Report. Washington, DC. Clin. Pharmacol. Ther. 39: 489–490.
Makuch, R.W. and Johnson, M.F. (1986). Some Young, M.J., Bresnitz, E.A., and Strom, B.L.
issues in the design and interpretation of (1983). Sample size nomograms for
“negative” clinical trials. Arch. Intern. Med. interpreting negative clinical studies. Ann.
146: 986–989. Intern. Med. 99: 248–251.
47
Basic Principles of Clinical Pharmacology Relevant

to Pharmacoepidemiologic Studies
Jeffrey S. Barrett
Critical Path Institute, Tuscon, AZ, USA
Introduction Clinical pharmacology is an important bridg-

ing discipline that includes knowledge about
Pharmacology deals with the study of drugs the relationships between: dose and exposure
while clinical pharmacology deals with the at the site of action (pharmacokinetics); expo-
study of drugs in humans. More specifically, sure at the site of action and clinical response
clinical pharmacology evaluates the character- (pharmacodynamics); and between clinical
istics, effects, properties, reactions, and uses of response and outcomes. In the process, it
drugs, particularly their therapeutic value in defines the therapeutic window (the dosage of
humans, including their toxicology, safety, a medication between the minimum amount
pharmacodynamics, and pharmacokinetics. that gives a desired effect and the minimum
While the foundation of the discipline is amount that gives more adverse effects than
underpinned by basic pharmacology (the study desired effects) of a drug in various patient
of the interactions that occur between a living populations. Likewise, clinical pharmacology
organism and exogenous chemicals that alter also guides dose modifications in various
normal biochemical function), the important patient subpopulations (e.g. pediatrics, preg-
emphasis of clinical pharmacology is the appli- nancy, elderly, and organ impairment) and/or
cation of pharmacologic principles and meth- dose adjustments for various lifestyle factors
ods in the care of patients. From the discovery (e.g. food, time of day, drug interactions).
of new target molecules and molecular targets The discovery and development of new med-
to the evaluation of clinical utility in specific icines is reliant upon clinical pharmacologic
populations, clinical pharmacology bridges the research. Scientists in academic, regulatory,
gap between laboratory science and medical and industrial settings participate in this
practice. The main objective is to promote the research as part of the overall drug develop-
safe and effective use of drugs, maximizing the ment process. The output from clinical phar-
beneficial drug effects while minimizing harm- macologic investigation appears in the drug
ful side effects. It is important that caregivers monograph or package insert of all new medi-
are skilled in the areas of drug information, cines and forms the basis of how drug dosing
medication safety, and other aspects of phar- information is communicated to healthcare
macy practice related to clinical pharmacology. providers.
48 4 Basic Principles of Clinical Pharmacology Relevant to Pharmacoepidemiologic Studies
linical Pharmacology
C asics of Clinical
B
and Pharmacoepidemiology Pharmacology
Pharmacoepidemiology is the study of the use Clinical pharmacology encompasses drug

and effects of drugs in large numbers of composition, drug properties, interactions,
people. toxicology, and effects (both desirable and
To accomplish this, pharmacoepidemiol- undesirable) that can be used in pharmaco-
ogy borrows from both clinical pharmacology therapy of diseases. Underlying the discipline
and epidemiology. Thus, pharmacoepidemi- of clinical pharmacology are the fields of
ology can also be called a bridging science. pharmacokinetics and pharmacodynamics,
Part of the task of clinical pharmacology is to and each of these disciplines can be further
provide risk–benefit assessment for the effect defined by specific subprocesses (absorp-
of drugs in patients. Studies that estimate the tion, distribution, metabolism, elimination).
probability and magnitude of beneficial Clinical pharmacology is essential to our
effects in populations, or the probability and understanding of how drugs work as well
magnitude of adverse effects in populations, as how to guide their administration.
will benefit from epidemiologic methodol- Pharmacotherapy can be challenging because
ogy. Pharmacoepidemiology then can also be of physiologic factors that may alter drug
defined as the application of epidemiologic kinetics (age, size, etc.), pathophysiologic dif-
methods to the content area of clinical phar- ferences that may alter pharmacodynamics,
macology. Figure 4.1 illustrates the relation- disease etiologies in studied patients that may
ship between clinical pharmacology and differ from those present in the general popu-
pharmacoepidemiology as well as some of lation, and other factors that may result in
the specific research areas reliant on both great variation in safety and efficacy out-
disciplines. comes. The challenge is more difficult in
PK Pharmacoepidemiology borrows from

P both clinical pharmacology and
epidemiology. Thus, pharmacoepide-
Clinical
miology can also be called a bridging
pharmacology Epidemiology
science spanning both clinical
E Paharmacology and epidemiology.
PD Part of the task of clinical pharmacol-
ogy is to provide a risk-benefit
assessment for the effect of drugs in
patients.
• Drug interactions
• Global trends in prescribing
• Generic vs reference utilization
CP + PE • Management of ADRs
Overlap • Screening studies (drug
development)
• Lifestyle effects on drug therapy
• Special population drug therapy
• Equivalence testing
• Spontaneous reporting of safety
Figure 4.1 Relationship between clinical pharmacology and pharmacoepidemiology, illustrating the
overlapping areas of interest.
Pharmacokinetic 49
the critically ill given the paucity of rather a conceptual parameter. It relates the
well-controlled clinical trials in vulnerable total amount of drug in the body to the concen-
populations. tration of drug (C) in the blood or plasma:
Vd = Drug/C..
Figure 4.2 represents the fate of a drug after
Pharmacokinetics intravenous administration. After administra-
tion, maximal plasma concentration is
Pharmacokinetics refers to the study of the achieved, and the drug is distributed. The
absorption and distribution of an administered plasma concentration then decreases over
drug, the chemical changes of the substance in time. This initial alpha (α) phase of drug distri-
the body (metabolism), and the effects and bution indicates the decline in plasma concen-
routes of excretion of the metabolites of the tration due to the distribution of the drug.
drug (elimination). Once a drug is distributed, it undergoes metab-
olism and elimination. The second beta (β)
phase indicates the decline in plasma concen-
Absorption
tration due to drug metabolism and clearance.
Absorption is the process of drug transfer from The terms A and B are intercepts with the ver-
its site of administration to the blood stream. tical axis. The extrapolation of the β phase
The rate and efficiency of absorption depend defines B. The dotted line is generated by sub-
on the route of administration. For intrave- tracting the extrapolated line from the original
nous administration, absorption is complete; concentration line. This second line defines α
the total dose reaches the systemic circulation. and A. The plasma concentration can be esti-
Drugs administered enterally may be absorbed mated using the formula: C = Ae − αt + Be − βt.
by either passive diffusion or active transport. The distribution and elimination half lives can
The bioavailability (F) of a drug is the fraction be determined by: t1/2α = 0.693/α and
of the administered dose that reaches the sys- t1/2β = 0.693/β, respectively.
temic circulation. If a drug is administered For drugs in which distribution is homoge-
intravenously, then bioavailability is 100% and nous across the various physiologic spaces,
F = 1.0. When drugs are administered by the distinction between the alpha and beta
routes other than intravenous, the bioavailabil- phase may be subtle and essentially a single
ity is usually less. Bioavailability is reduced by phase best describes the decline in drug
incomplete absorption, first-pass metabolism concentration.
(defined below), and distribution into other
tissues.
Metabolism
The metabolism of drugs is catalyzed by
Volume of Distribution
enzymes, and most reactions follow Michaelis
The apparent volume of distribution (Vd) is a Menten kinetics: V (rate of drug metabo-
hypothetical volume of fluid through which a lism) = [((Vmax) (C)/Km) + (C)], where C is the
drug is dispersed. A drug rarely disperses solely drug concentration, Vmax is the maximum rate
into the water compartments of the body. of metabolism in units of amount of product
Instead, the majority of drugs disperse to sev- over time, typically μmol/min, and Km is the
eral compartments, including adipose tissue Michaelis Menten constant (substrate concen-
and plasma proteins. The total volume into tration at which the rate of conversion is half
which a drug disperses if it were only fluid is of Vmax) also in units of concentration. In most
called the apparent volume of distribution. situations, the drug concentration is much less
This volume is not a physiologic space, but than Km and the equation simplifies to:
1000
A
α phase
slope = –α
B
100
Concentration
β phase
slope = –β
10
1
0 2 4 6 8 10 12
Time (hours)
Figure 4.2 Semi-logarithmic plot of concentration vs time after an intravenous administration of a drug
that follows two-compartment pharmacokinetics.
V = (Vmax)(C)/Km. In this case, the rate of drug absorbed intact from the small intestine and
metabolism is directly proportional to the con- transported to the liver via the portal system,
centration of free drug, and follows first order where they are metabolized. This process is
kinetic theory. A constant percentage of the called first pass metabolism, and may greatly
drug is metabolized per unit time, and the limit the bioavailability of orally administered
absolute amount of drug eliminated per unit drugs. In general, all metabolic reactions can
time is proportional to the amount of drug in be classified as either phase I or phase II bio-
the body. transformations. Phase I reactions usually con-
Most drugs used in the clinical setting are vert the parent drug to a polar metabolite by
eliminated in this manner. A few drugs, such introducing or unmasking a more polar site. If
as aspirin, ethanol, and phenytoin, are used in phase I metabolites are sufficiently polar, they
higher doses, resulting in higher plasma con- may be readily excreted. However, many phase
centrations. In these situations, C is much I metabolites undergo a subsequent reaction in
greater than Km, and the Michaelis Menten which endogenous substances such as glucu-
equation reduces to: V(rate of drug metabo- ronic acid, sulfuric acid, or an amino acid com-
lism) = (Vmax)(C)/(C) = Vmax. The enzyme sys- bine with the metabolite to form a highly polar
tem becomes saturated by a high free drug conjugate. Many drugs undergo these sequen-
concentration, and the rate of metabolism is tial reactions.
constant over time. This is called zero-order Phase I reactions are usually catalyzed by
kinetics, and a constant amount of drug is enzymes of the cytochrome P450 system.
metabolized per unit time. For drugs that fol- These drug-metabolizing enzymes are located
low zero-order elimination, a large increase in in the lipophilic membranes of the endoplas-
serum concentration can result from a small mic reticulum of the liver and other tissues.
increase in dose. Three gene families, CYP1, CYP2, and CYP3,
The liver is the principal organ of drug are responsible for most drug biotransforma-
metabolism. Other organs that display consid- tions. The CYP3A subfamily accounts for
erable metabolic activity include the gastroin- greater than 50% of phase I drug metabolism,
testinal tract, lungs, skin, and kidneys. predominantly by the CYP3A4 sub-type.
Following oral administration, many drugs are CYP3A4 is responsible for the metabolism of
Pharmacokinetic 51
drugs commonly used in the intensive care set- as CYP 2E1 inducers, nor are they substrates
ting, including acetaminophen, cyclosporine, for that system. Phenytoin in fact may be
methadone, midazolam, and tacrolimus. Most hepato-protective because it is an inducer of
other drug biotransformations are performed the glucuronidation metabolic pathway for
by CYP2D6 (e.g. clozapine, codeine, flecainide, acetaminophen, thus shunting metabolism
haloperidol, oxycodone), CYP2C9 (e.g. pheny- away from NAPQI production.
toin, S-warfarin), CYP2C19 (e.g. diazepam,
omeprazole, propranolol), CYP2E1 (e.g. aceta-
Elimination
minophen, enflurane, halothane), and
CYP1A2 (e.g. acetaminophen, theophylline, Elimination is the process by which drug is
warfarin). removed or “cleared” from the body. Clearance
Biotransformation reactions may be (CL) is the amount of blood from which all
enhanced or impaired by multiple factors, drug is removed per unit time (volume/time).
including age, enzyme induction or inhibi- The primary organs responsible for drug clear-
tion, pharmacogenetics, and the effects of ance are the kidneys and liver. The total body
other disease states. Approximately 95% of the clearance is equal to the sum of individual
metabolism occurs via conjugation to glucuro- clearances from all mechanisms. Typically, this
nide (50–60%) and sulfate (25–35%). Most of is partitioned into renal and nonrenal clear-
the remainder of acetaminophen is metabo- ance. Most elimination by the kidneys is
lized via the cytochrome P450 forming accomplished by glomerular filtration. The
N-acetyl-p-benzoquinone imine (NAPQI) amount of drug filtered is determined by glo-
thought to be responsible for hepatotoxicity. merular integrity, the size and charge of the
This minor but important pathway is cata- drug, water solubility, and the extent of protein
lyzed by CYP 2E1, and to a lesser extent, binding. Highly protein-bound drugs are not
CYP 1A2 and CYP 3A4. NAPQI is detoxified readily filtered. Therefore, estimation of the
by reacting with either glutathione directly or glomerular filtration rate (GFR) has tradition-
through a glutathione transferase catalyzed ally served as an approximation of renal
reaction. When hepatic synthesis of glu- function.
tathione is overwhelmed, manifestations of In addition to glomerular filtration, drugs
toxicity appear, producing centrilobular may be eliminated from the kidneys via active
necrosis. In the presence of a potent secretion. Secretion occurs predominantly at
CYP 2E1 inhibitor, disulfiram, a 69% reduc- the proximal tubule of the nephron, where
tion in the urinary excretion of these 2E1 met- active transport systems secrete primarily
abolic products is observed, supporting the organic acids and bases. Organic acids include
major role for 2E1 in the formation of NAPQI. most cephalosporins, loop diuretics, metho-
CYP 2E1 is unique among the CYP gene fami- trexate, nonsteroidal anti-inflammatories, pen-
lies in an ability to produce reactive oxygen icillins, and thiazide diuretics. Organic bases
radicals through a reduction of O2 and is the include ranitidine and morphine. As drugs
only CYP system strongly induced (drug mol- move toward the distal convoluting tubule,
ecule initiates or enhances the expression of concentration increases. High urine flow rates
an enzyme) by alcohol which is itself a sub- decrease drug concentration in the distal
strate (a molecule upon which an enzyme tubule, decreasing the likelihood of diffusion
acts). In addition to alcohol, isoniazid acts as from the lumen. For both weak acids and
inducer and a substrate. Ketoconazole and bases, the nonionized form is reabsorbed more
other imidazole compounds are inducers but readily. Altering pH can minimize reabsorp-
not substrates. Barbiturates and phenytoin, tion, by placing a charge on the drug and pre-
which are nonspecific inducers, have no role venting diffusion. For example, salicylate is a
weak acid. In case of salicylate toxicity, urine impairment in the function of the many regu-
alkalinization places a charge on the molecule, latory processes that provide functional inte-
and increases its elimination. The liver also gration between cells and organs. Cardiac
contributes to elimination through metabo- structure and function, renal and gastrointesti-
lism or excretion into the bile. After the drug is nal systems, and body composition are the
secreted in bile, it may then be either excreted physiologic systems most often implicated
into the feces or reabsorbed via enterohepatic when pharmacokinetic or pharmacodynamic
recirculation. The half-life of elimination is the differences are observed between an elderly
time it takes to clear half of the drug from and young population. Table 4.1 lists the pri-
plasma. It is directly proportional to the Vd, mary physiologic factors affected by aging.
and inversely proportional to CL: t1/2β = (0.693) With respect to absorption, the impact of age
(Vd)/CL. is unclear and many conflicting results exist.
While many studies have not shown signifi-
cant age-related differences in absorption
Special Populations rates, the absorption of vitamin B12, iron, and
calcium is slower through reduced active
The term “special populations” as applied to transport mechanisms. A reduction in first-
drug development refers to discussions in the pass metabolism is associated with aging, most
early 1990s by industry, academic, and regula- likely due to a reduction in liver mass and
tory scientists struggling with the then current blood flow. Likewise, drugs undergoing signifi-
practice that early drug development was cant first-pass metabolism experience an
focused predominantly on young, Caucasian, increase in bioavailability with age. This is the
male populations. Representatives from the case for propranolol and labetalol. Conversely,
US, Europe, and Japan jointly issued regula- prodrugs requiring activation in the liver (e.g.
tory requirements for drug testing and labeling ACE inhibitors enalapril and perindopril) are
in “special populations” (namely the elderly) likely to experience reduction in this phase and
in 1993. In later discussions, this generaliza- therefore reduced exposure to the active
tion was expanded to include four major demo- species.
graphic segments (women, elderly, pediatric, Based on age-related changes in body com-
and major ethnic groups); despite the size of position, polar drugs that are primarily water
each of these subpopulations, pharmaceutical soluble often exhibit smaller volumes of distri-
research had been limited. More importantly, bution, resulting in higher plasma concentra-
these “special populations” represent diverse tions in older patients. This is the case for
subpopulations of patients in whom dosing agents including ethanol, theophylline,
guidance is often needed and likewise targeted digoxin, and gentamicin. Conversely, nonpolar
clinical pharmacologic research is essential. compounds are often lipid soluble and exhibit
larger volumes of distribution in older patients.
The impact of the larger Vd is prolongation of
Elderly
half-life with age. This is the case for drugs
Physical signs consistent with aging include such as chlormethiazole and thiopentone.
wrinkles, change of hair color to gray or white, Conflicting results have been reported with
hair loss, lessened hearing, diminished eye- respect to age effects on protein binding, mak-
sight, slower reaction times, and decreased ing generalizations difficult.
agility. We are generally more concerned with Several drug classes including water-
how aging affects physiologic processes that soluble antibiotics, diuretics, water-soluble
dictate drug pharmacokinetics and pharmaco- beta-adrenoceptor blockers, and nonsteroidal
dynamics. Advancing age is characterized by anti-inflammatory drugs exhibit changes in
Special Population 53
Table 4.1 Physiologic systems affected during aging that influence drug pharmacokinetic and/or
pharmacodynamic behavior.
Physiologic system Impact of aging
Cardiac structure ●● Reduced elasticity and compliance of the aorta and great arteries (higher systolic
and function arterial pressure, increased impedance to left ventricular hypertrophy and
interstitial fibrosis)
●● Decrease in rate of myocardial relaxation
●● Left ventricle stiffens and takes longer to relax and fill in diastole
●● Isotonic contraction is prolonged and velocity of shortening reduced
●● Reduction in intrinsic heart rate and increased sinoatrial node conduction time
Renal system ●● Renal mass decreases (reduction in number of nephrons)
●● Reduced blood flow in the afferent arterioles in the cortex
●● Renal plasma flow and glomerular filtration rate decline
●● Decrease in ability to concentrate the urine during water deprivation
●● Impaired response to water loading
Gastrointestinal ●● Secretion of hydrochloric acid and pepsin is decreased under basal conditions
system ●● Reduced absorption of several substances in the small intestine including sugar,
calcium and iron
●● Decrease in lipase and trypsin secretion in the pancreas
●● Progressive reduction in liver volume and liver blood flow
Body composition ●● Progressive reduction in total body water and lean body mass, resulting in a
relative increase in body fat
c learance with age because of declining renal Pediatrics

function. With respect to hepatic metabolism,
As children develop and grow changes in body
studies have shown that significant reductions
composition, development of metabolizing
in clearance with age are observed for phase I
enzymes, and maturation of renal and liver
pathways in the liver.
function, all affect drug disposition.
From the standpoint of a clinical trial, age
categories are necessary to define the inclusion
and exclusion criteria for the population tar- Renal
geted for enrollment. Pharmaceutical sponsors Renal function in the premature and full-term
are increasingly encouraged to include a neonate, both glomerular filtration and tubu-
broader range of ages in their pivotal trials or lar secretion, is significantly reduced, as com-
specifically target an elderly subpopulation in pared to older children. Maturation of renal
a separate study, consistent with FDA guid- function is a dynamic process that begins dur-
ance. The FDA guideline for studies in the ing fetal life and is complete by early child-
elderly is directed principally toward new hood. Maturation of tubular function is slower
molecular entities likely to have significant use than that of glomerular filtration. The GFR is
in the elderly, either because the disease approximately 2–4 ml/minute/1.73 m2 in full
intended to be treated is characteristically a term neonates, but it may be as low as 0.6–
disease of aging (e.g. Alzheimer’s disease) or 0.8 ml/minute/1.73 m2 in preterm neonates.
because the population to be treated is known The GFR increases rapidly during the first two
to include substantial numbers of geriatric weeks of life and continues to rise until adult
patients (e.g. hypertension). values are reached at 8–12 months of age. For
drugs that are renally eliminated, impaired motility are the primary determinants of the
renal function decreases clearance, increasing rate at which drugs are presented to and dis-
the half-life. Therefore, for drugs that are pri- persed along the mucosal surface of the small
marily eliminated by the kidney, dosing should intestine. At birth, the coordination of antral
be performed in an age-appropriate fashion contractions improves, resulting in a marked
that takes into account both maturational increase in gastric emptying during the first
changes in kidney function. week of life. Similarly, intestinal motor activity
matures throughout early infancy, with conse-
Hepatic quent increases in the frequency, amplitude,
Hepatic biotransformation reactions are sub- and duration of propagating contractions.
stantially reduced in the neonatal period. At Changes in the intraluminal pH in different
birth, the cytochrome p450 system is 28% that segments of the gastrointestinal tract can
of the adult. The expression of phase I enzymes directly affect both the stability and the degree
such as the P-450 cytochromes changes mark- of ionization of a drug, thus influencing the
edly during development. CYP3A7, the pre- relative amount of drug available for absorp-
dominant CYP isoform expressed in fetal liver, tion. During the neonatal period, intragastric
peaks shortly after birth and then declines pH is relatively elevated (>4). Thus, oral
rapidly to levels that are undetectable in most administration of acid-labile compounds such
adults. Within hours after birth, CYP2E1 as penicillin G produces greater bioavailability
activity increases, and CYP2D6 becomes in neonates than in older infants and children.
detectable soon thereafter. CYP3A4 and In contrast, drugs that are weak acids, such as
CYP2C appear during the first week of life, phenobarbital, may require larger oral doses in
whereas CYP1A2 is the last hepatic CYP to the very young in order to achieve therapeutic
appear, at one to three months of life. The plasma levels. Other factors that impact the
ontogeny of phase II enzymes is less well rate of absorption include age-associated
established than the ontogeny of reactions development of villi, splanchnic blood flow,
involving phase I enzymes. Available data changes in intestinal microflora, and intestinal
indicate that the individual isoforms of glucu- surface area.
ronosyltransferase (UGT) have unique matu-
rational profiles with pharmacokinetic Body Composition
consequences. For example, the glucuronida- Age-dependent changes in body composition
tion of acetaminophen (a substrate for alter the physiologic spaces into which a drug
UGT1A6 and, to a lesser extent, UGT1A9) is may be distributed. The percent of total body
decreased in newborns and young children as water drops from about 85% in premature
compared with adolescents and adults. infants to 75% in full-term infants to 60% in the
Glucuronidation of morphine (a UGT2B7 sub- adult. Extracellular water decreases from 45%
strate) can be detected in premature infants as in the infant to 25% in the adult. Total body fat
young as 24 weeks of gestational age. in the premature infant can be as low as 1%, as
compared to 15% in the normal, term infant.
Gastrointestinal Many drugs are less bound to plasma proteins
Overall, the rate at which most drugs are in the neonate and infant than in the older
absorbed is slower in neonates and young child. Limited data in neonates suggest that
infants than in older children. As a result, the the passive diffusion of drugs into the central
time required to achieve maximal plasma lev- nervous system is age dependent, as reflected
els is longer in the very young. The effect of age by the progressive increase in the ratios of
on enteral absorption is not uniform and is dif- brain phenobarbital to plasma phenobarbital
ficult to predict. Gastric emptying and intestinal from 28 to 39 weeks of gestational age,
Special Population 55
emonstrating the increased transport of phe-

d drugs. Likewise, changes in the volume of dis-
nobarbital into the brain. tribution and clearance during pregnancy can
cause increases or decreases in the terminal
elimination half-life. Renal excretion of
Pregnancy
unchanged drugs is increased during preg-
The FDA classifies drugs into five categories of nancy and hence these agents may require
safety for use during pregnancy (normal preg- dose increases with pregnancy. Likewise, the
nancy, labor, and delivery) as shown in metabolism of drugs via select P450-mediated
Table 4.2. Few well-controlled studies of thera- pathways (3A4, 2D6, and 2C9) and UGT isoen-
peutic drugs have been conducted in pregnant zymes are increased during pregnancy, neces-
women. Most information about drug safety sitating increased dosages of drugs metabolized
during pregnancy is derived from animal stud- by these pathways. In contrast, CYP1A2 and
ies and uncontrolled assessments (e.g. post- CYP2C19 activity is decreased during preg-
marketing reports). nancy, suggesting dosing reductions for agents
Observational studies have documented that metabolized via these pathways. The effect of
pregnant women take a variety of medicines pregnancy on transport proteins is unknown.
during pregnancy. While changes in drug Data are limited; more clinical studies to deter-
exposure during pregnancy are well docu- mine the effect of pregnancy on the pharma-
mented, a mechanistic understanding of these cokinetics and pharmacodynamics of drugs
effects is not clear. The few studies conducted commonly used in pregnancy are sorely
suggest that bioavailability is not altered dur- needed.
ing pregnancy, though increased plasma vol-
ume and protein binding changes can affect
Organ Impairment
the apparent volume of distribution of some
Renal Dysfunction
Table 4.2 FDA categories of drug safety during Renal failure can influence the pharmacoki-
pregnancy. netics of drugs. In renal failure, the binding of
acidic drugs to albumin is decreased, because
Category Description of competition with accumulated organic acids
and uremia-induced structural changes in
A Controlled human studies show no albumin which decrease drug binding affinity,
fetal risks; these drugs are the safest.
altering the Vd. Drugs that are more than 30%
B Animal studies show no risk to the eliminated unchanged in the urine are likely to
fetus and no controlled human studies
have been conducted, or animal have significantly diminished CL in the pres-
studies show a risk to the fetus but ence of renal insufficiency.
well-controlled human studies do not.
C No adequate animal or human studies Hepatic Dysfunction
have been conducted, or adverse fetal Drugs that undergo extensive first-pass metab-
effects have been shown in animals,
olism may have a significantly higher oral bio-
but no human data are available.
availability in patients with liver failure than
D Evidence of human fetal risk exists,
but benefits may outweigh risks in
in normal subjects. Gut hypomotility may
certain situations (e.g. life-threatening delay the peak response to enterally
disorders, serious disorders for which administered drugs in these patients.
safer drugs cannot be used or are Hypoalbuminemia or altered glycoprotein lev-
ineffective).
els may affect the fractional protein binding of
X Proven fetal risks outweigh any acidic or basic drugs, respectively. Altered
possible benefit.
plasma protein concentrations may affect the
extent of tissue distribution of drugs that nor- which enzyme activity is increased by exposure
mally are highly protein-bound. The presence to a certain drug, resulting in an increase in
of significant edema and ascites may alter the metabolism of other drugs and lower plasma
Vd of highly water-soluble agents, such as concentrations. Common inducers include bar-
aminoglycoside antibiotics. The capacity of biturates, carbamezapine, isoniazid, and
the liver to metabolize drugs depends on rifampin. In contrast, inhibition is the process
hepatic blood flow and liver enzyme activity, by which enzyme activity is decreased by expo-
both of which can be affected by liver disease. sure to a certain drug, resulting in a decrease in
In addition, some P450 isoforms are more sus- the metabolism of other drugs, and subsequent
ceptible than others to liver disease, impairing higher plasma concentrations. Common
drug metabolism. enzyme inhibitors include ciprofloxacin, flu-
conazole, metronidazole, quinidine, and valp-
Cardiac Dysfunction roic acid. Inducers and inhibitors of phase II
Circulatory failure, or shock, can alter the enzymes have been less extensively character-
pharmacokinetics of drugs frequently used in ized, but some clinical applications of this
the intensive care setting. Drug absorption information have emerged, including the use of
may be impaired because of bowel wall edema. phenobarbital to induce glucuronyl transferase
Passive hepatic congestion may impede first- activity in icteric neonates. Water-soluble drugs
pass metabolism, resulting in higher plasma are eliminated unchanged in the kidneys. The
concentrations. Peripheral edema inhibits clearance of drugs that are excreted entirely by
absorption by intramuscular parenteral routes. glomerular filtration is unlikely to be affected
The balance of tissue hypoperfusion versus by other drugs. Organic acids and bases are
increased total body water with edema may renally secreted, and can compete with one
unpredictably alter Vd. In addition, liver another for elimination, resulting in unpredict-
hypoperfusion may alter drug-metabolizing able drug disposition.
enzyme function, especially flow-dependent
drugs such as lidocaine.
Pharmacodynamics
Drug Interactions
Pharmacodynamics characterizes what the
Patients are often treated with more than one drug does to the body (i.e. the effects or
(often many) drug, increasing the chance of a response to drug therapy). Pharmacodynamic
drug–drug interaction. Pharmacokinetic modeling constructs quantitative relationships
interactions can alter absorption, distribu- of measured, physiological parameters before
tion, metabolism, and clearance. Drug inter- and after drug administration, with effects
actions can affect absorption through the defined as the changes in a physiological
formation of drug–drug complexes, altera- parameter relative to its pre-dose or baseline
tions in gastric pH, and changes in gastroin- value. Baseline refers to un-dosed state and
testinal motility. This can have a substantial may be complicated in certain situations due
impact on the bioavailability of enterally to diurnal variations. Efficacy can be defined
administered agents. The volume of distribu- numerically as the expected sum of all benefi-
tion may be altered with competitive plasma cial effects following treatment. In this case,
protein binding and subsequent changes in we refer to clinical and not necessarily eco-
free drug concentrations. nomic benefits. Similarly, toxicity can be char-
Biotransformation reactions vary greatly acterized either by the time course of a specific
among individuals and are susceptible to drug– toxic event or the composite of toxic responses
drug interactions. Induction is the process by attributed to a common toxicity.
Pharmacodynamic 57
Overview effect results only after other physiologic or

pharmacologic conditions are satisfied.
Pharmacodynamic response to drug therapy,
Direct effect relationships are easily observed
i.e. the concentration-effect relationship,
with some cardiovascular agents, whose site of
evolves only after active drug molecules reach
action is the vascular space. Pharmacologic
their intended site(s) of action. Hence, the link
effects such as blood pressure, ACE-inhibition,
between pharmacokinetic and pharmacody-
and inhibition of platelet aggregation can be
namic processes is implicit. Likewise, the
characterized by direct response relationships.
respective factors that influence various sub-
Such relationships can usually be defined by
processes (absorption, distribution, tolerance,
three typical patterns– linear, hyperbolic
etc.) are relevant and may necessitate separate
(Emax), and sigmoid Emax functions. These are
study. Differences among drug entities in phar-
shown in Figure 4.3.
macodynamic time course can be considered
In each case, the plasma concentration and
as being direct or indirect. A direct effect is
drug concentration at the effect site are propor-
directly proportional to concentration at the
tional. Likewise, the concentration–effect rela-
site of measurement, usually the plasma. An
tionship is assumed to be independent of time.
indirect effect exhibits some type of temporal
Other drugs exhibit an indirect relationship
delay, either because of differences between
between concentration and response. In this
site of action and measurement or because the
(a) (b)
Effect
Effect
Emax × Conc
Effect =
Effect = S × Conc EC50 × Conc
Drug Concentration Drug Concentration
(c)
Effect
Emax × Concn
Effect =
EC50n × Concn
Drug Concentration
Figure 4.3 Representative pharmacodynamic relationships for drugs which exhibit direct responses: linear,
hyperbolic and Sigmoid-Emax relationships shown. S is the slope of the linear response; Emax refers to the
maximum effect observed; EC50 refers to the concentration at which 50% of the maximal response is
achieved, and n is the degree of sigmoidicity or shape factor (sometimes referred to as the Hill coefficient).
case, the concentration–effect relationship is site concentrations (Ce) are used as opposed to
time-dependent. One explanation for such the plasma concentrations (Cp). In this situa-
effects is hysteresis. Hysteresis refers to the tion, a hysteresis loop is observed when plot-
phenomenon where there is a time-lapse ting Ce versus Cp.
between the cause and its effect. With respect More complicated models (indirect-response
to pharmacodynamics, this most often indi- models) have been used to express the same
cates a situation in which there is a delay in observations but typically necessitate a greater
equilibrium between plasma drug concentra- understanding of the underlying physiologic
tion and the concentration of active substance process (e.g. cell trafficking, enzyme recruit-
at the effect site (e.g. thiopental, fentanyl). ment, etc.). The salient point is that pharmaco-
Three conditions are predominantly responsi- dynamic characterization and likewise dosing
ble for hysteresis: the biophase (actual site of guidance derived from such investigation
drug action) is not in the central compartment stands to be more informative than drug con-
(i.e. plasma or blood compartment); the mech- centrations alone.
anism of action involves protein synthesis; Likewise, pharmacodynamics may be the
and/or active metabolites are present. One can discriminating characteristic that defines dose
conceptualize a hypothetical effect compart- adjustment in special populations. This is the
ment (a physical space where drug concentra- case for the observed markedly enhanced
tions are directly correlated with drug actions) sensitivity in infants compared with older
such that the relationships defined in children and adults with respect to
Figure 4.4 are only observed when the effect immunosuppressive effects of cyclosporine,
PK-PD LINK MODEL:

(a) 1st-order abs. single dose, 1CPM (b)
400
300
EFFECT SITE (Ce, ng/mL)
350
Concentration (ng/mL)
Cp (ng/mL) 250
300
Ce (ng/mL)
250 200
200 150
150
100
100
50 50
0 0
0 20 40 60 80 0 100 200 300 400
Time (hours) PLASMA (Cp, ng/mL)
(c)
120
100
80
Effect (%)
60
40
20
0
0 50 100 150 200 250 300
Ce (ng/mL)
Figure 4.4 Concentration-time, hysteresis, and effect-concentration plots (clockwise order) illustrating the
use of an effect compartment to explain observed hysteresis.
Model-Informed Drug Developmen 59
and calcium channel blocking effects on the in the implementation of model-informed

PR interval in the elderly. drug development (MIDD) principles by many
pharmaceutical and Biotech companies. The
approach is also endorsed by the global regula-
Pharmacogenomics tory community including the EMA and FDA.
The use of modeling and simulation
Pharmacogenomics is the study of how an approaches to de-risk decision making in drug
individual’s genetic inheritance affects the development is not new but the systematic
body’s response to drugs. Pharmacogenomics integration of the unique model assets in an
holds the promise that drugs might one day be evolving computing environment that expands
tailored to individuals and adapted to each per- with knowledge about candidate molecules
son’s own genetic makeup. Environment, diet, and/or vaccines is still a work in progress for
age, lifestyle, and state of health all can influ- many pharmaceutical sponsors though feed-
ence a person’s response to medicines, but back from early adopters suggests that the
understanding an individual’s genetic compo- approach can reduce both time and cost in
sition is thought to be the key to creating per- drug development when conducted in an
sonalized drugs with greater efficacy and appropriate manner. Figure 4.5 highlights
safety. Pharmacogenomics combines tradi- many of the common early drug development
tional pharmaceutical sciences, such as bio- decision milestones in conjunction with the
chemistry, with comprehensive knowledge of various model types and methodologies that
genes, proteins, and single nucleotide poly- represent key stage gate milestones.
morphisms. Genetic variations, or SNPs (sin- Many of these milestones represent contri-
gle nucleotide polymorphisms), in the human butions from Clinical Pharmacology and the
genome can be a diagnostic tool to predict a supportive quantitative disciplines which col-
person’s drug response. For SNPs to be used in laborate in the MIDD effort (e.g. bioinformat-
this way, a person’s DNA must be sequenced ics, system pharmacology, DMPK,
for the presence of specific SNPs. SNP screen- pharmacometrics, and biostatistics). As the
ings will benefit drug development; those peo- figure suggests, many of these milestones are
ple whose pharmacogenomic screening shows not only critical to the progression of drug
that the drug being tested would be harmful or and/or vaccine candidates but they also repre-
ineffective for them would be excluded from sent critical go/no go criteria requiring quanti-
clinical trials. Prescreening clinical trial sub- tative definition around the pace of potential
jects might also allow clinical trials to be outcomes. The MIDD likewise is effective at
smaller, faster, and therefore less expensive. generating scenarios which explore the space
Finally, the ability to assess an individual’s of potential outcomes either through direct
reaction to a drug before it is prescribed will experimentation or model-based projection
increase confidence in prescribing the drug (i.e. simulation).
and the patient’s confidence in taking the drug, In addition to the utility of MIDD in the
which in turn should encourage the develop- decision-making process, the MIDD imple-
ment of new drugs tested in a like manner. mentation generates modeling assets which
can be used in the later stages of drug develop-
ment. These can represent inputs to epidemio-
odel-Informed Drug
M logic modeling and simulation exercises which
Development explore the utility of projecting candidate
attributes on target populations of interest and
One of the more recent developments in the also accommodate the complexity of the exist-
evolution of Clinical Pharmacology in the ing standard of care, population, and subpopu-
facilitation of early stage drug development is lation differences influenced by socioeconomic
Clinical
Pharmacology Pharmacology
Decision Points
Screening FTIM Therapeutic Dose Pivotal Trial Drug

Criteria Dosing Window Selection Design Monograph
MIDD Outputs
Systems Biology PK/PD Pop-PK/PD Disease Progression
Modeling PBPK Modeling Modeling Modeling CTS
Integrated Models
Figure 4.5 The MIDD approach learns and confirms key characteristics of new molecular entities in a
quantitative manner, with goal of providing explicit, reproducible, and predictive evidence for optimizing
drug development plans and enabling critical decision.
and lifestyle factors. This represents a new govern how drugs behave in humans. In an era
frontier for these disciplines to further interact when more holistic approaches for the care of
and inform each other. patients are sought to maintain a healthy over-
all well-being and avoid chronic and severe
disease, clinical strategies are likely to engage
Conclusion more preventative approaches. Likewise, clini-
cal pharmacology and pharmacoepidemiology
Clinical pharmacology serves an important will be essential disciplines that discriminate
role in the development of new drugs and the strategies that are truly beneficial from those
management of pharmacotherapy. In the con- that are not or are even harmful.
text of pharmacoepidemiologic investigations,
clinical pharmacology also provides a funda-
mental backbone for understanding the Key Points
expected associations between drug therapy
and clinical benefit as well as potential toxicity. ●● There is a great need for caregivers to be
The pharmacoepidemiologist must also have skilled in the areas of drug information,
intimate knowledge of clinical pharmacology medication safety, and other aspects of
as the impact (clinical and economic) of a new pharmacy practice related to clinical
drug once available to the marketplace can pharmacology.
often be forecast based on how the agent’s clin- ●● Clinical pharmacology defines the therapeutic
ical pharmacologic attributes compare to exist- window (the dosage of a medication between
ing therapies. The connection between the minimum amount that gives a desired
utilization, compliance, the complexities of effect and the minimum amount that gives
multimodal therapy, and the associations of more adverse effects than desired effects) of a
drug behavior with disease- or population- drug in various patient populations and guides
specific indices must be defined relative to the dose modifications in various patient subpop-
known clinical pharmacologic principles that ulations (e.g. pediatrics, pregnancy, elderly,
Further Readin 61
and organ impairment) and/or dose adjust- ●● There are many factors that affect an individ-
ments for various lifestyle factors (e.g. food, ual’s response to a drug. These factors include
time of day, drug interactions). sex, age, health conditions, concomitant med-
●● Clinical pharmacology comprises all aspects ications, and genetic makeup. An important
of the scientific study of medicinal drugs in goal of pharmacoepidemiology is to use pop-
humans. It can be divided into pharmacoki- ulation research methods to characterize fac-
netics (the relationship between the dose of tors that influence individual drug response.
a drug administered and the serum or blood ●● Factors that influence individual drug
level achieved) and pharmacodynamics (the response may do so via pharmacokinetic
study of the relationship between drug level mechanisms, pharmacodynamic mecha-
and effect). nisms, or both.
Further Reading
Avorn, J. (2007). In defense of Lehmann, D.F. (2000). Observation and

pharmacoepidemiology – embracing the yin experiment on the cusp of collaboration:
and yang of drug research. N. Engl. J. Med. 357 a parallel examination of clinical
(22): 2219–2221. pharmacology and
De, V.T.P. (1993). Presenting clinical pharmacoepidemiology. J. Clin.
pharmacology and therapeutics: a problem Pharmacol. 40 (9): 939–945.
based approach for choosing and prescribing Lehmann, D.F. (2001). Improving family ties: an
drugs. Br. J. Clin. Pharmacol. 35 (6): 581–586. examination of the complementary disciplines
Etminan, M., Gill, S. et al. (2006). Challenges of pharmacoepidemiology and clinical
and opportunities for pharmacoepidemiology pharmacology. Pharmacoepidemiol. Drug Saf.
in drug-therapy decision making. J. Clin. 10 (1): 63–68.
Pharmacol. 46 (1): 6–9. Luo, X., Cappelleri, J.C. et al. (2007). A
Etminan, M. and Samii, A. (2004). systematic review on the application of
Pharmacoepidemiology I : a review of pharmacoepidemiology in assessing
pharmacoepidemiologic study designs. prescription drug-related adverse events in
Pharmacotherapy 24 (8): 964–969. pediatrics. Curr. Med. Res. Opin. 23 (5):
Evans, S.J. (2012). An agenda for UK clinical 1015–1024.
pharmacology pharmacoepidemiology. Br. J. Royer, R.J. (1992). Clinical pharmacology
Clin. Pharmacol. 73 (6): 973–978. and pharmacoepidemiology: future
Guess, H.A. (1991). Pharmacoepidemiology in challenges for the European Community.
pre-approval clinical trial safety monitoring. J. Int. J. Clin. Pharmacol. Ther. Toxicol. 30
Clin. Epidemiol. 44 (8): 851–857. (11): 449–452.
Hartzema, A.G. (1992). Pharmacoepidemiology– Suissa, S. (1991). Statistical methods in
its relevance to clinical practice. J. Clin. pharmacoepidemiology. Principles in
Pharm. Ther. 17 (2): 73–74. managing error. Drug Saf. 6 (5):
Jones, J.K. (1992). Clinical pharmacology and 381–389.
pharmacoepidemiology: synergistic Theodore, W.H. (1990). Basic principles of
interactions. Int. J. Clin. Pharmacol. Ther. clinical pharmacology. Neurol. Clin. 8 (1):
Toxicol. 30 (11): 421–424. 1–13.
Leake, C.D. (1948). Current pharmacology; Tilson, H.H. (1990). Major advances in
general principles in practical clinical international pharmacoepidemiology. Ann.
application. JAMA 138 (10): 730–737. Epidemiol. 1 (2): 205–212.
62
When Should One Perform Pharmacoepidemiologic Studies?

Brian L. Strom
Introduction case, decision-making involves weighing the

costs and risks of a therapy against its
As discussed in the previous chapters, pharma- benefits.
coepidemiologic studies apply the techniques The main costs of a pharmacoepidemiologic
of epidemiology to the content area of clinical study are obviously the costs (monetary, effort,
pharmacology. This chapter will review when time) of conducting the study itself. These
pharmacoepidemiologic studies should be per- costs clearly will vary, depending on the ques-
formed. It will begin with a discussion of the tions posed and the approach chosen to answer
various reasons why one might perform phar- them. Generally, the cost per patient in a post-
macoepidemiologic studies. Central to many marketing study, with the exception of post-
of these is one’s willingness to tolerate risk. marketing randomized clinical trials, is likely
Whether one’s perspective is that of a manu- to be at least an order of magnitude less than
facturer, regulator, academician, or clinician, the cost of a premarketing study. Other costs to
one needs to consider the risk of adverse reac- consider are the opportunity costs of other
tions, which one considers tolerable. Thus, this research that might be left undone if this
chapter will continue with a discussion of the research is performed.
difference between safety and risk. It will con- One risk of conducting a pharmacoepidemi-
clude with a discussion of the determinants of ologic study is the possibility that it could iden-
one’s tolerance of risk. tify an adverse outcome as associated with the
drug under investigation when in fact the drug
does not cause this adverse outcome. Another
risk is that it could provide false reassurances
easons to Perform
R about a drug’s safety. Both these risks can be
Pharmacoepidemiologic minimized by appropriate study designs,
Studies skilled researchers, and appropriate and
responsible interpretation of the results
The decision to conduct a pharmacoepidemio- obtained.
logic study can be viewed as similar to the reg- The benefits of pharmacoepidemiologic
ulatory decision about whether to approve a studies could be conceptualized in four differ-
drug for marketing or the clinical decision ent categories: regulatory, marketing, clinical,
about whether to prescribe a drug. In each and legal (see Table 5.1). Each will be of
Reasons to Perform Pharmacoepidemiologic Studie 63
Table 5.1 Reasons to perform i mportance to different organizations and indi-

pharmacoepidemiologic studies. viduals involved in deciding whether to initiate
a study. Any given study will usually be per-
A) Regulatory
formed for several of these reasons. Each will
1) Required be discussed in turn.
2) To obtain earlier approval for marketing
3) As a response to question by regulatory
agency Regulatory
4) To assist application for approval for Perhaps the most obvious and compelling
marketing elsewhere reason to perform a postmarketing pharma-
B) Marketing coepidemiologic study is regulatory: a plan
1) To assist market penetration by for a postmarketing pharmacoepidemiologic
documenting the safety of the drug study is required before the drug will be
2) To increase name recognition approved for marketing. Requirements for
3) To assist in re-positioning the drug postmarketing research have become pro-
a) Different outcomes, e.g. quality of life gressively more frequent in recent years. For
and economic example, in the 1970s the FDA required post-
b) Different types of patients, e.g. the elderly marketing research at the time of approval
c) New indications for about one third of drugs, a requirement
d) Less restrictive labeling which increased to over 70% in the 1990s.
4) To protect the drug from accusations Many of these required studies have been
about adverse effects randomized clinical trials, designed to clarify
C) Legal residual questions about a drug’s efficacy.
1) In anticipation of future product liability
Others focus on questions of drug toxicity.
litigation Often it is unclear whether the pharmacoepi-
D) Clinical demiologic study was undertaken in response
to a regulatory requirement or in response to
1) Hypothesis testing
merely a “suggestion” by the regulator, but
a) Problem hypothesized on the basis of
drug structure the effect is essentially the same. Early exam-
ples of studies conducted to address regula-
b) Problem suspected on the basis of
preclinical or premarketing human data tory questions include the “Phase IV” cohort
c) Problem suspected on the basis of
studies performed of cimetidine and prazo-
spontaneous reports sin. These are discussed more in Chapter 1.
d) Need to better quantitate the frequency Now that FDA has the authority to require
of adverse reactions such studies, such requirements are becom-
2) Hypothesis generating–need depends on ing more common.
whether: Sometimes a manufacturer may offer to
a) it is a new chemical entity perform a pharmacoepidemiologic study with
b) the safety profile of the class the hope that the regulatory agency might
c) the relative safety of the drug within its
thereby expedite drug approval. If the agency
class believed that any new serious problem would
d) the formulation be detected rapidly and reliably after market-
ing, it could feel more comfortable about
e) the disease to be treated, including
releasing the drug sooner. Although it is dif-
i) its duration
ficult to assess the impact of volunteered
ii) its prevalence
postmarketing studies on regulatory deci-
iii) its severity sions, the very large economic impact of an
iv) whether alternative therapies are earlier approval has motivated some manu-
available facturers to initiate such studies. In addition,
64 5 When Should One Perform Pharmacoepidemiologic Studies?
in recent years regulatory authorities have studies is the potential marketing impact of
occasionally released a particularly important those answers. In fact, some companies make
drug after essentially only Phase II testing, the marketing branch of the company respon-
with the understanding that additional data sible for pharmacoepidemiology, rather than
would be gathered during postmarketing test- the medical branch.
ing. For example, zidovudine was released for Because of the known limitations in the
marketing after only limited testing, and only information available about the effects of a
later were additional data gathered on both drug at the time of its initial marketing, many
safety and efficacy, data which indicated, physicians are appropriately hesitant to pre-
among other things, that the doses initially scribe a drug until a substantial amount of
recommended were too large. experience in its use has been gathered. A for-
Some postmarketing studies of drugs arise mal postmarketing surveillance study can
in response to case reports of adverse reac- speed that process, as well as clarify advan-
tions reported to the regulatory agency. One tages or disadvantages a drug has compared to
response to such a report might be to suggest its competitors.
a labeling change. Often a more appropriate A pharmacoepidemiologic study can also
response, clinically and commercially, would be useful to improve product name recogni-
be to propose a pharmacoepidemiologic tion. The fact that a study is underway will
study. This study would explore whether this often be known to prescribers, as will its
adverse event in fact occurs more often in results once it is publicly presented and pub-
those exposed to the drug than would have lished. This increased name recognition will
been expected in the absence of the drug and, presumably help sales. An increase in a prod-
if so, how large is the increased risk of the uct’s name recognition is likely to result par-
disease. As an example, a Medicaid database ticularly from pharmacoepidemiologic
was used to study hypersensitivity reactions studies that recruit subjects for the study via
to tolmetin, following reports about this prescribers. However, while this technique
problem to the FDA’s Spontaneous Reporting can be useful in selected situations, it is
System. extremely expensive and less likely to be pro-
Finally, drugs are obviously marketed at ductive of scientifically useful information
different times in different countries. A post- than most other alternatives available. In par-
marketing pharmacoepidemiologic study con- ticular, the conduct of a purely marketing
ducted in a country which marketed a drug exercise under the guise of a postmarketing
relatively early could be useful in demonstrat- surveillance study, not designed to collect
ing the safety of the drug to regulatory agen- useful scientific information, is to be con-
cies in countries which have not yet permitted demned. It is misleading and could endanger
the marketing of the drug. This is becoming the performance of future scientifically useful
increasingly feasible, as both the industry and studies, by resulting in prescribers who are
the field of pharmacoepidemiology are becom- disillusioned and, thereby, reluctant to par-
ing more international, and regulators are col- ticipate in future studies.
laborating more. Pharmacoepidemiologic studies can also be
useful to reposition a drug that is already on
the market, i.e. to develop new markets for the
Marketing
drug. One could explore different types of out-
As will be discussed below, pharmacoepide- comes resulting from the use of the drug for
miologic studies are performed primarily to the approved indication, for example, the
obtain the answers to clinical questions. impact of the drug on the cost of medical care
However, it is clear that a major underlying (see Chapter 18) and on the patients’ quality-
reason for some pharmacoepidemiologic of-life (see Chapter 19). One could also explore
Reasons to Perform Pharmacoepidemiologic Studie 65
the use of the drug for the approved indication not as fortunate when questions were raised
in types of patients other than those included about anaphylactic reactions caused by zome-
in premarketing studies, for example in chil- pirac. If they had the data available at the time
dren, in the elderly, or in patients with multi- of the crisis which they were eventually able to
ple comorbidities and/or taking many have, they might not have removed the drug
concomitant medications. By exploring unin- from the market. Later, Syntex recognized the
tended beneficial effects, or even drug efficacy, potential benefit, and the risk, associated with
one could obtain clues and supporting infor- the marketing of parenteral ketorolac, and
mation for new indications for drug use. chose to initiate a postmarketing surveillance
Finally, whether because of questions about cohort study at the time of the drug’s launch.
efficacy or questions about toxicity, drugs are Indeed, the drug was accused of multiple dif-
sometimes approved for initial marketing with ferent adverse outcomes, and it was only the
restrictive labeling. For example, bretylium existence of this study, and its subsequently
was initially approved for marketing in the US published results, that saved the drug in its
only for the treatment of life threatening major markets.
arrhythmias. Approval for more widespread
use requires additional data. These data can
Legal
often be obtained from pharmacoepidemio-
logic studies. Postmarketing surveillance studies can theo-
Finally, and perhaps most importantly, phar- retically be useful as legal prophylaxis, in
macoepidemiologic studies can be useful to anticipation of eventually having to defend
protect the major investment made in develop- against product liability suits (see Chapter 6).
ing and testing a new drug. When a question One often hears the phrase “What you don’t
arises about a drug’s toxicity, it often needs an know, won’t hurt you.” However, in pharma-
immediate answer, or else the drug may lose coepidemiology this view is shortsighted and,
market share or even be removed from the in fact, very wrong. All drugs cause adverse
market. Immediate answers are often unavail- effects; the regulatory decision to approve a
able, unless the manufacturer had the fore- drug and the clinical decision to prescribe a
sight to perform pharmacoepidemiologic drug both depend on a judgment about the
studies in anticipation of this problem. relative balance between the benefits of a drug
Sometimes these problems can be specifically and its risks. From a legal perspective, to win a
foreseen and addressed. More commonly, they product liability suit using a legal theory of
are not. However, the availability of an existing negligence, a plaintiff must prove causation,
cohort of exposed patients and a control group damages, and negligence. A pharmaceutical
will often allow a much more rapid answer manufacturer that is a defendant in such a suit
than would have been possible if the study had cannot change whether its drug causes an
to be conducted de novo. One example of this adverse effect. If the drug does, this will
is provided by the experience of Pfizer presumably be detected at some point. Also,
Pharmaceuticals, when the question arose the manufacturer cannot change whether the
about whether piroxicam (Feldene) was more plaintiff suffered legal damages from the
likely to cause deaths in the elderly from gas- adverse effect, that is whether the plaintiff suf-
trointestinal bleeding than the other nonsteroi- fered a disability or incurred expenses result-
dal anti-inflammatory drugs. Although Pfizer ing from a need for medical attention. However,
did not fund studies in anticipation of such a even if the drug did cause the adverse outcome
question, it was fortunate that several pharma- in question, a manufacturer certainly can doc-
coepidemiologic research groups had data ument that it was performing state-of-the-art
available on this question because of other studies to attempt to detect whatever toxic
studies that they had performed. McNeil was effects the drug had. In addition, such studies
could make easier the defense of totally the lack of experience with related drugs
groundless suits, in which a drug is blamed for makes it more likely that the new drug has pos-
producing adverse reactions it does not cause. sibly important unsuspected effects.
The safety profile of the class of drugs should
also be important to the decision about
Clinical
whether to conduct a formal screening post-
Hypothesis Testing marketing surveillance study for a new drug.
The major reason for most pharmacoepidemi- Previous experience with other drugs in the
ologic studies is hypothesis testing. The same class can be a useful predictor of what
hypotheses to be tested can be based on the the experience with the new drug in question
structure or the chemical class of a drug. For is likely to be. For example, with the finding
example, the cimetidine study mentioned that troglitazone had an increased risk of liver
above was conducted because cimetidine was disease, that became a concern as well with the
chemically related to metiamide, which had later thiazolidinediones, i.e. pioglitazone and
been removed from the market in Europe rosiglitazone. Similarly, with the finding that
because it caused agranulocytosis. rofecoxib was associated with myocardial
Alternatively, hypotheses can also be based on infarction, that became a concern as well with
premarketing or postmarketing animal or clin- celecoxib.
ical findings. For example, the hypotheses can The relative safety of the drug within its
come from spontaneous reports of adverse class can also be helpful. A drug that has been
events experienced by patients taking the drug studied in large numbers of patients before
in question. The tolmetin, piroxicam, zome- marketing and appears safe relative to other
pirac, and ketorolac questions mentioned drugs within its class is less likely to need sup-
above are all examples of this. Finally, an plementary postmarketing surveillance stud-
adverse effect may clearly be due to a drug, but ies. An extension of this approach, of course, is
a study may be needed to quantitate its fre- comparative effectiveness research.
quency. An example would be the postmarket- The formulation of the drug can be consid-
ing surveillance study of prazosin, performed ered a determinant of the need for formal
to quantitate the frequency of first dose syn- screening pharmacoepidemiologic studies. A
cope. Of course, the hypotheses to be tested drug that will, because of its formulation, be
can involve beneficial drug effects as well as used mainly in institutions, where there is
harmful drug effects, subject to some impor- close supervision, may be less likely to need
tant methodologic limitations. such a study. When a drug is used under these
conditions, any serious adverse effect is likely
Hypothesis Generating to be detected, even without any formal study.
Hypothesis generating studies are intended to The disease to be treated is an important
screen for previously unknown and unsus- determinant of whether a drug needs addi-
pected drug effects. In principle, all drugs tional postmarketing surveillance studies.
could, and perhaps should, be subjected to Drugs used to treat chronic illnesses are likely
such studies. However, some drugs may to be used for a long period of time. As such, it
require these studies more than others. This is important to know their long-term effects.
has been the focus of a formal study, which This cannot be addressed adequately in the
surveyed experts in Pharmacoepidemiology. relatively brief time available for each premar-
For example, it is generally agreed that new keting study. Also, drugs used to treat common
chemical entities are more in need of study diseases are important to study, as many
than so-called “me too” drugs. This is because patients are likely to be exposed to these drugs.
Risk Toleranc 67
Drugs used to treat mild or self-limited dis- most of this book. The latter is the focus of the
eases also need careful study, because serious following discussion.
toxicity is less acceptable. This is especially
true for drugs used by healthy individuals,
such as contraceptives. On the other hand, Risk Tolerance
when one is using a drug to treat individuals
who are very ill, one is more tolerant of toxic- Whether or not to conduct a postmarketing
ity, assuming the drug is efficacious. surveillance pharmacoepidemiologic study
Finally, it is also important to know whether also depends on one’s willingness to tolerate
alternative therapies are available. If a new risk. From a manufacturer’s perspective, one
drug is not a major therapeutic advance, since can consider this risk in terms of the risk of a
it will be used to treat patients who would have potential regulatory or legal problem that may
been treated with the old drug, one needs to be arise. Whether one’s perspective is that of a
more certain of its relative advantages and dis- manufacturer, regulator, academician, or clini-
advantages. The presence of significant adverse cian, one needs to consider the risk of adverse
effects, or the absence of beneficial effects, is reactions that one is willing to accept as toler-
less likely to be tolerated for a drug that does able. There are several factors that can affect
not represent a major therapeutic advance. one’s willingness to tolerate the risk of adverse
effects from drugs (see Table 5.2). Some of
these factors are related to the adverse out-
Safety Versus Risk come being studied. Others are related to the
exposure and the setting in which the adverse
Clinical pharmacologists are used to thinking outcome occurs.
about drug “safety”: the statutory standard that
must be met before a drug is approved for mar-
keting in the US is that it needs to be proven to Table 5.2 Factors affecting the acceptability
of risks.
be “safe and effective under conditions of
intended use.” It is important, however, to dif- A) Features of the adverse outcome
ferentiate safety from risk. Virtually nothing is
1) Severity
without some risks. Even staying in bed is asso-
2) Reversibility
ciated with a risk of acquiring bed sores!
Certainly no drug is completely safe. Yet, the 3) Frequency
unfortunate misperception by the public per- 4) “Dread disease”
sists that drugs mostly are and should be with- 5) Immediate versus delayed
out any risk at all. Use of a “safe” drug, however, 6) Occurs in all people versus just in sensitive
still carries some risk. It would be better to people
think in terms of degrees of safety. Specifically, a 7) Known with certainty or not
drug “is safe if its risks are judged to be accept- B) Characteristics of the exposure
able.” Measuring risk is an objective but proba- 1) Essential versus optional
bilistic pursuit. A judgment about safety is a 2) Present versus absent
personal and/or social value judgment about
3) Alternatives available
the acceptability of that risk. Thus, assessing
4) Risk assumed voluntarily
safety requires two extremely different kinds of
activities: measuring risk and judging the 5) Drug use will be as intended versus misuse
is likely
acceptability of those risks. The former is the
C) Perceptions of the evaluator
focus of much of pharmacoepidemiology and
Features of the Adverse Outcome between now and then, which could make that
delayed effect irrelevant or, at least, mitigate its
The severity and reversibility of the adverse
impact. Thus, a delayed adverse event may be
reaction in question are of paramount impor-
worth incurring if it can bring about beneficial
tance to its tolerability. An adverse reaction
effects today.
that is severe is much less tolerable than one
It is also important whether the adverse out-
that is mild, even at the same incidence. This is
come is a Type A reaction or a Type B reaction.
especially true for adverse reactions that result
As described in Chapter 1, Type A reactions
in permanent harm, for example birth defects
are the result of an exaggerated but otherwise
or death.
usual pharmacological effect of a drug. Type A
Another critical factor that affects the toler-
reactions tend to be common, but they are
ability of an adverse outcome is the frequency
dose-related, predictable, and less serious. In
of the adverse outcome in those who are
contrast, Type B reactions are aberrant effects
exposed. Notably, this is not a question of the
of a drug. Type B reactions tend to be uncom-
relative risk of the disease due to the exposure,
mon, are not related to dose, and are poten-
but a question of the excess risk attributed to
tially more serious. They may be due to
the drug of interest. Use of tampons is extraor-
hypersensitivity reactions, immunologic reac-
dinarily strongly linked to toxic shock: prior
tions, or some other idiosyncratic reaction to
studies have shown relative risks between 10
the drug. Regardless, Type B reactions are the
and 20. However, toxic shock is sufficiently
more difficult to predict or even detect. If one
uncommon, that even a 10-to 20-fold increase
can predict an adverse effect, then one can
in the risk of the disease still contributes an
attempt to prevent it. For example, in order to
extraordinarily small excess risk of the toxic
prevent aminophylline-induced arrhythmias
shock syndrome in those who use tampons.
and seizures, one can begin therapy at lower
In addition, the particular disease caused by
doses and follow serum levels carefully. For
the drug is important to one’s tolerance of its
this reason, all other things being equal, Type
risks. Certain diseases are considered by the
B reactions are usually considered less
public to be so-called “dread diseases,” dis-
tolerable.
eases that generate more fear and emotion
Finally, the acceptability of a risk also
than other diseases. Examples are AIDS and
varies according to how well established it is.
cancer. It is less likely that the risk of a drug
The same adverse effect is obviously less tol-
will be considered acceptable if it causes one of
erable if one knows with certainty that it is
these diseases.
caused by a drug than if it is only a remote
Another relevant factor is whether the
possibility.
adverse outcome is immediate or delayed.
Most individuals are less concerned about
Characteristics of the Exposure
delayed risks than immediate risks. This is one
of the factors that have probably slowed the The acceptability of a risk is very different,
success of anti-smoking efforts. In part, this is depending upon whether an exposure is essen-
a function of denial; delayed risks seem as if tial or optional. Major adverse effects are much
they may never occur. In addition, an eco- more acceptable when one is using a therapy
nomic concept of “discounting” plays a role that can save or prolong life, such as chemo-
here. An adverse event in the future is less bad therapy for malignancies. On the other hand,
than the same event today, and a beneficial therapy for self-limited illnesses must have a
effect today is better than the same beneficial low risk to be acceptable. Pharmaceutical
effect in the future. Something else may occur products intended for use in healthy
Risk Toleranc 69
i ndividuals, such as vaccines and contracep- attendant risk voluntarily. Some people even
tives, must be exceedingly low in risk to be accept the enormous risks of death from
considered acceptable. tobacco-related disease, but would object
The acceptability of a risk is also dependent strongly to being given a drug that was a small
on whether the risk is from the presence of a fraction as toxic. In general, it is agreed that
treatment or its absence. One could conceptu- patients should be made aware of the possibly
alize deaths from a disease that can be treated toxic effects of drugs that they are prescribed.
by a drug that is not yet on the market as an When a risk is higher than it is with the usual
adverse effect from the absence of treatment. therapeutic use of a drug, as with an invasive
For example, the six-year delay in introducing procedure or an investigational drug, one usu-
beta-blockers into the US market has been ally asks the patient for formal informed con-
blamed for resulting in more deaths than all sent. The fact that fetuses cannot make
recent adverse drug reactions combined. As a voluntary choices about whether or not to take
society, we are much more willing to accept a drug contributes to the unacceptability of
risks of this type than risks from the use of a drug-induced birth defects.
drug that has been marketed prematurely. Finally, from a societal perspective, one also
Physicians are taught primum non nocere – needs to be concerned about whether a drug
first do no harm. This is somewhat analogous will be and is used as intended or whether
to our willingness to allow patients with termi- misuse is likely. Misuse, in and of itself, can
nal illnesses to die from these illnesses without represent a risk of the drug. For example, a
intervention, while it would be considered drug is considered less acceptable if it is
unethical and probably illegal to perform addicting and, so, is likely to be abused. In
euthanasia. In general, we are much more addition, the potential for over-prescribing by
tolerant of sins of omission than sins of physicians can also decrease the acceptability
commission. of the drug. For example, in the controversy
Whether any alternative treatments are about birth defects from isotretinoin, there
available is another determinant of the accept- was no question that the drug was a powerful
ability of risks. If a drug is the only available teratogen, and that it was a very effective ther-
treatment for a disease, particularly a serious apy for serious cystic acne refractory to other
disease, then greater risks will be considered treatments. There also was no question about
acceptable. This was the reason zidovudine its effectiveness for less severe acne. However,
was allowed to be marketed for treatment of that effectiveness led to its widespread use,
AIDS, despite its toxicity and the limited test- including in individuals who could have been
ing which had been performed. Analogously, treated with less toxic therapies, and a larger
studies of toxic shock syndrome associated number of pregnancy exposures, abortions,
with the use of tampons were of public health and birth defects than otherwise would have
importance, despite the infrequency of the dis- occurred.
ease, because consumers could choose among
other available tampons that were shown to
Perceptions of the Evaluator
carry different risks.
Whether a risk is assumed voluntarily is also Finally, much depends ultimately upon the
important to its acceptability. We are willing to perceptions of the individuals who are mak-
accept the risk of death in automobile acci- ing the decision about whether a risk is
dents more than the much smaller risk of acceptable. In the US, there have been more
death in airline accidents, because we control than a million deaths from traffic accidents
and understand the former and accept the over the past 30 years; tobacco-related
iseases kill the equivalent of three jumbo jet

d Table 5.3 Annual risks of death from some
loads every day; and 3000 children are born selected hazards.
each year with embryopathy from their moth-
ers’ use of alcohol in pregnancy. Yet, these Annual death rate
(per 100 000 exposed
deaths are accepted with little concern, while Hazard individuals)
the uncommon risk of an airplane crash or
being struck by lightning generate fear. The Heart disease (US, 1985) 261.4
decision about whether to allow isotretinoin Sport parachuting 190
to remain on the market hinged on whether Cancer (US, 1985) 170.5
the efficacy of the drug for a small number of
Cigarette smoking (age 35) 167
people who had a disease which was disfigur-
Hang gliding (UK) 150
ing but not life-threatening was worth the
birth defects that would result in some other Motorcycling (US) 100
individuals. There is no way to remove this Power boat racing (US) 80
subjective component from the decision Cerebrovascular disease 51.0
about the acceptability of risks. Indeed, much (US, 1985)
more research is needed to elucidate patients’ Scuba diving (US) 42
preferences in these matters. However, this Scuba diving (UK) 22
subjective component is part of what makes Influenza (UK) 20
informed consent so important. Most people Passenger in motor vehicle 16.7
feel that the final subjective judgment about (US)
whether an individual should assume the risk Suicide (US, 1985) 11.2
of ingesting a drug should be made by that Homicide (US, 1985) 7.5
individual, after education by their physician. Cave exploration (US) 4.5
However, as an attempt to assist that judg-
Oral contraceptive user (age 4.3
ment, it is useful to have some quantitative 25–34)
information about the risks inherent in some
Pedestrian (US) 3.8
other activities. Some such information is
Bicycling (US) 1.1
presented in Table 5.3.
Tornados (US) 0.2
Lightning (US) 0.05
Conclusion Source: data derived from O’Brien (1986), Silverberg

and Lubera (1988), and Urquhart and Heilmann
(1984).
This chapter reviewed when pharmacoepide-
miologic studies should be performed. After
beginning with a discussion of the various rea-
sons why one might perform pharmacoepide- Key Points
miologic studies, it reviewed the difference
between safety and risk. It concluded with a ●● The decision to conduct a pharmacoepide-
discussion of the determinants of one’s toler- miologic study can be viewed as similar to
ance of risk. Now that it is hopefully clear the regulatory decision about whether to
when one might want to perform a pharma- approve a drug for marketing or the clinical
coepidemiologic study, the next section of this decision about whether to prescribe a drug.
book will provide perspectives on pharma- In each case, decision making involves
coepidemiology from some of the different weighing the costs and risks of a therapy
fields that use it. against its benefits.
Further Reading 71
●● The main costs of a pharmacoepidemiologic different categories: regulatory, marketing,

study are: the costs (monetary, effort, time) legal, and clinical. Each will be of impor-
of conducting the study itself, the opportu- tance to different organizations and individ-
nity costs of other research that might be left uals involved in deciding whether to initiate
undone if this research is performed, the a study. Any given study will usually be per-
possibility that it could identify an adverse formed for several of these reasons.
outcome as associated with the drug under ●● There are several factors that can affect one’s
investigation when in fact the drug does not willingness to tolerate the risk of adverse
cause this adverse outcome, and that it could effects from drugs. Some of these factors are
provide false reassurances about a drug’s related to the adverse outcome being stud-
safety. ied. Others are related to the exposure and
●● The benefits of pharmacoepidemiologic the setting in which the adverse outcome
studies could be conceptualized in four occurs.
Further Reading
Binns, T.B. (1987). Therapeutic risks in Mattison, N. and Richard, B.W. (1987).
perspective. Lancet 2: 208–209. Postapproval research requested by the FDA
O’Brien, B. (1986). “What Are My Chances at the time of NCE approval, 1970–1984. Drug
Doctor?”–A Review of Clinical Risks. London: Inf. J. 21: 309–329.
Office of Health Economics. Rogers, A.S., Porta, M., and Tilson, H.H. (1990).
Bortnichak, E.A. and Sachs, R.M. (1986). Guidelines for decision making in
Piroxicam in recent epidemiologic studies. postmarketing surveillance of drugs. J. Clin.
Am. J. Med. 81: 44–48. Res. Pharmacol. 4: 241–251.
Feldman, H.I., Kinman, J.L., Berlin, J.A. et al. Rossi, A.C. and Knapp, D.E. (1982). Tolmetin-
(1997). Parenteral ketorolac: the risk for induced anaphylactoid reactions. N. Engl. J.
acute renal failure. Ann. Intern. Med. 126: Med. 307: 499–500.
193–199. Silverberg, E. and Lubera, J.A. (1988).
Hennessy, S., Kinman, J.L., Berlin, J.A. Cancer statistics. CA Cancer J. Clin.
et al. (1997). Lack of hepatotoxic effects 38: 5–22.
of parenteral ketorolac in the hospital Stallones, R.A. (1982). A review of the
setting. Arch. Intern. Med. 157: epidemiologic studies of toxic shock
2510–2514. syndrome. Ann. Intern. Med.
Humphries, T.J., Myerson, R.M., Gifford, L.M. 96: 917–920.
et al. (1984). A unique postmarket outpatient Strom, B.L., Berlin, J.A., Kinman, J.L. et al.
surveillance program of cimetidine: report on (1996). Parenteral ketorolac and risk of
phase II and final summary. Am. J. gastrointestinal and operative site bleeding: a
Gastroenterol. 79: 593–596. postmarketing surveillance study. JAMA 275:
Joint Commission on Prescription Drug Use 376–382.
(1980) Final Report. Washington, DC. Strom, B.L., Carson, J.L., Morse, M.L. et al.
Lowrance, W.W. (1976). Of Acceptable Risk. (1987). The effect of indication on
Los Altos, CA: William Kaufmann. hypersensitivity reactions associated with
Marwick, C. (1988). FDA ponders approaches to zomepirac sodium and other nonsteroidal
curbing adverse effects of drug used against antiinflammatory drugs. Arthritis Rheum. 30:
cystic acne. JAMA 259: 3225. 1142–1148.
Strom, B.L., Carson, J.L., Schinnar, R. et al. postmarketing drug surveillance studies for
(1988). The effect of indication on the risk of promotional purposes. Clin. Pharmacol. Ther.
hypersensitivity reactions associated with 48: 598.
tolmetin sodium vs. other nonsteroidal Urquhart, J. and Heilmann, K. (1984). Risk
antiinflammatory drugs. J. Rheumatol. 15: Watch–The Odds of Life. New York: Facts on
695–699. File.
Strom, B.L. and members of the ASCPT Young, F.E. (1988). The role of the FDA in the
Pharmacoepidemiology Section (1990). effort against AIDS. Public Health Rep. 103:
Position paper on the use of purported 242–245.
73
Views from Academia, Industry, Regulatory Agencies,

and the Legal System
Joshua J. Gagne1,2, Jerry Avorn1, Nicolle M. Gatto3, Jingping Mo4, Gerald J. Dal Pan5,
June Raine6, Shinobu Uzu7, Aaron S. Kesselheim1, and Kerstin N. Vokinger 1,8
1
Harvard Medical School and Brigham and Women’s Hospital, Boston, MA, USA
2
Johnson and Johnson, New Brunswick, NJ, USA
3
Aetion Inc., New York, NY, USA
4
Epidemiology, Worldwide Research & Development, Pfizer Inc., New York, NY, USA
5
Office of Surveillance and Epidemiology, Center for Drug Evaluation and Research, US Food and Drug Administration, Silver Spring,
MD, USA
6
Vigilance and Risk Management of Medicine, Medicines and Healthcare Products Regulatory Agency, London, UK
7
Pharmaceuticals and Medical Devices Agency, Tokyo, Japan
8
Institute for Law University of Zurich, Zurich, Switzerland
health care system, and particularly in aca-

The View from Academia demic medical centers.
Once a drug is approved for marketing, it
Introduction
enters a complex health care system in which
Every year prescribers and patients have more its prescription, its use by patients, and its out-
medications at their disposal, each with its comes often go largely unassessed. Until
own efficacy, side effects, and cost. When a recently, scant attention has been paid to sys-
new drug is introduced, its benefit-to-risk rela- tematic surveillance of these actions, except
tionship is often understood in only a prelimi- for the atypical settings of some integrated
nary way, as is its cost-effectiveness. This health care delivery systems. The prevailing
provides a limited perspective on how it ideally view has been that after the US Food and Drug
should be used. High-profile withdrawals of Administration (FDA) or comparable national
drugs for safety reasons, along with prominent authority approves a drug, it is used at the dis-
warnings about widely-used medications that cretion of the clinician, with little formal fol-
remain on the market, have caused physicians, low-up of the appropriateness or consequences
patients, and policymakers to become more of such decisions. The problem is made more
aware of drug safety concerns. At the same acute by the fact that many regulatory agencies
time, health care systems all over the globe purposely (and often by statute) do not base
are struggling with how to provide the most their approval decisions on a medication’s clin-
appropriate care in the face of rising costs ical or economic value compared to similar
and increasingly tight fiscal constraints. products; often superiority over placebo is suf-
Pharmacoepidemiology can serve as a key tool ficient for a drug to be approved. In addition, it
for helping to address all of these concerns. is generally no one’s responsibility (other than
These issues are growing throughout the the harried prescriber) to determine how
74 6 Views from Academia, Industry, Regulatory Agencies, and the Legal System
f aithfully patients are adhering to the pre- edical care delivery and health services
m
scribed regimen. Increasingly, more attention research, from the perspective of academia.
is being paid to assessing the outcomes of med-
ication use on a population level, considering
The Drug Approval Process
what its useful and harmful outcomes are
when it is taken by hundreds, thousands, or Each national health care system must grapple
even millions of patients rather than by single with the following inherent paradox of phar-
individuals in a clinical trial or in routine prac- macology: A new therapy must be evaluated
tice. It is now widely appreciated that some for approval when the available data on its
adverse events can be identified and their risk benefits and harms are still modest. Yet wait-
quantified only by observing a drug’s use in ing until “all the evidence is in” can pose its
large numbers of patients. The best perspective own public health threat, if this prevents an
on the impact of a medication on the health of important new treatment from being used by
the public requires measuring those outcomes patients who need it. Since any medication
in the health care system itself, rather than one that is effective is bound to have some adverse
person at a time. It is here that the insights of effect in some organ system in some patients at
pharmacoepidemiology are playing an increas- some doses, any approval must by definition be
ingly central role. based on a judgment that a drug’s efficacy is
Driven by the pressures noted above, this “worth it” in light of the known risks of the
situation is evolving, with growing apprecia- treatment. However, the trials conducted by a
tion of several important problems, each of given drug manufacturer to win approval are
which can be informed by the methods and often powered statistically (see Chapter 3) to
tools of pharmacoepidemiology: (i) medica- demonstrate success for that single product in
tions that seem acceptably safe on approval achieving a pre-specified therapeutic end-
may prove to have important risks which were point. Especially when this is demonstration of
under-appreciated at the time of approval; (ii) superiority over placebo, and/or when the
in typical practice, clinicians often make pre- required endpoint is reaching a surrogate out-
scribing decisions that do not reflect the best come (e.g. a change in a laboratory test such as
evidence-base or guideline recommendations; hemoglobin A1c or low density lipoprotein
(iii) even this evidence base is often thinner [LDL] cholesterol), the number of subjects
than it should be because head-to-head com- required for these exercises, and the duration
parisons of drug effectiveness or safety – either of the studies, are often inadequate to reveal
trial-based or observational – have not been important safety problems if they are present.
done; (iv) as a result, inadequate information This is exacerbated by the extensive exclusion
is available to inform decisions about which criteria for study participation, a particular
drugs work best, or most cost-effectively, for problem for high-risk populations such as the
specific indications; and (v) patients fre- frail elderly, pregnant women, and children
quently fail to take their medications as (see also Chapter 23).
directed. As a result, additional methods need to be
Pharmacoepidemiology is the core discipline applied even to pre-approval data to aggregate
required for a rigorous understanding of each adverse events from multiple study popula-
of these areas, and to guide the development tions to provide the power needed to assess
and evaluation of programs to address them. safety. Meta-analysis (see Chapter 20) of
Many of these topics are discussed in detail in adverse effects data from multiple pre-approval
the chapters that follow; this chapter provides trials represents the first opportunity to use
an overview of how the field and its methods these tools to inform the appropriate use of
can contribute to these larger themes in medications. This makes it possible to combine
The View from Academi 75
findings from different smaller studies – many physicians, and when they have, they are often
of them conducted before the drug in question not prescribed an adequate regimen to control
was approved – to produce evidence of poten- their risks, or even any regimen at all.
tial harm for drugs. Pharmacoepidemiology makes it possible to
These shortcomings of pre-marketing trials identify more clearly when and how prescrib-
are likely to become even more salient as regu- ing falls short, insights which can then be used
lators move toward alternative drug approval to shape programs to improve care (see below).
processes. The European Medicines Agency When medications are used, there is good
(EMA) recently completed a pilot project to evidence that clinicians frequently do not pre-
explore approval via adaptive pathways, which scribe regimens that are optimal, based on the
are intended to provide earlier and progressive available clinical evidence, or prescribe medi-
patient access to new drugs. The US FDA cations that may interact with other drugs a
maintains several expedited regulatory path- patient takes, or choose more expensive drugs
ways. Post-approval monitoring of drugs when comparable generic preparations would
approved through these pathways is even more work as well and be much more affordable.
critical, because safety information is more Pharmacoepidemiology makes it possible to
likely to emerge in the post-marketing setting assess the distribution of drugs used for a
for drugs approved by these pathways as com- given indication by clinician, practice, or sys-
pared to conventional pathways. In 2016 the tem and can account for contraindications
US Congress enacted the 21st Century Cures and compelling indications related to specific
Act which included, among other sections, drug choices, to refine the assessment of the
provisions that modify the data required for appropriateness of prescribing in an entire
FDA approval. In particular, the law promotes health care system, or for individual clini-
the use of biomarkers and surrogate measures cians. One study assessed all hypertension-
to support the approval of new drugs as well as related medication use and diagnoses in one
“real world evidence” from observational data large state-funded program of medications for
to support supplemental indications for exist- the elderly. The availability of clinical infor-
ing products. As these new provisions are mation made it possible to determine how
implemented, pharmacoepidemiology will well the regimen of each patient conformed to
have an even greater role in generating the guideline recommendations. This study found
“real world evidence” for supplemental indica- that a substantial proportion of treated hyper-
tions and will be increasingly relied upon to tensive patients were not receiving a regiment
evaluate the impact on clinical outcomes, both consistent with guidelines. Often, such subop-
beneficial and harmful, of new drugs approved timal prescribing involved omissions of an
on less rigorous evidence. indicated class (e.g. angiotensin converting
enzyme inhibitors in patients with diabetes
mellitus), or use of a calcium channel blocker
Prescribing Practices
when a beta-blocker would have been more
Once a drug has entered the health care deliv- appropriate (e.g. in a hypertensive patient who
ery system, the tools of pharmacoepidemiol- has had a myocardial infarction). Another
ogy can be used to document areas in which analysis reviewed all clinical encounters of
prescribing falls short of existing knowledge. patients who had filled prescriptions for clopi-
First is the issue of underprescribing. Studies of dogrel and found that about half did not have
many important chronic diseases such as any evidence of conditions (such as coronary
hypertension, hypercholesterolemia, and dia- artery stenting) for which the drug had an
betes reveal that many patients with these con- approved indication, or any other evidence-
ditions have not been diagnosed by their based reasons for its use.
More sophisticated health records systems evolved to define the two aspects of this prob-
are becoming available each year that integrate lem: secondary non-adherence refers to the fail-
pharmacy data with information from clinical ure by a patient to continue to use a medication
laboratories, electronic health records, regis- already begun; primary non-adherence occurs
tries, and sources of patient-reported informa- when a clinician writes a prescription that the
tion to measure the adequacy of use of patient does not even fill once. The magnitude
cholesterol-lowering agents, diabetes drugs, of both primary and secondary non-adherence
antihypertensives, and other drugs. This makes is substantial, and varies by drug class as well
it possible to assess the effectiveness of pre- as country.
scribing outcomes for a given clinician (or
practice or system), by measuring how well
Assessment of the Quality
target metrics such as normotension or goal
and Outcomes of Medication Use
LDL cholesterol or hemoglobin A1c are being
in Populations
achieved. In all these analyses, pharmacoepi-
demiology makes it possible to evaluate the Much attention is now being paid to the assess-
appropriateness of medication use in selected ment of the outcomes of medication use in
populations, even if it cannot with certainty typical “real-world” populations. This perspec-
determine whether a given prescription in a tive is based on the difference between efficacy,
particular patient was the best choice. the effect of a medication in the rigorous but
idealized setting of a clinical trial, compared to
its effectiveness, a measure of its outcomes in
Evaluation of Patients’ Use
typical practice settings. These often differ. For
of Drugs in the Health Care System
example, one important conventional rand-
Even when a medication is appropriately pre- omized trial demonstrated convincingly that
scribed, patients may underuse it or use it in addition of spironolactone to the regimen of
unsafe ways. Underuse of needed drugs by patients with congestive heart failure substan-
patients is one of the most common tially improved their clinical status and
medication-related problems, and one that can reduced mortality. However, a population-
be readily identified by pharmacoepidemiol- based analysis later found that when these
ogy (see also Chapter 21). findings were applied in routine practice by
Because many assessments of underuse are typical physicians treating a much larger num-
based on pharmacy-generated data on filled ber of typical patients, there was a significant
prescriptions, it is sometimes difficult to know increase in hyperkalemia-associated morbidity
whether non-use of an indicated drug was the and mortality. By contrast, an analysis of pre-
result of a failure of the patient to fill a prescribers’ response to a different study that pro-
scription, or the failure of the clinician to write vided new evidence about optimal management
it. Electronic prescribing makes it possible to of atrial fibrillation demonstrated a more posi-
define this problem more precisely. One large tive change in practice.
study found that a fourth of initial prescrip- Other analyses document that despite over-
tions written electronically were never picked whelming randomized trial evidence showing
up at the pharmacy. As a result, the approxi- the efficacy of warfarin use in preventing
mately 50% rate of non-adherence seen over stroke in patients with atrial fibrillation,
time in pharmacoepidemiologic datasets based population-based studies of older patients liv-
on filled prescriptions is a best-case scenario, ing in nursing homes revealed a surprisingly
as it does not even take into account the addi- low prevalence of use of this therapy. Such
tional millions of regimens that are not even underuse was found to be associated with phy-
initiated by the patient. Terminology has sicians’ recent experience with adverse events
caused by the drug, as well as by their percep- cost of increased morbidity and healthcare uti-
tions and attitudes about risks and benefits. lization. However, a careful time-series analy-
This kind of real-world population research sis of all medication use, physician visits, and
can lay the foundation for enlightened inter- hospital care in the province before and after
ventions to address such non-use, by taking on policy implementation provided compelling
its underlying causes. evidence that the new reimbursement system
Pharmacoepidemiologic methods can also produced no important clinical downsides, but
be used to track the diffusion of new medica- did achieve substantial savings for the provin-
tion classes into practice, as well as the reac- cial health-care budget. Such observational
tion of practitioners in various settings to new methods have also been combined with
information about the comparative benefits population-based randomized policy trials,
and risks of drug. Acknowledging the gap and were found to yield similar results.
between the characteristics of clinical trial par-
ticipants and those who often use a medication
Interventional
in practice, methods are also being developed
Pharmacoepidemiology
and applied to generalize trial results to more
typical patient populations. For example, the Although epidemiology is traditionally seen as
newer oral anticoagulant, dabigatran, was a merely observational discipline, it can also be
approved on the basis of a large randomized used for what might be called “interventional
trial comparing it to warfarin, an older oral epidemiology” – in this case, using the tools of
anticoagulant. A simulation-based approach pharmacoepidemiology to define baseline
was used to assess how the comparative bene- medication use, to direct the implementation
fits and risks would translate to cohorts of of programs to improve such use, and then to
patients who use these drugs outside of the employ the same rigorous ascertainment of
randomized trial and in usual routine care. practice patterns and clinical events to evalu-
Such an approach preserves the strengths of ate the effectiveness of those interventions.
the randomized trial but makes the results One example of such interventional phar-
more useful to patients and clinicians. macoepidemiology has been the development,
testing, and widespread deployment of the
form of educational outreach known as “aca-
Policy Analysis
demic detailing.” The academic detailing
Usually, policy changes are implemented in approach uses the engaging interactive out-
the health care system with no systematic reach of the pharmaceutical industry, but puts
plans for their evaluation, and no follow-up in the service of transmitting messages based
studies of their impact. Such changes in bene- solely on evidence-based recommendations of
fit design are often applied to medication use. optimal prescribing, developed by academic
However, even if a policy is changed in a way physicians. Building on pharmacoepidemio-
that does not anticipate an evaluation, logic assessment of overall prescribing pat-
population-based observational studies after terns in a given area, the method was then
the fact can still yield important conclusions tested in several population-based randomized
concerning its effects, both good and bad. For trials in which it was shown to be effective in
example, when the Canadian province of improving prescribing, as well as in reducing
British Columbia implemented a reference- unnecessary medication expenditures.
pricing policy for antihypertensive medica- As computerized drug and medical data
tions in which it reimbursed only the cost of an have matured, their role has expanded to sup-
effective generic drug in several classes, critics port large-scale, multi-center, pragmatic rand-
charged that any savings would come at the omized trials of medications themselves. In
the Salford Lung Study, investigators rand- (IOM) has been promoting the idea of a
omized over 4000 typical patients with asthma Learning Healthcare System in which the data
to receive an inhaled combination of a beta- generated within a medical center are ana-
agonist and a corticosteroid or to usual care. lyzed and used to improve the delivery of care
The trial was conducted in more than six dozen within the system. The science of pharmacoep-
general practice clinics in the UK, using an idemiology is central to the collection, analy-
integrated electronic health record system that sis, and interpretation of the data generated
enabled the investigators to collect study data and used in this continuous feedback loop for
during the course of the trial with little addi- several reasons, including its capacity to rigor-
tional interaction required between patients ously specify treatment exposures and out-
and trial staff. comes, and its perspective that takes into
account the concept of “population at risk.”
For academic medical centers that evolve in
Economic Assessment
the coming years to become the hubs of com-
of Medication-Related Issues
prehensive accountable care organizations, the
Using population-based datasets that contain availability of such data and investigator teams
information on expenditures as well as utiliza- will make it possible to use these epidemio-
tion makes it possible to assess the economic logic tools to study – and improve – the pat-
impact of such prescribing issues as well (see terns of use and outcomes of medications
Chapter 18). The above study of patients across the entire inpatient-outpatient contin-
treated for hypertension, for example, found uum of care.
that better adherence to the guideline recom-
mendations would not only have led to more
Consortia of Academic Medical
evidence-based prescribing (and therefore bet-
Center Programs
ter clinical outcomes), it would also have
for Pharmacoepidemiologic
resulted in savings of $1.2 billion annually if
Research
the findings were projected nationally.
Similarly, the clopidogrel-use study suggested As the field of pharmacoepidemiology matures,
that if aspirin had been substituted in patients new collaborations are emerging to enhance the
who lacked an evidence-based or FDA- capacity of the health care delivery system and
approved indication for use of the more costly of academic centers to address important ques-
drug, it would have saved $1.5 billion at a tions in medication use. Such collaborations
national level. can bring together large groups of patients for
study, increasing the size of populations availa-
ble for research, as well as their diversity and
The Academic Medical Center
representativeness. Equally important, such
The academic medical center represents a spe- consortia can bring together the expertise of
cial case of inquiry for pharmacoepidemiology, several groups whose skills may be complemen-
and one where the field can make particularly tary in addressing the difficult methodologic
useful contributions. These centers are the issues inherent in observational studies of drug
home base for many researchers in the field, use and outcomes. The EMA has created
and such settings are more likely than many ENCePP, the European Network of Centres for
routine practices to have available the elec- PharmacoepidemiologyandPharmacovigilance.
tronic datasets that make such analyses possi- The project has developed an inventory of
ble. In recent years, the Institute of Medicine European research centers and data sources in
pharmacoepidemiology and pharmacovigi- The Future

lance, and provides a public index of such
The continuing evolution of health care sys-
resources. ENCePP has also developed an elec-
tems in both the industrialized and the devel-
tronic register of studies that provides a publicly
oping worlds will bring about a growing role
accessible means of identifying all registered
for pharmacoepidemiology in multiple set-
ongoing projects in pharmacoepidemiology and
tings. Many new medications have novel effi-
pharmacovigilance. In order to be registered
cacy but also daunting risks of toxicity, and
and receive formal ENCePP approval, study
often enormous costs. Health care systems all
investigators must agree to a Code of Conduct,
over the world face pressures to provide only
which sets forth a set of principles for such stud-
those interventions that have the best effi-
ies concerning methodologic practices and
cacy and safety, but also at a price. To accom-
transparency; they must also agree to adhere to
plish this will require relying on more than
a checklist of methodologic standards.
manufacturers’ assessments of the utility,
Examples in the US include the FDA’s
safety, or economic value of their own prod-
Sentinel system, the Centers for Disease
ucts, and more than clinicians’ received wis-
Control and Prevention’s (CDC’s) Vaccine
dom or traditional prescribing habits. Nor
Safety Datalink (VSD), and the Patient-
will the interest of some insurers in promot-
Centered Outcome Research Institute’s
ing use of the most inexpensive medications
(PCORI) PCORnet. Sentinel is FDA’s national
necessarily lead to optimal outcomes
monitoring system that brings together a large
clinically, economically, or ethically.
number of electronic health care data providers
Pharmacoepidemiology (and its related disci-
and academic investigators to conduct post-
pline, pharmacoeconomics) can provide the
approval safety surveillance of FDA-regulated
tools for rigorous assessment of the good and
medical products. The VSD, which is a collabo-
harm that specific medications provide, and
rative project between the CDC and health care
hold the promise of applying science to ther-
organizations that provide data and scientific
apeutic decisions that are still too dominated
expertise, is a precursor to Sentinel that focuses
by other forces.
on vaccine safety surveillance, such as monitor-
ing the safety of the seasonal influenza vaccine.
A product of the healthcare reform program
enacted in 2010, PCORI was designed to pro-
Summary Points for the View
vide funding for comparative effectiveness
from Academia
research (CER), which was to include the study
of medications, often by means of observa- ●● Pharmacoepidemiology has a growing role
tional studies. PCORnet is PCORI’s collabora- in providing insights into therapeutic deci-
tive network of health systems, clinicians, sions as routinely collected health care data
researchers, patients, and data intended to fos- can be analyzed by increasingly powerful
ter patient-centered research across various software and hardware, combined with
health systems. However, those who expected emerging sophisticated epidemiologic
PCORI to function as a CER resource that methods.
would fund trial or observational studies com- ●● Beyond defining the benefits and risks of
paring relevant treatment options head-to-head therapeutics, these tools can point to how
have been surprised at what a small proportion best to maximize benefits, reduce risks, and
of its activities have supported such work. contain costs.
Case Example 6.1

The view from academia: The role of aca- harmacoepidemiology with evidence-
p
demia in developing, implementing, and based medicine expertise and social mar-
evaluating innovative programs to improve keting methods to create interventions
medication use (Avorn 2011). that bridge the gap between therapeutics
Background knowledge and practice.
●● Practitioners have difficulty keeping up
●● Such programs can help offset their costs
with important new findings about drug by improving outcomes and reducing use
risks and benefits. As a result, there is a of therapies that are not cost-effective.
knowledge gap between what is known Limitations
and what is practiced. ●● Requires cooperation among stakeholders
●● Pharmacoepidemiology can play a more with different goals and perspectives
prominent role in defining prescribing within a fragmented health care system.
patterns and their outcomes. ●● Lack of cooperation among different pay-
Question ers can reduce incentives to operate such

How can researchers in academia improve programs and sustain their costs.
the use of therapeutics to maximize patient Key Points
benefit and minimize the likelihood of harm? ●● As health care organizations become more
Approach centered on enhancing outcomes and

●● Academic researchers have expertise in value rather than on maximizing volume,
analyzing population-based data to identify opportunities are increasing for academic
patterns of medication use. They also have researchers to “push out” innovative user-
a broad grasp of the overall clinical litera- friendly educational programs to improve
ture; train future practitioners and research- medication use.
ers; develop and evaluate therapies; deliver ●● Pharmacoepidemiology can play a central
care to patients; and include thought lead- role in such work, by defining patterns of
ers who influence national policy. medication use, assessing risks and bene-
●● Such expertise can be used to understand fits, and evaluating changes in prescribing
the outcomes of medications and improve and outcomes in populations of patients.
their use. ●● Health care organizations face pressing
●● Programs of “academic detailing” devised needs to develop data repositories and

by academic researchers can adopt the local expertise in the analysis of medica-
effective outreach methods of pharmaceu- tion use and outcomes. Academic medical
tical companies for noncommercial pur- centers can set the example by organizing
poses by visiting physicians in their offices their own data and providing access to it to
to promote optimal prescribing practices. improve quality and therapeutic strategies.
●● Academic medical centers can leverage their
Results combined missions of education, health care
Many randomized controlled trials have dem- delivery, and research to move national prac-
onstrated that this approach is effective in tice toward better use of therapeutics.
providing prescribers with better information ●● Academic researchers are well suited to
about benefits, risks and appropriate use of translate medical research findings on
therapeutics, and changing their prescribing. drug benefits, risks, and cost-effectiveness
Strengths into educational outreach programs that
●●This approach enables those in academia can be acted upon by policy-makers, prac-
to combine the analytic tools of titioners and the public.
The View from Industr 81
The View from Industry Regulatory and Industry Focus on Risk

Management and Epidemiology
Note Biopharmaceutical risk management (see also
The views expressed are those of the authors, Chapter 23) is fundamentally concerned with
which are not necessarily those of Aetion Inc. or preserving an appropriate benefit–risk balance
Pfizer Inc. among patients using a medicine, vaccine, or
device. There are many tools by which this goal
Introduction
can be achieved, but risk assessment and risk
Epidemiology is recognized as a key compo- mitigation are the two primary components of
nent of risk management and safety assess- risk management. Epidemiologists play a vital
ment activities during pre- and post-approval role in the quantification and interpretation of
drug development. In addition to risk manage- risk. Pre-approval, they contextualize risks
ment, epidemiology contributes to several emerging from clinical studies by understand-
other important functions within a biophar- ing the background rates of clinical outcomes
maceutical company, including portfolio of interest in the indicated population. Post-
development, the commercialization of drugs, approval, they assess the safety of drugs as
and benefit–risk assessments. The use of epi- used in actual clinical practice. Epidemiologists’
demiology to support the appropriate market- training in observational research, data analy-
ing of drugs, including health economics, and sis and interpretation, and survey and program
benefit–risk assessment, are discussed else- design also contributes to effective risk mitiga-
where in this book (see Chapters 18 and 23). tion program planning and assessment.
The most consistent contribution of epidemi-
ology in the biopharmaceutical industry is
arguably drug safety evaluation, including the The Evolution of Biopharmaceutical
contextualization and refinement of safety sig- Risk Management
nals, and examination of specific research The guidance and regulations around risk
hypotheses. To meet these aims, epidemiolo- management have evolved since the 1990s.
gists design and implement background epide- Public pressure to speed drug approvals for
miology studies among indicated populations, HIV and cancer drugs led to the Prescription
risk management interventions and evalua- Drug User Fee Act (PDUFA) in the US. Ten
tions, and post-approval safety studies. years later, concern that speed might come at
Additionally, epidemiologists contribute con- the expense of fully evaluating safety led to
tent, expertise, and strategy to regulatory docu- the inclusion of a risk management frame-
ments such as global Risk Management Plans work for safety assessment in PDUFA III in
(RMP), Pediatric Investigation Plans (PIP), and 2002. For the first time, dedicated funding
orphan drug applications, and are key contrib- was provided to the FDA for risk manage-
utors in interactions with regulatory authori- ment resources. In response to this regula-
ties. This section discusses the specific tion, the FDA issued three guidance
application of pharmacoepidemiology to safety documents in 2005: (i) Pre-Marketing Risk
assessment throughout the development lifecy- Assessment, (ii) Pharmacovigilance and
cle from the perspective of epidemiologists Pharmacoepidemiology, and (iii) Risk
working within the biopharmaceutical indus- Minimization Action Plans (RiskMAPs).
try. At the end of this section, we include a case After a number of widely used drugs
example of the epidemiology strategy imple- were withdrawn in 2004 and 2005 for safety
mented to support the development, approval, reasons, the public questioned the effective-
and post-approval activities for Xeljanz® (tofaci- ness of the FDA’s methods to assess and
tinib), a Janus kinase (JAK) inhibitor for treat- approve drugs. The IOM was tasked with eval-
ment of rheumatoid arthritis (RA). uating the US drug safety system and making
r ecommendations for improvements to risk harmacovigilance legislation introduced a

p
assessment, safety surveillance, and the safer pharmacovigilance system master file (PSMF),
use of drugs. The IOM committee made required RMPs for all new products, enhanced
numerous recommendations, several of which post-authorization measures with legally bind-
pertained to epidemiologists, including: pro- ing post-authorization safety studies (PASS),
vide the FDA additional funding and staff; including evaluation of the effectiveness of
improve communications on drug safety, additional risk minimization measures
including a larger role for the drug safety staff; (aRMMs), and post-authorization efficacy
and most importantly, be given additional studies (PAES). The new EU pharmacovigi-
authority and enforcement tools. lance legislation also introduced clarity in the
As a result of the IOM report and other stake- oversight by the authorities for non-
holder research and advocacy, Congress passed interventional studies: the national competent
the Food and Drug Administration Amendment authority is responsible for nationally author-
Act (FDAAA) in 2007, which further strength- ized products, EMA and its Pharmacovigilance
ened the FDA’s oversight of risk management and Risk Assessment Committee (PRAC) has
activities. With FDAAA, the FDA was granted the oversight responsibility when more than
the ability to mandate post-approval studies one member state is involved. To facilitate the
(Postmarketing Requirements, or PMR) and performance of pharmacovigilance in accord-
Risk Mitigation Evaluation Strategies (“REMS”; ance with the new EU legislation, the EMA
see section later in this chapter for further infor- developed good pharmacovigilance practices
mation) by imposing substantial fines for non- (GVP) modules. The modules that are most rel-
compliance or denial/revocation of drug evant to epidemiologists are Module VIII PASS
approval. FDAAA also allowed for voluntarily and Module XVI Risk minimization measures:
post-marketing commitments (PMC), i.e. stud- selection of tools and effectiveness.
ies that may not necessarily be required but Besides US and EU, regulations on risk man-
could provide important public health informa- agement planning, including post-approval
tion. Observational studies could be either safety studies, are evolving in other parts of the
PMRs or PMCs, and are further described in the world, such as Asia and Latin America. In
FDA Guidance for Industry Postmarketing Japan, post-marketing surveillance (PMS) is
Studies and Clinical Trials. required for newly approved medicine and must
Europe passed similar legislation in 2005, be conducted in according with the good post-
The Rules Governing Medicinal Products in marketing study practice (GPSP), a set of stand-
the European Union-Volume 9A, which pro- ards unique to Japan. The GPSP ordinance
vide guidelines on pharmacovigilance and risk mandates PMS studies, commonly known as
management between companies and the drug use results surveys (DURSs), and defines
EMA. EU law requires companies to submit a the approach for DURS conduct. There is little
formal RMP with each marketing authoriza- flexibility in the design and format; protocol
tion application (MAA). Following a review of finalization and approval are usually stream-
the European system of safety monitoring as lined processes. Japan’s Pharmaceuticals and
well as extensive public consultation, a Medical Devices Agency (PMDA) has been
Directive and Regulation (also called new EU working to strengthen its drug safety assess-
pharmacovigilance legislation) were adopted ment framework. The Medical Information for
by the European Parliament and Council of Risk Assessment Initiative (MIHARI) project
Ministers in December 2010, which became was initiated in 2009 with the aim of utilizing
effective in July 2012, bringing about signifi- large-scale electronic health information data-
cant changes in the safety monitoring of bases as novel information sources for pharma-
medicines across the EU. The new EU coepidemiologic drug safety assessments in
Japan. After conducting extensive pilot studies, (see Chapter 7). Clinical trials and spontaneous
the framework was implemented into practical reports are useful and have a unique place in
applications in 2014 and is expected to play an assessing drug safety (e.g. signal detection).
important role in Japan’s pharmacovigilance However, both sources have limitations that
and risk management in the future. can be addressed, in part, by the proper use of
In China, policies for post-approval safety observational epidemiology. Epidemiologic
studies, known as intensive monitoring, are still studies complement these two sources of data
evolving, and the available guidance and over- to refine the safety signals they generate and to
all approach are not as comprehensive as in provide a more comprehensive and pragmatic
the US, EU, or Japan. However, the basis for picture of the safety profile of a drug as it is
intensive monitoring has evolved over the past used in clinical practice.
decade, and provisions for the ideas of post-
marketing re-evaluation and re-registration Contributions of Pre-approval
are delineated in China’s FDA regulations. In Epidemiology
Mexico, Federal Commission for Protection Before evaluation of a potential medicine can
against Health Risks (COFEPRIS) is the regu- begin, extensive pre-clinical research is con-
latory authority for pharmaceuticals. The ducted, involving lengthy in vitro and in vivo
National Center of Pharmacovigilance within testing. Preclinical safety studies evaluate and
COFEPRIS is responsible for the oversight of identify potential toxic effects of the drug,
all pharmacovigilance activities, in addition to which include assessing whether a medicine is
setting local policies in line with national and carcinogenic, mutagenic, or teratogenic.
international pharmacovigilance guidelines. Although the information generated from pre-
The main standard guideline governing phar- clinical studies provides guidance on the selec-
macovigilance, including PMS studies, in tion of a safe starting dose for the first
Mexico is the Installation and Operation of administration-to-human study, the limited
Pharmacovigilance. predictability of animal studies to the toxicity
Epidemiology has become increasingly of drugs in human is well-recognized.
important to risk management over the last However, these studies can provide important
three decades with evolving pharmacovigi- information about hypothetical drug risks.
lance regulation globally, which has further Randomized clinical trials (RCTs) provide
solidified epidemiology’s role in informing the abundant high-quality data about identified
benefit–risk of medicines throughout the and hypothetical risks, but have limitations.
development lifecycle. Pre-approval RCTs typically involve highly
selected subjects followed for a short period of
time and, in the aggregate, include at most a
Epidemiology in Drug Safety
few thousand patients. These studies are gen-
Evaluation
erally sufficiently large to provide evidence of
Background a beneficial clinical effect, exclude large
The safety profile of any drug reflects an evolv- increases in risk of common adverse events,
ing body of knowledge extending from pre- and identify the most common and acutely
clinical investigations through the post-approval occurring adverse events. However, they are
lifecycle of the product. Drug manufacturers rarely large enough to detect small differences
traditionally relied on two major sources for in the risk of common adverse events or to reli-
information on the safety of drugs: the clinical ably estimate the risk of rare events. Using the
trials supporting the New Drug Application “rule of three,” where the sample size needed
(NDA) and, once the drug was marketed, spon- is roughly three times the reciprocal of the fre-
taneous reports received throughout the world quency of the event, at least 300 patients would
be required in a trial in order to observe at least descriptive epidemiologic studies, and stand-
one adverse event that occurs at a rate of 1/100. ing cohorts (i.e. open cohorts of indicated pop-
Likewise, a sample of 3000 is needed to observe ulations which are updated over time and
at least one adverse event with 95% probability queried for incidence of safety events and
if the frequency of the event is 1/1000 (see other data as needed) to support regulatory fil-
Chapter 3 for more discussion of the sample ings and to complete epidemiology sections of
sizes needed for studies). Increasingly, preap- key regulatory documents (e.g. RMP and PIP,
proval studies – particularly in the rare dis- orphan drug applications). These background
eases or where long-term placebo treatments epidemiology data can also be a key compo-
are unethical – include unbalanced randomi- nent for internal decision making such as trial
zation or treatment arms with short duration design, data monitoring committee decisions
of placebo or active-comparator, or use non- to stop/continue trials, decisions to move/not
contemporaneous controls. While clinical tri- move to the next phase of development, risk
als are not intended or designed to address all management decisions, and risk mitigation
potential safety issues related to a particular planning.
drug, like preclinical studies, they often give During development, in addition to summa-
rise to signals that cannot be adequately rizing the existing relevant literature and
addressed from trial data alone. designing and executing background epidemi-
Pre-approval epidemiology complements ology studies, industry epidemiologists are
safety data from preclinical and clinical studies often involved in safety signal evaluation,
and provides a context for signals arising from observational analyses of RCT data (e.g. as-
clinical trials. Comprehensive reviews of the treated or observed versus expected analyses),
epidemiologic literature are complemented by and designing post-approval epidemiology
epidemiologic studies to establish the back- studies and risk minimization planning.
ground epidemiology (e.g. incidence, preva- Planning for successful post-approval epidemi-
lence, mortality) of the indication among ology studies often begins well before approval.
patients expected to use the new medication During the peri-approval phase, epidemiolo-
(i.e. indicated populations); expected preva- gists may conduct feasibility assessments for
lence/incidence of risk factors, co-morbidities planned post-approval studies, start key opera-
and complications; patterns of health care uti- tional aspects of post-approval studies (e.g.
lization and prescribing of currently approved identifying key external partners such as con-
treatments; and background rates of mortality tract research organizations and scientific
and serious nonfatal events. steering committee members for the design
Epidemiologic studies conducted before or and conduct of the study), and contribute to
during the clinical development program are regulatory submissions, responses, and nego-
often critical to place the incidence of adverse tiations (e.g. responding to regulatory inquiries
events observed in clinical trials in perspective. related to epidemiology and participating in
Data are often lacking on the expected rates of regulatory meetings).
events in the population likely to be treated. There are several other areas where epidemi-
For example, studies examining the risk fac- ologists are increasingly providing their exper-
tors for and rates of sudden unexplained death tise to support pre-approval development. In
among people with epilepsy were able to pro- the context of risk minimization planning, the
vide reassurance that the rates observed in a epidemiologist may conduct research to test
clinical development program were within the the comprehension and utility of educational
expected range for individuals with a compara- materials, evaluate the proposed risk minimi-
bly severe disease. Epidemiologists use infor- zation tools and processes to assess their bur-
mation from the published literature, den on the healthcare system and patients,
pilot and/or user test assessment materials at scientific conferences and peer-reviewed
such as surveys, and generally contribute to publications.
the design and implementation of these pro- Spontaneous reporting systems are the most
grams. Furthermore, many regulatory agencies commonly used pharmacovigilance method to
are utilizing various benefit–risk assessment generate signals on new or rare adverse events
frameworks in their reviews. Epidemiologists not discovered in clinical trials (see Chapter 7).
can provide inputs or lead both quantitative However, there are several important limita-
and qualitative benefit-risk assessments such tions in interpreting spontaneous report data.
as multi-criteria decision analysis (MCDA), Due to the lack of complete numerator (num-
stochastic multicriteria acceptability analysis ber of cases) and the need to estimate the
(SMAA), and the PhRMA Benefit-Risk Action denominator (total number of patients actu-
Team (BRAT) framework, among others (see ally exposed to the drug) data, it is not possible
Chapter 23). Lastly, several accelerated/condi- to determine the incidence of a particular
tional approval pathways and regulations exist event from spontaneous reports. Further eval-
in the EU, and are anticipated for the US and uation of an apparent association between a
other regions, which have requirements for drug and an adverse reaction usually requires
Real World Data and Evidence (RWD/RWE) to post-approval epidemiologic studies.
complement the incomplete or uncertain data Likewise, the nature of pre-approval clinical
from abbreviated development programs in trials often necessitates further safety evalua-
areas of high unmet needs. Epidemiologists’ tion through post-approval epidemiology. In
expertise in regulatory-quality RWD/RWE addition to the limited sample size and length
generation is often critical to the success of of follow-up of pre-approval RCTs, with
these accelerated options. respect to drug safety, an additional limitation
of these studies is the common strict inclu-
Contributions of Post-approval sion/exclusion criteria. Patients included in
Epidemiology pre-approval clinical studies may be the
The need for a post-approval epidemiology healthiest segment of that patient population.
study can be known and devised pre-approval Special groups such as the elderly, pregnant
or can arise once a new drug is marketed. Post- women, or children are frequently excluded
approval signals may come from clinical trial from trials. Patients in clinical trials also tend
extension data (e.g. long-term extension [LTE] to be treated for well-defined indications, have
studies), spontaneous reports, published case- limited and well-monitored concomitant drug
series, or signal detection of electronic health- use, and are closely followed for early signs
care data. Once a drug is marketed, and symptoms of adverse events which may be
epidemiologists execute post-approval com- reversed with proper treatment.
mitments (e.g. epidemiology studies, active In contrast, once a drug is marketed, it is
surveillance studies, other registries, REMS/ used in a “real-world” clinical context.
aRMM evaluations, PIP observational studies, Patients using the drug may have multiple
etc.); conduct studies evaluating the effective- co-morbidities for which they are being
ness of risk mitigation activities; perform sig- treated simultaneously. Patients may also be
nal detection in existing cohorts (e.g. via taking over-the-counter medications, “natu-
claims or electronic patient record data); and ral” remedies, or illicit drugs unbeknownst to
design and implement new studies as addi- the prescribing physician. The interactions of
tional signals arise (e.g. from spontaneous various drugs and treatments may result in a
reports, signal detection or other sources). particular drug having a different safety pro-
Epidemiologists also communicate scientific file in a post-marketing setting compared to
findings through oral and poster presentations the controlled premarketing environment.
An example is the drug mibefradil, which effects if the severity of the underlying ill-
was voluntarily withdrawn from the market ness (i.e. any conditions specified as labeled
by the manufacturer after less than a year indications or contraindications, or included
as a result of new information about multi- in the precautions or warnings) can be val-
ple potentially serious drug interactions. idly measured (see Chapter 22). Confounding
Adherence to medications also often differs by indication is more of an issue when a par-
between closely monitored trials and general ticular property of the drug is very likely to
post-approval use, as is the case with affect the type of patient it is used by or pre-
antihypertensives. scribed to. In these cases, studies using rand-
Because of the logistical complexity, high omization to treatment may be necessary. A
cost, and low external validity, large controlled pragmatic clinical trial (PCT) is an RCT with
trials have not been widely used for the post- one or more pragmatic elements and an LST
marketing evaluation of drugs. Regulators and is a type of PCT that combines randomiza-
the medical community have communicated a tion to treatment with observational follow-
desire for safety data from the populations that up of patients. The characteristics of LSTs
actually use the drugs in “real-world” clinical are further described in Chapter 17.
practice. This has led to a greater emphasis on
the use of observational methods to under-
Epidemiology in Evaluation of Risk
stand the safety profile of new medications
Mitigation Interventions
after they are marketed.
In addition to typical epidemiologic designs, Epidemiology not only plays an important role
depending on the specific safety research in evaluation of the drug safety profile pre-and
hypothesis, epidemiologists design and imple- post-approval but, as noted earlier, also makes
ment active surveillance studies, pragmatic significant contributions to the evaluation of
trials (including the most naturalistic version, the effectiveness of risk mitigation interven-
the large simple trial [LST]), and self- tion measures (see also Section 23.6). This
controlled designs such as the case-crossover component of biopharmaceutical risk man-
study and self-controlled case series. Active agement has grown considerably in the last
surveillance studies can be defined as descrip- decade, with the US, EU, Taiwan, Egypt,
tive studies intended to solicit information on Australia, and a number of other countries
adverse events among a specified population, implementing legislation that supports risk
such that the numerator and denominator are mitigation interventions.
as complete as possible, potentially allowing Under FDAAA, the FDA can require a
calculation of incidence. sponsor to submit a proposed Risk
Purely observational epidemiologic stud- Evaluation and Mitigation Strategies
ies may not always be the most appropriate (REMS) as part of its initial application if
method of evaluating safety signals or com- the FDA finds that a REMS is necessary to
paring the safety profile of different medica- ensure the benefits of the drug or biological
tions, especially when there are concerns of product outweigh the risks. The FDA may
confounding by indication. Confounding by also require a REMS post-approval based
indication occurs when the risk of an adverse upon new safety information. FDAAA has
event is related to the indication for medica- defined this as any information obtained
tion use such that in the absence of the med- during the initial review process, as a result
ication, those actually exposed are at higher of post-approval studies, or spontaneous
or lower risk of the adverse event than those reports. REMS are intended to be utilized to
unexposed. As with any other form of con- reduce known or hypothetical risks when
founding, one can, in theory, control for its traditional minimization approaches (i.e.
the product label) are insufficient. These uestionnaires that measure these concepts
q
tools generally fall into three categories: do not exist.
enhanced education, i.e. patient labeling The implementation of the REMS and
(including medication guides) or communi- aRMM legislation has highlighted a number
cation plans such as prescriber training proof difficulties. The mandated assessments may
grams; Elements to Assure Safe Use be difficult to achieve, or achieve within the
(ETASU), e.g. requiring documentation of US legislative timelines, for many reasons: the
laboratory tests before each prescription or need to develop and pilot knowledge/compre-
restricting distribution only to those who hension surveys unique to each drug subject
are certified prescribers; and an implemen- to a REMS; to design, implement, and assess
tation system to monitor and evaluate the complex safe use programs; the scarcity of
ETASU. A critical addition to this legislation patients treated with the drug of interest; or
that was particularly relevant to epidemiolo- difficulties in identifying them through auto-
gists within industry was the requirement to mated channels. The fractured healthcare and
perform assessments of the effectiveness of prescription delivery system in the US and the
these risk minimization tools and to submit wide variety of health systems, legal and pri-
these to the Agency for review at prescribed vacy requirements, and attitudes toward these
time points, generally at 18 months, 3 years, programs and research participation across
and 7 years. The EU has similar legislation Europe present a barrier to efficient distribu-
to require sponsors to implement aRMM tion of educational materials, to the imple-
where necessary to ensure the benefits out- mentation of many safe use elements, and to
weigh the risks, and a similar requirement the scientifically valid evaluation of these pro-
to assess the effectiveness of the aRMM, grams overall. Unfortunately, there is rela-
although without defined timelines. These tively little scholarly work published on how
aRMM programs and assessments are best to assess, how best to define success, and
described in the EU-RMP. where necessary, how best to improve these
Epidemiologists play a critical role in the often burdensome but important risk mitiga-
design and implementation of these assess- tion programs. Knowledge in these areas con-
ments because of their expertise in observa- tinues to mature as more companies and the
tional study design, survey design, data regulatory agencies garner additional experi-
analysis, and program evaluation. For exam- ence, and we expect that existing guidance
ple, using an automated healthcare or claims will evolve. Risk mitigation evaluation is thus
database, assessments may measure compli- still an emerging area for epidemiologists in
ance with monitoring guidelines or to meas- industry but one that complements our spe-
ure whether a contraindicated population is cialized training and expertise.
prescribed the drug. Assessments may also
examine the frequency of occurrence of an
Conclusion
adverse event of interest before and after
implementation of the risk minimization Epidemiology makes a significant contribu-
tool. Most commonly, however, assessments tion to the development and marketing of
measure prescriber, pharmacist, or patient safe and effective biopharmaceutical prod-
comprehension of risk information or self- ucts worldwide. It facilitates the regulatory
reported adherence to risk minimization process and provides a rational basis for drug
behaviors, and require the epidemiologist to safety evaluation, particularly in the post-
craft cross-sectional surveys specific for approval phase, and evaluation of risk miti-
each recipient, drug, and associated unique gation interventions. Like any other
risk profile, as standardized or validated discipline, it must be properly understood
and appropriately utilized. Industry has an Summary Points for the View
opportunity to contribute to the development from Industry
of the field and the responsibility to do so in
●● The safety profile of any drug reflects an
a manner that expands resources while
evolving body of knowledge extending from
assuring scientific validity. With the passage
preclinical investigations to the first use of
of the 2007 FDAAA legislation and the 2010
the agent in humans and through the post-
EMA Regulation on Pharmacovigilance, the
approval life cycle of the product.
need for scientists with training and research
●● Results from clinical trials, spontaneous
experience in pharmacoepidemiology has
reports, epidemiologic studies, and where
never been greater. To best support drug
relevant, preclinical datasets, should all be
safety evaluation, epidemiology strategies
evaluated for their potential to address
must: (i) begin early in development, (ii)
safety questions, with close consideration
continue throughout the lifecycle of the
given to the unique strengths and limita-
drug, (iii) evolve as new safety information
tions of the study designs and data collec-
becomes available, and (iv) be innovative,
tion methods used.
requiring epidemiologists to be aware of
●● Epidemiology plays a central role in drug
new methodologies and methods specific to
safety assessment and risk management
the disease area. Epidemiologists within
activities within the pharmaceutical indus-
industry have an opportunity to build on the
try, whether through studies of the natural
successes of the last 40 years by collabo-
history of disease, disease progression/treat-
rating with academics, non-p rofit organi-
ment pathways, and morbidity and mortality
zations and regulators to advance the
patterns, or in the design and implementa-
methods of drug safety evaluation and risk
tion of post-approval safety studies or risk
management.
minimization programs.
Case Example 6.2

View from industry: Tofacitinib pre- latency. Furthermore, LTE studies, while
approval and post-approval epidemiology providing greater exposure in patients
strategies. taking tofacitinib, lacked control groups to
allow a comparative risk assessment.
Background
Evidence of the expected rates in a con-
●● Tofacitinib is a JAK inhibitor for the treat-
current and directly comparable patient
ment of rheumatoid arthritis (RA) for
population was also lacking.
adults with an inadequate response to
methotrexate. Question
●● RCTs provide high-quality data about iden- How were epidemiologic studies strategi-
tified and hypothesized risk but have limi- cally implemented to support the safety pro-
tations. Even though all the RCTs in the file of tofacitinib during the entire product
tofacitinib RA development program lifecycle from pre-approval to post-approval
included at least one placebo or active activities?
control group, the size of the control
Approach
groups and duration of treatment prohib-
●● Pre-
approval epidemiological strategy con-
ited precise comparative assessments for
sisted of several distinct but complemen-
adverse events with low frequency or long
tary efforts for risk characterization for
Case Example 6.2 (Continued)
tofacitinib, including literature reviews, of the REMS assessment (i.e. surveys at

meta- analyses, and a standing cohort 18 months, 3 years, and 7 years). The EU
within a US-based registry of patients with aRMMs were implemented within the
RA. Indirect comparative methods via EU as a condition of approval within the
external patient cohorts were also used to EU consisting of an educational program
provide evidence of expected adverse intended to enhance the communication
event rates, while taking into account key of the risks and risk minimization prac-
potential differences in the populations. tices to patients and health care profes-
The output of the analyses were used to sionals (HCP).
assess the rates of identified and potential
risks of interest in the tofacitinib clinical Results
●● Pre-approval epidemiology strategy: The
program compared with those from
cohorts of RA patients treated with bio- data from pre- approval epidemiological
logic disease modifying antirheumatic analyses were articulated within regula-
drugs (bDMARDs). tory documents including the NDA for FDA
●● Post-approval epidemiology strategy con- and the briefing document (i.e. summary of
sisted of several approaches including (i) clinical safety (SCS), clinical overview (CO),
safety studies to characterize the safety of etc.) for EMA and also presented at the
tofacitinib within the real-world or clinical FDA Advisory Committee Meeting for
practice setting, (ii) surveillance program tofacitinib in May 2012. The collective
is to evaluate any excess risk in the occur- body of evidence (including interim data
rence of known or potential adverse from the US-based registry study) provided
events, after accounting for confounding substantial additional context to rates of
factors including disease severity and con- selected adverse events observed in the
comitant therapy, and (iii) risk mitigation tofacitinib clinical trial program, and
activities. therefore addressed the uncertainties
⚪⚪ In the US, two observational post- related to the potential risks previously
approval safety studies were initiated: (i) expressed by the EMA.
●● Tofacitinib was approved in the US in
an active surveillance study within the
Corrona RA registry, and (ii) a pregnancy November 2012 and in the EU in March
outcome study within the Organization 2017.
●● Post-approval epidemiology strategy: Several
of Teratology Information Specialists
Registry (OTIS) registry as a condition for regulatory commitments, including post-
approval of tofacitinib. approval safety studies, ongoing pharma-
⚪⚪ In the EU, four 5-year surveillance stud- covigilance (i.e. spontaneous reports), and
ies embedded within existing registries REMS and aRMM related work served to
(ARTIS, BIOBADASER, BSRBR, and address regulatory concerns regarding
RABBIT) were proposed as a commit- drug safety among US and European
ment to EMA to assess the safety of patients receiving care in the real-world.
⚪⚪ In the US, the FDA deemed the 18- month
tofacitinib.
⚪⚪ The US REMS for tofacitinib consisted of
REMS epidemiological assessment ade-
a Medication Guide, a communication quate, citing that the survey data dem-
plan, and a timetable for the submission onstrated that patients understood the
(Continued)
risks associated with therapy and deter- ●● Purely observational epidemiologic stud-
mined that maintaining the Medication ies may not always be the most appropri-
Guide as part of the approved labeling ate method of evaluating safety signals or
was sufficient to address safety-related comparing the safety profile of different
concerns. medications, especially when there are
⚪⚪ In the EU, the elements of the aRMM concerns of confounding by indication.
included Xeljanz HCP Brochure, Xeljanz
HCP Treatment Initiation Checklist, Key Points
Xeljanz HCP Treatment Maintenance ●● Pre-approval epidemiology strategy com-
Checklist, an educational website, and plements safety data from preclinical and
Xeljanz Patient Alert Card. clinical studies and provides a context for
signals arising from clinical trials.
Strengths
●● Post-approval epidemiology strategies
●● Data from multiple sources (i.e. observa-
may require industry epidemiologists to
tional studies, RCTs with other agents, and
address regulatory-
mandated post-
cohorts of patients stratified by RCT inclu-
approval commitments, such as epidemi-
sion/exclusion criteria) were used to pro-
ology studies, active surveillance studies,
vide indirect comparisons, drawing from
registries, REMS/aRMM evaluations, and
the strengths of each data source while
PIP observational studies or address post-
balancing their weaknesses.
approval safety signals arising from clini-
●● Epidemiological evidence from the entire
cal trial extension data, spontaneous
lifecycle of tofacitinib allowed regulatory
reports, published case series, or signal
agencies in the US and EU to assess and
detection from electronic healthcare data.
ensure the safety of the product.
●● To support drug safety evaluations, epide-
Limitations miologic strategies must initiate early in

●●All data sources have unique limitations. development, continue throughout the
Limitations in clinical trials and spontane- lifecycle of the drug, evolve as new safety
ous reports can be addressed, in part, by information becomes available, and may
the proper use of observational require innovative methods specific to the
epidemiology. disease area.
he View from Regulatory

T edicines that are effective, acceptably safe,
m
Agencies and of high quality. A wide range of regulatory
activities spans the entire lifecycle of a medi-
Note cine, involving laboratory-based understand-
The views expressed herein are those of the ing of pharmacological action, animal testing,
authors, and not necessarily of the US FDA, providing scientific and regulatory input to
the MHRA, or the Japanese PMDA. drug development programs, protection of
human subjects during clinical trials, assuring
the integrity of the manufacturing process,
Introduction
reviewing the dossier to support product
The regulation of pharmaceuticals aims at approval or licensure, monitoring the safety of
ensuring that the public has access to medicines after they enter the market, and
The View from Regulatory Agencie 91
many other activities. These regulatory activi- c areful understanding of background rates can
ties are firmly rooted in science, have a strong be important.
public health focus, and are executed within a
legal and regulatory framework.
Orphan Drugs
In the last decade, there has been substantial
Assessing the Need for Medicines
activity and progress in the development of
Pharmacoepidemiology, along with other drugs for rare diseases. Orphan drug programs
areas of medical epidemiology, can be used in are designed to provide incentives to pharma-
drug development long before a medicine is ceutical manufacturers who develop medi-
licensed or even tested in humans. cines for rare conditions, known as “orphan
Pharmacoepidemiologic approaches can be drugs.” In the United States, an orphan drug
used to examine patterns of utilization of exist- designation is given to a drug or biologic that
ing disease treatments in order to identify and has shown promise as a therapy intended to
characterize disease populations and sub- treat a disease affecting fewer than 200 000 per-
populations for which unmet medical needs sons in the United States. In Japan, orphan
exist. In some cases, no available therapies may designation is granted for drugs or medical
exist. In other cases, available therapy may be devices if they are intended for use in less than
ineffective for or poorly tolerated by certain 50 000 patients in Japan and for which there is
patients. In these cases, pharmacoepidemio- a high medical need. In the European Union, a
logic approaches can be used to characterize prevalence rate of five per 10 000 persons in the
the patients who experience a suboptimal EU is used. When all rare diseases are taken
response to the medicine, thus defining the together, their public health impact is signifi-
target population for a drug development pro- cant; approximately 25 million people in North
gram. For example, population-based data- America are affected by these diseases.
bases can be used to characterize the frequency Medical epidemiology is central to the desig-
and distribution of characteristics of patients nation of a product as an orphan drug product,
with a specific disease, so that relevant popula- as determination of prevalence is the basis for
tions can be included in the developmental such designation. Data sources for determin-
clinical trials. Healthcare databases can be ing prevalence can include administrative
used to estimate the frequency of co-morbid healthcare databases, electronic medical
conditions in the setting of the specific under- record systems, registries, and surveys. In
lying disease to be treated, so that relevant many cases, combining data from multiple
background rates can be derived to place sources will be necessary. In most cases, data
potential adverse events that arise during from these sources, even when combined, will
development in context. This is especially use- not cover the entire jurisdiction for which the
ful for clinical events that are seen more fre- orphan designation applies. Thus, some form
quently in patients with the disease for which of extrapolation must be performed to deter-
the new treatment is being tested, but which mine if the relevant population prevalence has
could also represent an adverse drug reaction. been exceeded. Most orphan drug designations
This situation, known as confounding by indi- are for diseases or conditions whose preva-
cation, is a well-known methodological prob- lence is much lower than the 200 000 preva-
lem in observational pharmacoepidemiologic lence threshold in the United States. A review
studies, but can also complicate the interpreta- of 25 years’ experience with the orphan drug
tion of adverse events in clinical trials, espe- program in the United States covering 1892
cially if the trial is not designed or powered to orphan designations found that the median
analyze these events. In these situations, prevalence was 39 000; the most common
patient prevalence was 10 000 or fewer patients, shortened by allowing the use of surrogate
with relatively few prevalence rates near the markers rather than clinical markers.
200 000 threshold. For estimates of population Examples of programs to optimize drug
prevalence near the threshold, care must be develop include breakthrough therapy des-
taken to ensure that the most rigorous meth- ignation in the United States, the PRIME
ods have been used to estimate the population (PRIority MEdicines) initiative in the
prevalence of a rare disease. The closer the European Union, and the SAKIGAKE review
estimated prevalence is to the threshold, the program in Japan. “SAKIGAKE” is a Japanese
greater the precision needed to characterize word meaning “frontrunner” or “pioneer.” In
the prevalence. each of these situations, pharmacoepidemio-
logic analyses can aid in the comparison of
new treatments to existing treatments, espe-
Planning Drug Development
cially when data on existing treatments are
Programs
derived from clinical experience and not from
Despite the availability of an increasing num- formal clinical trials.
ber of medicines, there remain a substantial The goal of a drug development program is to
number of unmet medical needs. Advances in demonstrate that a medicine has a beneficial
understanding the molecular pathogenesis of and meaningful effect on a clinically important
cancers, rare diseases, and infectious diseases outcome, generally a measure of how the
have led to a rapid rise in the number of drug patient feels, functions, or survives. Clinical tri-
development programs targeting these condi- als whose primary endpoint is a direct measure
tions. At the same time, the aging population of a clinically important outcome may be very
across the globe has led to a need for improved long and delay access by patients to effective
treatments for widespread diseases such as therapies. To allow patient access as rapidly as
diabetes, hypertension, ischemic heart dis- is feasible, and to assure that definitive evi-
ease, chronic obstructive pulmonary disease, dence of effectiveness is obtained, an alterna-
Alzheimer disease, other neurodegenerative tive approach allows marketing approval for a
disorders, and many others. The continuing new drug product on the basis of adequate and
emergence of antibacterial resistance and the well-controlled clinical trials establishing that
threat of new viral illnesses prompt the need the drug product has a beneficial effect on a
for new antimicrobial agents. Infections with surrogate endpoint. A surrogate endpoint is an
Mycobacteria tuberculosis, Plasmodium falci- outcome measure that is used in place of a
parum, human immunodeficiency virus, direct measure of a clinically meaningful out-
endemic parasitic diseases, and other agents come when the effect of treatment on the sur-
contribute substantially to the global burden rogate endpoint is expected to reflect changes
of disease and require new treatments. in the clinically meaningful outcome. In the
Regulatory agencies have responded to this context of drug development, a validated sur-
demand with a variety of regulatory programs rogate endpoint is one for which evidence exists
and pathways designed to promote efficient that the effect of treatment on the surrogate
development of medicines and to reduce drug endpoint predicts the effect of treatment on the
development time so that these unfulfilled clinical outcome of interest. For example, sys-
medical needs can be met. Some of these pro- tolic blood pressure is used as a surrogate end-
grams seek to optimize drug development by point in clinical trials of antihypertensive
providing timely consultation between the agents because it predicts the risk of occurrence
regulator and the company developing a of stroke. Similarly, human immunodeficiency
drug to clarify scientific requirements; other virus viral load is used as a surrogate endpoint
programs allow clinical development to be in clinical trials of antiretroviral agents because
it predicts the development of an acquired early approval program has applied to drugs
immunodeficiency syndrome diagnosis. The offering high efficacy and clinical usefulness in
use of validated surrogate markers to support the treatment of serious diseases, drugs for
approval of medicines is widely employed. which conducting confirmatory studies is
There are, however, many serious and life- impracticable, and other designated drugs.
threatening conditions for which there are no One prerequisite for approval will be a com-
validated surrogate markers, yet there is still an mitment to complete postmarketing studies as
urgent need to bring effective therapies to necessary in order to reconfirm product safety
patients in a timely way. For this latter situa- and efficacy.
tion, the concept of “accelerated approval” has Understanding the relationship between a
been developed. Under this framework, the US surrogate endpoint and a clinically relevant
FDA may grant approval to a medicine intended endpoint, as well as validation of the surrogate
to treat a serious or life-threatening disease endpoint, are opportunities for pharmacoepi-
based on an unvalidated surrogate endpoint demiologists to contribute to drug develop-
that is reasonably likely, based on epidemio- ment. Pharmacoepidemiologists can use
logic, therapeutic, pathophysiologic, or other principles of epidemiology to distinguish sim-
evidence, to predict clinical benefit on the basis ple correlation between a potential endpoint
of an effect on a clinical endpoint other than and a clinically meaningful outcome, on the
survival or irreversible morbidity. In these one hand, from a true surrogate marker. For
cases, postmarketing studies must be con- example, a marker of disease status used in
ducted to demonstrate the actual clinical bene- natural history studies may not be an adequate
fit of the medicine. surrogate endpoint in a clinical trial because it
A key regulatory tool in the EU to fulfill is not related to the disease mechanisms that
unmet medical needs is the conditional mar- give rise to symptoms, morbidity, and
keting authorization, which has reduced data mortality.
requirements linked to a one-year time-limited
authorization where the authorization’s
Pre-approval Review of Clinical
renewal is linked to further data submission.
Safety Data
Under the applicable regulations, manufactur-
ers must study the drug further once it is While the traditional role of pharmacoepide-
approved, to verify and describe its clinical miology, from a regulatory standpoint, has
benefit, where there is uncertainty as to the been the assessment of the safety of medicines
relationship of the surrogate endpoint to clini- in the post-licensing period, pharmacoepide-
cal benefit, or of the observed clinical benefit miology can play an important role during the
to ultimate outcome. At the time of approval, pre-licensing review of safety data. The limita-
postmarketing studies would usually already tions of pre-licensing clinical trials in defining
be underway. the full scope of adverse drug reactions are
In Japan, the Pharmaceuticals, Medical mainly related to the fact that clinical trials are
Devices and Other Therapeutics Products Act relatively small in size, compared to the popu-
(“PMD Act”) established a system of condi- lation of patients that will ultimately take the
tional and time-limited approval for regenera- medicine once it is marketed. Patients who
tive medicines based on probable benefit from participate in clinical trials may have fewer
early clinical trials. After obtaining such an comorbidities and take fewer concomitant
approval, the marketing authorization holder medications than those treated in actual prac-
is required to submit a standard marketing tice. Pre-licensing clinical trials generally pro-
application with additional data on safety and vide relatively little data, or no data at all, in
efficacy. Similarly, since 2017 a conditional certain populations such as children, the
elderly, and pregnant women, or at-risk groups effectiveness of the product. In these situations,
such as immunosuppressed patients. These pharmacoepidemiologists can be involved in
groups, however, are treated with the medi- studies assessing the validity of the surrogate
cines in the course of clinical practice once the marker.
medicine is licensed. Second, pharmacoepidemiologists can be
The analytic methods of clinical trials are best involved in the design and interpretation of
suited for data arising from randomized, con- postmarketing studies designed to assess the
trolled, comparative trials. Many clinical trials impact of new formulations of medicines
of medicines intended for chronic or long-term developed to have a more favorable safety pro-
use, including those trials in pre-approval drug file than earlier versions.
development programs, may have single-arm, Third, pharmacoepidemiologists can be
open-label extensions after participants have involved in planning postmarketing studies
completed the randomized portion of the trial. when safety signals are detected prior to
For data generated from this portion of the clini- approval.
cal trial, the techniques of observational phar-
macoepidemiology may be appropriate. In
Monitoring Post-approval Safety
addition to tallying the frequencies of specific
adverse events, data from long-term extension For the regulator, the postmarketing assess-
studies can be examined to characterize pat- ment of the safety of medicines involves both a
terns of adverse event onset over time. If appro- proactive approach and, of necessity, a reactive
priate, analyses based on person-time can be approach. The International Council for
performed. In this setting, the interpretations of Harmonisation of Technical Requirements for
adverse events must take into account the prior Pharmaceuticals for Human Use (ICH) has
treatment received during the randomized por- developed a useful and practical framework
tion of the trial, the duration of treatment, the that summarizes the known safety issues of a
underlying frequency of medical outcomes in product and can form the basis of ongoing
the population with the disease being treated, monitoring and, as needed, specific studies.
and other factors. Pharmacoepidemiology can The ICH framework characterizes important
inform this approach. identified risks, important potential risks, and
important missing information. This frame-
work allows pharmacoepidemiologists and
Planning for Post-approval Studies
others to devise proactive strategies to design
At the time a medicine is approved, there are observational studies or clinical trials to
uncertainties and unknowns regarding the address unanswered questions about the safety
safety profile of the medicine. In many cases, profile of a medicine. The identification of
the nature of the safety issues that will unfold knowledge gaps can occur at any time in the
post-approval cannot be predicted at the time life cycle of a medicine, and can be based on
the product is brought to market. In some cases, data from clinical trials or observational stud-
however, a careful review of the clinical data at ies of the medicine, or safety findings from
the time of approval can lead to a proactive other medicines in the same class. In these
approach to obtaining more safety information. cases, careful review of the available data can
Pharmacoepidemiology can play an impor- allow the regulator, often working with the
tant role in several specific situations. First, developer, to develop a thoughtful and rational
drug development programs based on the use approach to drug safety issues in the post-
of unvalidated surrogate markers, as described approval period.
above, generally require postmarketing stud- Reactive approaches are also needed in regu-
ies to demonstrate definitively the clinical latory pharmacoepidemiology because the
adverse effects of medicine can become recog- intended public health outcomes. For serious
nized at any time, sometimes many years, after safety issues, it is not enough simply to add a
approval. Because not all drug safety issues can warning to a product label. Such an action is
be predicted, regulators continue to need reac- in itself an intervention, and it is thus impor-
tive approaches. These approaches require the tant to understand its impact. Recognizing the
efficient review of the existing data, careful and fundamental importance of the need for such
timely assessment of the need for immediate or assessments, the Pharmacovigilance Risk
near-term regulatory action, and interaction Assessment Committee of the EMA developed
with the product’s manufacturer to plan further a formal strategy to measure the impact of
study. Reactive approaches become necessary, pharmacovigilance activities. The strategy is
for example, when new safety issues are identi- aimed both at informing the review of indi-
fied from spontaneously reported suspected vidual medicines that have been the subject of
adverse drug reactions, when drug safety find- major risk minimization efforts and at deter-
ings are published by independent groups, and mining which activities are successful and
when events such as manufacturing-related which are not, in order to optimize the phar-
product recalls result in a large number of macovigilance system. Pharmacoepidemiology
adverse event reports that need to be reviewed is critical to this endeavor, as it can relate regu-
in a short period of time. From the regulator’s latory activities to the outcomes which those
point of view, the scientific studies that form activities are intended to impact.
the basis of regulatory actions must be as sound Pharmacoepidemiologic thinking and meth-
and robust as possible. odologies underpin the EMA strategy.
One domain of assessment of the impact of
regulatory activities is an understanding of the
Assessing Actual Use Patterns of a
effectiveness of regulatory agencies’ communi-
Medicine
cations about the risks of medicines. A study
Regulators are interested not only in whether a that examined the extent to which patients
medicine meets the relevant regulatory stand- understand important information about a
ards for approval, but also in how a medicine is serious risk of a medicine that they are taking
actually used in clinical practice. Because the examined the results of patient-directed
harms of medicines can result not only from knowledge surveys for 66 medicines for which
their intrinsic pharmacological properties but patients were supposed to have received a
also from how they are used, or misused, in Medication Guide, a type of patient-directed
actual practice, understanding the actual usage labeling. For each Medication Guide, accepta-
allows regulators to assess the degree to which ble knowledge defined was 80% or more of
the medicine is used in ways that are consist- patients correctly answering questions about
ent with the safe use of the medicine as the medicine’s primary risk. The study found
described in the label or marketing authoriza- that only 20 Medication Guides (30.3%) met
tion. To do so, regulators can use a variety of the 80% threshold, a finding that underscores
pharmacoepidemiologic techniques, including the need for improved patient-directed
administrative claims data, electronic medical information.
records, or other public health databases. Analysis of the impact of regulatory actions
is not limited to the assessment of actions
related to individual medicines. Rather, it can
Assessing Impact of Regulatory
look broadly at how the functioning of a drug
Actions
regulatory system contributes to the system’s
Because of its public health focus, drug regu- public health mission. Because of the rapidly
lation must ensure that its actions lead to the changing and expanding data that inform
pharmacoepidemiologic studies, studies that that is a distributed database compiling elec-

examine overall performance are important tronic medical records, insurance claims
because they can lead to system-wide data, and DPC-compatible inpatient care
improvements. data under a common data model. The use of
the MID-NET® system will be opened to rele-
vant members of industry and academic
Advancing the Science
researchers for use in pharmacoepidemio-
logic studies.
Pharmacoepidemiology is a complex, FDA’s Sentinel Initiative, as noted earlier,
dynamic, and changing field. It relies on the has created a linked system of electronic
integration of epidemiology, clinical pharma- healthcare databases to investigate safety
cology, pharmacy, medicine, statistics, and questions about FDA-regulated medical prod-
other disciplines, for its full execution. ucts, which was developed after FDA sought
Increasingly, the rapid advances in the avail- extensive stakeholder input as it worked with
ability of large, diverse, and relevant datasets outside organizations to develop Sentinel to
have made informatics an important contrib- address important issues of governance, pri-
uting discipline to pharmacoepidemiologic vacy, data standards, and public disclosure of
efforts. Acquiring expertise in pharmacoepi- results.
demiology thus requires an environment that Pharmacoepidemiologic efforts such as
provides access to experts in all the relevant ENCePP, MID-NET ®, and Sentinel all make
disciplines. Furthermore, this discipline use of various traditional sources of health-
relies on population-based healthcare data care data derived from existing sources that
and thus an understanding of the healthcare reflect current clinical practice and actual
system in which the data were generated, patient experiences. While these contempo-
which experts in the above fields may not rary systems often rely on large datasets
have. As more and more drug safety ques- and, at times, integration of datasets through
tions arise that require expertise in pharma- networks, it is important to note that phar-
coepidemiology as well as appropriate data, it macoepidemiologic research has a decades-
is crucial that there be sufficient capacity, long tradition of using observational data
both in the form of well-trained pharmacoep- recorded at the point of care to describe the
idemiologists and in the availability of sys- effects of medicines in populations, though
tems, such as networks that combine relevant the scope of this work has largely focused on
data with scientific expertise, that can be safety issues. The emergence in recent years
used for pharmacoepidemiologic studies. of additional digital health-related data,
Because pharmacoepidemiology is a multi- such as data generated from wearable
disciplinary effort, there must also be appro- devices and health-related applications, has
priate mechanisms for collaboration. given rise to the notion that the expanding
Regulatory agencies play a role in facilitating variety of electronic healthcare data can be
the reaching of these goals. used to study the effects of medicines
To strengthen the monitoring of marketed beyond those related to safety. As methods
medicines, as noted above, EMA developed to develop real-world evidence move for-
ENCePP, a network of centers with the capac- ward, pharmacoepidemiologists, including
ity to perform post-authorization studies focus- those in regulatory agencies, will have an
ing on safety and benefit–risk. important and growing role to play, particu-
In Japan, PMDA launched the Medical larly in supporting timely and robust public
Information Database NETwork (MID-NET®) health decisions.
Conclusion Summary Points for the View

from Regulatory Agencies
Pharmacoepidemiology is an essential disci-
pline in the activities of a drug regulatory ●● Drug regulation has a public-health focus, is
agency and is used throughout the lifecycle of based in science, and is executed in the con-
a product. Clear, robust, and transparent meth- text of applicable laws and regulations.
ods of integrating data from multiple sources ●● Pharmacoepidemiology plays an important
to arrive at sound, evidence-based conclusions role in drug regulation.
are critically important. Methodologic ●● Pharmacoepidemiology is important across
advances in the field are needed to analyze the the lifecycle of a medicine.
increasingly wide range of large data sources. ●● Synthesis of data from multiple sources is
These efforts depend on collaborations among critical for sound regulatory decisions.
regulatory agencies, academia, industry, and ●● Advancing the science of pharmacoepidemi-
other stakeholders. ology is essential.
Case Example 6.3

The view from regulatory agencies: Duration (85%) varied from 1 to 90 days, yet 15% of
of the use of metoclopramide (Kaplan S., the patients appeared to have received pre-
Staffa, J.A., Dal Pan, G.J. 2007). scriptions for metoclopramide for a period
longer than 90 days. Cumulative therapy for
Background longer than 90 days was recorded for almost
The use of long-term treatment with meto- 20% of the patients. These data indicate that
clopramide is a known risk factor for tardive a substantial percentage of patients were
dyskinesia. The product label in the United taking metoclopramide for longer than the
States recommended a treatment duration recommended duration of treatment. The
of no longer than 12 weeks. manufacturer was subsequently required to
Issue add an additional warning to the product’s
The extent of use beyond the recommended label, cautioning against prolonged use.
12 weeks of treatment had not been Strengths
quantified. The data were drawn from a reasonably
large population.
Approach
Prescription claims data were used to esti- Limitations
mate duration of therapy and the extent of The data did not include information on
therapy beyond the maximum time period of diagnoses, so the outcome of tardive dyski-
12 weeks evaluated in the clinical trials and nesia could not be ascertained.
recommended in the label.
Key Points
Results ●● Drug safety problems can emerge not only
During the study period, almost 80% of from problematic drugs, but problematic
approximately 200 000 persons who had drug use
received a prescription for metoclopramide ●● Studies of the appropriateness of drug use
had only one episode of therapy. The length in populations can identify poor drug use
of the longest episode for most patients and lead to regulatory intervention
The View from the Legal System However, some products contained a high
enough inherent risk of harm that courts
Introduction decided they should be held to a different legal
standard. Starting in the early 1960s, judges
Pharmacoepidemiologists in their daily work started applying the theory of strict liability to
encounter many different aspects of the law. certain product liability cases, which merely
Three of the most important intersections of requires demonstration that the dangerous
pharmacoepidemiology and the law involve product caused the injury. As distinguished
product liability law, contract law, and intellec- from negligence, the question of whether the
tual property law. The basic legal rules in these defendants followed customary practices or
subject areas, and practical and ethical impli- exercised reasonable precautions is moot. For
cations for pharmacoepidemiology, will be dis- example, the product could have a “manufac-
cussed in turn. turing defect,” meaning that the product did
not comply with the manufacturer’s own stand-
ards, or a “design defect,” meaning that the
Tort Law and Product Liability
product was designed in a way that conferred
Lawsuits
inherently unreasonable risk for the consumer.
Individuals harmed by a drug may seek dam- However, courts also generally agreed that if
ages from its manufacturer. A basic under- the manufacturer of a dangerous product ade-
standing of product liability law is essential for quately warned about the known risks of its
pharmacoepidemiologists, even for those who product, those warnings were sufficient to insu-
might never find themselves in a courtroom, late the manufacturer from liability. Thus, strict
because such lawsuits also exert substantial product liability also allowed plaintiffs to bring
influence on the field itself. Tort litigation causes of action against manufacturers based
brought by government agencies and individ- on a third principle: a “warning defect” (also
ual patients can help uncover previously una- called a “failure to warn”).
vailable data on adverse effects, questionable In the pharmaceutical field, product liability
practices by manufacturers, and flaws in drug cases alleging a manufacturing defect are rare,
regulatory systems. in part because of strict regulatory oversight of
drug manufacturing plants. Also, cases based
The Legal Theory of Product Liability on a design defect theory are also difficult to
Product liability law is a variation of tort law win because most courts agree that all pre-
that covers the principles under which con- scription drugs have some inherent risks that
sumers harmed by products sold in interstate must be weighed against their substantial ben-
commerce may seek redress for their injuries. efits. Rather, the most common bases for litiga-
Originally, consumers were required to prove tion over pharmaceutical products are warning
four elements to make out a claim for negli- defects about the adverse event at issue. The
gence against manufacturers for creating a ultimate disposition of failure to warn cases
dangerous product: (i) that defendants had a turns on the question of whether the warning
duty to exercise reasonable care, (ii) that is reasonable.
defendants’ conduct diverged from customary
practices that would be followed by other man- Failure-to-Warn Claims
ufacturers or members of the industry, (iii) A failure-to-warn product liability action
that there was a causal link between the includes three main contentions: knowledge
defendants’ lack of care and the outcome at of the drug risk by the manufacturer, improper
issue, and (iv) that the preceding three factors warning of the drug risk, and causation of
led to damages. damages.
The View from the Legal Syste 99
Knowledge rhabdomyolysis, a potentially fatal kidney dis-

The plaintiff must demonstrate that a pharma- ease. The manufacturer, Bayer, was found to
ceutical manufacturer knew, or should have possess several reports from as early as 1999
known, of the risk. A manufacturer of a phar- suggesting a 10-fold risk of rhabdomyolysis
maceutical product cannot generally be held relative to other medications in its class, but it
accountable for risks about which it could not allegedly did not process these reports and pass
have known. For example, in one case, a plain- them along to patients or regulators. In some
tiff brought a lawsuit claiming that her oral lawsuits, Bayer was charged with having con-
contraceptive medication led to her having a structive knowledge of these concerns by 1999,
cerebrovascular accident. The jury found that because the company should have processed
the particular risk the plaintiff claimed could the reports and acted on them by that time. A
not have been known at the time the drug was common legal standard used in these situa-
prescribed, based in part on the testimony of tions is what a reasonably prudent company
the expert pharmacoepidemiologist who with expertise in this area would have done.
reported that “new techniques to measure
these clotting effects had not then been devel- Warning
oped” at the time of the injury. According to If a manufacturer has the duty to provide a
the court, “The warnings contained in the warning about adverse events associated with
package inserts were adequate or that the its product, then the next question is whether
statements contained therein were a fair repre- an adequate warning was provided. A proper
sentation of the medical and scientific knowl- warning is relevant, timely, and accurate. For
edge available at the time the drug was taken example, a relevant warning about an adverse
by the plaintiff.” effect is commensurate with the scope and
Knowledge can be actual or constructive. extent of dangers associated with the drug.
Actual knowledge is defined as literal aware- Warnings must not be subject to undue delay.
ness. Actual knowledge can be demonstrated A manufacturer must keep up with emerging
by showing that the manufacturer was cogni- scientific data and patient reports, and warn of
zant of reasonable information suggesting a new side effects discovered after initial
particular risk that it did not pass on to con- approval. In the case of rosiglitazone (Avandia),
sumers. In the case of selective serotonin reup- a 2007 meta-analysis linked the drug to life-
take inhibitors (SSRIs), used to treat depression, threatening cardiovascular adverse events.
various manufacturers were found to have However, after a review of internal company
conducted clinical trials that showed an documents, a US Senate Finance Committee
increased risk of suicidal ideation in adoles- report suggested that the manufacturer knew
cent patients taking the drug. Plaintiffs brought about these risks but delayed publicly warning
lawsuits charging that these findings were about them and sought to limit their dissemi-
delayed for lengthy periods of time, not nation. Thus, a primary question in lawsuits
released, or the concerns not fairly arising from use of rosiglitazone is whether
represented. these tactics inappropriately delayed reasona-
Constructive knowledge is sometimes called ble warnings about the adverse effect. Finally,
“legal knowledge,” because it is knowledge warnings must be of appropriately urgent tone.
that the law assumes should be present, even if In the case of rofecoxib (Vioxx), lawsuits
it is not. This knowledge could have been alleged that the warning was insufficiently
acquired by the exercise of reasonable care. For urgent because the risk of cardiovascular
example, the cholesterol-lowering drug ceriv- events was described in vague terms and
astatin (Baycol) was removed from the mar- placed in the less prominent “precautions”
ket in 2001 after it was linked to cases of section of the label.
The plaintiff must also demonstrate that the The legal standard for causation is therefore
inadequate warnings about the adverse effect challenged by product liability cases, in which
were relevant to the plaintiff’s receiving the probabilistic evidence links drugs to injuries.
drug. If a defendant can demonstrate that even Courts struggle with the question of legal cau-
an adequate warning would have made no dif- sation in these cases on two distinct levels:
ference in the decision to prescribe the drug, or general and specific causation.
to monitor the patient postprescription, the case General causation addresses whether a prod-
may be dismissed for lack of a proximate cause. uct is capable of causing a particular injury in
According to the “learned intermediary” rule, the population of patients like the plaintiff.
pharmaceutical manufacturers fulfill their The basic common law standard to prove gen-
duties to warn by providing accurate and ade- eral causation is that a particular product
quate warnings to prescribing physicians. If the “more likely than not” caused the damages.
manufacturer imparts an appropriate warning Some courts have held that legal causation
that physicians can sufficiently grasp, then the must be demonstrated by more than an asso-
manufacturer can be insulated from liability. ciation and a mere possibility of causation,
Therefore, warnings do not have to be offered even though causal hypotheses based on such
about risks that should be obvious or are gener- considerations are common in the scientific lit-
ally known to skilled medical practitioners. erature. A few courts have even gone further
However, when the information given to physi- and defined “more likely than not” as having a
cians omits, underemphasizes, misstates, or relative risk of greater than 2.0, no matter how
obfuscates dangers, this deficiency is legally tight the confidence intervals are around a sta-
transferred to the patient, who maintains a right tistically significant finding of association
of redress against the manufacturer if those between 1.0 and 2.0. Presumably this is based
dangers materialize and cause injury. on the calculation of attributable risk in the
In some special situations, pharmaceutical exposed group exceeding 50%, when the rela-
manufacturers may lose the ability to invoke tive risk exceeds 2.0. This standard has been
the learned intermediary defense. If a manu- replicated in the Federal Judicial Center’s
facturer markets its product very aggressively “Reference Manual on Scientific Evidence”
and without sufficient attention to certain and employed in some cases to exclude epide-
risks, courts may rule that it has essentially miologic evidence with weaker associations.
undone the physician-patient prescribing However, all courts do not adhere rigidly to
relationship. Direct-to-consumer advertising the relative risk = 2.0 rule for general causa-
(DTCA) is one modality that can undercut tion. Both clinical trials and epidemiologic
the assumption that patients are largely igno- studies of the product at issue can establish
rant of prescription drug risks and manufac- general causation between a pharmaceutical
turers lack means of interacting with patients product and an outcome. Animal studies,
other than through physicians. The New meta-analyses, case reports/case series, and
Jersey Supreme Court has ruled that DTCA secondary source materials (such as internal
created a limited exception to the learned company documents) have also been used in
intermediary defense, and in 2007 the West court to help support establishing a causal
Virginia Supreme Court rejected the learned link. Since pharmacoepidemiologic studies
intermediary defense in its entirety on this tend to assess the presence of an association,
basis. Nonetheless, in most jurisdictions, the rather than directly addressing causation,
learned intermediary rule still stands. courts sometimes apply the Bradford Hill
criteria to build the bridge between an associa-
Causation tion and general causation (see Table 6.1).
Legal causation usually requires a clear causal To demonstrate specific causation, a plaintiff
chain from event to outcome, in an individual. must show that the product in question caused
Table 6.1 Bradford Hill criteria. information was acted upon appropriately.
Experts usually describe the current state of
1) Strength of association knowledge about the adverse event at issue,
2) Consistency and replication of findings and may analyze available data to present
3) Specificity with respect to both the substance before the court.
and injury at issue As federal Circuit Court Judge Richard Posner
4) Temporal relationship has explained, “the courtroom is not the place
5) Biological gradient and evidence of a for scientific guesswork, even of the inspired
dose–response relationship sort.” Pharmacoepidemiologists seeking to pre-
6) Plausibility sent expert evidence in litigation will routinely
7) Coherence face judicial inquiry to determine whether they
8) Experimental removal of exposure are fit to serve in that role. Traditionally, the
9) Consideration of alternative explanation
judge evaluated whether expert witnesses lack
qualifications or espouse scientific theories out
of step with accepted knowledge. In the 1993
the alleged injury in the individual plaintiff. In case of Daubert v. Merrell Dow, the US Supreme
some cases, like instantaneous allergic reac- Court outlined a number of markers for review-
tions, the causal link is clear. For more suba- ing the appropriateness of expert witness testi-
cute or later onset responses, however, specific mony, including whether the theory was current
causation may be hard to demonstrate. For and whether it had been tested or subjected to
example, in one case against Merck brought by peer review and publication. A subsequent case
a plaintiff who suffered a myocardial infarc- applied these rules and further refined them in
tion shortly after starting rofecoxib, the manu- evaluating a debate over the admissibility of
facturer argued that the outcome was expert testimony suggesting that polychlorin-
attributable to the plaintiff’s prior existing cor- ated biphenyls (PCBs) can cause lung cancer.
onary artery disease. The plaintiff countered The research was excluded because the experts
with the fact that he was in a state of stable car- did not validate their conclusions – the epide-
diovascular health prior to initiation of miologic studies did not report a statistically sig-
rofecoxib and that he simultaneously devel- nificant causal link between PCBs and lung
oped two coronary artery clots after the drug’s cancer, lacked proper controls, and examined
initiation (a rare presentation for ischemic substances other than PCBs. In the US, some
heart disease). While the trial court held for state courts have embraced the Daubert guide-
the plaintiff, the decision was reversed on lines, which have also been taken up by revised
appeal; the appeals court ruled, “although Federal Rules of Evidence; others adhere to a
plaintiffs were not required to establish spe- more basic doctrine that excludes testimony
cific causation in terms of medical certainty, containing theories that do not enjoy “general
nor to conclusively exclude every other reason- acceptance in the relevant scientific
able hypothesis, because [the plaintiff’s] preex- community.”
isting cardiovascular disease was another
plausible cause of his death, the plaintiffs were Intersection Between Drug Regulation
required to offer evidence excluding that cause and Product Liability Litigation
with reasonable certainty.” In most countries, when the government regu-
latory authority charged with overseeing sales
Pharmacoepidemiologic Expertise and of pharmaceutical products approves a drug
Daubert for widespread use, the drug comes with an
In product liability cases, pharmacoepidemi- official drug labeling. The labeling presents a
ologists serve as expert witness, helping explain description of drug’s efficacy, including the tri-
data about drugs and determine whether risk als performed in the premarket period, as well
as safety concerns that have emerged during the following contentions: (i) defective prod-
this period of testing. In the US, the labeling is uct, (ii) causation of damage, and (iii) no exclu-
usually written by the manufacturer and sion of liability. A product is defective if it does
approved by the FDA. not provide the safety that a person is entitled
In the US, the label takes on particular legal to expect, taking all circumstances in account,
significance. The FDA requires the manufac- including the presentation of the product, the
turer to mention important warnings that are use reasonably expected of the product, and
in the official labeling when marketing its the time when the product was put into circu-
product, but does not require manufacturers to lation. Because the product liability rule is a
mention warnings that are not in the labeling. directive, member states retain some flexibility
Recently, there has been controversy over the in implementing aspects of it, such as whether
intersection between the drug label and prod- they permit compensation for non-economic
uct liability lawsuits. For example, in one case, damages (e.g. pain and suffering) or which
a man was prescribed the antidepressant ser- manufacturer defenses they seek to incorpo-
traline (Zoloft) and immediately started expe- rate. As a result of this flexibility, there is sub-
riencing agitation, confusion, and suicidal stantial diversity across EU countries in how
thinking, ultimately leading him to take his product liability cases are adjudicated.
own life one week later. The plaintiffs claimed Country-specific laws outside the EU set up
that the manufacturer failed to warn appropri- similar legal regimes. Like in the US, most
ately about the risks of suicidal behaviors. The product liability lawsuits in EU and non-EU
manufacturer contended that such a claim countries in Europe are based on failure-to-
could not be brought because the FDA had not warn claims about the adverse event at issue,
included such a warning in the official label. rather than design defects. One of the exclu-
That is, the claim was “preempted” by the sions of liability, as in US, is the learned inter-
FDA’s regulatory action. However, this view mediary defense. However, while similar
was overturned by the US Supreme Court in product liability rules apply in Europe, fewer
the seminal case of Wyeth v. Levine, which cases are brought to court and damage com-
held, “It has remained a central premise of pensation is lower.
drug regulation that the manufacturer bears
responsibility for the content of its label at all
times.” The brand-name drug manufacturer
and Contract Law
can therefore strengthen the label at its own
discretion by adding warnings to it without Many studies in the field of pharmacoepidemi-
first notifying the FDA and receiving approval ology emerge from collaborations between
to do so. Notably, if the FDA does review all the individuals at different institutions.
data surrounding a particular safety issue and Cooperative work can allow more complex
makes a specific statement that a strong warn- research to be performed and help advance the
ing is not necessary, such an action can still field of pharmacoepidemiology. One type of
preempt a failure-to-warn lawsuit. The collaborative work of particular public health
Supreme Court has also held that the responsi- importance is contract research. Contract
bility to proactively update the label does not research is undertaken by an individual, aca-
extend to generic drug manufacturers, which demic, or nonprofit investigator supported by a
only must have labels that match their brand- sponsor (usually an industry or governmental
name counterparts. agency). The contract classically represents the
Product liability law in Europe is in many full outline of the agreement between the par-
ways similar to the US. A liability action aris- ties. In countless cases, contract research in
ing under the controlling EU directive includes pharmacoepidemiology has led to important
public health findings and changes in health For researchers based in academic medical
care delivery. centers, institutional research administration
However, contract research may pose vari- offices usually handle the details of contract
ous potential concerns, generally centering negotiation with research sponsors. However,
around: (i) trial design, (ii) access to data and surveys of academic medical centers have
data analysis, and (iii) publication of results. found that academic institutions routinely
Investigators should be wary of performing engage in industry-sponsored research with-
contract research in which the sponsor has out sufficient protection for investigators. For
the right to unduly influence the design of the example, improper contracts can pass through
trial. Many sponsors prefer to retain control such offices that allow contract provisions per-
of the data and insert their own statistical mitting the research sponsor to insert its own
analyses. They argue that such efforts guard statistical analyses and draft the manuscript,
against “investigators [who] want to take the while prohibiting investigators from sharing
data beyond where the data should go,” while data with third parties after a trial had ended.
investigators argue that this arrangement pro- Whether or not they receive support from
vides the company with an opportunity to research administration offices, pharmacoepi-
“provide the spin on the data that favors demiologists must thoroughly evaluate con-
them.” Examples from both government and tracts guiding research for inappropriate
industry abound. In the case of rosiglitazone, language regarding control of design of the
a clinical trial organized by the manufacturer trial, access to data, and reporting of results
sought to compare the product against other (see Table 6.2). Problematic language includes
treatment options for diabetes, and an inde- overly broad confidentiality clauses, clauses
pendent academic steering committee was that define and assign ownership of intellec-
organized to oversee the data analysis. tual property, and clauses that require approval
Company documents suggest that the clinical from a sponsor prior to publication. It may be
trial database was exclusively controlled by reasonable to allow sponsors a limited amount
the company, which provided limited access of time to review proposed publications for
to the investigators. When members of the inadvertent release of proprietary company
steering committee questioned the presenta- information or to contribute suggestions based
tion of the results, their concerns were largely on their expertise. However, researchers have
overlooked. an ethical obligation to ensure that contracts
There have also been numerous conflicts do not unreasonably delay the publication of
over so-called “gag clauses” that prevent con- potentially important results. Poorly written
tract investigators from publishing their ulti- contracts can lead to inappropriate secrecy of
mate results. For example, after a University results, which can have public health ramifica-
of Toronto physician identified safety issues tions, as well as result in litigation against
related to an experimental drug used to researcher.
treat iron overload in transfusion-dependent
patients with thalassemia, she was not granted
permission to publish her results. When she
and Intellectual Property Law
ultimately exposed her findings, she was the
subject of a breach of contract lawsuit from the A patent is a formal grant of market exclusiv-
sponsor on the basis that her research contract ity authorized by the federal government,
provided that the published work-product was lasting for 20 years. Patents can be issued for
“secret and confidential” and could not be dis- any process, machine, manufacture, or com-
closed except with the manufacturer’s “prior position of matter. To be worthy of a patent,
written consent.” an innovation must be useful, novel, and
Table 6.2 Potentially objectionable language in research contracts for pharmacoepidemiologists.
Category Contractual terms Critique

Control over “____ shall provide confidential information to Broad definition of
investigator work CONSULTANT for the purpose of conducting “confidential information”
product the CONSULTANT’S professional services. All seems to cover all information.
information whether written or verbal provided Researcher’s work product
by, or developed for ______, and all data becomes sponsor’s confidential
collected during the performance of this information.
Agreement is deemed to be the Confidential
Information of ______.”
Gag clauses “No information regarding this Agreement or Prevents disclosure of existence
the interest of ____ or Client in the subject of the contract as a financial
matter hereof shall be disclosed to any third source in publication.
party without the prior written consent of
_____”
Opportunity to Client “shall not present or publish, nor submit Contract allows sponsor to
influence outcome for publication, any work resulting from the quash publication unless it
Services without _____ prior written approval.” approves analyses.
All examples adapted from actual contracts offered to engage in sponsored research.
onobvious. These criteria ensure that patents

n coepidemiology, including investigating char-
cannot be awarded for inventions that already acteristics of drug use and adverse events. The
exist, or small improvements on those inven- US Supreme Court has held that patentable
tions that are obvious to a person of ordinary processes may not include fundamental princi-
skill in the field. ples such as “laws of nature, natural phenom-
A patent is classically thought of as a “quid ena, or abstract ideas,” or purely mental
pro quo” between inventors and society. The process. By contrast, applications of laws of
goal of a patent is to encourage inventors to nature to a particular process may still be
invest in the development of their ideas, patentable. For example, a well-known case
because it gives them a competition-free period involved a patent over a method of curing syn-
in which to market a successful invention. At thetic rubber that used the Arrhenius Equation
the same time, in filing for a patent, an inven- to calculate the optimal cure time. The process
tor must fully disclose the content of the was found to be patentable because the for-
claimed invention in a patent document. The mula was a part of a larger inventive process
government provides its police power to pro- for curing rubber.
tect an inventor’s intellectual property for a set There are important ethical and legal con-
length of time and, in exchange, inventors cerns related to patenting processes that pro-
make their inventions available to the public vide exclusive control over various aspects of
and fully describes it, so that others can use it the conduct of pharmacoepidemiology and
and potentially improve on it in creating subse- pharmacovigilance research. In one case, an
quent innovation. HIV researcher at Stanford has faced a patent-
Patents have become increasingly visible in infringement lawsuit over a publicly-available
the practice of pharmacoepidemiology. Most database he created to help guide antiretroviral
fall into the “process” category, such as meth- therapy based on the resistance characteristics
ods of analyzing claims data and comparing of the disease, because searching this database
outcomes to identify adverse events. In recent may involve a similar process to one previously
years, numerous patents have been obtained patented (but never implemented) by a for-
on methods and techniques used in pharma- profit company.
Recently, the US Supreme Court laid down a Pharmacoepidemiologists may be involved in

new, strict standard for patentability of pro- product liability cases brought by individuals
cesses, excluding those that simply describe a against drug manufacturers, either as expert
correlation or instruct people to “gather data witnesses or on the basis of academic work
from which they may draw an inference.” A they undertake. These cases traditionally
patentable process must involve an inventive involve a claim of a failure to warn, which
and novel application of a law of nature beyond requires proof that the manufacturer knew of
“well-understood, routine, conventional activ- the safety issue, that any provided warnings
ity, previously engaged in by those in the field.” were insufficient, and that the injury received
One way to operationalize this definition is was directly caused by use of the drug.
using the machine-or-transformation test. Manufacturers can invoke a “learned interme-
That is, a process is likely to be patentable if it diary” defense to deflect responsibility onto the
can be tied to a particular machine or appara- treating physician. Pharmacoepidemiologists
tus, or if it transforms an object into a different may also be involved in contract research, but
state or thing. Notably, as pertaining to phar- should carefully consider contractual require-
macoepidemiologic patents, gathering data ments related to ownership of the work prod-
may not constitute a “transformation” because uct and withholding publication. Finally,
every algorithm inherently requires the gather- pharmacoepidemiologists may decide to try to
ing of data inputs. patent their research methods, but should
In Europe, the European Patent Convention weigh the risks and benefits of this form of
(EPC) provides the legal framework under intellectual property.
which patents are granted. Patent lengths are
the same and US and European standards with
Summary Points for the View
regard to the patentability for methods and
from the Legal System
techniques are also close. There are three pos-
sible routes for obtaining patent protection in ●● Product liability is the term for the set of
Europe. One can apply for a patent directly to principles under which consumers harmed
the national patent office of a particular coun- by products sold in interstate commerce seek
try (national patent); one can apply for a patent redress for their injuries.
to the EPO and designate specific EU member ●● A product liability case against a manufac-
states where patent protection is wanted (“clas- turer alleging failure to warn about a drug
sical” European patent); or – as part of a new risk includes three main contentions: (i)
pathway intended to start in 2018 – one can actual or constructive knowledge of the risk,
apply for a patent to the EPO with the designa- (ii) lack of a warning, or a warning that is
tion of a unitary patent that will be applicable not relevant, timely, and accurate, and (iii)
for all of the EU member states where the gov- causation of damages.
ernment has ratified the Agreement on a ●● Product liability cases involve two types of
Unified Patent Court. Decisions about which causation: general causation, which addresses
pathway is appropriate are applicant-specific whether the product can cause the alleged
and could involve considerations such as the injury in patients like the plaintiff, and spe-
need for broad geographical coverage vs pro- cific causation, which addresses whether the
tection in one (or a few) Member States. product caused the alleged injury in the indi-
vidual plaintiff. The standard for general cau-
sation is usually that the product “more likely
Conclusion
than not” caused the damages, which some
Legal issues intersect with the practice of phar- courts have interpreted as a relative risk of
macoepidemiology in a number of ways. greater than 2.0. The Bradford Hill criteria
can build the bridge between an association confidentiality clauses, clauses that define
and general causation. and assign ownership of intellectual prop-
●● Pharmaceutical manufacturers can fulfill erty, and clauses that require approval from
their duties to warn by providing accurate a sponsor prior to publication.
and adequate warnings to the prescribing ●● Patents offer 20-year period of government-
physician (the “learned intermediary” enforced market exclusivity for novel and
defense). Warnings do not have to be offered nonobvious processes or products. A patent-
about obvious risks or risks generally known able process must involve an inventive and
to skilled medical practitioners. novel application of a natural law or correla-
●● A regulatory authority’s drug label repre- tion. For example, a natural correlation may
sents its best judgment about risks that war- be patentable if it can be tied to a particular
rant disclosure and how to describe those machine, or if it can transform an object into
risks. In the US, the drug label does not a different state.
preempt manufacturers’ responsibility to ●● Patented processes that provide exclusive
monitor emerging data about adverse effects control over the conduct of pharmacoepide-
and update the label as needed. miology and pharmacovigilance research
●● Contract research is central to effective phar- can hurt the public health if they prevent
macoepidemiologic collaborations but prob- sharing of data or technologies necessary for
lematic contract terms include overly broad research into drug outcomes and effects.
Case Example 6.4

The view from the legal system By contrast, applications of laws of nature to
a particular process may still be patentable.
Background
An inventor seeks to patent a method of Question
using adverse event data regarding vaccine Is the inventor’s patent valid, or does it
administration to inform subsequent health improperly claim a “natural law”?
care delivery. The patent claims, “A method of
Approach
determining whether an immunization
The recent Supreme Court case of Mayo
schedule affects the incidence or severity of
Collaborative Services v. Prometheus Laboratories
a chronic immune- mediated disorder in a
holds that processes cannot be patentable if
treatment group of mammals, relative to a
they restate a basic scientific discovery and
control group of mammals, which comprises
then instruct physicians to gather data from
immunizing mammals in the treatment
which they may draw an inference. That is, a
group of mammals with one or more doses
patentable method cannot amount to “nothing
of one or more immunogens, according to
significantly more than an instruction to doc-
said immunization schedule, and comparing
tors to apply the applicable laws when treat-
the incidence, prevalence, frequency, or
ing their patients.”
severity of said chronic immune- mediated
disorder or the level of a marker of such a Results
disorder, in the treatment group, with that in Under the Prometheus reasoning, the method
the control group.” Patentable methods may described above likely would not reach the
not include fundamental principles such as level of a patentable invention. The inventor
“laws of nature, natural phenomena, or here has uncovered a potentially important
abstract ideas,” or purely mental processes. correlation between immunization s chedules
Further Readin 107
and patient outcomes, but these natural cor- patents on new pharmacoepidemiologic
relations are not patentable by themselves. methods and discoveries may be essential
for recouping the costs of innovation.
Strengths
●● Government patent offices, which are often
The Prometheus principle prevents patents
underfunded and understaffed, will now
from being issued on certain fundamental
face the task of distinguishing the devel-
discoveries related to pharmacoepidemiol-
opment and characterization of patentable
ogy. Patents in this field that are sufficiently
methods from other methods that simply
broad could prevent others from conducting
describe natural correlations.
necessary research into drug outcomes and
effects. Key Points
Limitations: A patentable process needs to involve an
●● Excluding certain discoveries from the inventive and novel application of a law of
possibility of patenting has led some nature or natural correlation beyond “well-
observers to worry about the implications understood, routine, conventional activity,
for private investment. The prospect of previously engaged in by those in the field.”
Further Reading
The View from Academia Clancy, C. and Collins, F.S. (2010). Patient-
centered outcomes research institute: the
Avorn, J. and Soumerai, S.B. (1983). Improving
intersection of science and health care. Sci.
drug-therapy decisions through educational
Transl. Med. 2: 37.
outreach. A randomized controlled trial of
Cutrona, S.L., Choudhry, N.K., Stedman, M. et
academically based “detailing.”. N. Engl. J.
al. (2010). Physician effectiveness in
Med. 308: 1457–1463.
interventions to improve cardiovascular
Avorn, J. (2005). Powerful Medicines: The
medication adherence: a systematic review. J.
Benefits, Risks, and Costs of Prescription Drugs.
Gen. Intern. Med. 25: 1090–1096.
New York: Knopf.
Avorn, J. (2011). Teaching clinicians about Fischer, M.A. and Avorn, J. (2004). Economic
drugs – 50 years later, whose job is it? N. Engl. implications of evidence-based prescribing for
J. Med. 364: 1185–1187. hypertension: can better care cost less? JAMA
O’Brien, M.A., Rogers, S., Jamtvedt, G. et al. 291: 1850–1856.
(2007). Educational outreach visits: effects on Fischer, M.A., Choudhry, N.K., and
professional practice and health care Winkelmayer, W.C. (2007). Impact of
outcomes. Cochrane Database Syst. Rev. (4): Medicaid prior authorization on angiotensin-
CD000409. https://doi.org/10.1002/14651858. receptor blockers: can policy promote rational
CD000409.pub2. prescribing? Health Aff. (Millwood) 26:
Choudhry, N.K., Avorn, J., Glynn, R.J. et al. 800–807.
(2011). Post-myocardial infarction free Rx Fischer, M.A., Morris, C.A., Winkelmayer, W.C.,
event and economic evaluation (MI FREEE) and Avorn, J. (2007). Nononcologic use of
trial. Full coverage for preventive medications human recombinant erythropoietin therapy in
after myocardial infarction. N. Engl. J. Med. hospitalized patients. Arch. Intern. Med. 167:
365: 2088–2097. 840–846.
Fischer, M.A., Stedman, M.R., Lii, J. et al. (2010). treatment and/or comparative effectiveness:
Primary medication non-adherence: analysis recommendations from the Joint ISPOR-ISPE
of 195, 930 electronic prescriptions. J. Gen. Special Task Force on Real-World Evidence in
Intern. Med. 25: 284–290. Healthcare Decision-Making.
Jackevicius, C.A., Li, P., and Tu, J.V. (2008a). Pharmacoepidemiol. Drug Saf. 26 (9):
Prevalence, predictors, and outcomes of 1033–1039.
primary nonadherence after acute myocardial Coplan, P., Noel, R., Levitan, B. et al. (2011).
infarction. Circulation 117: 1028–1036. Development of a framework for enhancing
Jackevicius, C.A., Tu, J.V., Demers, V. et al. the transparency, reproducibility and
(2008b). Cardiovascular outcomes after a communication of the benefit–risk balance of
change in prescription policy for clopidogrel. N. medicines. Clin. Pharmacol. Ther. 89: 312–315.
Engl. J. Med. 359: 1802–1810. European Medicines Agency. Guideline on the
Juurlink, D.N., Mamdani, M.M., Lee, D.S. et al. Exposure to Medicinal Products During
(2004). Rates of hyperkalemia after Pregnancy: Need for Post-Authorisation Data
publication of the randomized Aldactone London: Committee for Medicinal Products
evaluation study. N. Engl. J. Med. 351: for Human Use (CHMP); 2005. http://www.
543–551. ema.europa.eu/docs/en_GB/document_
Shah, N.D., Montori, V.M., Krumholz, H.M. et library/Regulatory_and_procedural_
al. (2010). Responding to an FDA warning – guideline/2009/11/WC500011303.pdf.
geographic variation in the use of European Medicines Agency, Committee for
rosiglitazone. N. Engl. J. Med. 363: Medicinal Products for Human Use (CHMP).
2081–2084. Guideline on risk assessment of medicinal
Shrank, W.H., Hoang, T., Ettner, S.L. et al. products on human reproduction and
(2006). The implications of choice: prescribing lactation: From data to labelling. 2008.
generic or preferred pharmaceuticals The Rules Governing Medicinal Products in the
improves medication adherence for chronic European Union-Volume 9A.
conditions. Arch. Intern. Med. 166: 332–337. European Medicines Agency. Guideline on
Solomon, D.H., Finkelstein, J.S., Katz, J.N. et al. good pharmacovigilance practices (GVP)
(2003). Underuse of osteoporosis medications Module V – Risk management systems (Rev
in elderly patients with fractures. Am. J. Med. 2). 2017a.
115: 398–400. European Medicines Agency. Guideline on good
Solomon, D.H., Van Houten, L., Glynn, R.J. et al. pharmacovigilance practices (GVP). Module
(2001). Academic detailing to improve use of VIII – Post-authorisation safety studies (Rev
broad-spectrum antibiotics at an academic 3). 2017b.
medical center. Arch. Intern. Med. 161: European Medicines Agency. Guideline on good
1897–1902. pharmacovigilance practices (GVP). Module
XVI – Risk minimisation measures: selection
of tools and effectiveness indicators (Rev 2).
The View from Industry
2017c.
AsPEN collaborators AM, Bergman, U., Choi, European Network of Centres for
N.K. et al. (2013). The Asian Pharmacoepidemiology and
Pharmacoepidemiology Network (AsPEN): Pharmacovigilance (ENCePP). Guide on
promoting multinational collaboration for Methodological Standards in
pharmacoepidemiologic research in Asia. Pharmacoepidemiology (Revision 6) 2017.
Pharmacoepidemiol. Drug Saf. 22 (7): 700–704. http://www.encepp.eu/standards_and_
Berger, M.L., Sox, H., and Willke, R. (2017). guidances/documents/ENCePPGuideofMeth
Good practices for real-world data studies of StandardsinPE_Rev6.pdf.
Further Readin 109
Food and Drug Administration. Guidance for development of a new model using multi-
Industry Postmarketing Studies and Clinical criteria decision analysis. Pharmacoepidemiol.
Trials 2011. https://www.federalregister.gov/ Drug Saf. 16: S2–S15.
documents/2011/04/01/2011-7707/ Nyeland, M.E.L.M. and Callréus, T. (2017).
guidance-for-industry-on-postmarketing- Evaluating the effectiveness of risk
studies-and-clinical-trials-implementation-of- minimisation measures: the application of a
section-505o3. conceptual framework to Danish real-world
Food and Drug Administration. Guidance for dabigatran data. Pharmacoepidemiol. Drug
Industry and FDA Staff: Best Practices for Saf. 26 (6): 607–614.
Conducting and Reporting Platt, R., Wilson, M., Chan, K.A., et al. (2009).
Pharmacoepidemiologic Safety Studies Using The new sentinel network--improving the
Electronic Healthcare Data. 2013. evidence of medical-product safety. N. Engl. J.
Food and Drug Administration. Guidance for Med. 361 (7): 645–647.
industry: Establishing pregnancy exposure Suissa, S., Henry, D., Caetano, P., et al. (2012).
registries Rockville 2002. http://www.fda.gov/ CNODES: the Canadian network for
downloads/Drugs/GuidanceComplianceRegul observational drug effect studies. Open Med. 6
atoryInformation/Guidances/ucm071639.pdf. (4): e134–e140.
Food and Drug Administration. Draft Guidance Wang, S.V., Schneeweiss, S., and On behalf of
for Industry: FDA’s Application of Statutory the joint ISPE-ISPOR Special Task Force on
Factors in Determining When a REMS Is Real World Evidence in Health Care Decision
Necessary. 2016. Making (2017). Reporting to improve
Institute of Medicine (IOM). The Future of Drug reproducibility and facilitate validity
Safety: Promoting and Protecting the Health assessment for healthcare database studies
of the Public 2006. http://www.iom.edu/ V1.0. Pharmacoepidemiol. Drug Saf. 26 (9):
Reports/2006/The-Future-of-Drug-Safety- 1018–1032.
Promoting-and-Protecting-the-Health-of-the-
Public.aspx.
The View from Regulatory Agencies
International Society for Pharmacoepidemiology
(2008). Guidelines for good Braun, M.M., Farag-El-Massah, S., Xu, K., and
pharmacoepidemiology practices (GPP). Cote, T.R. (2010). Emergence of orphan drugs
Pharmacoepidemiol. Drug Saf. 17 (2): 200–208. in the United States: a quantitative assessment
Ishiguro, C.T.Y., Uyama, Y., and Tawaragi, T. of the first 25 years. Nat. Rev. Drug Discov. 9:
(2016). The MIHARI project: establishing a 519–522. https://doi.org/10.1038/nrd3160.
new framework for pharmacoepidemiological Fleming, T.R. and Powers, J.H. (2012).
drug safety assessments by the Biomarkers and surrogate endpoints in
Pharmaceuticals and Medical Devices Agency clinical trials. Stat. Med. 10: 2973–2984.
of Japan. Pharmacoepidemiol. Drug Saf. 25: https://doi.org/10.1002/sim.5403.
854–859. Goedecke, T., Morales, D., Pacarariu, A., and
Klungel, O.H., Kurz, X., de Groot, M.C.H. et al. Kurz, X. (2017 Nov 5). Measuring the impact
(2016). Multi-centre, multi-database studies of medicines regulatory interventions –
with common protocols: lessons learnt from systematic review and methodological
the IMI PROTECT project. considerations. Br. J. Clin. Pharmacol. https://
Pharmacoepidemiol. Drug Saf. 25 (Suppl 1): doi.org/10.1111/bcp.13469 [Epub ahead of
156–165. print].
Mussen, F., Salek, S., and Walker, S. (2007). A Hunter, N.L., Rao, G.R., and Sherman, R.E.
quantitative approach to benefit–risk (2017). Flexibility in the FDA approach to
assessment of medicines – part 1: the orphan drug development. Nat. Rev. Drug
Discov. 16: 737–738. https://doi.org/10.1038/ Litigation. Philadelphia, PA: University of

nrd.2017.151. Pennsylvania Press.
Kaplan, S., Staffa, J.A., and Dal Pan, G.J. (2007). Hill, A.B. (1965). The environment and disease:
Duration of therapy with metoclopramide: a association or causation? Proc. R. Soc. Med. 58:
prescription claims data study. 295–300.
Pharmacoepidemiol Drug Saf. Aug. 16 (8): Kesselheim, A.S. and Avorn, J. (2007). The role
878–881. of litigation in defining drug risks. JAMA 297:
Knox, C., Hampp, C., Willy, M. et al. (2015). 308–311.
Patient understanding of drug risks: an Kessler, D.A. and Vladeck, D.C. (2008). A critical
evaluation of medication guide assessments. examination of the FDA’s efforts to preempt
Pharmacoepidemiol. Drug Saf. 24: 518–525. failure-to-warn claims. Georgetown Law J. 96:
https://doi.org/10.1002/pds.3762. 461–495.
Schieppati, A., Henter, J.I., Daina, E., and Shapo, M.S. (2008). Experimenting with the
Aperia, A. (2008). Why rare diseases are an Consumer: The Mass Testing of Risky Products
important medical and social issue. Lancet on the American Public. Westport, CT: Praeger.
371: 2039–2041. https://doi.org/10.1016/
S0140-6736(08)60872-7.
Contract-Related Issues
Sherman, R.E., Anderson, S.A., Dal Pan, G.J.
in Pharmacoepidemiology
et al. (2016). Real-world evidence – what is it
and what can it tell us? N. Engl. J. Med. 375: Bodenheimer, T. (2000). Uneasy alliance –
2293–2297. https://doi.org/10.1056/ clinical investigators and the pharmaceutical
NEJMsb1609216. industry. New Engl. J. Med. 342: 1539–1544.
Temple, R.J. (1995). A regulatory authority’s Eichler, H.G., Kong, S.X., and Grégoire, J.P.
opinion about surrogate endpoints. In: (2006). Outcomes research collaborations
Clinical Measurement in Drug Evaluation (eds. between third-party payers, academia, and
W.S. Nimmo and G.T. Tucker). New York: pharmaceutical manufacturers: what can we
Wiley. learn from clinical research? Eur. J. Health
Econ. 7 (2): 129–135.
Kong, S.X. and Wertheimer, A.I. (1998).
The View from the Legal System
Outcomes research: collaboration among
Drug Safety and Tort Law academic researchers, managed care
Angell, M. (1997). Science on Trial: The Clash of organizations, and pharmaceutical
Medical Evidence and the Law in the Breast manufacturers. Am. J. Manag. Care 4 (1):
Implant Case. New York: W. W. Norton & Co. 28–34.
Avorn, J. (2004). Powerful Medicines: The Washburn, J. and University Inc (2006). The
Benefits, Risks and Costs of Prescription Drugs. Corporate Corruption of Higher Education.
New York, NY: Alfred A Knopf. New York: Basic Books.
Brennan, T.A. (1988). Causal chains and
statistical links: the role of scientific
Patent Law and
uncertainty in hazardous-substance litigation.
Cornell Law Rev. 73: 469–533.
Federal Judicial Center (2000) Reference Manual Heller, M.A. and Eisenberg, R.S. (1998). Can
on Scientific Evidence (2nd edn). http://www. patents deter innovation: the anticommons in
fjc.gov/public/pdf.nsf/lookup/sciman00. biomedical research. Science 280: 698–701.
pdf/$file/sciman00.pdf Jaffe, A.B. and Lerner, J. (2004). Innovation and
Green, M.D. (1996). Bendectin and Birth Defects: Its Discontents: How Our Broken Patent System
The Challenges of Mass Toxic Substances Is Endangering Innovation and Progress, and
Further Readin 111
What To Do About It. Princeton, NJ: Princeton National Research Council (2004). A Patent
University Press. System for the 21st Century (eds. S.A. Merrill,
Kesselheim, A.S. and Karlawish, J. (2012). R.C. Levin and M.B. Myers). Washington, DC:
Biomarkers unbound – the supreme Court’s The National Academies Press.
ruling on diagnostic-test patents. N. Engl. J. Walker, A.M. (2006). More lawyers, more
Med. 366: 2338–2340. bureaucrats, less information on drug
Nard, C.A. and Wagner, R.P. (2007). Patent Law: safety. Pharmacoepidemiol. Drug Saf. 15:
Concepts and Insights. St. Paul, MN: 394–395.
Foundation Press.
113
Part II
Sources of Pharmacoepidemiology Data

115
Postmarketing Spontaneous Pharmacovigilance Reporting

Systems
Gerald J. Dal Pan1, Marie Lindquist2, and Kate Gelperin1
1
Office of Surveillance and Epidemiology, Center for Drug Evaluation and Research, US Food and Drug Administration, Silver Spring, MD, USA
2
Uppsala Monitoring Centre, WHO Collaborating Centre for International Drug Monitoring, Uppsala, Sweden
The views expressed herein are those of the action when needed to protect the public from
authors, and not necessarily of the US Food unnecessary risks or harms. At times, a causal
and Drug Administration. association between a drug and an AE may
seem clear due to strong temporal association
between exposure to the drug and onset of an
Introduction adverse effect, or when there is confirmation of
positive re-challenge (i.e. signs or symptoms
Potential signals for adverse drug reactions resolve when exposure is stopped but recur
(ADRs) or adverse drug effects for marketed when re-introduced). But more often, causality
products most often arise from postmarketing assessment is challenging, and well-designed
spontaneous case reports, which are collated pharmacoepidemiology or other clinical stud-
and analyzed by drug safety experts, evaluated ies are needed to assess the signal and quantify
as clinical case series, and considered for poten- the risk.
tial regulatory action. These efforts are not pos- In the United States, the Food and Drug
sible without input from dedicated health Administration (FDA) issues Drug Safety
professionals and other concerned stakehold- Communications (DSCs) to alert the public
ers. Adverse events (AEs) thought to be poten- about emerging safety issues, such as investi-
tially drug-related may be reported by a gations into safety signals that may have a clin-
consumer or a health professional to a drug’s ically important impact on a product’s safety
manufacturer, or they may be reported directly profile. Recently, FDA launched a new web
to a health authority through programs such as portal that enables the public to view summary
MedWatch or Eudravigilance. In addition, case charts and listings of de-identified cases from
reports and case series with valuable clinical the FDA Adverse Event Reporting System
details may be published in a peer-reviewed (FAERS), a compilation of all postmarketing
journal. Concerned stakeholders – health pro- AE reports received by FDA.
fessionals as well as patients and consumers – The term “pharmacovigilance” is widely
are the source of the signals that can trigger used to denote postmarketing safety activities,
further investigation and appropriate regulatory and is defined by the World Health
116 7 Postmarketing Spontaneous Pharmacovigilance Reporting Systems
Organization (WHO) as “the science and activ- outside of the approved range. For these rea-
ities relating to the detection, assessment, sons, patients treated in clinical practice are
understanding and prevention of adverse more diverse than those treated in clinical tri-
effects or any other possible drug-related als. A postmarketing drug pharmacovigilance
problems.” reporting system is therefore necessary.
Monitoring and understanding the safety of
drug and therapeutic biologic products is a
process that proceeds throughout the product’s
life cycle, spanning the period prior to first
Description
administration to humans through the entire
Adverse Events and Adverse Drug
marketing life of the product. Throughout the
Reactions
product lifecycle, astute clinical observations
made at the point of care constitute an impor- A key concept in pharmacovigilance is the dis-
tant source of information. While new tech- tinction between an adverse event and an
nologies have enabled more thorough adverse drug reaction. The International
knowledge of a drug’s actions, and computer- Conference on Harmonization of Technical
ized databases have enabled large-scale, Requirements for Registration of
population-based analyses of drug safety inves- Pharmaceuticals for Human Use (ICH) E2D
tigations, these advancements are adjuncts to, guideline on Post-Approval Safety Data
and not substitutes for, careful, well thought- Management: Definitions and Standards for
out clinical observations. Expedited Reporting, defines an AE as follows:
Though the preapproval testing of a drug is
typically rigorous, and the review of the data is An adverse event (AE) is any untoward
thorough, there are still inevitable uncertain- medical occurrence in a patient admin-
ties about the complete safety profile of a drug istered a medicinal product and which
when it is brought to market. Several factors does not necessarily have to have a
contribute to these uncertainties. First, the causal relationship with this treatment.
number of patients treated with the drug prior An adverse event can therefore be any
to approval is limited, generally from several unfavorable and unintended sign (for
hundred to a few thousand. Second, patients in example, an abnormal laboratory find-
clinical trials tend to be carefully selected for ing), symptom, or disease temporally
inclusion in these trials, and are thus more associated with the use of a medicinal
clinically homogeneous than patients treated product, whether or not considered
in the course of clinical practice once a drug is related to this medicinal product.
marketed. Compared to patients in clinical tri-
als, patients treated in clinical practice may The same guideline describes an ADR as
have a broader range of co-morbidities, take a follows:
wider variety of concomitant medications, and
have a wider clinical severity spectrum of the All noxious and unintended responses
underlying disease being treated. Third, addi- to a medicinal product related to any
tional populations of patients, such as children dose should be considered adverse drug
or older adults, who may not have been studied reactions.
in large numbers in premarketing clinical tri-
als, may be treated with the product once it is The phrase “responses to a medicinal product”
marketed. In addition, marketed drug prod- means that a causal relationship between a
ucts are often used for diseases or conditions medicinal product and an AE is at least a
for which they are not indicated, or at doses possibility.
Descriptio 117
A reaction, in contrast to an event, is character- The scope of pharmacovigilance is broad.

ized by the fact that a causal relationship The core activity is usually the identification of
between the drug and the occurrence is sus- previously unrecognized AEs/ADRs with the
pected. If an event is spontaneously reported, use of the drug. However, it is not sufficient
even if the relationship is unknown or simply to note that use of a drug can lead to an
unstated, it meets the definition of an adverse AE/ADR. Rather, an investigation into not
drug reaction. only the potential causal role of the drug in the
The principal difference between an AE and development of the AE/ADR, but also into
an ADR is that a causal relationship is sus- the conditions leading to the occurrence of the
pected for the latter, but is not required for the AE/ADR in one person or population and not
former. In this framework, ADRs are a subset in others must be the focus of any postmarket
of AEs. In some countries, postmarketing drug safety effort. Factors such as dose–
pharmacovigilance reporting systems are response relationships, drug–drug interac-
focused on ADRs, while in others data on AEs tions, drug–disease interactions, drug–food
are collected. In the United States, for example, interactions, and the possibility of medication
the scope of reporting requirements is “[a]ny errors must be carefully considered. A full
adverse event associated with the use of a drug understanding of the factors that can lead to an
in humans, whether or not considered drug AE/ADR may yield ideas for effective interven-
related . . .” tions to minimize the severity or occurrence of
While many of the principles discussed in the AE/ADR, and thus enhance the safe use of
this chapter apply equally to AEs and ADRs, the drug. For this reason, the approach to
it is important to understand the distinction detecting and understanding clinically impor-
between these two concepts. Specifically, tant AEs/ADRs in the postmarketing period
some databases may contain only ADRs, must be as comprehensive as possible.
while others may contain AEs. These data- The identification of a new safety issue with
bases may behave differently when used for a medicinal product often begins with a single
data mining. However, because many of the observation. In the postmarketing period, such
principles of drug safety surveillance apply observations are usually clinical observations,
to both AEs and ADRs, we will use the term often made at the point of care in the course of
“AE/ADR” to refer to these two terms col- clinical practice. A practitioner or patient notes
lectively in this chapter, for convenience. the development of symptoms or signs that
When needed, we will use the individual were not present, or were present in less severe
terms if a distinction between the two is form, prior to the patient’s using the medicine.
required. If this sign or symptom is not listed in the
product’s approved labeling, patients and
healthcare professionals may not think to
Overview of Pharmacovigilance
attribute it to the medicine. If further evalua-
Reporting Systems
tion reveals a clinically significant process (e.g.
The goal of a postmarketing, or post- liver injury, rhabdomyolysis, agranulocytosis),
approval, safety program is to identify drug- it is important to keep in mind the possibility
related AEs or ADRs that were not identified of a side effect due to a medication in the dif-
prior to a drug’s approval, to refine knowl- ferential diagnosis of the event. If a medication
edge of the known adverse effects of a drug, side effect is not included in the list of possible
and to understand better the conditions conditions or diseases that could be causing
under which the safe use of a drug can be the observed problem, the patient may not be
assured. treated appropriately.
In the postmarketing period, the investiga- whether to report an AE and what events to
tion of AEs/ADRs is a multidisciplinary one. report. Sometimes, spontaneous reporting sys-
The analysis of a complex AE/ADR can involve tems are also labeled as “passive,” based on the
the fields of medicine, pharmacology, epidemi- argument that the reporting center or the man-
ology, statistics, pharmacy, toxicology, and oth- ufacturer passively receives this information,
ers. There are several methods of clinical rather than actively seeking it out. However,
postmarketing safety assessment. These this term does not do justice to the proactive
include the review of case reports and case way in which many pharmacovigilance centers
series from spontaneous reporting systems and seek to operate, even if resource constraints
published medical literature, a wide variety of often limit the ability to interact adequately
types of observational epidemiologic studies, with reporters. Moreover, “spontaneous
and clinical trials. This chapter will focus on reporting” does not fit well with the reporting
spontaneous pharmacovigilance reporting sys- situation of today, when most countries have
tems. No one method is a priori better than introduced or enacted legislation which man-
another in all settings. Rather, the choice of dates reporting from pharmaceutical compa-
methods depends on the particular safety nies. For marketed products, companies often
question to be answered. conduct planned, structured interactions with
patients and practitioners (e.g. patient support
programs) that have the potential to generate
The Concept of Spontaneous AE/
AE reports that may not otherwise have been
ADR Reporting
communicated. Reporting may also include
A core aspect of pharmacovigilance is the vol- canvassed or stimulated reporting of suspected
untary reporting of AEs/ADRs either directly reactions of particular interest (see also below,
to established national or regional centers, or in the section “National pharmacovigilance
alternatively to pharmaceutical manufactur- systems”).
ers, who in turn are obligated to report the Underlying the concept of a spontaneous
information to regulators. National reporting postmarketing AE/ADR pharmacovigilance
systems are typically run by regulatory agen- reporting system is the notion that clinical
cies (e.g. the US FDA runs the MedWatch pro- observations made at the point of care are
gram) or by centers designated by the health often valuable pieces of information in further
ministry or the drug regulatory authority. In a refining the knowledge of a drug’s safety pro-
few countries, the national pharmacovigilance file. This is an important, though frequently
center is run by a university or other scientific underemphasized, idea. After approval, when
body. In the United States, AEs/ADRs in indi- formal study often ends and marketing of the
vidual patients are generally identified at the medicine begins, there is often no further sys-
point of care. Patients, physicians, nurses, tematic way to continue the study of a medi-
pharmacists, or anyone else who suspects that cine’s safety, or even to generate drug safety
there may be an association between an AE/ hypotheses. While scientific advances and
ADR and a drug or therapeutic biologic prod- access to new data sources (e.g. electronic
uct are encouraged to, but are generally not healthcare records) may provide some oppor-
required to, report the case to either the manu- tunity to monitor the safety of a marketed
facturer or to the FDA. medicine, these alternative approaches to
This system of AE/ADR reporting is often safety signal detection remain unproven.
referred to as a spontaneous reporting system; When healthcare professionals, patients,
“spontaneous” because the person who ini- and consumers want to make a notification of
tially reports the AE/ADR to either the report- a potentially adverse effect of a medication, it is
ing center or to the manufacturer chooses useful for this information to be systematically
Descriptio 119
organized, stored, and analyzed. A reporting ranged from 5.0% (atorvastatin) to 31.2%
system fills this need. If such information were (cerivastatin). Further analysis revealed that
not systematically collected, potentially valua- the high proportion of reporting of cerivasta-
ble data about medicines would be lost. This tin cases was driven by reports received after
system implies an important role for health- the dissemination of a Dear Healthcare
care professionals in postmarketing safety Professional letter notifying physicians of the
assessment, as the quality of reports is always risks of cerivastatin-associated rhabdomyoly-
dependent on the details provided by health- sis. The estimated extent of reporting was
care professionals. 14.8% before the letter and rose to 35.0% after.
Because most AE/ADR reporting systems It is important to note that the results of this
rely on healthcare professionals, patients, and study apply only to reporting cases of statin-
consumers to submit reports voluntarily, it is associated rhabdomyolysis. The extent of
generally recognized that there is substantial reporting for different drug-adverse pairs will
underreporting of AEs/ADRs via current be different, and cannot be estimated from the
reporting systems. Two survey-based studies results of this study.
conducted in the US in the 1980s, one in Once case reports are received by national
Maryland and the other in Rhode Island, pharmacovigilance centers, they are entered
examined physician reporting to FDA, and into AE/ADR databases. These databases can
concluded that fewer than 10% of AEs/ADRs then be inspected for drug safety signals, which
were reported to FDA. These studies were conform the basis of further study, necessary regu-
ducted prior to the development of the current latory action, or both.
MedWatch program in 1993, and do not con-
sider the contribution of reporting from
Report Characteristics
sources other than physicians. Calculating the
proportion of AE reports that a reporting sys- The individual case report is the fundamental
tem actually receives requires that the true unit of a postmarketing pharmacovigilance
number of AEs/ADRs in the population be reporting system. The extent to which such a
known. For most AEs/ADRs, this number is reporting system can address specific drug
not known or readily available. In some cases, safety questions depends, in large part, on the
however, data are available that allow an esti- characteristics and quality of the individual
mate of the extent of reporting to be calculated. reports. Specific report formats differ across
For example, the extent of reporting to FDA of jurisdictions, though many countries and
cases of hospitalized rhabdomyolysis associ- regions collect information compatible with
ated with statin use was estimated using a pro- the ICH E2B format. The standard is designed
jected estimate of the number of such cases in to work with a variety of national and interna-
the United States and comparing it to the num- tional systems and incorporates endorsement
ber of reports of statin-associated hospitalized of standards by participating Standards
rhabdomyolysis in FAERS, a database that Development Organizations such as the
houses FDA’s postmarketing AE reports. The International Standards Organization (ISO),
projected national estimate was obtained by Health Level Seven (HL7), European
using incidence rates from a population-based Committee for Standardization (CEN), and
cohort study, and applying those incidence Clinical Data Interchange Standards
rates to national estimates of statin use. Across Consortium (CDISC) to enable wider inter-
four statins (atorvastatin, cerivastatin, pravas- operability across the regulatory and health-
tatin, and simvastatin), the estimated overall care communities. Although comprehensive
extent of AE reporting was 17.7%. For individ- in scope, the format also allows for limited
ual statins, the estimated extent of reporting data to be submitted. The principal domains of
case information in the ICH E2B standard on many years of reviewing case reports, is that
include: (i) patient characteristics, (ii) reaction(s) while a substantial amount of useful clinical
or event(s), (iii) results of tests and procedures information can be written in a succinct narra-
relevant to the investigation of the patient, (iv) tive, most narratives are incomplete, many to
drug(s) information, and (v) a narrative case the extent that they are uninterpretable. While
summary and further information. follow-up with the reporter is sometimes feasi-
Regardless of the specific formatting require- ble for drug safety analysts to perform during
ments across jurisdictions, there are some fun- case review, this has been the exception not
damental components of an individual safety the rule, often due to resource constraints.
report that are important for a thorough Incomplete and uninterpretable case reports
review. limit the effectiveness of postmarket pharma-
Product identification, in as much detail as covigilance reporting systems. Attempts to
possible, is essential for an assessment of a improve the systems will need to address the
case report. For pharmaceuticals, the identifi- problem of poor case report quality rather than
cation of the active ingredient(s) is critical to merely increasing the number of reports.
product identification. However, other factors Unfortunately, it is not unusual for FDA to
can also be important, depending on the spe- receive potentially important spontaneous
cific safety question. For example, the formula- reports which cannot be evaluated because of
tion of the product can be important, as certain missing key information. For instance, many
active ingredients may be present in a variety spontaneous reports of hypersensitivity AEs/
of formulations. For example, many opioid ADRs associated with heparin administration
agents are available in oral, injectable, and during an investigation of tainted heparin
transdermal formulations. Because the phar- were excluded from an FDA analysis published
macokinetic and other pharmaceutical proper- in 2010 because they lacked “basic or critical
ties can differ across these formulations, clinical information.”
information about the formulation is impor- Information on product use should include
tant in determining if there are formulation the start date(s), stop date(s), dose(s), fre-
specific effects, including those that may result quency of use, and indication for use. Dosage
from medication errors. Additionally, if the information is important in exploring dose-
drug safety question involves the assessment event relationships. Duration of use is impor-
of an AE/ADR related to a product quality tant for characterizing the time course of AEs/
defect, information on both manufacturer and ADRs relative to initiation of product use.
lot/batch number can be very important, as Indication for use is also an important piece of
product quality problems typically involve spe- information, as many products are used for
cific lots from an individual manufacturer. more than one indication (either on-label or
Reports describing medication errors, or the off-label).
potential for medication errors, ideally contain Patient information should include age, gen-
information on the product involved, the der, medical history, and concomitant medica-
sequence of events leading up to the error, the tion usage. The presence of factors that could
work environment in which the error occurred, confound the relationship of the drug to the
and the type of error that occurred. AE/ADR, especially elements of the medical
Characteristics of a good quality case report history and concomitant medication usage, are
include adequate information on product use, critical to the interpretation of individual case
patient characteristics, medical history, and safety reports.
concomitant treatments, and a description of A description of the AE/ADR that allows for
the AE/ADR, including response to treatments independent medical assessment is critical. A
and clinical outcome. Our experience, based narrative of the event that includes the
Descriptio 121
t emporal relationship of drug usage to the (https://web-radr.eu/) to develop new techni-

development of the AE/ADR, the clinical and cal tools to facilitate the detection and analysis
diagnostic features, the clinical course, any of potential ADRs in social media sites. It also
measures instituted to treat the AE/ADR, the aimed to develop a mobile phone app for the
response to these measures, and the clinical reporting of suspected ADRs to regulatory
outcome are all essential components of a high authorities in the European Union (in the con-
quality case report. Results of laboratory tests, text of traditional AE reporting). One of sev-
imaging studies, and pathology results facili- eral planned outgrowths of these efforts is the
tate an independent interpretation of the establishment of a regulatory framework for
report. Information on de-challenge (the social media mining for ADRs. Preliminary
course of the AE/ADR when the medication is recommendations from IMI WEB-RADR for a
withdrawn) and re-challenge (the determina- regulatory framework note that data from
tion of whether the AE/ADR recurs when the social media should be treated as a “secondary
drug is re-introduced), if available, can be use of data,” the use of social media for signal
invaluable. detection and validation should be optional,
reporting of individual case safety reports of
ADRs from social media sites should not be
Social Media
required, and that follow-up with social media
Social media are a range of computer-based users should not be required. Rather, drug
technologies that allow the creation and shar- manufacturers should include insights gained
ing of information, ideas, photographs and from social media regarding the safety of their
other messages via electronic communication. products in the product’s periodic safety
User-generated content is a defining feature of update report or risk management plan. More
social media. This content can be made avail- work is needed to refine the methods of
able to others via computer-based networks extracting and analyzing data from social
that connect one user with other users or media for detection of ADRs. Results of the
groups to form social networks. Depending on WEB-RADR project yielded a realistic and
privacy settings, which in some cases may be cautionary appraisal of the current ability of
chosen by the user, content may be widely social media to provide primary pharmacovig-
available to other users or it may be restricted ilance information.
to only certain users or groups. Given the
widespread use of the internet and, to a lesser
National Pharmacovigilance
degree, of social media, for health-related top-
Systems
ics, there is interest in whether social media
can be a source of drug safety signals or other- The organization of postmarketing safety
wise shed light on ADRs. Because social media reporting systems and national pharmacovigi-
posts describe individual experiences, they lance systems varies around the world. The
can, in theory, describe adverse reactions to fundamental feature is that health profession-
medicines. The use of social media for phar- als, and in some cases patients or consumers,
macovigilance presents both opportunities are encouraged to send reports of AEs/ADRs
and challenges, and is an area of active to one or more specified locations. These loca-
research. In 2014, the Innovative Medicines tions can be the drug regulatory authority, an
Initiative (IMI), a public-private partnership academic or hospital-based pharmacovigilance
between the European Union and the center (often working with or on behalf of a
European Federation for Pharmaceutical drug regulatory authority), or the drug manu-
Industries and Associations, launched WEB- facturer. The roles of these institutions vary
RADR: Recognising Adverse Drug Reactions from country to country, and depend greatly
on the regulatory and national drug monitor- The WHO is working together with major
ing system in the country. donor organizations to address the urgent need
In low-and middle-income countries, with for capacity building in low- and middle-
varying regulatory infrastructure, the focus income countries. The strategy is focused on
in pharmacovigilance has been different sustainable development, covering not only
from that in the more affluent parts of the the implementation of reporting systems,
world. Reports can result from counterfeit technical support, and training of healthcare
and sub-standard drugs, known ADRs and professionals, but also improvements in gov-
drug interactions of concern to reporters, and ernance and infrastructure to support pharma-
ADRs resulting from medical error. In some covigilance activities in the broader context of
countries, responding to queries about regulatory systems strengthening.
adverse reaction incidence, diagnosis, and The perceived responsibility of healthcare
management are a major part of the work of professionals to report AEs/ADRs often varies
pharmacovigilance centers. In developing around the world. Because the largest gaps in
countries, there are often deficiencies in drug safety knowledge are believed to be for
access to up-to-date information on drug recently approved medicines, most countries
safety that need remedying. On the other emphasize the need to report AEs/ADRs, even
hand, large donations of new drugs to com- less serious ones, for this group of medicines.
bat the endemic scourges of malaria, HIV/ For example, in the United Kingdom, recently
AIDS, tuberculosis, infestations, and other approved drugs containing new active ingre-
diseases, along with vaccines, have led to the dients are marked in the British National
high priority of monitoring their use for both Formulary with a black triangle, a symbol used
safety and efficacy. to denote a drug product whose active ingredi-
However, in many low- and middle-income ent has been newly licensed for use in the UK.
countries there is currently not enough capac- In some cases, drug products meeting certain
ity for effective drug safety monitoring, and additional criteria are also marked with a black
the improved access to new medicines adds triangle, even if the active ingredient has been
additional strain on already overburdened or previously approved. The aim of the black tri-
non-existent pharmacovigilance systems. A angle program is to prompt health profession-
survey from 2010 of pharmacovigilance sys- als to report all suspected adverse reactions
tems in low- and middle-income countries associated with the use of these products. In
found that seven of 55 responding countries some countries, it is mandatory for physicians
indicated that they had no designated system and dentists to report cases of suspected ADRs
in place, and fewer than half of the respond- to the regulatory authority. Most countries,
ents had a budget for pharmacovigilance. however, do not have such specific programs
Consequently, lack of funding was mentioned or requirements, but health professionals are
as a hindrance to the development of pharma- encouraged to report and the national report-
covigilance, together with lack of training and ing centers provide general advice to health
a culture that does not promote AE/ADR professionals on what events to report.
reporting. Suggested key developments In a majority of countries, including coun-
included: training for health workers and tries in the ICH regions, other high income
pharmacovigilance program managers; active countries, and 33 of 55 low- and middle-
surveillance methods, sentinel sites and regis- income countries responding to a 2008 survey,
tries; and better collaboration between phar- pharmaceutical companies that hold market-
macovigilance centers and public health ing authorizations are obligated to report AEs
programs, with a designated budget for phar- or ADRs to the regulatory authority. In some
macovigilance included in the latter. countries, the event is reportable only if an
Descriptio 123
attribution of causality has been made. In of conduct of physicians in Germany, all ADRs
other countries, the event is reportable even if should be reported to the Drug Commission.
no attribution has been made. For example, in In the Netherlands, the practical responsibility
the United States, pharmaceutical companies for post-marketing surveillance is shared
are required by law to submit spontaneous between the Medicines Evaluation Board
reports of AEs/ADRs, regardless of attribution (MEB) and the Netherlands Pharmacovigilance
of causality, on an expedited basis if they are Centre (Lareb). The MEB handles communica-
serious and unexpected. The AE/ADR is con- tions with market authorization holders; the
sidered serious when the patient outcome is: role of Lareb is to process and analyze reports
death; life-threatening; hospitalization (initial from health professionals and patients.
or prolonged); disability; congenital anomaly; Decentralized drug monitoring systems exist
or, requires intervention to prevent permanent both within and outside the ICH region. In
impairment or damage. Periodic reporting of France, the French Medicines Agency coordi-
other types of AEs/ADRs, such as those con- nates the network of 31 regional centers that
sidered serious and expected (labeled), or non- are connected to major university hospitals. In
serious, is typically required as well. The the United Kingdom, there are four regional
periodicity of such aggregate reports is deter- centers connected to university hospitals,
mined by the length of time the drug has been which have a special function of encouraging
marketed, with increased frequency for newly reporting in their regions. Since 2018, the
approved drugs, and decreased (e.g. annual) pharmacovigilance system in China is regu-
with older drugs. lated by the National Medical Products
While spontaneous reports of AEs/ADRs Administration (NMPA), within the State
usually originate initially from the point of Administration for Market Regulation. In
care, the more proximal source of reports com- India, the Pharmacovigilance Programme of
ing into the national pharmacovigilance cent- India has been in operation since 2010, with
ers may vary, with reports most often being the Indian Pharmacopoeia Commission (IPC)
received from drug manufacturers or from running the National Coordinating Centre.
healthcare professionals and patients. Some The system is now operating nationwide, with
countries restrict reports to only those origi- 250 local monitoring centers in medical
nating from physicians. Other countries also institutes.
accept reports from pharmacists, nurses, and
patients. There is a current trend toward
National and International
encouraging direct patient or consumer report-
Postmarketing Adverse Event
ing, replacing the notion held by many in the
Databases
past that such reports would not be a reliable
and useful source of information. Once submitted to the national drug safety
In most countries, the national pharma- monitoring program, individual case safety
covigilance center is part of the drug regula- reports are stored in computerized postmarket-
tory authority; in some, the monitoring is ing AE databases. Examples of national report-
carried out jointly by the drug regulatory ing systems and databases include the “Blue
authority/Ministry of Health and an independ- Card” system (Australia), Canada Vigilance
ent institution. In Germany, the Federal (Canada), the Canadian Adverse Events
Institute for Drugs and Medical Devices Following Immunization Surveillance System
(BfArM) maintains a joint database for record- (CAEFISS) database (Canada), the French
ing reported ADRs, together with the Drug Pharmacovigilance Spontaneous Reporting
Commission of the German Medical System database (France), the Adverse Drug
Profession. According to the professional code Reaction Information Management System of
the Pharmaceutical and Medication Devices database include reports only from health pro-
Agency, Ministry of Health, Labor, and Welfare fessionals, or does it also include reports from
(Japan), the Lareb database (Netherlands), the patients and consumers? Third, what is the
BISI database (Sweden), the MHRA ADR data- range of medical products included in the
base (United Kingdom), the FAERS database database – drugs, biologicals, blood, blood
(United States), and the Vaccine Adverse Event products, vaccines, dietary supplements?
Reporting System (VAERS) database (United Fourth, does the database include reports from
States). In addition, there are two international only one country or region, or does it include
reporting and database systems: EudraVigilance reports from regions outside the jurisdiction of
in the European Union (run by the European the regulatory authority? Fifth, does the data-
Medicines Agency, EMA); and VigiBase, pool- base include both “nonserious” and “serious”
ing data from more than 130 member countries AEs/ ADRs; if so, what proportion of the
of the WHO International Drug Monitoring reports have been classified by the health
Programme (run by the Uppsala Monitoring authority (or other database manager) as seri-
Centre, UMC). VigiBase is also the database ous? Sixth, does the database include all AEs
model used as the national database by around (i.e. events which may or may not be judged to
70 pharmacovigilance centers around the be causally related to a medicine) or does it
world; reports are entered, stored and managed include only ADRs (i.e. events for which a
in an internet-based data management system, likely causal relationship has been determined
VigiFlow, from which the data is seamlessly prior to entering the report into the database)?
transferred to VigiBase, and can be analyzed Seventh, how many individual case safety
remotely through a search and analytical inter- reports are in the database? Each of these fac-
face, VigiLyze. tors is important in determining the utility of a
To understand the results of an analysis of particular database in answering a specific
individual case safety reports from a postmar- drug safety question.
keting database, it is necessary to understand
the unique features of the database, as each
Detecting Signals from a
large postmarketing AE database differs from
Postmarketing Adverse Event
the others. It is necessary to understand if, and
Database
how, the data are coded. Many databases code
drugs according to a local or national standard Identifying potential associations of AEs/
drug dictionary, while others use a standard ADRs to drugs using only information within
international dictionary, such as WHODrug. the database involves the detection of signals.
Similarly, many databases code individual AE/ According to the WHO, a signal is “reported
ADR reporter verbatim terms which describe information on a possible causal relationship
the AE/ADR according to a standard medical between an adverse event and a drug, the
dictionary, such as the Medical Dictionary for relationship being unknown or incompletely
Regulatory Activities (MedDRA). In the ICH documented previously.” While there have
regions, (Europe, Japan, and the United States) been many definitions of a signal put forth
use of MedDRA is mandatory for coding of over the years, the important underlying prin-
AEs/ADRs. ciple is that a signal is a hypothesis that calls
Beyond coding, several other features of the for further work to be performed to evaluate
database are important to understand. First, that hypothesis. Signal detection is the act of
does the database include only reports from looking for or identifying signals from any
postmarketing systems, or does it include source.
reports from other sources, such as the medical In the setting of a relatively small number of
literature or clinical trials? Second, does the reports, review of groups of reports or periodic
Descriptio 125
summaries of reports has been a standard Second, these methods rely exclusively on
method of signal detection. For example, one reports within the database; no external data
could look at a list of all reports in which the are needed. For this reason, understanding the
outcome was “death” to see if this outcome characteristics of the database, as discussed
was reported more frequently for some drugs above, is important. This feature has several
than others. Summaries based on specific consequences. Because the expected number
organ class toxicities could be reviewed to of reports of a specific AE/ADR for a given
examine if reports in one system organ class drug (and thus the disproportionality of the
were proportionately more frequent for one drug–event pair) depend on the reports within
drug than others. These methods depend on the individual database, the degree of dispro-
the ability of a drug safety specialist to recog- portionality for a given drug–event pair may
nize new or unusual patterns of case reports. vary from one database to the next. In the
While an astute specialist can identify signals extreme, a given drug–event pair may have a
using this method, this manual review is often strong signal of disproportionality in one data-
neither practical nor reproducible for detecting base, and no such signal in another. A second
signals from large postmarketing AE data- consequence is that as the background infor-
bases, some of which contain several million mation for all drugs in the database changes,
records. so does the expected number of reports of a
In an effort to address this challenge, data specific AE/ADR for a given drug (and again
mining techniques have been applied to phar- the disproportionality of the drug–event pair).
macovigilance AE/ADR databases. In broad Third, a signal of disproportionality is a
terms, data mining refers to a process of ana- measure of a statistical association within a
lyzing data to find patterns. In the case of AE/ collection of AE/ADR reports (rather than in a
ADR databases, most of these patterns would population), and it is not a measure of causal-
not be visible without the use of statistically ity. In this regard, it is important to underscore
based, computerized algorithms. There are a that the use of data mining is for signal detec-
variety of specific algorithms that have been tion – that is, for hypothesis generation – and
applied to safety signal detection in AE/ADR that further work is needed to evaluate the
databases. The fundamental feature of data signal.
mining techniques used to analyze AE data- Fourth, the absence of a signal of dispropor-
bases is that each is based on finding “dispro- tionality in a postmarketing AE database is not
portionalities” in data, which in this context is evidence that an important AE/ADR is not
the finding that a given AE/ADR is reported associated with a particular drug.
for a particular drug more often than would be Some of the data mining techniques used
expected based on the number of reports of in pharmacovigilance have included the pro-
that AE/ADR for all other drugs in the data- portional reporting ratio, the reporting odds
base. Several features of these methods are ratio, the Bayesian Confidence Propagation
worth noting. Neural Network (BCPNN), and the Empirical
First, the methods are transparent. While the Bayes method (also known as the Gamma
number of reports received for a drug varies Poisson Shrinker or the Multi-item Gamma
over time (and may be highest in the first few Poisson Shrinker). Data mining is sometimes
years of reporting), this temporal trend will not done using a subset of an AE/ADR database.
necessarily alter the proportion of specific For example, a portion of the database lim-
reactions for the drug. Thus, a given reaction ited to a specific class of drugs might be used
may still be found to be disproportionately to find relative differences in the frequencies
reported even as the total number of reports of specific AEs/ADRs across the class. As
for the drug changes. part of the IMI, a public-private partnership
in Europe, the EMA established the the question at hand), or to certain indications
Pharmacoepidemiological Research on for use (e.g. limit the case series to case reports
Outcomes of Therapeutics by a European in which the medicine was used for a certain
Consortium (IMI PROTECT) with a goal of off-label indication, if such exclusion is appro-
conducting research to develop and test new priate to the question at hand). Exclusion cri-
tools for the benefit–risk assessment of teria for a case series must be carefully
marketed drugs. A range of signal detection considered so that potentially relevant cases
algorithms were compared across seven are not excluded, and all available information
spontaneous reporting databases with no is fully assessed. In general, if the purpose of
method found to be better than the others. the case series is to examine the relationship
Findings were inconsistent across databases. between a medicine and a suspected AE/ADR
The choice of signaling criteria had a greater that has not been previously associated with
impact on signal detection performance than the medicine, it is best to err on the side of
the choice of disproportionality methods. inclusion to avoid missing clinically relevant,
though incomplete, information about cases
of interest.
Review of Individual Case Safety
Once the case series has been developed, it is
Reports
next necessary to review each case report indi-
The review of individual case safety reports of vidually in order to determine if there is a plau-
AEs/ADRs is a complex process. It typically sible causal relationship between the medicine
begins by identifying one or more case reports and the AE. At the level of the individual case
with the outcome of interest. Because the case safety report, it is often difficult to establish
reports that form a case series often come from with certainty that the medicine caused the AE
disparate sources, it is usually necessary to of interest. For example, if the AE/ADR of
develop a case definition. The case definition interest is one that is already common in the
centers on the clinical characteristics of the population that takes the medication, estab-
event of interest, without regard to the causal lishing a causal role for the medicine in the
role of the medicine whose relationship to the development of the condition is generally not
AE is being investigated. Once a case defini- feasible using individual case safety reports or
tion is established, each report is reviewed to case series. For example, the incidence of
determine if the event meets the case defini- Parkinson disease is much higher in persons
tion and if the report is to be included in the over the age of 60 years than it is in persons
case series. Depending on the specific below that age. In this situation, review of a
question(s) to be answered by the case series, report describing a myocardial infarction in a
other exclusion criteria may also apply. For 70-year-old patient on an anti-Parkinsonian
example, one would always exclude a case in agent will generally not be informative in
which the report suggests that the patient determining if the agent played a causal role in
never took the medicine of interest. In other the development of the myocardial infarction,
cases, one may restrict the case series to only as myocardial infarction occurs commonly in
certain formulations of the medicine (e.g. this age group. Similarly, review of a case
include case reports in which an intravenous report is not likely to shed light on the causal
formulation, but not an oral formulation, was relationship between a medicine and an AE/
used, if such exclusion is appropriate for the ADR when the AE/ADR is a manifestation of
question at hand), or to certain age groups the underlying illness which the medicine is
(e.g. limit the case series to only case reports treating. For example, a review of case reports
describing the suspected AEs in pediatric of worsening asthma in patients taking an
patients, if such exclusion is appropriate for anti-asthma medication is not likely to be
Descriptio 127
s ufficient to establish a causal link between the is it heterogeneous? Are certain co-morbidities
worsening asthma and the medication. Review or concomitant medications more likely to be
of a case series to establish a causal relation- present in patients with the event? In the
ship between a drug and a AE/ADR is most review of a case series, there are no prespeci-
straightforward when the suspected AE/ADR: fied answers to these questions that establish
(i) is rare in the population when the medica- or exclude the possibility that the drug led to
tion is not used, (ii) is not a manifestation of the AE/ADR. Rather, the characteristics of the
the underlying disease, (iii) has a strong tem- individual cases, taken together with the pat-
poral association with drug administration, terns observed in the case series itself, can lead
and (iv) is biologically plausible as a drug reac- the analyst to determine if the medication has
tion or is generally the result of a drug reaction a reasonable possibility of causing the condi-
based on other clinical experience. Examples tion of interest.
of AEs/ADRs that often meet these criteria are
acute hepatic failure, aplastic anemia, agranu-
Reporting Ratios
locytosis, rhabdomyolysis, serious skin reac-
tions such as Stevens-Johnson syndrome and Because postmarketing safety reporting sys-
toxic epidermal necrolysis, and certain tems do not capture all cases of an event of
arrhythmias, such as torsades de pointes. interest, it is not possible to calculate an inci-
The approach to assessing the causal role of dence rate for a particular drug–event pair.
a medicine in the development of an AE/ADR However, analysis of AEs/ADRs based simply
has evolved over the past four decades. In gen- on numbers of reports, even after thorough
eral, the approach relies on a systematic review analysis of these reports, does not in itself put
of each case report to ascertain the temporal these reports into the context of how widely a
relationship between drug intake and the medicine is used.
development of the adverse reaction, an assess- To adjust for the extent of drug utilization in
ment of any co-existing diseases or medica- a population in the analysis of AE/ADR
tions that could confound the relationship reports, a reporting ratio can be used. A report-
between the medicine and the AE/ADR, the ing ratio is defined as the number of cases of a
clinical course after withdrawing the drug particular AE/ADR reported to an AE data-
(“de-challenge”), and the clinical course after base during a specific time period divided by
re-introduction of the drug (re-challenge), some measure of drug utilization in the same
when applicable. Naranjo and colleagues time period. Across drugs in an AE database,
described a method based on these general the reporting ratios measure the relative fre-
principles for estimating the likelihood that a quency of the AE/ADR reports adjusting for
drug caused an adverse clinical event. The differences in the level of drug utilization in
WHO has developed a qualitative scale for cat- that database. The numerator is derived from
egorizing causality assessments. counts of AE/ADR reports associated with the
In the development of a case series, once the drug of interest that are recorded in the post-
individual cases are reviewed, it is important to marketing AE database during a specified
integrate the findings across the cases to deter- time period. In the past, the denominator typi-
mine patterns that may point to a relationship cally consisted of the number of dispensed
between the drug and the AE/ADR. For exam- prescriptions, used as a surrogate measure of
ple, does the AE/ADR appear at some doses, drug exposure in the population over that
but not at others? Does the AE/ADR appear same time period, and often estimated from
after one or a few doses, or does it appear only proprietary drug utilization databases. The
after a more prolonged exposure? Is the spec- number of dispensed prescriptions was used
trum of severity of the event homogeneous or because data on the number of unique
i ndividuals using the drug in a specified time Strengths

period was generally not available. More
recently, such data have become available, and Signal Detection
reporting ratios based on persons using the
medication, and not prescriptions, are being The principal strength – and, arguably, the
calculated. In some cases, information is avail- principal purpose – of a postmarketing safety
able on not only the number of persons receiv- reporting system is that it allows for signal
ing the drug or the number of prescriptions detection, the further exploration of drug
dispensed, but also on the duration of use. safety hypotheses, and appropriate regula-
When such data are available, the denomina- tory decision-making and action when neces-
tor for the reporting ratio may be expressed in sary. As noted earlier in this chapter, signals
person-time. When using denominators based can be detected by data mining methods,
on person-time, it is important to be mindful reviews of individual case safety reports, or
of the assumptions of the person-time method, assessment of case series. In many instances,
especially the assumption that events in the further work is needed to determine with
numerator occur uniformly over time. Because more certainty the relationship of the drug to
AEs/ADRs may not occur uniformly over time the AE/ADR. The capability for timely and
after a drug is started, this assumption does effective signal detection is a key strength of
not always hold. a postmarketing pharmacovigilance report-
Because the reporting ratio (sometimes ing system.
referred to as “reporting rate”) is not a measure Another key strength of a well designed
of incidence or prevalence, it must be inter- and effectively utilized postmarketing phar-
preted cautiously. For AEs/ADRs that are rare macovigilance reporting system is that, in cer-
in the general population (e.g., aplastic ane- tain cases, the relationship of a drug to an
mia), reporting ratios are sometimes compared AE/ADR can be established with sufficient
to the background rate (incidence or preva- confidence, usually by a case series, that nec-
lence) of that event in a defined population. In essary regulatory action can be taken. AEs/
other situations, individual reporting ratios of ADRs for which the relationship to a drug can
a particular AE/ADR across different drugs be established with reasonable certainty are
used for a similar indication or within the generally those that have a strong temporal
same class are calculated and the magnitude of association with drug administration, a low
the differences in reporting ratios is compared. or near absent frequency in the underlying
Interpretation of the comparison of reporting population, are not part of the underlying ill-
ratios across drugs must be made with caution, ness being treated, are generally the result of
since such comparisons are highly sensitive to exposure to a drug or other toxin, and have no
variation in AE/ADR reporting and thus it is other likely explanation. Aplastic anemia,
necessary to consider the differential underre- agranulocytosis, acute liver failure, rhabdo-
porting of AEs in the postmarketing safety myolysis, certain arrhythmias such as tor-
reporting system. The underlying assumption sades de pointes, and serious skin reactions
in estimating reporting ratios for comparison such as Stevens-Johnson syndrome are exam-
across a group of drug products is that each of ples of AEs whose relationship to a drug can
the respective manufacturer’s reporting prac- often be established by case series. However,
tices for the drug of interest are similar over relative to all signals detected in a postmar-
the reporting period. However, this assump- keting safety reporting system, those about
tion may not hold true in some cases, and a which a reasonably firm conclusion can be
comparison of reporting ratios across drugs made on the basis of AE/ADR reports alone
may not be valid. are few in number.
Limitation 129
Opportunity for the Public receive reports of AEs/ADRs occurring in any

to Report AEs/ADRs member of the population. Because it need
not restrict the reports it receives, it can
Postmarketing safety reporting systems allow
receive AE/ADR reports throughout a medi-
healthcare professionals to report suspected
cine’s marketed lifecycle. Thus, AEs/ADRs
AEs/ADRs to national pharmacovigilance
recognized late in a product’s lifecycle, such as
centers, drug regulatory authorities, and/or
those resulting from prolonged exposure to a
manufacturers. Such systems allow for direct
medicine, can, in theory, be ascertained. In
engagement of healthcare professionals in the
practice, such ascertainment is difficult to
drug safety monitoring system. The advantage
achieve, because healthcare professionals may
of this involvement is that it allows for careful
be less likely to ascribe an AE/ADR not known
clinical observations, made at the point of care,
to be associated with a medicine that has been
to inform drug safety surveillance. Clinicians
marketed for several years. In addition,
can provide succinct but detailed accounts of
patients who take a medicine for several years
relevant symptoms, signs, diagnostic test
may also receive other treatments during that
results, past medical history, concomitant
time, making it difficult to conclude that there
medications, and clinical course of an AE/
is an association between the medicine and
ADR, including information on de-challenge
the AE/ADR.
and re-challenge. Such a synthesis of clinical
Despite this broad scope, a postmarketing
information is generally not available from
spontaneous reporting system can be relatively
automated data sources. For those AEs/ADRs
inexpensive. Most of these pharmacovigilance
that are serious, rare, and often the result of a
systems rely on voluntary reporting, and those
medication exposure, the ability to obtain
who report AEs/ADRs are generally not paid
detailed information directly from the point of
for doing so. Thus, information collection is
care is an essential feature of postmarketing
not expensive from the perspective of an effec-
pharmacovigilance reporting systems.
tive pharmacovigilance, given that the system
Postmarketing safety reporting systems also
has the capacity to handle all medicines and all
can accept reports from consumers and
outcomes. This is in contrast to other data used
patients, though this practice is not a feature of
to study drug safety questions, such as data
all such reporting systems. In the United
from clinical trials, registries, and electronic
States, where consumers and patients can
healthcare data, each of which is relatively
report either to the manufacturer or directly to
expensive to operate.
the FDA, the percentage of reports in 2016 that
originated from consumers was about 50%.
When consumer and patient-generated reports
do not contain sufficient medical detail for Limitations
meaningful review, subsequent follow up with
health professionals may be possible in poten- Quality of Reports
tially important cases, so that more complete Perhaps the major potential limitation of a
clinical information (e.g. hospital discharge spontaneous postmarketing safety reporting
summary) can be obtained. system is that it depends quite heavily on the
quality of individual reports. Although data
mining and other informatics methods can
Scope
detect signals using coded bioinformatics
The scope of a postmarketing safety reporting terms in AE databases, each individual case
system is quite broad. The system can cover all safety report must still be carefully reviewed by
medicines used in the population, and it can a clinical analyst to determine if there is a
plausible relationship between the medicine after approval and decline thereafter, even
and the development of the AE/ADR. The though the drug may be used more widely.
quality of the report, as described earlier in this This phenomenon, known as the Weber effect,
chapter, is critical for an informative and was originally described in relation to non-
meaningful review of the individual case safety steroidal anti-inflammatory medicines. A
report. Report quality depends on the care, more recent analysis of reporting patterns for
effort, and judgment of the person submitting the angiotensin II receptor blocker class of
the report, as well as the diligence of the per- medicines revealed no discernible trend when
son receiving and/or transmitting the report to the number of reports over time was exam-
the health authority. Reports without suffi- ined. Specifically, this analysis did not confirm
cient information for an independent determi- that the number of reports increased toward
nation of the relationship between the the end of the second year and declined there-
medicine and the AE/ADR are problematic for after. Rather, the analysis indicated that addi-
drug safety surveillance. However, with suc- tional factors, such as the approval of additional
cessful follow up, sometimes even such defi- indications and modifications of the firms’
cient reports can yield useful information. reporting requirements affected the total num-
ber of reports received. However, when the
number of reports in a year was adjusted for
Underreporting
the number of prescriptions dispensed in that
Another well-recognized limitation of sponta- year’s period, it was found that the adjusted
neous postmarketing reporting systems is number of reports was highest in the first years
underreporting. Because most systems are vol- after approval and declined thereafter. The fre-
untary, not all AEs/ADRs are reported. A con- quency of AE/ADR reports per estimated unit
sequence of underreporting of AEs/ADRs is of drug utilization may not be constant over
that population-based rates of AEs/ADRs can- time.
not be calculated, because all such occurrences Second, publicity about an important new
in the population are not reported, and the AE/ADR often gives rise to a large number of
extent of underreporting for any individual reports shortly after the publicity, with a
AE/ADR is not known. Reporting ratios, dis- decline in the number of reports shortly there-
cussed earlier in this chapter, allow the after. This phenomenon is known as stimu-
reported number of AEs/ADRs to be put into lated reporting, and was observed, for example,
the context of drug utilization, though this in the reporting pattern of statin-induced hos-
measure is not an incidence rate. pitalized rhabdomyolysis after there was pub-
licity of this risk. For these reasons, changes in
the number of AE/ADR reports for a given
Non-uniform Temporal Trends
drug–event pair cannot reliably be interpreted
in Reporting
as a change in the population-based frequency
Another limitation of spontaneous reporting of the AE/ADR.
systems is that temporal trends in the number Another limitation of a postmarketing
of AE/ADR reports for a drug–event combina- reporting system is that it is usually not well
tion may not reflect actual population-based suited to ascertaining the relationship of a
trends for the drug–event combination. This is medicine to an AE/ADR that is common in the
because multiple factors can affect the number treated population, especially if the condition
of AE/ADR reports received for a given drug– is a manifestation of the underlying illness. In
event pair. such cases, the combined effect of confound-
First, the number of reports for a medicine ing of patient factors and indication make cau-
has been thought to peak in the second year sality assessment of individual cases difficult.
Particular Application 131
Finally, duplicate reports of the same AE/ triage algorithm (data mining) and are reported
ADR may be received by drug manufacturers quarterly in the Signal document, which is cir-
and health authorities, and if undetected as culated in restricted fashion to national phar-
duplicates, may be entered into the database as macovigilance centers for the purpose of
multiple occurrences of the same event. communicating the results of UMC evalua-
Algorithms have been developed, and various tions of potential data mining signals. The dis-
methods can be used to identify such reports; proportionality measure used by the UMC is
nonetheless, this issue is a potential source of the Information Component (IC), originally
bias and limits the utility of data mining or introduced through the BCPNN, which is a
other calculations which rely on “crude” case logarithmic measure of the disproportionality
counts which have not been “de-duplicated.” between the observed and expected reporting
of a drug-ADR pair. A positive IC value means
that a particular drug–event pair is reported
Particular Applications more often than expected, based on all the
reports in the database. The following is an
Case Series and Reporting Rates example of a signal identified by data mining
techniques applied to VigiBase regarding the
Spontaneous AE/ADR reports have at times
occurrence of glaucoma with topiramate.
served as a necessary and sufficient basis for
Topiramate was approved in the US in 1996 as
regulatory actions including product with-
an anticonvulsant drug. In the second quarter
drawals. For instance, in August 2001 the man-
of 2000, reports of topiramate and glaucoma in
ufacturer of cerivastatin withdrew that drug
VigiBase reached the threshold of an “associa-
from marketing based on “a markedly
tion” (i.e. the lower limit of a 95% Bayesian
increased reporting rate of fatal rhabdomyoly-
confidence interval for the IC exceeded zero).
sis” compared to the other drugs in the statin
When potential signals are identified, the
class. Additional confirmation of the unac-
available information is reviewed by the UMC
ceptably high risk of rhabdomyolysis with
staff and an expert review panel. At the time,
cerivastatin was eventually available three
there were six cases reported to VigiBase. After
years later when results of a well-designed epi-
review, a summary of the findings was circu-
demiologic study were published. Clearly, that
lated in the Signal document in April 2001 to
timeframe would have been far too long to
all national pharmacovigilance centers in the
delay decisive action, which in retrospect was
WHO Programme. Later the same year, the
soundly based on the signal from spontaneous
Market Authorization Holder issued a Dear
reports. The timely detection of this signal
Healthcare Professional letter warning about
would not have happened without the efforts
“an ocular syndrome that has occurred in
of the point-of-care clinicians who took the
patients receiving topiramate. This syndrome
time to report rhabdomyolysis when it
is characterized by acute myopia and second-
occurred in their patients.
ary angle closure glaucoma.” At the time, there
were 23 reported cases according to the com-
pany. FDA issued a warning in the revised
Data Mining Signals
labeling October 1, 2001.
According to the UMC glossary of pharma-
covigilance terms, a signal is “a hypothesis of a
Signals from Developing Countries
risk with a medicine, with various levels of evi-
dence and arguments to support it.” Signals are At the annual meetings of the WHO
identified by UMC analysts from the WHO Programme members, country representatives
database (VigiBase) by applying a predefined are invited to share problems of current
i nterest in their countries. Below is an example in current AE/ADR reports. The challenge of
illustrating the kind of issues that have been such a system will be to encourage reporters to
investigated in developing countries, presented routinely provide a clinically meaningful nar-
at the 2017 meeting in Uganda: rative that concisely explains the clinical
course of the AE/ADR and its relationship to
Blindness and Retinal Disorder medication usage. There is also interest in
Associated with Clomifene Citrate: using modern informatics techniques to facili-
Case Series Assessment tate a review of AE reports, especially in large
A case of retinal detachment with the use of AE databases. For example, the use of natural
clomifene citrate that caused irreversible language processing techniques is being
blindness triggered an assessment by the explored to determine if it can identify indi-
Eritrean Pharmacovigilance Centre. A search vidual case safety reports that warrant further
of VigiBase identified 24 cases of blindness evaluation, or if it can identify individual case
and retinal disorder. All cases were evaluated reports that suggest a causal association
using Austin Bradford Hill considerations to between a medicine and an AE. Postmarketing
assess the causal relation. In all cases, clo- safety reporting systems depend on the involve-
mifene was reported as the sole suspected drug ment of healthcare professionals and, in some
and in all but three cases no concomitant drugs areas, consumers and patients as well, for high
were reported. There were two cases of blind- quality AE/ADR reports. As new medicines
ness in which the reaction abated with seque- become available, it will be increasingly neces-
lae following withdrawal of clomifene. The sary to monitor postmarketing safety. Efficient
conclusion was that the findings support a safety reporting infrastructure is particularly
causal relationship and warrant further inves- important in situations where compassionate
tigation to substantiate the signal. use programs are utilized. Pharmacovigilance
reporting systems will continue to be the cor-
nerstone of this effort, because of their unique
The Future advantages. As social media, active surveil-
lance, and the use of large healthcare data-
Spontaneous AE/ADR reporting is an impor- bases begin to play a role in drug safety
tant component of drug safety surveillance. surveillance, demonstrate their utility, and
The widespread availability of electronic realize their potential, they could become valu-
healthcare data may, at first, seem to under- able adjuncts to existing pharmacovigilance
mine the importance of AE/ADR reporting. reporting systems worldwide.
This is not likely to be the case. Because careful
observation at the point of care is an essential
component of pharmacovigilance, electronic Key Points
systems may be able to facilitate AE/ADR
reporting in the future, but will not replace it. ●● AEs and ADRs are closely related, but none-
It is technologically and administratively feasi- theless distinct, concepts.
ble for carefully designed systems to allow cli- ●● Spontaneous reporting systems are based on
nicians to report AEs/ADRs directly from the notion that clinical observations made at
electronic medical record systems. If designed the point of care are often valuable pieces of
properly, these systems could allow for the information in further refining knowledge
accurate, complete, and efficient inclusion of of a drug’s safety.
laboratory, radiologic, and other diagnostic test ●● Spontaneous reporting systems can be used
results, information which is often incomplete to describe adverse drug events, adverse
Key Point 133
device events, medication errors, or a combi- ●● Interpretation of spontaneous reports always

nation of these. requires careful analysis, thought, and clear
●● The characteristics and quality of individual communication of results, conclusions, and
case safety reports determine the extent to limitations.
which the reports can address a drug safety ●● Because postmarketing safety reporting sys-
question. tems do not capture all cases of an event of
●● The organization of national pharmacovigi- interest, it is not possible to calculate an inci-
lance centers can vary from one country to dence rate for a particular drug–event pair.
the next. ●● Reporting ratios help put the number of
●● Data mining of spontaneous reporting data- reports of a drug-AE pair into the context of
bases, which allows relationships and pat- how widely a medicine is used.
terns to be seen in the data which would ●● The capability for timely and effective signal
otherwise be missed, can be used to detect detection is a key strength of a postmarket-
drug safety signals and generate hypotheses. ing pharmacovigilance reporting system.
Case Example 7.1

Spontaneous pharmacovigilance reporting based on a careful review of the temporal
systems: Felbamate relationship of the AE to use of the drug,
the patients’ past medical history, and con-
Background:
comitant medication usage.
Felbamate is an anticonvulsant agent
●● An estimated 100 000 patients had taken
approved for use in the United States on July
felbamate during this time. While the true
29, 1993. Pre-approval studies showed no
incidence of aplastic anemia in patients
evidence of significant, nonreversible hema-
taking felbamate cannot be calculated
tologic abnormalities.
because case ascertainment is likely
Question incomplete, the minimum rate is
Can spontaneous postmarketing reports 20/100000/year, or 200/million/year.
identify a signal for a rare event such as ●● In contrast, the population background
aplastic anemia and can safety surveillance rate of aplastic anemia is low, about two
result in a regulatory decision and labeling per million per year. Thus, the observed
change that supports the safe use of this cases of aplastic anemia suggest that
product? aplastic anemia is at least 100 times more
frequent in patients taking felbamate than
Approach
in the general population.
Within about one year of approval, cases of
aplastic anemia were reported to the manu-
Outcome
facturer and to the US FDA. This finding
●● Based on this finding, the FDA and the
prompted a search for and comprehensive
manufacturer recommended that patients
review of all case reports of aplastic anemia
not be treated with felbamate unless the
in persons taking felbamate.
benefits of the drug were judged to out-
Results weigh the risk of aplastic anemia.
●●Twenty cases of aplastic anemia, three of ●● A subsequent review of 31 case reports of
them fatal, had been reported in the aplastic anemia in patients taking fel-
United States. Review of the case reports bamate using the criteria of the
suggested a causal role for felbamate, International Agranulocytosis and Aplastic
(Continued)
Anemia Study (IAAAS), established that fel- reporting rate can be compared to a back-
bamate was the only plausible cause in ground incidence rate to demonstrate that
three cases, and the most likely cause in the incidence of the AE in patients exposed
11 cases. For the remaining nine cases, to the drug is higher than would be
there was at least one other plausible expected in the absence of the drug.
cause. The authors concluded that the
Limitations
“most probable” incidence of aplastic ane-
Formal incidence rates cannot be calculated
mia was estimated to be 127 per million.
from spontaneous reports.
Strengths
Key Points
●● Spontaneous reports of drug AEs can be
Because aplastic anemia is uncommon in the
used to identify ADRs that are rare, serious,
population and because it is generally the
and generally the result of a drug or toxin
result of a medication or other toxin, a care-
exposure.
ful analysis of a case series can establish the
●● While reporting rates cannot be used to
relationship of a drug to aplastic anemia.
calculate incidence rates, in certain cases a
Further Reading
Almenoff, J., Tonning, J., Gould, A. et al. (2005). ICH Harmonized Tripartite Guideline (2003).
Perspectives on the use of data mining in Post-approval safety data management:
pharmacovigilance. Drug Saf. 28 (11): definitions and standards for expedited
981–1007. reporting. https://www.fda.gov/media/71228/
Bohn, J., Kortepeter, C., Muñoz, M. et al. (2015). download.
Patterns in spontaneous adverse event Kaufman, D., Kelly, J., Anderson, T. et al. (1997).
reporting among branded and generic Evaluation of case reports of aplastic anemia
antiepileptic drugs. Clin. Pharmacol. Ther. 97 among patients treated with felbamate.
(5): 508–517. Epilepsia 38 (12): 1265–1269.
Edwards, I.R. (2017). Causality assessment in McAdams, M., Governale, L., Swartz, L. et al.
pharmacovigilance: still a challenge. Drug Saf. (2008). Identifying patterns of adverse event
40 (5): 365–372. reporting for four members of the Angiotensin
Edwards, I.R. and Lindquist, M. (2011). Social II receptor blockers class of drugs: revisiting
media and networks in pharmacovigilance: the weber effect. Pharmacoepidemiol. Drug
boon or bane? Drug Saf. 34 (4): 267–271. Saf. 17: 882–889.
EMA. EudraVigilance. http://www.ema.europa. McAdams, M., Staffa, J., and Dal Pan, G. (2008).
eu/ema/index.jsp?curl=pages/regulation/ Estimating the extent of reporting to FDA: a
general/general_content_000679.jsp case study of statin-associated
ICH Harmonized Tripartite Guideline (1994). rhabdomyolysis. Pharmacoepidemiol. Drug
Clinical Safety Data Management: Definitions Saf. 17 (3): 229–239.
and Standards for Expedited Reporting E2A. McMahon, A., Pratt, R., Hammad, T. et al.
Current Step 4 version, dated 27 October (2010). Description of hypersensitivity adverse
1994. http://www.ich.org/LOB/media/ events following administration of heparin
MEDIA436.pdf that was potentially contaminated with
Further Readin 135
oversulfated chrondroitin sulfate in early RegulatoryInformation/Guidances/

2008. Pharmacoepidemiol. Drug Saf. 19 (9): ucm129428.htm.
921–933. US Food and Drug Administration (2005b).
Meyboom, R.H.B., Hekster, Y.A., Egberts, A.C.G. Guidance for Industry: Good
et al. (1997). Causal or casual? The role of Pharmacovigilance Practices and
causality assessment in pharmacovigilance. Pharmacoepidemiologic Assessment. http://
Drug Saf. 17 (6): 374–389. www.fda.gov/downloads/
Nightingale, S. (1994). Recommendation to RegulatoryInformationGuidances/
immediately withdraw patients from UCM126834.pdf
treatment with felbamate. J. Am. Med. Assoc. US Food and Drug Administration (2009) 21
272 (13): 995. CFR 314.80. http://www.accessdata.fda.gov/
Russmann, S. (2006). Case reports of suspected scripts/cdrh/cfdocs/cfcfr/CFRSearch.
adverse drug reactions: case reports generate cfm?fr=314.80 (accessed 4 January 2009).
signals efficiently. BMJ 332 (7539): 488. US Food and Drug Administration (2009). The
Sloane, R., Osanlou, O., Lewis, D. et al. (2015). ICH Guideline on Clinical Safety Data
Social media and pharmacovigilance: a review Management: Data Elements for Transmission
of the opportunities and challenges. Br. J. Clin. of Individual Case Safety Reports.
Pharmacol. 80 (4): 910–920. US Food and Drug Administration (2020).
Staffa, J., Chang, J., and Green, L. (2002). MedWatch: The FDA Safety Information and
Cerivastatin and reports of fatal rhabdomyolysis. Adverse Events Reporting System. https://
N. Engl. J. Med. 346 (7): 539–540. www.fda.gov/safety/
van Stekelenborg, J., Ellenius, J., Maskell, S. medwatch-fda-safety-information-and-
et al. (2019). Recommendations for the use of adverse-event-reporting-program.
social media in pharmacovigilance: lessons US Food and Drug Administration (2020). Drug
from IMI WEB-RADR. Drug Saf. 42 (12): Safety Communications. https://www.fda.gov/
1393–1407. Drugs/DrugSafety/ucm199082.htm
US Food and Drug Administration (2005a). World Health Organization (2002). The
Guidance for Industry–E2B(M): Data Importance of Pharmacovigilance. Safety
Elements for Transmission of Individual Case Monitoring of Medicinal Products. Geneva:
Safety Reports. http://www.fda.gov/ WHO.
136
Overview of Electronic Databases in Pharmacoepidemiology

Brian L. Strom
Introduction databases,” as potential data sources for phar-

macoepidemiologic studies.
Once hypotheses about drug-induced adverse Large electronic databases can often meet
effects are generated, usually from spontane- the need for a cost-effective and efficient
ous reporting systems (see Chapter 7), tech- means of conducting postmarketing surveil-
niques are needed to test these hypotheses. lance studies. To meet the needs of pharma-
Usually between 500 and 3000 patients are coepidemiology, the ideal database would
exposed to the drug during Phase III testing, include records from inpatient and outpatient
even if drug efficacy can be demonstrated with care, emergency care, mental health care, all
much smaller numbers of patients. Studies of laboratory and radiological tests (including
this size would be expected to observe a single pharmacogenomic tests that may not have
case of outcomes with an incidence of one per been performed as part of clinical care), func-
1000 to six per 1000 (see Chapter 3). Given this tional assessments, and all prescribed and
context, postmarketing studies of drug effects over-the-counter medications, as well as alter-
must then generally include at least 10 000 native therapies. The population covered by
exposed persons in a cohort study, or enroll the database would be large enough to permit
diseased patients from a population of equiva- discovery of rare events for the drug(s) in ques-
lent size for a case–control study. Given a study tion, and the population would be stable over
of this size, the upper 95% confidence limit for its lifetime. Although it is generally preferable
the incidence of any event that is not identified for the population included in the database to
would be three per 10 000 (see Chapter 3). be representative of the general population
However, prospective studies this large are from which it is drawn, it may sometimes be
expensive, and difficult to perform. Yet, such advantageous to emphasize the more disad-
studies often need to be conducted quickly, to vantaged groups that may have been absent
address acute and serious regulatory, commer- from premarketing testing. The drug(s) under
cial, and/or public health crises. For all of investigation must of course be present in the
these reasons, the past decades have seen a formulary and must be prescribed in sufficient
growing use of electronic databases containing quantity to provide adequate power for
health care data, sometimes called “automated analyses.
Descriptio 137
Other requirements of an ideal database are for use by researchers in Europe, and similar
that all parts are easily linked by means of a databases have been developed in the US more
patient’s unique identifier, that the records are recently.
updated on a regular basis, and that the records
are verifiable and are reliable. The ability to
Claims and Other Administrative
conduct medical chart review to confirm out-
Databases
comes is also a necessity for most studies
(unless validated algorithms for the study out- Claims data arise from billable interactions
come already exist), as diagnoses entered into between patients and the health care system
an electronic database may include rule-out (see Figure 8.1). When a patient goes to a
diagnoses or interim diagnoses and recurrent/ pharmacy and gets a drug dispensed, the phar-
chronic, as opposed to acute, events. macy bills the insurance carrier for the cost of
Information on potential confounders, such as that drug, and has to identify which medica-
smoking and alcohol consumption, may only tion was dispensed, the milligrams per tablet,
be available through chart review or, more con- number of tablets, etc. Analogously, if a
sistently, through patient interviews. With patient goes to a hospital or to a physician for
appropriate permissions and confidentiality medical care, the providers of care bill the
safeguards in place, access to patients is some- insurance carrier for the cost of the medical
times possible and useful for assessing compli- care, and have to justify the bill with a diagno-
ance with the medication regimen as well as sis. If there is a common patient identification
for obtaining biosamples or information on number for both the pharmacy and the medi-
other factors that may relate to drug effects. cal care claims, these elements could be
Information on drugs taken intermittently for linked, and analyzed as a longitudinal medical
symptom relief, over-the-counter drugs, and record.
drugs not on the formulary must also be Since drug identity and the amount of drug
obtained directly from the patient. dispensed affect reimbursement, and because
These automated databases are the focus of the filing of an incorrect claim about drugs dis-
this section of the book. Of course, no single pensed is fraud, claims are often closely
database is ideal for all questions. In the cur- audited, e.g. by Medicaid. Indeed, there have
rent chapter, we will introduce these resources, also been numerous validity checks on the
presenting some of the general principles that drug data in claims files that showed that the
apply to them all. In Chapters 9 and 10, we will drug data are of extremely high quality, i.e.
present more detailed descriptions of those confirming that the patient was dispensed
databases that have been used in a substantial exactly what the claim showed was dispensed,
amount of published research, along with the according to the pharmacy record. In fact,
strengths and weaknesses of each. claims data of this type provide some of the
best data on drug exposure in pharmacoepide-
miology (see Chapter 13).
Description
So-called automated databases have existed Claims Databases: Sources of Data

and been used for pharmacoepidemiologic Provider: Pharmacy
research in North America since 1980, and are Data
Provider: Hospital Payor
primarily administrative in origin, generated User
by the request for payments, or claims, for clin- Provider: Physician
ical services and therapies. In contrast, elec-
tronic health record databases were developed Figure 8.1 Sources of claims data.
138 8 Overview of Electronic Databases in Pharmacoepidemiology
The quality of disease data in these data- the increasing use of computerization in med-
bases is somewhat less perfect. If a patient is ical care. Initially, computers were used in
admitted to a US hospital, the hospital charges Medicine primarily as a tool for literature
for the care and justifies that charge by assign- searches. Then, they were used for billing.
ing diagnosis codes (until recently International Now, however, there is increasing use of com-
Classification of Diseases-Ninth Revision- puters to record medical information at the
Clinical Modification [ICD-9-CM] codes) and point of care. In most instances, this is replac-
a Diagnosis Related Group (DRG). Hospital ing the paper medical record as the primary
diagnosis codes are reasonably accurate diag- medical record. As medical practices increas-
noses that are used for clinical purposes, based ingly become electronic, this opens up a
primarily on the discharge diagnoses assigned unique opportunity for pharmacoepidemiol-
by the patient’s attending physician (of course, ogy, as larger and larger numbers of patients
this does not guarantee that the physician’s are available in such systems. The best-known
diagnosis is correct). The amount paid by the and most widely used example of this
insurer to the hospital is based on the DRG, so approach is the UK Clinical Practice Research
there is no financial incentive to provide incor- Database (CPRD), along with the newer data-
rect diagnosis codes. In fact, most hospitals base, The Health Improvement Network
have mapped each set of diagnosis codes into (THIN), both described in Chapter 10. As gen-
the DRG code that generates the largest eral practice databases, these contain primar-
payment. ily outpatient data. In addition, recently there
In contrast, however, outpatient diagnoses are inpatient electronic health record data-
are assigned by the practitioners themselves, bases available.
or by their office staff. Once again, reimburse- Electronic health record databases have
ment in the US does not usually depend on unique advantages. Importantly among them
the actual diagnosis, but rather on the visit is that the validity of the diagnosis data in
intensity during the outpatient medical these databases is probably better than that in
encounter, and the resulting procedure codes claims databases, as these data are being used
indicate the intensity of the services provided. to document medical care rather than just for
Thus, there is no incentive for the practitioner billing purposes. When performing a pharma-
to provide incorrect diagnosis codes, but there coepidemiologic study using these databases,
is also no incentive for them to be particularly there is no purpose in validating the data
careful or complete about the diagnoses against the actual medical record, since one is
provided. For these reasons, the outpatient analyzing the data from the actual medical
diagnoses are the weakest link in claims record. However, there are also unique issues
databases. one needs to be concerned about, especially
Some other databases are not made up of the uncertain completeness of the data from
actual claims, but derive from other adminis- other physicians and sites of care. Any given
trative processes, e.g. data from US Health practitioner provides only a piece of the care a
Maintenance Organizations or other data patient receives, and inpatient and outpatient
sources. The characteristics of these data are care are unlikely to be recorded in a common
similar in many ways to those of claims data, medical record.
and they are discussed together as encounter-
based databases in Chapter 9.
Strengths
Electronic Health Record Databases
Computerized databases have several impor-
In contrast, electronic health record databases tant advantages. These include their potential
are a more recent development, arising out of for providing a very large sample size. This is
Weaknesse 139
especially important in the field of pharma- laboratory results data to these resources can
coepidemiology, where achieving an adequate assist in diagnosis validity, as well.
sample size is uniquely problematic. In addi- In addition, such databases can lack infor-
tion, these databases are relatively inexpensive mation on some potential confounding varia-
to use, especially given the available sample bles. For example, in claims databases there
size, as they are by-products of existing admin- are no data on date of menopause, and
istrative systems. Studies using these data sys- diagnosis-based algorithms to identify smok-
tems do not need to incur the considerable cost ing and alcohol abuse may have poor sensitiv-
of data collection, other than for those subsets ity, all of which can be of great importance to
of the populations for whom medical records selected research questions. This argues that
are abstracted and/or interviews are con- one either needs access to patients or access to
ducted. The data can be complete, i.e. for physician records if these contain the data in
claims databases, information is available on question, or one needs to be selective about the
all medical care provided for covered services, research questions that one seeks to answer
regardless of who the provider was. As indi- through these databases, avoiding questions
cated above, this can be a problem though for that require data on variables which may be
electronic health record databases, especially important potential confounders that must be
in the US, where primary care providers often controlled for.
do not serve as gatekeepers to specialty care. In A major other disadvantage of administra-
addition, these databases can be population- tive data is the instability of the population due
based, they can include outpatient drugs and to job changes, employers’ changes of health
diseases, and there is no opportunity for recall plans, and changes in coverage for specific
and interviewer bias, as they do not rely on employees and their family members. The
patient recall or interviewers to obtain their opportunity for longitudinal analyses is
data. Another advantage is that these data- thereby hindered by the continual enrollment
bases can potentially be linked to external and dis-enrollment of plan members. Another
other electronic databases (e.g. death records, source of instability of the population is when
maternal-child records, police accident patients transfer out of the system due to death
records), to expand the capabilities and scope or relocation. The effect of this is an inflated
of research. This requires using common iden- list with patients no longer seeking medical
tification elements (e.g. name and date of care. This will invalidate calculations of
birth) and standardized semantics to allow patient-time in studies of disease incidence, for
communication across databases. example, because the denominator is inflated.
The challenge for the investigator is to be crea-
tive in devising strategies to guard or correct
Weaknesses for this incomplete information in the data-
base (e.g. by performing sensitivity analysis
The major weakness of such data systems is censoring follow-up one or two years after the
the uncertain validity of diagnosis data. This is patient’s last recorded entry in the database).
especially true for claims databases, and for Alternatively, strategies can be adopted for
outpatient data. For these databases, access to selecting stable populations within a particular
medical record data for validation purposes is database, and for example, by examining pat-
usually needed. This issue is less problematic terns of prescription refills for chronically used
for electronic health record databases; how- medications and by restricting the study popu-
ever, the validity of medication data from elec- lation to include only continuously enrolled
tronic health record databases in the United patients. Of course, the largest such data sys-
States is less certain than pharmacy dispensing tem, i.e. US Medicare, suffers much less from
data from claims databases. The addition of this problem, since it covers the elderly, so
140 8 Overview of Electronic Databases in Pharmacoepidemiology
eople never lose eligibility. Even there, how-

p sample size, (ii) when a denominator is needed
ever, patients can switch between fee-for- to calculate incidence rates, (iii) when one is
service plans and managed care plans, and the studying short-term drug effects (especially
latter may not record all health care that is pro- when the effects require specific drug or surgi-
vided (see Chapter 9). cal therapy that can be used as validation of
Further, by definition, such databases only the diagnosis), (iv) when one is studying objec-
include illnesses severe enough to come to tive, laboratory-driven diagnoses, (v) when
medical attention. In general, this is not a recall or interviewer bias could influence the
problem, since illnesses that are not serious association, (vi) when time is limited, and (vii)
enough to come to medical attention and yet when the budget is limited.
are uncommon enough for one to seek to study Uniquely problematic situations include: (i)
them in such databases, are generally of lower illnesses that do not reliably come to medical
importance. attention, (ii) inpatient drug exposures that are
Some results from studies that utilize these not included in some of these databases, (iii)
databases may not be generalizable, e.g. on outcomes that are poorly captured by the cod-
health care utilization. This is especially rele- ing system, such as Stevens-Johnson Syndrome,
vant for databases created by data from a popu- (iv) descriptive studies, if the population stud-
lation that is atypical in some way, e.g. US ied is non-representative, (v) delayed drug
Medicaid data. effects, wherein patients can lose eligibility in
Finally, as noted briefly above, as an increas- the interim, and (vi) important confounders
ing number of electronic health record data- about which information cannot be obtained
bases emerge in the US, to date all are without accessing the patients, such as ciga-
problematic in that they do not include com- rette smoking, occupation, menarche, meno-
plete data on a defined population. In the US pause, etc.
health system, unlike other countries, patients
can, and often do, seek medical care from a
variety of different health care providers at The Future
unaffiliated institutions with a non-linked
electronic health record system. Thus, provid- Given the frequent use of these data resources
ers’ electronic health records are inherently for pharmacoepidemiologic research in the
incomplete, and need to be linked to adminis- recent past, we have already learned much
trative data in order to be useful for quality about their appropriate role. Inasmuch as it
research. This is different from the situation in, appears that these uses will be increasing, we
for example, the UK, where electronic health are likely to continue to gain more insight in
record databases are much more likely to be the coming years, especially with the access in
complete given the general practitioner gate- the US to Medicare data, and the advent in the
keeper paradigm and unique patient identifier US of FDA’s Sentinel system, exceeding
for all healthcare services. 170 million individuals. However, care must be
taken to ensure that all potential confounding
factors of interest are available in the system or
Particular Applications addressed in some other way, that diagnoses
under study are chosen carefully, and that
Based on these characteristics, one can identify medical records can be obtained when needed
particular situations when these databases are to validate the diagnoses. In this section of
uniquely useful or uniquely problematic for the book, Chapters 9 and 10, we will review
pharmacoepidemiologic research. These data- the details of a number of these databases. The
bases are useful in situations: (i) when looking databases selected for detailed review have
for uncommon outcomes because of a large been chosen because they have been the most
Further Readin 141
widely used for published research. They are ●● Claims data arise from a person’s use of the
also good examples of the different types of health care system, and the submission of
data that are available. There are multiple oth- claims to insurance companies for payment.
ers like each of them and undoubtedly many While claims data provide some of the best
more will emerge over the ensuing years. Each data on drug exposure in pharmacoepidemi-
has its advantages and disadvantages, but each ology, the quality of disease data in these
has proven it can be useful in pharmacoepide- databases can be more problematic.
miologic studies. ●● Medical record databases are a more recent
development, arising out of the increasing
use of computerization in medical care. The
Key Points validity of the diagnosis data in these data-
bases is better than that in claims databases,
●● The past three decades have seen a growing as these data are being used for medical
use of computerized databases containing care. However, the completeness of the data
medical care data, so-called “automated from other physicians and sites of care is
databases,” as potential data sources for uncertain.
pharmacoepidemiology studies.
Further Reading
Ray, W.A. and Griffin, M.R. (1989). Use of Strom, B.L. and Carson, J.L. (1990). Use of
Medicaid data for pharmacoepidemiology. automated databases for
Am. J. Epidemiol. 129: 837–849. pharmacoepidemiology research. Epidemiol.
Strom, B.L. and Carson, J.L. (1989). Automated Rev. 12: 87–107.
data bases used for pharmacoepidemiology
research. Clin. Pharmacol. Ther. 46: 390–394.
142
Encounter Databases
Tobias Gerhard1, Yola Moride2, Anton Pottegård3, and Nicole Pratt4
1
Ernest Mario School of Pharmacy, Rutgers Biomedical and Health Sciences, Piscataway, NJ, USA
2
Center for Pharmacoepidemiology and Treatment Science, Rutgers Biomedical and Health Sciences, New Brunswick, NJ, USA
3
Clinical Pharmacology and Pharmacy, Department of Public Health, University of Southern Denmark, Odense, Denmark
4
Quality Use of Medicines and Pharmacy Research Centre, Clinical and Health Sciences, University of South Australia, Adelaide, South
Australia, Australia
Introduction to provide an encyclopedic description of avail-

able databases, this chapter focuses on the
Encounter databases contain electronic description of key commonalities and distinc-
records of healthcare encounters for large, tions across encounter databases, illustrated
defined populations. They capture information with selected examples and supplemented by
on patient characteristics, prescription fills, references to more comprehensive resources in
and medical services, as part of the routine the literature. The chapter also includes a dedi-
administration or reimbursement of health- cated discussion of the considerations faced by
care. This is in contrast to electronic health researchers when evaluating the appropriate-
record (EHR) databases, described in detail in ness of a specific encounter database or decid-
Chapter 10, which are primarily intended and ing among multiple encounter databases for
maintained to support patient care. Encounter their research question.
data may contain records at various levels of
granularity ranging from records of individual
services (fee-for-service claims) to aggregate Description
records of care episodes (hospital discharge
records). Encounter databases exist in many Encounter data arise as part of the routine
countries and within a number of vastly differ- administration of a person’s interactions with
ent healthcare systems. An increasing number various sectors of the healthcare system.
are available for research and consequently, When combined, these data can be used to
encounter databases have become a corner- infer a longitudinal picture of a person’s medi-
stone of pharmacoepidemiologic research. cal and treatment history. The quality of that
Although they vary markedly in their specific picture, i.e. its usefulness for pharmacoepide-
characteristics, encounter databases share a miologic and other research, depends on the
number of defining features that warrant their completeness and validity of the information
discussion as a group. Rather than attempting available.
Descriptio 143
The essential attribute of all encounter data- to prevent fraud and thus assure high accuracy
bases useful for pharmacoepidemiologic of the data in instances where the information
research is a defined population for which is directly relevant to the processing of the cor-
healthcare services are recorded regardless of rect payment amount (e.g. quantity and dose
the provider or location of care received (e.g. of medications dispensed by a community
outpatient, inpatient, emergency department). pharmacy or type of procedure performed dur-
Such databases are considered population- ing an outpatient physician visit). In contrast,
based (see Chapter 12). Precise definition of data elements that are not directly tied to the
the database population avoids various forms payment, for example, the specific diagnoses
of selection bias common in non-population- associated with an outpatient visit or proce-
based studies (e.g. biased control selection in dure may be recorded with lower accuracy. In
hospital-based case–control studies). Complete purely administrative databases, data charac-
capture of all relevant healthcare services teristics depend on the specific data collection
avoids bias from incomplete and potentially and quality assurance processes in place for
differential measurement of healthcare ser- each of the data elements.
vices (e.g. incomplete ascertainment of hospi- While an ideal encounter database would
talizations occurring in a non-participating capture all types of healthcare services, in
healthcare system). Although representative- practice, individual databases often lack cover-
ness of a geographic region or the general pop- age of certain service types depending on the
ulation is often desirable, it is not necessary, as purpose of the database and the nature of the
long as the database population is accurately data collection process. The completeness of
defined. While encounter databases ideally information captured in a database is a func-
capture all healthcare services, in practice spe- tion of the types of healthcare services (data
cific service types may not be captured due to domains) included, as well as of the compre-
the nature of the data collection process (most hensiveness of data capture within each
often due to lack of reimbursement). However, domain. Encounter databases useful for phar-
accurate qualitative description of the specific macoepidemiologic research typically contain
service types with lack of coverage or incom- the following core data domains: (i) healthcare
plete capture is critically important to allow plan eligibility and basic demographic infor-
evaluation of the appropriateness of the data- mation, (ii) outpatient pharmacy dispensing,
base for a given research question. and (iii) medical services (typically including
Encounter databases are maintained by a hospitalizations; commonly also including
number of different entities including govern- outpatient health services).
ment agencies, insurance companies, health Data domains may be maintained in sepa-
plans, and information services companies. rate files within a single integrated database
The primary purpose of encounter databases is (e.g. US private and governmental databases),
often the reimbursement of fee-for-service or in multiple autonomous databases that
payment claims, and such encounter data are together function as a federated virtual data-
often referred to as claims data. In some base (e.g. Nordic healthcare databases),
instances, however, for example in US health depending on whether the data are collected
plans with staff model delivery systems or cap- and maintained by a single or by multiple enti-
itated payment models, the purpose is purely ties. Both integrated and federated databases
administrative with no processing of payments require reliable linkage of an individual’s
for individual services. This distinction can be records over time and between data domains.
important as the accuracy and validity of data Table 9.1 summarizes commonly available
correspond to the purpose of the record. For data elements within the core data domains.
example, claims records are routinely audited The content of the core data domains often
144 9 Encounter Databases
Table 9.1 Core data domains in encounter databasesa.
Membership Patient identifier, sex, age/date of birth, race/ethnicity (not universally available),
zip code, dates of enrollment and disenrollment, benefits package/eligibility
category (if applicable).
Medical
Outpatient Services Patient identifier, encounter date, service location (physician office, hospital
outpatient, etc.), procedure codes (e.g. CPT, HCPCS), primary and secondary
diagnosis codes (e.g. ICD-10-CM), provider identifier, provider profession/specialty.
Inpatient Services Patient identifier, primary diagnosis, secondary diagnoses, admission and
discharge dates, length of stay, patient destination, hospital identifier. Inpatient
data generally do not include information on in-hospital medication use and
typically represent summaries for an entire hospital stay, resulting in some lack of
detail.
Pharmacy Patient identifier, unique drug identifier (e.g. US-NDC, Nordic article number)
which identifies generic name, brand name, dosage form, and strength (crosswalks
may be needed for some databases while others include the individual data
elements coded by the unique identifier), date dispensed, quantity dispensed,
prescription duration/days supply.
Typically, not recorded: Indication for the prescription, inpatient drug use,
over-the-counter drugs.
a
adapted from Chapter 13, Strom et al. (2012); CPT, Current Procedural Terminology; HCPCS, Healthcare
Common Procedure Coding System; ICD, International Classification of Diseases; NDC, National Drug Code.
varies across individual databases in terms of complex encounter data-based algorithms, is

which types of healthcare services are cap- critical for rigorous pharmacoepidemiologic
tured. While some databases are limited to research with encounter databases (see
hospital discharge data, many also capture Chapter 13). Validation necessitates the ability
data on outpatient office-based physician vis- to reliably link an individual’s encounter data
its, outpatient clinic visits, long-term care facil- to non-encounter data sources that serve as the
ities, dental, and vision services. Another external gold standard, such as electronic or
example for incomplete data capture within a paper medical records, disease registries, or
data domain, is incomplete or lack of record- survey data. Furthermore, linkage with com-
ing of over-the-counter (OTC) medication fills plementary non-encounter data resources or
in prescription databases. Similar variability ad hoc data collection (see Chapter 11) is also
across databases exists in terms of access to commonly implemented in order to supple-
non-encounter data, such as EHRs, laboratory ment an encounter database with variables
test results, diagnostic examinations, provider that are required to answer a specific research
specialty/characteristics, vital statistics, or dis- question but are not recorded in the database,
ease registries. Lastly, profound differences such as lifestyle factors or disease severity.
also exist in data structure and coding Because of their size, population-based
systems. nature, comprehensive capture of the full spec-
Because the primary purpose of encounter trum of healthcare encounters, and ability to
data is administrative, any inferences about a rapidly assemble cohorts and identify outcomes
patient’s medical history made from these data among them, encounter databases represent a
have to be carefully evaluated. Validation of tremendous resource for pharmacoepidemio-
encounter data, ranging from the validation logic studies. For some research questions,
of individual data elements to the validation of encounter data may be sufficient on their own,
Descriptio 145
particularly when the outcome of interest has encounter databases. A large study population
been previously validated and data on all is generally necessary to ensure adequate sta-
important confounders are available within the tistical power when exposures or outcomes are
database. In many instances however, valida- rare (particularly when both are rare), effect
tion of outcomes and supplementation with sizes are small, and when subgroup effects or
external data are necessary. In these cases, the treatment effect heterogeneity are of interest.
encounter databases provide the study founda- In addition, some common study designs and
tion (population base and comprehensive cap- analytic methods may further increase the size
ture of healthcare interactions) with certain of the database necessary to achieve adequate
data elements critical to the study question statistical power. For example, the new-user
fleshed out through linkage with additional active comparator design results in study pop-
data resources. ulations that often represent only a small frac-
tion of the total number of users of a drug of
interest during the study period; restriction, a
Attributes of Encounter Databases
common approach to reduce confounding, can
Although encounter databases share a basic substantially decrease the size of study cohorts;
set of defining characteristics, they differ in and instrumental variable methods are statisti-
numerous attributes that deserve considera- cally inefficient compared to standard regres-
tion when evaluating the fit of a database to sion approaches (see Chapter 22).
address a specific research question. In addition to the size of the database, the
Importantly, in some databases, such as the US characteristics of the database population have
commercial insurance databases, these attrib- to be carefully considered. As a general rule,
utes can be heterogeneous across individual the population covered by an encounter data-
people, as availability of supplemental data base is a function of the underlying healthcare
(e.g. laboratory test results or ability to retrieve system in the respective country during the
medical records) or even core data domains study period. Knowledge of these systems is a
(e.g. pharmacy dispensing data) may be prerequisite for informed consideration and
restricted to subsets of the full database popu- use of databases for pharmacoepidemiologic
lation. In these instances, the suitability of the research. Databases in countries or regions
database (e.g. in terms of sample size and rep- with universal single-payer coverage, such as
resentativeness) should be evaluated based on Taiwan, South Korea, and the Northern
the subset of the population in a given data- European countries, generally include the
base for which the attributes required to entire population and do not impose eligibility
address the question under study (i.e. key restrictions. All individuals are included and
study variables) are available rather than the membership is maintained throughout a per-
database population as a whole. son’s life regardless of qualifying factors such
as age, employment, or financial situation.
Population and Coverage Period Therefore, the characteristics of the popula-
The population captured is a critically impor- tion included in these databases are stable over
tant consideration when examining the suita- time and closely track the characteristics of the
bility of an encounter database for the study of population of the respective country or region
a specific research question of interest. The as a whole. In contrast, database populations
size of the database is typically one of the key in countries or regions with less complete or
criteria when considering an encounter data- more fragmented coverage, including the US,
base for a specific research question; both in are heterogeneous and far more complicated.
comparison to electronic medical record data- The fragmentation of the US healthcare sys-
bases as well as in comparison to alternative tem, in particular, leads to a complex landscape
for encounter databases with different data- Drugs administered during hospital stays or in
bases covering distinctly different subsets of long-term care units, in the emergency room,
the US population (discussed in more detail or in outpatient physician office settings are
below). Furthermore, individuals may be typically not included. The latter, however, can
included in different databases at different in some instances be captured through drug-
points in time based on their personal situa- specific outpatient procedure codes (e.g. drug-
tion (e.g. employment and state of residence), specific procedure codes for injection
resulting in short average enrollment periods administration). OTC drug use is generally not
(dwell times) in any specific database environ- recorded, unless OTC drugs are prescribed and
ment. Dwell time is an important consideration specifically covered by the insurance or health
particularly when the research question of system. In databases for which data capture
interest involves studying a long-term effect of depends on a reimbursement mechanism,
a medicine. Similarly, when dwell time is drug dispensing may also be missing in cases
short, it becomes increasing difficult to study where drugs are paid for entirely out of pocket
new users of medicines as a lag time at the (i.e. because the cash price is lower than the
start of an individual’s data capture is required required copayment), or for non-reimbursable
to differentiate incident from prevalent medi- drugs (benzodiazepines, for example, were
cation exposure. excluded from reimbursement by Medicare
Lastly, the time period covered by a database Part D prior to 2013).
often determines its usefulness for a given Lastly, drug formularies, stepped therapy
study question, depending on the start of data requirements, and prior authorization pro-
collection and recency of the latest available grams may impose restrictions on availability
data. Studies examining trends in drug utiliza- and copayments and thus have a significant
tion over time or studies on the long-term impact on use rates of individual medications
effects of drugs, such as studies with cancer as and medication classes. Individual formularies
an outcome, are best served by databases with may apply to an entire database population or
long coverage periods and a stable population. vary widely across individuals depending on
Studies of newly approved medications pri- the underlying healthcare system.
marily require the most current data available. Encounter databases also vary substantially
The US Medicaid Analytic Extracts (MAX) in terms of which medical services are included
data, for example, are generally not appropri- and importantly what information is captured
ate for studies of recently approved drugs, due about these services. Most widely used encoun-
to a two to three year lag in data availability. ter databases capture hospital services includ-
Importantly, when studying long-term utiliza- ing emergency departments. Hospital services
tion trends or long-term drug effects, it is are generally recorded as hospital discharge
important to be cognizant of any changes over data that summarize information for an entire
time in health service reimbursement and hospital or emergency department stay rather
administration and appreciate their impact on than provide documentation of individual ser-
drug utilization. vices. Differences, however, exist in the granu-
larity of these data, such as number of
Services Covered and Data diagnosis fields and availability of procedure
Completeness codes.
For obvious reasons, medication data are a pre- Even greater variation between databases
requisite for all encounter databases used for exists in the capture of outpatient services. For
pharmacoepidemiologic research. Generally, example, in contrast to databases in the US,
these data are limited to information on medi- Canada, Taiwan, and South Korea, the Nordic
cations dispensed by community pharmacies. countries do not maintain a database of
Descriptio 147
utpatient office-based physician visits, though

o data, biobanks, or survey data. Linkage of
visits to outpatient hospital/specialty clinics encounter data to complementary data sources
are captured. Therefore, Nordic database stud- serves two distinct purposes: (i) validation of
ies of outcomes that do not result in hospitali- encounter-based information against an exter-
zations or that require outpatient office-based nal gold-standard, and (ii) provision of supple-
diagnoses for adjustment of confounding have mentary data not available in the encounter
to rely on medication use as a proxy for outpa- database, such as disease-specific data and
tient office-based diagnoses. Capture of other symptoms. Linkage to an external gold stand-
service types, such as dental, vision, or long- ard, ideally the medical record, for a sample of
term care also depend on the database and the cases is particularly critical in order to facili-
patient’s specific insurance coverage. Lastly, tate outcome validation and calculation of
particularly in the US, specific benefits such as positive predictive values (PPVs) of encounter
mental health or other specialty services may data-based algorithms. In absence of the medi-
be excluded (“carved out”) in certain benefits cal record, validation may be performed
packages and thus not captured for individuals against disease registries or patient self-report/
covered under these plans. For many data- survey. The validity of pharmacoepidemiologic
bases, it is thus important to evaluate the avail- drug and diagnosis data as well as approaches
ability of data on specific service types not only to the conduct of validation studies are dis-
at the level of the database but at the individ- cussed in detail in Chapter 13. The ability to
ual level and over time using information on retrieve medical records for outcome valida-
each person’s benefit package. tion varies between databases and is often a
Finally, databases differ in the information critical factor in database selection.
available about the patient and service pro- Linkage to non-encounter data may also be
vider. For example, data on the patient’s race necessary to provide supplemental informa-
and ethnicity are generally not available in US tion on variables that are unmeasured or
administrative claims databases but available poorly measured in the encounter data but
in US governmental databases. Similarly, death necessary to adjust for confounding or appro-
is reliably recorded in some databases while priate restriction of the study population (e.g.
others require linkage to vital statistics. indication for drug prescribing, lifestyle fac-
Databases also differ in the availability of pro- tors, measures of disease severity).
vider specialty and identity for physician medi- Supplemental information such as laboratory
cal services as well as prescriber specialty and test results or autopsy records may also be
identity for dispensing data. required for outcome ascertainment (e.g.
HbA1c level as an outcome for a study on the
Linkage to Non-encounter Data comparative effectiveness of various hypogly-
Many pharmacoepidemiologic research ques- cemic agents).
tions cannot be answered with encounter data Due to privacy restrictions that prevent the
alone. Some questions will require randomized sharing or use of personal identifiers, retrieval
trials (see Chapter 17) or prospective primary of medical records or information obtained
data collection (see Chapter 11). However, through direct contact with physicians or
linkages to complementary sources of data patients is generally not performed by investiga-
may help to overcome inherent limitations of tors, but rather facilitated through third parties
encounter data. Commonly used sources for (e.g. retrieval of redacted medical records for US
non-encounter data include electronic or Medicaid and Medicare) or handled internally
paper medical records, laboratory test results, by employees of the participating health plans
cause of death registries or autopsy records, (e.g. in several US commercial insurance data-
disease or immunization registries, census bases). Depending on the database, encounter
and non-encounter data may be available under error in the linkage variables. However,
the same umbrella organization (e.g. linkage to probabilistic linkage is more accurate in error
EHRs in many US health plans) or require link- prone data. Although often challenging, vali-
age to outside entities (e.g. retrieval of hospital dation of linkage quality is critically impor-
medical records for US Medicaid beneficiaries tant as all linkage methods are susceptible to
for the purpose of outcome validation), which error. The Nordic prescription database net-
greatly affects the feasibility, efficiency, cost, works are examples of highly reliable linkage
and success rates of the linkage. between encounter data and disease registries
Healthcare data linkages are governed by with unique identifiers while the Dutch
both privacy restrictions and the availability PHARMO system uses probabilistic record
of common linkage variables in the respective linkage methods.
databases. Privacy regulations governing the
ability to link personal health information are Access
complex and vary between countries, data- Access regulations, costs, and feasibility con-
base owners, and over time. When these regu- siderations vary widely between encounter
lations do not preclude linkage, health databases and often have a major impact on
information databases can be linked using database choice. Access may, for example, be
either deterministic or probabilistic methods. restricted to certain researchers, such as those
Briefly, in deterministic linkage, a unique working in academia or governmental agen-
identifier or a combination of several non- cies. Some encounter databases facilitate direct
unique variables available in both databases access to either “off the shelf” or customized
must match exactly (though the match can be anonymized datasets which may be physically
implemented based on transformed versions transferred to the researcher’s institution or
of the variables, e.g. phonetic codes instead of accessed remotely (e.g. select US commercial
names to minimize the impact of spelling databases, US governmental databases, or the
errors). Deterministic linkage is most useful South Korean HIRA data), while others require
if reliable unique identifiers are available (e.g. in-house data analyses and thus necessitate
US social security number) but is also achiev- collaborative agreements with researchers
able with combinations of multiple non- employed by the database custodian or affili-
unique variables (e.g. birth dates, admission ated research institutes (e.g. US health plan
dates, and names). However, use of variables databases or Nordic prescription databases).
with low discriminative power and errors or Some databases are directly accessible in
missingness in the matching variable(s) will anonymized form but require in-house analy-
lead to a high number of overlooked (false sis performed by the database custodian when
negative) matches. Probabilistic linkage additional “custom” linkages that require per-
methods can reduce the number of over- sonal identifiers have to be implemented (e.g.
looked (false negative) matches by allowing Truven MarketScan). For studies conducted
imperfect matches due to partially inaccurate through the database custodian, it is important
or missing data but in turn may produce false to not only consider the attributes of the data-
positive matches. Choice of matching method base itself, but also the data analytic capacity
thus involves a trade-off between false nega- and track record of the in-house research col-
tive matches (i.e. missed matches) and false laborators. While complexity of database struc-
positive matches (i.e. incorrectly matched ture varies between databases and studies, all
records). Simulation studies have suggested work with large encounter databases requires
that deterministic linkage is an equally valid sophisticated programming skills as well as a
but less computationally intensive method for comprehensive understanding of database-
databases with low rates of missingness and specific details. The latter consideration can be
Descriptio 149
a major advantage of collaborative arrange- Encounter Databases in the United

ments that include researchers or program- States
mers from the database custodian. US encounter databases are arguably both the
Costs of data access vary across databases largest databases available and the most frag-
and often within databases depending on the mented worldwide. Unlike most industrial-
specific characteristics of the study in ques- ized nations, the US does not have a uniform
tion. Fees often vary by size (number of indi- health system or universal healthcare cover-
viduals) and complexity (number of files/data age resulting in databases with characteristics
sources/number of years of data) of the that differ markedly from databases in the rest
requested dataset as well as by funding source of the world. In 2018, 324 million people, or
(e.g. federal versus commercial funding). In- 92% of the US population had health insur-
house data analysis often imposes substantial ance coverage, with 27 million uninsured.
additional costs. 217 million people had coverage from private
Application processes vary widely as well. plans (67%), mostly employment-based plans
While all databases require compliance with (178 million, 55%). 111 million people (34%)
data privacy and security restrictions, some may had coverage from a governmental plan;
also impose scientific vetting of the research 58 million by Medicaid (18%), 58 million by
plan or a justification of the benefit of the Medicare (18%), and 3 million had coverage
research to the public. Particularly in projects from the Department of Veterans Affairs
that require custom linkage with identifiable Healthcare System (1%). Note that these cen-
patient or provider information, close collabora- sus data-based estimates show some inconsist-
tion with the database custodian is needed to encies with the reporting from the Centers for
obtain necessary approvals and maintain confi- Medicare and Medicaid Services (CMS) pre-
dentiality. In addition, the time required for the sented later in the chapter. Broadly speaking,
creation of study-specific data-cuts depends on most employed individuals and their depend-
the staffing resources and experience at the data- ents are covered by commercial insurance,
base custodian and the complexity of the adults 65 years and older and qualifying indi-
required dataset. As a result, the duration from viduals with disabilities are covered by
the beginning of the application process until Medicare, and the poor and other disadvan-
the start of the research can vary dramatically taged groups are covered by Medicaid.
between several weeks to multiple years. Furthermore, insurance coverage in the US is
In practice, while access considerations and not mutually exclusive. In 2018, 15% of the
familiarity with a given database are often population with health insurance had multi-
important drivers of database choice, it is vital ple coverage types either due to switches in
never to lose sight of the suitability of the data- coverage type or due to simultaneous coverage
base to the specific research question under to supplement their primary insurance type.
study (fit-for-purpose). With the exception of Medicaid programs,
which generally provide prescription drug cov-
erage for all beneficiaries, prescription drug
Selected Encounter Databases
insurance is typically provided separately from
A selection of widely used encounter data- medical insurance resulting in subgroups of
bases and database types with their basic char- patients in major databases for whom only
acteristics is presented in Table 9.2 and pharmacy or medical data are available.
discussed below. Databases will be discussed Although pharmacy claims are recorded with
by region and include US databases, Canadian high accuracy, medication dispensing can be
databases, European databases, and Asian incompletely captured in patients covered by
databases. multiple insurance programs or in instances
Table 9.2 Select database characteristics.
Government, Health System Commercial Insurance, Government, Government, Northern Government,

Type US Government, US Databases, US US Canada Europe Asia
Examples Medicare Medicaid Kaiser, HealthCore, Saskatchewan, Denmark, Norway, South Korea,
Analytic Extract Geisinger MarketScan, Optum, Quebec Sweden, Netherlands Taiwan
(MAX) Pharmetrics
Networks Sentinel Sentinel HCSRN, Sentinel, CNODES CNODES PROTECT AsPEN
Sentinel,
PCORnet, VSD
Population Province Country Country
Relative Size +++ +++ ++ +++ ++ ++ +++
Dwell time +++ + to ++ + to ++ + +++ ++++ ++++
Lag in availability 3–4 years 1–2 years <1/2 year <1/2 year Variable Up to 2 years Variable
Access Direct Direct In-house In-house < 1–2 years In-house Variable
Retrieval of Medical Yes Yes Yes Partial Noa Yes Yes for some
Records for Validation databases
Coding, Drug NDC NDC NDC NDC AHFS ATC ATC
Coding, Dx ICD-9-CM, ICD-9-CM, ICD-9-CM, ICD-9-CM, ICD-10-CM ICD-9-CM, ICD-8, −9 and − 10 ICD-10,
ICD-10-CM ICD-10-CM ICD-10-CM ICD-10-CM ICD-9
Validation +++ +++ ++++ + to +++ ++ ++ ++
Supplementation +++ ++ ++++ + to +++ ++ +++ +++
a
Apart from a few rare exceptions, one cannot retrieve medical charts of cases ascertained in a given study. However, can identify patients in medical records in institutions
and link back to the database.
Descriptio 151
where the copayment is greater than the cash has important implications for pharmacoepi-
price of the medication. In recent years, sev- demiologic research conducted in US data-
eral large US retailers have begun to offer low- bases. Despite the existence of crosswalks, the
cost generic medications for as little as $4 for a performance characteristics of encounter-data
monthly supply, considerably less than the based algorithms have to be demonstrated for
average tier 1 copayment ($11 in 2019). Since the new coding system and studies that span
there is no financial incentive, pharmacies the transition date will have to implement
may not submit insurance claims when multiple coding systems in a single study. Data
patients pay cash resulting in potential under- privacy and security of identifiable healthcare
ascertainment of low cost generic medications. data in the US is governed by the Health
To date, empirical studies examining the miss- Insurance Portability and Accountability Act
ingness of dispensing in claims databases have of 1996 (HIPAA).
reported a limited impact of such generic drug
discount programs. Payments rates and modal- US Private Insurance Databases
ities for medical services vary widely ranging Most healthcare in the US is covered through
from fee-for service to capitated arrangements private insurance, predominantly employer-
in which providers receive a fixed payment per based insurance. For-profit and not-for profit
patient per unit of time for the delivery of a insurance companies offer a wide range of
specified set of services. Detailed claims data plans that vary in characteristics such as pre-
are often not available for services or patients mium, copayment/coinsurance, deductibles,
covered by such capitated payment models as out of pocket limits, services covered, drug for-
the payment amount is independent from the mularies, as well as provider choice. Payment
specific services provided. systems and business models are complex and
Several large US encounter databases are undergo continuing change over time. Because
available and have been widely used for phar- most private insurance plans are associated
macoepidemiologic research. These databases with the employer, many patients frequently
include markedly different groups of the popu- change insurance plans due to changes in
lation and often individuals with heterogene- employment or when employers change their
ous healthcare coverage are included within contracted insurance portfolio. Although there
the same database. To complicate matters fur- are hundreds of health insurance companies
ther, significant mobility exists between data- in the US, a relatively small number of compa-
bases as changing life circumstances (loss of nies provide coverage for a majority of the pri-
employment, change in employer, disability, vately insured population. The great majority
reaching age 65/Medicare eligibility) result in of the privately insured population are covered
changes in insurance coverage. This is often by insurance systems that pay for the care pro-
referred to as “churning,” and substantially vided by others. Commercial insurance data-
affects the average dwell time of individuals in bases derived from these systems are some of
US encounter databases. the largest databases available for pharma-
US databases generally use the National coepidemiologic research. A smaller group is
Drug Code (NDC) for medication data, the covered by integrated, often not-for-profit
Current Procedural Terminology (CPT) coding healthcare delivery systems that assume
and Healthcare Common Procedure Coding responsibility for preventive and therapeutic
System (HCPCS) for procedures, and the health services to a defined population, often
International Classification of Diseases, employ group or staff model delivery systems,
Clinical Modification (ICD-CM) system for and frequently operate their own hospitals
diagnoses. The US transitioned from ICD- (e.g. Kaiser Permanente). Though typically
9-CM to ICD-10-CM on October 1, 2015, which smaller in size, the databases associated with
these healthcare systems offer extensive data procedures would have to be limited to the
resources that combine encounter-data with subset of beneficiaries with a dental benefit
detailed clinical data resources including during a specific time period. Completeness
EHRs and direct access to patients and and quality of the claims data also depends on
providers. the payment model employed by the respective
Commercial insurance databases are longitu- insurance products. As discussed earlier, com-
dinal collections of billable healthcare interac- pleteness and accuracy with which services are
tions. These databases are maintained by a captured may differ substantially depending
variety of entities. This includes large insur- on whether services are reimbursed through
ance companies, often through health data fee-for service payments or through capitated
analytics-focused subsidiaries (e.g. Optum arrangements. Such capitated arrangements
Clinformatics/UnitedHealth Group; may apply to all medical coverage or be limited
HealthCore Integrated Research Database/ to specific services (e.g. specialist visits or men-
Anthem, Comprehensive Health Insights tal healthcare services). The ability to validate
Outcomes Data/Humana), as well as health or supplement the claims data is also often lim-
information technology (IT) companies (e.g. ited to subgroups of members included in the
Truven Health MarketScan, IQVIA database. For example, for databases main-
PharmetricsPlus). Commercial insurance tained by subsidies of insurance companies,
databases typically include several millions to data validation and supplementation may not
tens of millions of individuals cross-sectionally be permitted for the (sometimes substantial)
and cumulatively often exceed 100 million proportion of individuals in “self-funded”
unique patients over the life span of the data- plans, where the employer assumes direct risk
base. Importantly however, the extremely large for payment and the insurance company only
sizes of these databases do not necessarily provides administrative services (ASO mem-
translate directly into the size of pharmacoepi- bers). Similarly, the ability to identify patients
demiologic study cohorts. Given the approxi- and validate or supplement patient data
mately 30% annual churn rate in commercial depends on the contractual arrangements with
insurance coverage and the fact that prescrip- the data sources (employers, health plans) and
tion drug coverage is often separately adminis- is generally restricted to a limited subset of the
tered or absent, only approximately 50%, 30%, full database populations. Given the substan-
and 15%, of beneficiaries with medical cover- tial heterogeneity in multiple data attributes
age have continuous medical and pharmacy within and between commercial databases,
coverage for 1, 2, and 4+ years, respectively. thoughtful consideration of detailed informa-
Another important and often underappreci- tion on members’ individual benefit packages
ated feature of commercial insurance data- is critical to facilitate restriction of the study
bases is the large within-database heterogeneity population to those for whom all necessary
in data availability, data completeness, data data elements and linkages are captured or
quality, and ability to link member data to non- available in the database.
encounter data. Within a typical commercial Several models exist to enable research
database, members are covered by a variety of access to commercial insurance databases.
insurance products (often from multiple insur- Some databases are directly available in their
ance companies) leading to substantial differ- entirety through licensing arrangements (e.g.
ences in services captured in the database. MarketScan), while others are solely accessible
Drug formularies, which determine coverage on a project by project basis via collaborative
and out-of-pocket costs for prescription drugs, arrangements involving in-house program-
for example, vary widely between plans. mers. Databases available for licensing are
Similarly, a study that requires data on dental de-identified, with all personal identifiers
Descriptio 153
removed and as such do not support external in the Health Care System Research Network
linkages. Studies that require such linkages for (HCSRN, formerly known as the HMO
validation or supplementation of the encoun- Research Network) and data partners for the
ter data typically require collaboration with Sentinel System.
researchers employed by the database custo-
dian. Such collaborations have the added US Government
advantage of tapping into the often substantial The US government funds healthcare services
experience of the custodian research team. through several major programs, including
Most major commercial insurance providers Medicare, and Medicaid as well as the
also participate as data partners for the Sentinel Department of Veterans Affairs Healthcare
System. System (VA). In contrast to the VA, which is a
Integrated healthcare delivery system data- large provider of healthcare services operating
bases differ from commercial insurance data- numerous hospitals, clinics, and nursing
bases in that they include a defined population homes, Medicaid and Medicare function as
whose entire spectrum of care is under the payers. Both programs pay directly for services
responsibility of and provided by the inte- using fee-for-service arrangements, but a large
grated delivery system. Similar to commercial and growing proportion of beneficiaries
insurance databases, the delivery system data- receives Medicaid (69% in 2017) or Medicare
bases include pharmacy dispensing data as (34% in 2018) coverage administered by private
well as encounter data on diagnoses and proce- insurance companies through capitated man-
dures from care delivered in both ambulatory aged care plans. For beneficiaries covered by
and inpatient settings. However, because all managed-care plans, encounter data for indi-
care is provided by the delivery system, these vidual services have historically been unavail-
databases also have access to full inpatient and able (Medicare) or were of mixed completeness
outpatient electronic and paper medical and quality (Medicaid). Research with
records, and have the ability to interact with Medicaid or Medicare data has thus histori-
providers and patients. Although the latter fea- cally been restricted to individuals with fee-
tures are also available for subsets of patients for-service coverage. Recent efforts have aimed
in many commercial insurance databases, the to improve availability and quality of data for
uniqueness of integrated delivery systems Medicare and Medicaid beneficiaries covered
databases lies in the fact that these linkages by managed-care plans. CMS administers
cover the entire care received by the patient Medicare and Medicaid data and facilitate
and is not limited to care received by specific access to research identifiable files for research
practices or hospitals. Since many EHR sys- purposes. Requests for these data files require
tems include information on drugs prescribed, a research protocol and data use agreement,
delivery system databases have often access to and are reviewed by CMS’s Privacy Board. The
both prescription and dispensing data, which application process is managed and supported
can be useful for a variety of research ques- by the Research Data Assistance Center
tions, such as questions of primary non- (ResDAC) at the University of Minnesota,
adherence. In addition, several integrated which provides technical assistance to
healthcare delivery systems include affiliated researchers interested in CMS Medicare and
research centers that maintain a variety of Medicaid data. Data access requires payment
additional data resources such as registries for of fees based on the requested population size
cancer, diabetes, or cardiovascular disease. as well as the number of data files requested,
Integrated health delivery systems have a long which can be provided through release of data
track record of pharmacoepidemiologic files to investigators or remotely via the CMS
research, and many are consortium members Virtual Research Data Center (VRDC). A
mechanism to obtain inpatient hospital and use. MAX and TAF data are both organized in
emergency department medical records corre- 5 file types: (i) Person Summary/Demographic
sponding to Medicare and Medicaid claims has and Eligibility (demographic characteristics
been described and implemented. Medicaid and enrollment information), (ii) Inpatient
and Medicare data for select states or popula- (inpatient hospital claims with one record per
tions are also available from commercial enti- stay; procedure and diagnosis codes), (iii)
ties (e.g. IBM Watson Health). Long-Term Care (e.g. nursing facility claims),
Medicaid is a joint state/federal program (iv) Prescription Drug/Pharmacy (outpatient
intended to provide health coverage for low- pharmacy data including NDC, quantity dis-
income individuals. It is administered sepa- pensed, days supply), and (v) other services
rately by each state and state-specific eligibility (e.g. claims for physician, outpatient hospital,
rules differ within federal regulations. and laboratory services). Data are based on
Traditionally, the program has provided cover- state-level data submitted through the
age limited to certain groups of low-income Transformed Medicaid Statistical Information
individuals, including pregnant women, low- System (T-MSIS, since 2011 for some states and
income families with children, the chronically since October 2015 for all states) and the
disabled, and the elderly. Following the pas- Medicaid Statistical Information System
sage of the Affordable Care Act in 2010, about (MSIS; 1999 through 2015). TAF and MAX
one half of US states have expanded coverage both are produced by CMS. MAX data under-
to all individuals under certain income thresh- went extensive editing and quality control at
olds. In 2018, the average monthly enrollment the federal level resulting in a substantial lag of
in Medicaid was 74.6 million (6.0 million aged, approximately three to four years between the
10.7 million blind/disabled, 28.5 million chil- end of a calendar year and data availability. For
dren, 15.8 million traditionally eligible adults, TAF, more of the responsibility for data quality
and 12.2 million adults eligible through control has been shifted to the individual states
Medicaid expansion). In 38 states 50% of ben- resulting in a reduced lag in data availability
eficiaries were covered through private (Two years or less). Once released, MAX and
managed-care plans. Medicaid coverage for TAF files are final. Compared to MAX, TAF
eligible individuals is generally comprehensive files add hundreds of additional data elements
although each state, within federally man- (third-party liability, provider, and managed
dated parameters, administers their Medicaid care plan data) as well as modifications of
program differently, resulting in variations in existing data elements.
Medicaid coverage across the country. One of the intentions of the transition from
Medicaid data files include enrollment and MSIS to T-MSIS was to improve capture and
claims data for all Medicaid enrollees in the 50 quality of encounter data for beneficiaries cov-
states and the District of Columbia as well as ered by managed care plans. Data for these
for the approximately 7.0 million (2016) enroll- beneficiaries have historically been considered
ees in the Children’s Health Insurance Program not to be up to research standards and typically
(CHIP) which serves uninsured children up to have been excluded from most pharmacoepi-
age 19 in families with incomes too high to demiologic research. Given that the majority
qualify them for Medicaid. Data files have been of Medicaid enrollees are now covered under
produced since 1999, available per state per managed-care plans, availability of research
year. From 1999 through 2013 exclusively as quality data for this population (after extensive
MAX files, from 2016 onward exclusively as quality checks and validation studies) would
T-MSIS Analytic Files (TAF), with a transition substantially increase the potential of Medicaid
period in 2014 and 2015, during which, data as a resource for pharmacoepidemiologic
depending on the state, both file types were in research. Gaps in data capture due to periods
Descriptio 155
of ineligibility are common as eligibility is typi- majority of these so called Medicare Advantage
cally determined monthly and changes with plans also include Part D benefits (i.e. prescrip-
income and life circumstances. This issue tion drug coverage). Part C plans are optional
affects individual eligibility groups differently and require premiums. In 2018, 33.6% of
with more stable enrollment for those qualify- Medicare beneficiaries received coverage
ing based on disability and less stable enroll- through Medicare Advantage plans.
ment for low-income adults. Exclusion of Importantly, encounter data for Medicare
beneficiaries without stable enrollment has Advantage beneficiaries have only recently
been implemented based on eligibility files as become available through CMS (to date solely
well as through requirements for Medicaid for service years 2015 and 2016). Part D pro-
encounters during specified periods before and vides outpatient prescription drug coverage.
after person time under study. Because Established in 2006, the program is adminis-
Medicaid is administered at the state-level, tered by private companies that provide cover-
state-specific policies (e.g. opioid quantity lim- age through hundreds (901 in 2019) of
its or prior approval requirements) have to be prescription drug plans (PDPs) that differ in
considered in the research design. Medicaid formulary coverage and cost sharing.
and Medicare data for dually eligible benefi- Enrollment in Part D is voluntary and requires
ciaries can be linked. Such linkage is impor- a monthly premium that varies between the
tant in studies of dual enrollees since Medicaid individual PDPs. Medicare Part D imposes a
or Medicare data alone fail to document the coverage gap (doughnut hole) that requires
full spectrum of care provided to such dual beneficiaries to pay a substantial percentage of
enrollees. Medicaid data for research are also the cost of their medications (35% and 44% for
available directly from individual states but brand name and generic drugs, respectively in
access is often limited to researchers with 2018) until they reach the out-of-pocket spend-
established ties to the specific state Medicaid ing limit ($5000 in 2018). A large proportion of
programs. Medicare beneficiaries have some type of sup-
Medicare is the federal program that pro- plemental coverage (employer sponsored,
vides healthcare coverage for almost all people Medicaid, so-called Medigap policies) to
65 years and over as well as for qualifying indi- reduce out-of-pocket costs from cost-sharing
viduals with permanent disabilities. Medicare requirements. In 2018, the average monthly
coverage consists of four parts: Medicare Part Medicare enrollment was 60.0 million
A (Hospital Insurance), Medicare Part B (51.3 million aged, 8.7 million disabled).
(Medical Insurance), Medicare Part C 20.2 million beneficiaries were covered
(Medicare Advantage), and Medicare Part D through Medicare Advantage and 44.2 million
(Medicare Prescription Drug Coverage). All had a Part D benefit, including 16 million
Parts of Medicare coverage require beneficiar- through Medicare Advantage plans.
ies to pay deductibles and some stipulate cost Medicare data are available in several file
sharing. Part A covers inpatient care in hospi- types that are linkable to each other, as well as
tals and skilled nursing facilities, as well as to Medicaid data for dually-enrolled benefi-
hospice. It is premium-free for the great major- ciaries. File types include: MBSFs (Master
ity of beneficiaries. Part B covers physician and Beneficiary Summary Files), which include
other outpatient services. It is an optional profiles on demographics and enrollment, chronic
gram that requires monthly premiums. conditions, and cost and utilization;
Approximately 90% of Medicare beneficiaries Institutional Claims, which include files on
enroll in Part B. Part C allows Medicare benefi- inpatient services, skilled nursing facilities,
ciaries to enroll in private health plans that and hospice; Non-institutional Claims, which
administer Part A and B benefits. The large include outpatient physician claims (Carrier
file) and claims for durable medical equip- Within each province, three encounter data-
ment; and the Part D event Data file, which bases are available: (i) beneficiary, (ii) medical
provides detailed prescription level outpatient services, and (iii) prescription drugs acquired
pharmacy claims. Supplementary files provide through the public drug plan (with the excep-
information on Part D plan characteristics, tion of British Columbia which collects drug
pharmacies, drugs (crosswalks from First dispensings acquired through both public and
DataBank), prescribers, and formularies. private drug plans, as well as out-of-pocket).
Since prescription drug data for Medicare These databases are linkable through a unique
have become available after the establish- patient identifier that remains unchanged
ment of Medicare Part D in 2006, Medicare, over time. Additional linkage capacities are
due to its large and stable population, has available to hospitalization databases, popula-
become one of the largest and most compre- tion health surveys or to province-specific dis-
hensive resources for pharmacoepidemio- ease registries. Linkage of hospital charts or
logic research. outpatient charts for validation of diagnoses
or collection of data that are not present in the
Encounter Databases in Canada databases requires approval from the provin-
Canada, with its population of approximately cial information access commissioner and
37.5 million, has a universal healthcare pro- may not be feasible in all provinces. A number
gram covering all residents regardless of age of validation studies of Canadian databases,
or income. Program administration is under primarily of diagnoses codes in the medical
the responsibility of each of its 10 provinces or services databases, can be found in the litera-
territories. Physician visits, diagnostic tests, ture, but validation data remain far from
procedures (in- or out-patient), and hospitali- comprehensive.
zations are provided without payment by the Each province maintains its own medical
patient at the point of care. Encounter data are services encounter database, which includes
transactional and consist of billings submitted all claims submitted by physicians regardless
by healthcare providers on a fee-for-service of setting (inpatient, outpatient, or emergency
basis. Similar to US encounter databases, department) as long as the physician is paid on
some in-hospital services are not billed indi- a fee-for-service basis. The nature of the infor-
vidually and a small number of physicians mation in the various provincial medical ser-
may have all or a portion of their activities vices databases is similar though differences
covered by salary; hence, the services they exist in coding systems, such as the ICD ver-
provide may not be included in the medical sion. For each medical service, the following
services databases. In contrast, public drug information is recorded: service (date, descrip-
coverage programs differ among provinces; tion, location, diagnosis, and cost), provider
programs have been available for varying (identifier and specialty). The vast majority of
lengths of time and differ with respect to eligi- claims are submitted electronically, and the
bility criteria as well as characteristics (i.e. resulting medical services claims databases are
copayments and deductibles). Some prov- populated in real-time. In a few provinces,
inces, such as Saskatchewan and Manitoba, such as Nova Scotia, Manitoba, and British
provide coverage for the entire population Columbia, mental health services, including
while in the others, public drug programs psychotherapy, are recorded in a distinct
restrict coverage to specific segments of the database.
population, such as the elderly, welfare recipi- Unlike the medical services databases, hos-
ents, youth (less than 25 years of age) or those pitalization databases are intended for the cre-
who do not have access to private insurance ation of health statistics rather than for
plans through their employers. reimbursement purposes. The databases
Descriptio 157
c ontain clinical data related to hospital dis- across provinces, inclusion of individual drugs
charges from acute or chronic care units, or in the formulary and type of listing (general or
rehabilitation centers, as well as day surgeries. restricted) may vary. For each patient, the year
With the exception of Quebec, which main- of entry and exit from the drug program are
tains its own hospital discharge database available in the beneficiary database. This is
(MED-ECHO), all provinces contribute to the important information for studies that include
Discharge Abstract Database (DAD) main- segments of the population whose member-
tained by the Canadian Institute for Health ship in the drug program may be transitory,
Information (CIHI). The information is there- such as membership based on income or
fore homogeneous across provinces. In the access to private insurance programs. Only 7
hospitalization databases, diagnosis was coded of the 10 Canadian provinces make prescrip-
with ICD-9-CM until 31 March 2006 and with tion data available for pharmacoepidemio-
ICD-10 thereafter. In the DAD database, infor- logic research. Approximately half of these
mation on mental health resources, cancer databases are accessible through custodians
staging, and reproductive history were added located in a university setting while the other
in 2009–2010. Hospitalization databases are half are accessible through provincial govern-
typically available six months after the end of ment agencies. In addition to the drug data-
the fiscal year (March 31st). bases, custodians also act as a repository for
Province-specific prescription drug data- other provincial databases and are responsible
bases record all prescription drugs dispensed for their linkage.
in an outpatient setting to individuals covered Database access varies across provinces.
by the public drug plan. Drugs acquired out- Some provinces (Saskatchewan, Quebec, and
of-pocket/OTC, in-hospital, long-term care Nova Scotia), provide raw anonymized data-
units, not included in the formulary, or cov- sets to researchers (from academic or industry
ered only by private insurance programs are settings) while others (Ontario, BC) require
not usually included in the database. One data to be analyzed in-house by specific
exception is PharmaNet in British Columbia research organizations. To maintain confiden-
that links all pharmacies to a central data sys- tiality of the data, no patient, healthcare pro-
tem. Every prescription dispensed in the out- vider (including pharmacist), or institution
patient setting is recorded regardless of identifiers are transmitted to researchers.
coverage; hence, it includes medications cov- Additional restrictions are in place in individ-
ered by the public drug plan, private insurance ual provinces. For example, in Quebec only a
programs, as well as those acquired out-of- random sample of approximately 75% of the
pocket. Drugs are coded according to the population eligible for a given study (capped at
Canadian-specific Drug Information Number a maximum of 125 000 eligible patients) may
(DIN) as well as the American Hospital be obtained, and no birthdates are transmitted.
Formulary Service (AHFS). For each dispens- Exceptions can be granted through a request to
ing, the following information is recorded: the Provincial Access to Information
drug (date of dispensing, drug name, dose per Commission, which substantially increases
unit, mode of administration, prescribed dura- the delay in data extraction. Although
tion [the latter is not recorded in Canadian encounter databases are much
Saskatchewan], cost including dispensing smaller than the US encounter databases, their
fees), pharmacist (identifier, pharmacy loca- greatest advantage is that they include a stable
tion), and prescriber (identifier, specialty). population, thereby allowing longer follow-up
Indication for a drug prescription is not periods. This is, for example, illustrated
recorded in any of the dispensing databases. through a study on benzodiazepines and
While data and coding systems are similar Alzheimer’s disease by Billioti de Gage and
colleagues, in which a 10 year follow-up was include but are not limited to hospital dis-
available. The time required for database charge databases, laboratory data including
extraction varies across provinces, ranging results, pathology databases, medical birth
from 10 to 20 weeks to 1 year, more if a request databases, cancer registries, and cause of death
to the Provincial Access to Information databases, as well as census data, health sur-
Commission is required. veys, biobanks, and patient records. Together,
Access to linked data (prescription, medical these databases create a federated database
services, hospitalizations) is possible on a pro- network that provides exposure information
vincial basis. At a national level, the CIHI from the prescription database as well as
houses encounter databases from multiple pre- patient and clinical outcome data from the
scription claims-level data, from public drug patient router file and multiple linked autono-
programs of British Columbia, Alberta, mous databases.
Saskatchewan, Manitoba, Ontario, New The prescription databases largely include
Brunswick, Nova Scotia, Prince Edward Island, similar data elements with slight variations
Newfoundland and Labrador, and Yukon. between countries. Besides a patient identifier
These data have been linked at CIHI to the (which also encodes birth year and sex), data
hospital discharge database (DAD), although includes drug data (dispensing date, Nordic
those multi-province linked data are not acces- article number, a unique identifier similar to
sible to external researchers. In addition, the the NDC code used in the US, Anatomical
IQVIA Private Drug Plan Database houses Therapeutic Chemical [ATC] classification,
adjudicated prescription claims from approxi- quantity dispensed in defined daily doses), a
mately 83% of the total private (direct-pay) prescriber identifier (which can be linked to
business in Canada. These data are useful for prescriber data such as basic demographics,
the conduct of drug utilization studies but are profession, specialty, practice site), and phar-
currently not linked to medical services or hos- macy data (name and location). OTC drugs are
pitalization data. not included unless they are obtained via pre-
scription. Importantly, some drugs that are
Encounter Databases in Europe also available OTC are used primarily via pre-
The Nordic Prescription Databases scription, to ensure reimbursement. Besides
The Nordic countries (Denmark, Iceland, the difference in the age of the databases, the
Norway, Sweden, and Finland) have tax- most noticeable difference is the fact that non-
supported universal health coverage. All citi- reimbursed drugs are not covered by the
zens (a combined population of over 25 million Finnish database.
people ranging from ~300 000 in Iceland to Outcome data are primarily based on
more than 9 million in Sweden) are provided national hospital discharge databases (regis-
with unrestricted access to health services tries). While comparable, some differences
including partial or complete reimbursement exist in the age of the patient databases with
of medications. Pharmacies electronically sub- the Finnish database dating back to 1969, fol-
mit information on dispensed prescriptions to lowed by the Danish (1977), Swedish (1987),
national databases without a requirement for and Norwegian registry (2008). Numerous
informed consent by the patient (available other databases including cancer, birth, and
since 1994 in Finland and Denmark, since death, together with pathology and laboratory
2004 in Norway, since 2005 in Sweden, and results further complement the dataset.
since 2006 in Iceland). Unique civil registra- Importantly, no large-scale data are available
tion codes facilitate unambiguous linkage to that provide details regarding general practice
various national databases using a central visits or other non-hospital health services.
patient router file. Linkable national databases This is often referred to as a lack of “outpatient”
Descriptio 159
data. However, this term can lead to misunder- on outpatient visits, dispensed medication,
standings in the context of the Nordic health procedures, chronic conditions, as well as hos-
care model. All hospital databases cover activi- pital admission diagnoses and procedures, and
ties within hospital outpatient clinics, and as date of death. Data access, however, is
such all specialized care is covered. However, complex.
in all Nordic countries, general practice physi-
cians serve as gate keepers to specialized care Encounter Databases in Asia
(including both hospital and private practicing There are many encounter databases available
specialists). Detailed data, e.g. diagnoses or across the Asia-Pacific Region. Many of these
laboratory data, are not available. However, are population-wide databases due to promi-
data on contacts (without specification for the nence of nationwide healthcare coverage in
reason for such contacts) can be obtained. these countries. For example, South Korea and
Rules governing data access vary between Taiwan both have single-payer, universal
the Nordic countries, but generally require col- government-run health insurance systems that
laboration with local researchers. Access to predominantly operate on a fee-for-service
Danish prescription data is particularly restric- basis and have established national research
tive. Consequently, data from the Danish databases. The National Health Insurance
National Prescription Registry cannot leave Databases of South Korea and Taiwan are the
the data havens provided by Danish authori- most well-established and widely used Asian
ties. For multinational studies involving encounter databases. Similar to encounter
Danish individual-level prescription data, databases in the US, Canada, and Europe,
pooled analyses require data to be transferred these databases capture patient demographic
to e.g. Statistics Denmark or meta-analysis information, medical (in- and outpatient) ser-
techniques to be applied to obtain pooled esti- vices and prescription and dispensing data.
mates. Other sources of Danish prescription Encounter databases also exist in Australia,
data are not restricted in the same way, but and Japan. In Australia, the commonwealth
only provide data on reimbursed prescriptions government maintains a dataset of dispensing
and either only offer local coverage or only of subsidized medicines under the
provide data on reimbursed prescriptions and Pharmaceutical Benefits Scheme (PBS) and
only cover more recent years. medical services under the Medicare Benefits
Schedule (MBS). A 10% sample of these data,
Other European Encounter Databases linked longitudinally, is available and has been
Pharmacy-based federated database networks used for research. Additionally, an encounter
also exist in the Netherlands (PHARMO) and database of services provided to Australian
Scotland (Tayside MEMO). These networks are Veterans is maintained by the Australian
limited to specific regions of their respective Department of Veterans Affairs (DVA). These
countries and provide the ability to link to a data include all prescriptions dispensed, medi-
number of databases that provide outcome and cal services claimed and hospital visits attended
confounder information similar to those in the by the veterans, their dependents and spouses.
Nordic countries. In addition, integrated The DVA data have been used widely for
encounter databases are available in France, research.
and select regions of Italy (Lombardy, One of the advantages of databases across
Tuscany). The French national claims data- the Asia Pacific region is the consistency of
base, SNIIRAM, captures data for more than coding systems. For example, encounter data-
66 million individuals (~98% of the French bases in South Korea, Taiwan and Australia
population) regardless of socioeconomic or all use ATC (Anatomical Therapeutic
employment status. It captures encounter data Chemical) codes to identify individual
edicines and all but Taiwan use ICD-10

m urine, newborns) as well as regular examina-
codes to identify diagnoses. This allows for tions in school children. Strict procedures for
comparisons of similar products across differ- data access and human subject review are in
ent countries without the need to map indi- place to assure protection of confidentiality
vidual country-specific codes. This has and data security.
allowed cross-national studies to be con-
ducted using a distributed network approach The South Korean Health Insurance Review
through the Asian Pharmacoepidemiology and Assessment (HIRA) Data
Network (AsPEN). Pharmacoepidemiologic South Korea provides universal health cover-
studies using Asian databases have histori- age since 1989. In 2000, all health insurance
cally been limited due to restrictions in the systems were integrated into a single national
accessibility of these data. Milea and col- system creating the National Health Insurance
leagues reported that of 54 encounter data- Service (NHIS) and the Health Insurance
bases across the Asia pacific region very few Review and Assessment Service (HIRA). All
allowed access to raw data. Databases in healthcare providers are covered under the
Australia, Taiwan and Japan for example NHIS and are, with a few exceptions, reim-
were considered as having a high level of data bursed on a fee-for-service basis. Claims are
accessibility, while South Korea had a electronically submitted by providers to HIRA
medium level and Thailand, China, Malaysia for reimbursement and are the basis for the
and Singapore had a low level of accessibility. HIRA database, which contains healthcare uti-
The level of accessibility can differ for indi- lization and prescribed medications for
vidual databases within the same country, as approximately 50 million individuals. Use of
some databases may require a local researcher the database for research was initially limited
to access data while others do not provide raw until it became publicly available for research
data with only summary level data available in 2009. HIRA research data include benefi-
for researchers. ciary ID, basic demographics, procedures,
diagnostic tests, all diagnosis received by the
The Taiwanese National Health Insurance Research beneficiary (coded in KCD6, the Korean
Database (NHIRD) Standard Classification of Disease Version 6,
Established in 1995, the National Health which is closely based on the ICD-10 system),
Insurance (NHI) program of Taiwan covers in- and outpatient prescriptions (including
approximately 23 million individuals, more brand name, generic name, prescription and
than 99% of the country’s population. The NHI dispensing date, duration, dose, and route of
maintains the National Health Insurance administration), as well as provider ID and
Research Database (NHIRD), which is acces- characteristics. Validity of diagnosis data in the
sible for research. NHIRD includes but is not HIRA database has been shown to vary accord-
limited to patient demographics, prescription ing to the severity of the condition (with
and dispensing data, outpatient visits, hospi- greater validity for more severe conditions)
talizations, and dental care. Data are updated and the care setting (with higher validity for
biannually. The NHIRD can be linked to a inpatient than for outpatient diagnoses). HIRA
number of external national databases through data are available to researchers in academia
a unique and universal personal identification and government agencies and for those in the
number. Databases available for linkage private sector such as pharmaceutical compa-
include numerous registries (birth, death, nies and medical device companies, but access
immunization, cancer, reportable infectious requires in person consultation at the HIRA
diseases, and suicide), population-based and submission of a study proposal. Once
screening programs (various cancers, myopia, approval is given, tailored data extracts with
Strength 161
encrypted ID information for protection of pri- ultiple databases within the Canadian
m
vacy are uploaded in a remote access system Network for Observational Drug Effect Studies
accessible only by the individual researcher for (CNODES). The respective sizes of study
the study. Importantly, HIRA data are cur- cohorts assembled using a common protocol
rently available only for a five year period and allowing multiple cohort entry dates for a
beginning from the current year although single patient were approximately 2.2 million
plans exist to expand this period to 10 years. for MarketScan, 1.5 million for the combined
In response to the COVID-19 pandemic, Canadian provincial databases, and 0.6 mil-
HIRA made available a de-identified lion for UK General Practice Research
COVID-19 nationwide patient dataset for Database (GPRD), the largest population-
researchers. The dataset contains a five year based EHR-database. The MarketScan cohort
history of insurance claims for all tested cases was more than 3.5 times larger than the GRPD
of COVID-19. Researchers are able to down- cohort, despite not including data on 65 year
load a sample datafile to develop analytic code olds who made up around 35% of the total
which can then be submitted to HIRA for exe- study population.
cution in the full dataset. Second, because encounter databases are
population-based and provide a comprehen-
sive capture of covered health care encounters
Strengths regardless of the provider, they can support the
full range of epidemiologic study designs
Encounter databases have a number of including cohort, nested-case control, and self-
strengths in comparison to other data sources controlled designs. While this strength is
for pharmacoepidemiologic research, which shared by a number of other population-based
explain their broad representation in the automated databases, it is a critical limitation
literature. to non-population based data sources such as
First, automated healthcare databases facili- EHR databases of individual institutions or
tate the rapid and cost-efficient assembly of health systems.
extremely large cohorts of patients and provide Third, many encounter databases facilitate
data on drug exposures, health outcomes, and systematic or ad-hoc linkage to non-encounter
potential confounding factors. Encounter data- data resources including electronic or paper
bases, in particular, are the largest available medical records, disease registries, laboratory
population-based healthcare databases. results, or patient and provider surveys. Such
Several of the databases discussed in this chap- linkages can support validation of study out-
ter cumulatively include more than 100 mil- comes and allow supplementation of
lion individuals and provide the ability to encounter-data with variables such as labora-
rapidly assemble cohorts that are substantially tory results or lifestyle data. In ideal circum-
larger than analogous cohorts from EHR- stances, such linkages thus provide the ability
databases or ad-hoc data collection. Encounter to take advantage of the size and population-
databases thus are uniquely able to address based nature of encounter data, while also
research questions that require the largest pos- accruing the advantages of higher data quality
sible study sizes. The following example illus- and greater clinical detail available from data
trates the differences in cohort sizes for the sources such as EHRs, disease registries, or
same study in select encounter and EHR data- patient and provider surveys. Importantly,
bases. Filion and colleagues examined proton however, linkage ability and quality vary sub-
pump inhibitors and the risk of hospitalization stantially between individual encounter data-
for community-acquired pneumonia among bases and has to be carefully considered for
new users of NSAIDs, aged 40 years in each study question.
Fourth, many large encounter-databases are greatly diminishing the likelihood of recall or
broadly representative of nations, regions, or assessment biases.
particular health systems. As such, these data-
bases can often serve an important role in facil-
itating health services and health policy Limitations
research. Many include very stable popula-
tions that facilitate assessment of long-term Encounter databases are primarily intended
safety effects and long-term trends in treat- and maintained for payment or other adminis-
ment practice and quality. Further, encounter trative purposes, and therefore are subject to
databases from countries or regions with uni- important limitations when used for research.
versal health coverage – by definition – are free First, one of the greatest concerns when using
from selection bias as inclusion in the database encounter databases for pharmacoepidemio-
is universal. logic research is the uncertain validity of diag-
Fifth, for encounter data generated from fee- nostic information (see Chapters 8 and 13).
for-service payment claims, data elements that While these concerns apply to all diagnostic
directly pertain to the payment amount are encounter data, they are amplified for diagno-
subject to auditing and considered highly ses recorded in the outpatient setting where
accurate. This is true for procedure claims diagnosis is typically not directly linked to a
(type of procedure performed) as well as for particular level of payment. It is thus critically
pharmacy claims (date, drug, and quantity dis- important for all encounter-based research to
pensed). Importantly however, the accuracy of validate diagnostic data (for both outcomes
procedure data primarily relates to the occur- and important confounders) against external
rence of the procedure billed while the accu- gold standards such as the medical record or
racy of the clinical indication associated with disease registries. These gold standards, of
the procedure may be substantially lower. For course, may not be correct either when com-
example, a validation study by Wysowski and pared to research grade diagnoses as employed
Baum that used specific surgical procedure by RCTs.
codes in Medicaid data as part of an algorithm Second, encounter data lack clinical detail
to identify cases of hip fracture found in medi- such as markers of disease severity (e.g. blood
cal record review that while all of the proce- pressure, ejection fraction) and lifestyle fac-
dures billed for were actually performed, some tors (tobacco and alcohol use, body mass
of the procedures were used to correct ortho- index, physical activity). Oftentimes data ele-
pedic conditions other than hip fracture. A fur- ments are available (e.g. diagnostic codes for
ther advantage of pharmacy data compared to obesity or smoking status) but of extremely
prescription data recorded in EHR databases low sensitivity. For example, a study by Lloyd
(see Chapter 10) is the fact that prescription and colleagues using data from the National
dispensings are one step closer to ingestion Health and Nutrition Examination Survey to
than what was prescribed and thus are subject validate diagnosis of obesity in Medicare
to a lesser degree of exposure misclassification. claims found that claims-based diagnostics
The accuracy of encounter data generated by codes fail to identify a great majority of
administrative processes not related to pay- patients with obesity (sensitivity of 18%).
ment is less well established and likely to vary Though still far from perfect, clinical detail
depending on the existence and rigor of quality such as disease severity and lifestyle factors
assurance processes. are generally better captured by paper or
Sixth and last, data capture processes in electronic medical records. Because such
encounter data are automated and independ- clinical detail is often critical for confound-
ent of the study question and hypothesis, ing adjustment, methods that minimize
unmeasured or residual confounding (self- Particular Applications

controlled designs, active comparator new-
user designs, instrumental variable analyses, Encounter databases have been used in thou-
propensity score calibration) are of great sands of pharmacoepidemiologic publications,
importance to encounter-based pharmacoep- many of which have shaped clinical medicine
idemiologic research (see Chapter 22). or regulatory decision-making. These data-
Third, while limitations of encounter databases have supported work across a wide spec-
bases can often be overcome by facilitating trum of areas including drug safety,
linkage to non-encounter data such as EHRs, comparative effectiveness, drug-utilization,
disease registries, or laboratory results, such and health services research, methods and val-
linkages are typically time consuming and idation, as well as pharmacoeconomics. This
costly and, in many cases, only available to section outlines some typical activities involved
subsets of the database population. Further, in encounter database studies and presents
when compared to population-based EHR some of the considerations in choosing the
databases the resulting linked/enriched optimal encounter database when multiple
encounter data typically remain less compre- options are available or assessing the suitabil-
hensive, and validation is often restricted to ity of a specific database for a given research
small samples often with poor response/ question.
retrieval rates.
Fourth, in certain situations medication
Typical Activities Involved
dispensing information may not capture data
in Studies Using Encounter
for specific drugs or drug classes. This may
Databases
include drugs excluded from reimbursement,
drugs that are primarily obtained OTC, as Although encounter databases vary in data
well as low-cost generic drugs that are paid structure, coding schemes, and numerous
for out of pocket because the cash price is other specifics, a number of activities are typi-
lower than the required copayment. This may cal across all such databases. Virtually all phar-
result in misclassification of exposure, such macoepidemiologic studies of encounter
that some patients will appear not to be databases require linkage of records between
exposed to a medicine when in fact they data files and over time. Records from different
were. Non-reimbursable drugs as well as low- data domains, such as membership, outpatient
cost generics are often better captured in services, inpatient services, and pharmacy, are
EHR databases, which contain information linked so that an individual’s entire set of
on all prescriptions written. However, the encounters over the study period can be avail-
disadvantage of prescription information is able for analysis. Another ubiquitous step in
that not all prescriptions will be dispensed the conduct of pharmacoepidemiologic studies
and will result in misclassification of expo- involves the aggregation of drug-, diagnosis-,
sure, such that some patients will appear to and procedure codes into meaningful study vari-
be exposed to a medicine when in fact they ables. Exposures, outcomes, potential con-
were not. founders, and in−/exclusion criteria for study
Fifth and last, due to the fragmentation are defined via code lists using drug-, diagno-
of the US healthcare system, many large sis-, and procedure codes, or combinations
US encounter databases lack representa- thereof. These code lists are typically study-
tiveness of the general population and fea- and database- specific using the coding
ture significant turn-o ver and short dwell schemes utilized by the respective database
times (e.g. US private insurance databases, and drugs approved and available for the study
MAX/TAF). population during the study period. It is often
desirable to use previously validated algo- study question. For example, Stroup and col-
rithms for the definition of study outcomes leagues aimed to examine the effectiveness of
and important confounding variables. Such initiating treatment with either clozapine or a
algorithms often combine diagnostic codes, standard antipsychotic among adults with evi-
drug codes, and procedure codes for more dence of treatment-resistant schizophrenia
accurate measures of disease (see Chapter 13). using national US Medicaid data. On first
Together with demographic information these glance, this might not be an obvious choice as
study-specific variables (e.g. drug classes, dis- the adult Medicaid population is highly selec-
ease states) as described above facilitate the tive and often transient. However, Medicaid
creation of the study population. Study popula- covers approximately two thirds of all US
tions often consist of (new) users of specific adults with schizophrenia because most
drugs or drug classes within individuals who patients with severe schizophrenia qualify for
meet specific inclusion and exclusion criteria disability. In addition, because these individu-
based on their encounter-derived medical his- als qualify for Medicaid because of disability
tory. Once the study population is identified in rather than because of their economic condi-
the dataset, analytic plans often specify the tion, they are typically stably enrolled without
construction of longitudinal histories. breaks in coverage. While non-US encounter
Exposure, occurrence of outcome events, and databases might have provided similarly large
presence of confounding factors are measured numbers of stably-enrolled patients with
over time, typically in temporal relation to the schizophrenia, the authors sought a US data-
study’s index date. This facilitates the assess- base because of the pronounced differences in
ment of exposure periods, person-time at risk, psychiatric treatment practice between US and
and allows calculation of incidence rates and most other countries. The study was conducted
measures of association. If additional data not as a 1 : 1 propensity score matched cohort study
available in the encounter database are and found that clozapine treated patients com-
required, complementary information may be pared to patients treated with a standard antip-
gathered through linkage to electronic medical sychotic had a decreased risk of psychiatric
records, data obtained directly from patients or hospital admission (hazard ratio = 0.78, 95%
their physicians from surveys, retrieval of CI = 0.69–0.88) but at an increased risk of dia-
paper medical records, or data routinely col- betes mellitus (hazard ratio = 1.63, 95%
lected in disease, immunization, or national CI = 0.98–2.70).
vital registries. Database size: The database should be large
enough to provide sufficient power to answer
the research question, e.g. to detect a mean-
Deciding Between Individual
ingful difference between treatment groups
Encounter Databases
(should a difference truly exist). This assess-
Database choice or evaluation of suitability of ment should be based not on the size of the
a single database should involve consideration overall database, but rather consider the size
of all database attributes relevant to the of the actual study cohort, i.e. the cohort after
research question under study. Some of the key exclusion of individuals for whom required
attributes that differentiate individual encoun- data elements are unavailable (e.g. after exclu-
ter databases are shown in Table 9.2 and dis- sion of individuals under capitated payment
cussed below. plans), and after application of inclusion and
Target population: The database should cap- exclusion criteria (e.g. sufficient uninter-
ture a large and representative sample of the rupted baseline period). A study by Shin et al.
target population (e.g. patients exposed to a aimed to determine the risk of cardiovascular
particular drug) to adequately address the conditions in children and adolescents with
attention deficit hyperactivity disorder any of the outcome events within any of the
(ADHD) associated with use of methylpheni- four databases. Other outcomes are notori-
date. As the outcome was rare, the South ously under-coded in encounter data and
Korean HIRA database of over 50 million par- require development of custom algorithms.
ticipants was used. From this large population For example, using data from Quebec, Moride
database, 144 258 patients aged less than et al. developed and validated a case detection
18 with a diagnosis of ADHD were retrieved. algorithm for suicide attempts in youth
Of these, 114 657 were new users of methyl- through a review of medical charts. The fol-
phenidate and 1224 had an incident cardio- lowing algorithm was used: diagnostic code of
vascular event. Due to the rare outcome a injury or intoxication with a location of service
self-controlled case series design was used in the ED, followed by a psychiatric consult or
which, compared to other designs, has the a psychiatric diagnosis (psychiatric diagnoses
advantage of requiring fewer patients for simi- consisting of depression, eating disorder,
lar power (see Chapter 22). schizophrenia, ADHD, substance abuse, oth-
Ability to validate outcomes: Because encoun- ers) within two days of the ED visit. This algo-
ter data are primarily collected for administra- rithm had a sensitivity of 70% and a specificity
tive purposes, the ability to validate or of 97.6%.
adjudicate outcome definitions derived from Availability of non-standard encounter data.
these data is essential for pharmacoepidemio- While all encounter databases provide infor-
logic studies. Outcome validation should gen- mation on medical services and prescription
erally be performed as part of any drugs, studies often require encounter data on
encounter-based study unless the outcome services that are not universally available in all
measures have previously been validated for databases. For example, Gupta and colleagues
the database. However, the ability to validate examined opioid prescribing practices among
outcomes, through reliable linkage to external US dentists from 2010 to 2015 using the
gold standards, such as the medical record or MarketScan database. Because dental services
disease registries, varies markedly between are not captured for all individuals in the data-
databases and is often a major consideration base, the study population was appropriately
for database selection. Lo Re and colleagues, restricted to those with simultaneous enroll-
for example, conducted a series of postauthori- ment in a medical and a dental plan.
zation safety studies to examine the safety Ability to supplement with non-encounter
(hospitalization for major adverse cardiovas- data: Studies using encounter data may require
cular events, acute kidney injury, acute liver clinical detail not available from encounter
failure, infections, and severe hypersensitivity data often for the purpose of confounding
events) of saxagliptin compared to other oral adjustment or to supplement outcome identifi-
antidiabetic drugs in patients with type 2 dia- cation. The ability to perform linkages that
betes. The studies were conducted separately allow enrichment of the dataset with non-
in two EHR databases (Clinical Practice encounter data is thus vital and often a deci-
Research Datalink, The Health Improvement sive consideration in choosing a study
Network) and two encounter databases database. For example, Huybrechts and col-
(Medicare, HealthCore Integrated Research leagues examined the comparative mortality
Database). One of the requirements for the risk of individual antipsychotics in elderly
choice of encounter databases in this study nursing home residents using data for US
was the ability to obtain inpatient medical nursing home residents dually eligible for
records for outcome adjudication. Using a Medicaid and Medicare. Clinical variables
new-user active comparator cohort design, the such as cognitive function or behavioral symp-
study found no evidence of increased risk of toms of dementia are important potential
c onfounders but poorly measured in encoun- media, web searches, and around the clock
ter databases. Linkage to the Minimum Data biometric information from wearables. In
Set (MDS, available from CMS), a federally addition, automated tools for data visualiza-
mandated health assessment tool used in nurs- tion and analysis of health data are becoming
ing homes that captures information on physi- more accessible. The potential for rapid devel-
cal, psychological, and psychosocial opment of progressively complex, detailed,
functioning, active clinical diagnoses, health and complete data resources is likely to be
conditions, treatments, and services, allowed counteracted by increasingly strict regulations
the inclusion of these important covariates governing data privacy. These regulations will
into the study. Using a propensity score vary substantially between countries and are
adjusted new-user cohort design, the authors likely subject to rapid change. Last, and maybe
showed that compared to initiators of risperi- most importantly, encounter-based data are a
done, initiators of haloperidol had an increased secondary byproduct of administrative sys-
mortality risk and initiators of quetiapine had tems, created to support the local healthcare
a decreased mortality risk. As another exam- system; research applications are secondary
ple, a Swedish–Danish study investigated the uses. As such, encounter-based healthcare
risks associated with being admitted to an data will continue to be subject to changes in
emergency department with suspected poison- the healthcare systems that generate the data.
ing, most often psychotropics or analgesics. Again, these changes are likely to vary drasti-
Leveraging the ability to link data on admis- cally between countries and over time.
sions and prescription fills to a dataset includ- For example, the US healthcare environ-
ing detailed ECGs on those admitted to the ment is undergoing enormous transformation.
hospital, they could estimate not only the Historically, healthcare providers in the US
occurrence of QTc prolongation within the have been paid using a fee-for-service
population but also to what extent QTc prolon- approach, where providers bill health insur-
gation as a marker was associated to 30-day ance companies for the cost of the services
mortality. they provide, generally justifying those bills
with diagnoses. These paid claims represent
the core of these encounter databases.
The Future However, the net result of this approach is that
the more providers do, the more they are paid
Pharmacoepidemiologic research with which may result in over servicing and wasted
encounter databases has become more and resources. The result has been a large incentive
more widely used and involves an increasing to increase utilization, and rapidly increasing
number of databases in a growing number of costs in the US for providing healthcare, made
regions of the world. This trend is expected to worse by an aging population. Under this
continue, particularly as encounter databases model the levels of expenditure are unsustain-
become available in regions for which cur- able. This has led to a shift from a fee-for-
rently no data are available. In addition, the service model to a “per patient per month”
following three major factors are likely to payment system, so-called “population
shape the future of encounter databases: (i) health,” which of course switches the incen-
advances in IT, (ii) privacy regulations, and tive to providing less care. In order to attempt
(iii) changing healthcare systems. Advances in to address that, there are incentives being put
IT will continue to expand the boundaries of in place to ensure that people are not receiving
data storage and processing, and increasingly too little care, referred to as “value health.” The
facilitate linkages with new and more complex US is in the middle of this transition now, vary-
sources of data including biomarkers, social ing greatly in different parts of the country.
Key Point 167
However, in response, there has been a remark- for the usefulness of encounter databases for
able consolidation of physician practices, hos- pharmacoepidemiologic research. Such
pitals, etc., in order to achieve sufficient scale databases are considered population-based.
to create the needed extensive and costly data ●● Fit-for-purpose evaluation of a given
infrastructure, and to assume the large risk encounter database should consider the fol-
associated with population health. Many other lowing characteristics: (i) population and
initiatives are underway as well, to limit the coverage period, (ii) services covered and
increasing costs of medical care. The results data completeness, (iii) ability to link to
will likely be large changes over the next few complimentary external data sources, and
years in the data as part of US encounter (iii) access.
databases. ●● Data accuracy for dispensed outpatient pre-
Encounter-based data are an important scription drugs is typically high.
resource for pharmacoepidemiologic ●● The accuracy of diagnostic codes is highly
research. These data are comprehensive and variable, depending on the specific condi-
often have a high level of quality as they are tion, the setting of care, and the purpose of
collected for payment purposes. As these data the encounter database. Use of validated
are generated for purposes other than outcome definitions or verification of diag-
research, careful consideration of their appli- noses using primary medical records is
cability, completeness, and generalizability strongly encouraged.
need to be carefully weighed against their ●● The US has a highly fragmented healthcare
convenience. As with any data source careful system and no single database captures a
consideration should be given to the issues of representative cross-section of its
bias and confounding (see Chapter 2) which population.
are not problems diminished by the increased ●● US Commercial insurance databases include
size of the database. some of the largest databases available.
These databases typically aggregate data for
members covered by a variety of insurance
Key Points products.
●● US integrated healthcare delivery system
●● Encounter databases are maintained in databases include defined populations
numerous countries, by a variety of entities whose entire spectrum of care is under the
(e.g. government agencies, insurance responsibility of and provided by the inte-
companies, health plans), and for a grated delivery system. They are substan-
variety of purposes (e.g. reimbursement, tially smaller than typical commercial
administration). insurance databases but provide a more
●● Encounter databases used in pharmacoepi- complete picture of the healthcare received
demiology typically contain the following by its members.
core data domains: (i) eligibility and basic ●● Medicare in the US provides healthcare cov-
demographic information, (ii) outpatient erage for almost all people 65 years and over
pharmacy dispensings, and (iii) medical ser- as well as for qualifying individuals with per-
vices (typically including hospitalizations manent disabilities. Prescription drug cover-
and emergency department services; com- age is optional and provided under Medicare
monly also including outpatient health Part D by private PDPs.
services). ●● Medicaid in the US provides comprehensive
●● A defined population for which healthcare hospital, medical, and prescription drug cov-
services are recorded regardless of the pro- erage for certain categories of disadvantaged
vider or location of care received is critical individuals.
●● In Canada, there is no single national linked ●● Multiple other countries throughout Europe
encounter database. Owing to the structure hold data sources covering sizeable propor-
of the healthcare system, each province and tions of the population, e.g. Italy, Germany,
territory maintains encounter databases that France, United Kingdom, and Holland.
include prescription and medical services ●● Many long established databases in the Asia-
claims of the public healthcare programs, Pacific region cover high proportions of the
which can be linked to hospital discharge population due to single-payer, nationwide
data. healthcare coverage.
●● All Canadian residents are covered for medi- ●● Despite variation in healthcare systems and
cal services regardless of age or income. languages used throughout the region many
However, drug coverage varies between of the databases conform to world standard
provinces and access to raw data is also medical vocabularies such as the WHO ATC
restricted in some provinces; these aspects classification system for medicines and ICD-
are important to consider in the selection of 10 codes for diagnoses.
a Canadian database. ●● Use of Asian databases have historically
●● Owing to coverage characteristics, patients been limited due to restrictions in the acces-
can be followed long term in Canadian sibility of these data. Access to databases
databases. may only be provided to local researchers
●● The Nordic countries provide nationwide while other databases do not provide direct
coverage of prescription fills with easy link- success to raw data with only summary level
age to other health registries on, e.g. data available for researchers.
hospitalizations.
Case Example 9.1 US Commercial Insurance Databases (Dave et al. 2019)
Background databases: IBM MarketScan and Optum

●● Sodium–glucose cotransporter- 2 inhibi- Clinformatics Datamart between March
tors (SGLT2i) are a class of antidiabetic 2013 to September 2015.
medications that reduce serum glucose by ●● Within each database, patients aged
inhibiting its reabsorption in the proximal 18 years or older with type 2 diabetes ini-
tubule. tiating SGLT2i versus dipeptidyl pepti-
●● Based on their mechanism of action, they dase-4 (DPP4i) were identified and
are thought to increase the risk of severe matched using 1 : 1 propensity score
Urinary Tract Infections (UTIs); however, matching using >90 baseline covariates.
findings from case- reports and meta- ●● The primary outcome was a severe UTI
analysis of randomized controlled trials event, defined as a composite outcome of
are inconclusive. hospitalization for primary UTI, sepsis with
UTI, or pyelonephritis. Hazard ratios (HR)
Question
were estimated using Cox regression.
Are type 2 diabetes patients initiating an
SGLT2i at a higher risk of developing severe
Results
UTIs compared to those initiating non-
●● After 1 : 1 propensity score matching, the
SGLT2i therapies?
study identified 61 876 patients in each
Approach group.
●●A retrospective cohort study was conducted ●● New initiators of SGLT2i had 61 severe
using the two US based commercial claims UTI events (incidence rate [IR] per 1000
Key Point 169
person-years, 1.76), compared with 57 diabetes. Although the study adjusted for
events in the DPP-4 inhibitor group (IR, more than 90 potential confounders, it could
1.77), corresponding to an adjusted HR of not directly control for important variables
0.98 [95% confidence interval, 0.68–1.41]). such as duration of diabetes or body mass
index. Hemoglobin A1c results were available
Strengths
for <15% of sample further limiting the study
●● This cohort study utilized a large, diverse
ability to adjust for diabetes severity.
cohort of patients with diabetes from two
commercial insurance databases. Key Points
●● The study was able to follow patients for ●● In this large population-based cohort
longitudinal exposures and important study of patients with type 2 diabetes,
clinical outcomes, including hospitaliza- SGLT2i was not associated with an
tions due to severe UTIs. increased risk of severe UTIs. On the basis
of our findings, other factors beyond the
Limitations risk of severe UTIs should be considered in
The small number of outcomes precluded decisions about whether to prescribe
detailed evaluation of the associations SGLT2i for patients with diabetes in rou-
between individual agents and the risk of tine care settings.
Case Example 9.2 Canada (Faure et al. 2020)
Background Questions
●● Despite the existence of clear clinical rec- ●● What are the patterns of secondary stroke
ommendations, previous studies have prevention treatments in the real-world
shown that 19–50% of patients who expe- clinical practice setting?
rienced an acute ischemic stroke (AIS) do ●● What is the long- term effect of secondary
not receive secondary prevention, consist- stroke prevention on the risk of death or
ing mainly of antiplatelet therapy (e.g. AIS recurrence?
dipyridamole, clopidogrel, acetylsalicylic
acid (ASA), or ticlopidine), or anticoagu- Approach
lants (e.g. vitamin K antagonist (VKA) or ●● Population-based cohort study of adult
nonvitamin K antagonist oral anticoagu- patients (age ≥ 18 years) who were dis-
lants (NOAC)). charged following a hospitalization for an
●● Although the efficacy of pharmacological incident AIS between January 1, 2011, and
secondary prevention has been demon- December 31, 2012 and who survived
strated in clinical trials, evidence gaps beyond the short-term period (30 days).
remain regarding the effectiveness of sec- Data sources were the linked prescription
ondary prevention in the real-world clini- and medical services claims databases of
cal practice setting; available data are Quebec (RAMQ databases).
either not recent, involved a short-term ●● The main study outcome was the compos-
follow-up (e.g. 30 days), or restricted to ite of death or AIS recurrence between

specific subpopulations. 31 and 365 days post discharge. Cox
(Continued)
roportional hazard models were used to

p ●● Compared with previous studies that were
compare the risk of death or AIS recur- mainly limited to the short-term period
rence across treatments using no pre- after a stroke, treatment effectiveness was
scribed treatment as the reference. In evaluated over a one-year period.
order to account for treatment switches ●● The use of a time-dependent exposure
and discontinuation during follow-up, variable accounted for treatment hetero-
exposure was time-dependent. geneity during follow-up.
Results Limitations
●● In the month after discharge, 44.3% of the ●● Information on drugs acquired OTC were
patients did not receive the recommended not available, which may have overesti-
treatment and > 20% did not have any mated the proportion of untreated patients
treatment. owing to the availability of ASA OTC.
●● Untreated patients were younger, had less However, for cost consideration, patients
comorbidities, and a more severe index AIS chronically treated favor acquiring ASA by
(using hospital length of stay as a proxy for prescription.
severity). ●● Absence of data on previous AIS events
●● Anticoagulants and antiplatelets were that predated the one year look back
associated with a lower risk of death or period may have resulted in misclassifica-
recurrence (hazard ratio [HR] 0.27; 95% tion of incident cases.
confidence interval [CI] 0.20–0.36 and HR
Key Points
0.25; 95% CI 0.16–0.38, respectively) com-
●● The risk of death or AIS recurrence was
pared with the untreated group.
reduced by 50–75% in patients receiving a
Strengths secondary prevention treatment when
●●This was a population-based study that compared with the untreated patients.
reflected current prescription patterns for ●● Findings confirm treatment benefits
secondary AIS prevention in the commu- shown in clinical trials and emphasize the
nity practice setting. importance of AIS secondary prevention.
Case Example 9.3 Encounter Databases in Asia (Roughead et al. 2015)
Background Questions
Thiazolidinediones (rosiglitazone and piogl- Do the risks of heart failure and edema asso-
itazone), are associated with heart failure ciated with the thiazolidinediones vary
and edema. Many of the published studies between populations located in Asia,
investigating these adverse events were Australia and Canada?
conducted in caucasian populations. Due to
the differences in metabolizing enzymes and Approach
pharmacodynamic-based variation in ●● Sequence symmetry analysis (SSA), a sig-
response to the thiazolidinediones there nal detection method for adverse drug
may be a difference in the prevalence of events utilizing administrative claims data,
adverse events across different ethnic was used to assess the association
groups. between the thiazolidinediones and
Further Reading 171
edema (as indicated by furosemide initia- ●● Standardized analytical code and stand-
tion) or heart failure hospitalization across ardized data variables were used to avoid
countries. differences due to coding
●● Incident dispensing of rosiglitazone, ●● While results differed across ethic popula-
pioglitazone, metformin, and furosemide tions, results were consistent within
and hospitalization for heart failure were populations
determined for each individual patient. Limitations
All incident dispensing that occurred ●● Differences in the underlying prevalence
within one year of each other for the same of polymorphisms in both metabolizing
person were included in the analysis. The enzymes and pharmacodynamic receptors
crude sequence ratio (SR) was calculated may have played a role; however, differ-
by dividing the number of persons with ences in other factors, including diet, phys-
furosemide initiated after rosiglitazone ical activity or healthcare practice may
initiation with the number of persons also be contributors.
with furosemide initiated prior to ●● Diagnostic information was not available
rosiglitazone. in all countries which limited the strength
●● The SSA method uses a within-person of conclusions that could be drawn regard-
design, making it robust toward confounding the more serious outcome of heart fail-
ers that are stable over time. ure hospitalization
●● Not all medicines were available in all
Results countries resulting in the exclusion of

When results were pooled across the some comparisons
Caucasian populations there was a signifi- Key Points
cantly increased risk of furosemide initiation ●● The risk of edema and heart failure associ-
for rosiglitazone and pioglitazone, while in ated with rosiglitazone was generally
the Asian populations, the pooled risk esti- lower in the Asian population than in the
mates were lower for rosiglitazone and not Caucasian population.
significant for pioglitazone. ●● There is potential for differences in
response to medicines by ethnicity and
Strengths these differences should be investigated
●●This was a large study across six different when considering whether regulatory
countries in eight different databases. action is required.
Further Reading
Hall, G.C., Sauer, B., Bourke, A. et al. (2012). Faure, M., Castilloux, A.M., Lillo-Le-Louet, A.
Guidelines for good database selection and et al. (2020). Secondary stroke prevention: a
use in pharmacoepidemiology research. population-based cohort study on
Pharmacoepidemiol. Drug Saf. 21 (1): 1–10. anticoagulation and antiplatelet treatments,
Strom, B.L., Kimmel, S.E., and Hennessy, S. and the risk of death or recurrence. Clin.
(eds.) (2012). Pharmacoepidemiology, Fifth Pharmacol. Ther.
Edition. John Wiley and Sons. 107: 443.
Roughead, E.E., Chan, E.W., Choi, N.K. et al. Dave, C.V., Schneeweiss, S., Kim, D. et al. (2019
(2015 Sep). Variation in association between Aug 20). Sodium-glucose cotransporter
Thiazolidinediones and heart failure across 2 inhibitors and the risk of severe urinary
ethnic groups: retrospective analysis of large tract infections. Ann Intern Med. 171 (4):
healthcare claims databases in six countries. 248–256.
Drug Saf. 38 (9): 823–831.
US Databases
Crystal, S., Akincigil, A., Bilder, S., and Walkup, antipsychotic drugs: population based cohort
J.T. (2007). Studying prescription drug use and study. BMJ 344: e977.
outcomes with Medicaid claims data: Lloyd, J.T., Blackwell, S.A., Wei, I.I. et al. (2015).
strengths, limitations, and strategies. Med. Validity of a claims-based diagnosis of obesity
Care 45 (10 Supl 2): S58–S65. among Medicare beneficiaries. Eval. Health
Gupta, N., Vujicic, M., and Blatz, A. (2018). Prof. 38 (4): 508–517.
Opioid prescribing practices from 2010 Lo Re, V., Carbonari, D.M., Saine, M.E. et al.
through 2015 among dentists in the United (2017). Postauthorization safety study of the
States: what do claims data tell us? J. Am. DPP.4 inhibitor saxagliptin: a large scale
Dent. Assoc. 149 (4): 237–245. e236. multinational family of cohort studies of five
Hennessy, S., Leonard, C.E., and Bilker, W.B. outcomes. BMJ Open Diabetes Res. Care 5 (1):
(2007). Researchers and HIPAA. Epidemiology e000400.
18 (4): 518. Stroup, T.S., Gerhard, T., Crystal, S. et al. (2016).
Hennessy, S., Bilker, W.B., Weber, A., and Strom, Comparative effectiveness of clozapine and
B.L. (2003). Descriptive analyses of the standard antipsychotic treatment in adults
integrity of a US Medicaid claims database. with schizophrenia. Am. J. Psychiatry 173 (2):
Pharmacoepidemiol. Drug Saf. 12 (2): 103–111. 166–173.
Huybrechts, K.F., Gerhard, T., Crystal, S. et al. Wysowski, D.K. and Baum, C. (1993). The
(2012). Differential risk of death in older validity of Medicaid diagnoses of hip fracture.
residents in nursing homes prescribed specific Am. J. Public Health 83 (5): 770.
European Databases
Bezin, J., Duong, M., Lassalle, R. et al. (2017). Schade Hansen, C., Pottegard, A., Ekelund, U.
The national healthcare system claims et al. (2018). Association between QTc
databases in France, SNIIRAM and EGB: prolongation and mortality in patients with
powerful tools for pharmacoepidemiology. suspected poisoning in the emergency
Pharmacoepidemiol. Drug Saf. 26 (8): 954–962. department: a transnational propensity score
Furu, K., Wettermark, B., Andersen, M. et al. matched cohort study. BMJ Open 8 (7):
(2010). The Nordic countries as a cohort for e020036.
pharmacoepidemiological research. Basic Schmidt, M., Schmidt, S.A., Sandegaard, J.L.
Clin. Pharmacol. Toxicol. 106 (2): 86–94. et al. (2015). The Danish National Patient
Pottegard, A., Schmidt, S.A.J., Wallach- Registry: a review of content, data quality, and
Kildemoes, H. et al. (2017). Data resource research potential. Clin. Epidemiol. 7:
profile: the Danish National Prescription 449–490.
Registry. Int. J. Epidemiol. 46 (3): 798.
Asian Databases 173
Canadian Databases
Billioti de Gage, S., Moride, Y., Ducruet, T. et al. pneumonia: replicated cohort studies with
(2014). Benzodiazepine use and risk of meta-analysis. Gut 63 (4): 552–558.
Alzheimer’s disease: case–control study. BMJ Moride, Y., Lynd, L.D., Ducruet, H. et al. (2014).
349: g5205. Antidepressants and risk of suicide or
Filion, K.B., Chateau, D., Targownik, L.E. et al. self-harm in Canadian youth: a study
(2014). Proton pump inhibitors and the risk of involving common data models in Quebec and
hospitalisation for community-acquired British Columbia. Pharmacoepidemiol. Drug
Saf. 23 (S1): S10–S11.
Asian Databases
Hsing, A.W. and Ioannidis, J.P. (2015). Park, B.J., Sung, J.H., Park, K.D. et al. (2003).
Nationwide population science: lessons from Report of the Evaluation for Validity of
the Taiwan National Health Insurance Discharged Diagnoses in Korean Health
Research Database. JAMA Intern. Med. 175 Insurance Database, 19–52. Seoul: Seoul
(9): 1527–1529. National University.
Kim, J.A., Yoon, S., Kim, L.Y., and Kim, D.S. Shin, J.Y., Roughead, E.E., Park, B.J., and
(2017). Towards actualizing the value potential Pratt, N.L. (2016). Cardiovascular safety of
of Korea Health Insurance Review and methylphenidateamong children and
Assessment (HIRA) data as a resource for young people with attention-deficit/
health research: strengths, limitations, hyperactivity disorder (ADHD): nationwide
applications, and strategies for optimal use of self controlled case series study. BMJ 353:
HIRA data. J. Korean Med. Sci. 32 (5): 718–728. i2550.
174
10
Electronic Health Record Databases

Daniel B. Horton1, Harshvinder Bhullar2, Francesca Cunningham3, Janet Sultana4,
and Gianluca Trifirò5
1
Rutgers Robert Wood Johnson Medical School, Rutgers Center for Pharmacoepidemiology and Treatment Science, Rutgers School of
Public Health, New Brunswick, NJ, USA
2
Independent Consultant, London, UK
3
US Department of Veterans Affairs, Hines, IL, USA
4
Mater Dei Hospital, Msida, Malta and Exeter College of Medicine and Health, University of Exeter, Exeter, UK
5
Department of Diagnostics and Public Health, University of Verona, Verona, Italy
Introduction including some from Europe, may lack data

from secondary care (e.g. hospitals and special-
Electronic health record (EHR) databases are ists) or linkage to these datasets. While EHR
longitudinal patient record databases that are databases usually contain data on prescribed
used by clinicians in caring for their patients outpatient drugs, many lack information on
and anonymized for the purpose of research. drug dispensing or inpatient medications.
They often record information unavailable in In this chapter, we focus on selected
administrative databases, such as symptoms of European primary care EHR databases and a
illness, historical data, family history, smoking national EHR database for United States
and alcohol use, vital signs (e.g. body mass Veterans. While there are similarities among
index [BMI]), and laboratory data. Further, EHR databases, there are also important differ-
EHR data recorded in the provision of patient ences (see Tables 10.1 and 10.2).
care may more accurately represent patients’
true clinical states than administrative claims
data, which are maintained primarily for bill-
Description
ing or administrative purposes.
Overview of Health Care Systems
Despite their advantages, EHR databases
and Populations
have certain limitations. Certain available data
may have high rates of missingness (e.g. race, Europe: Many European nations (e.g. Italy,
disease severity, smoking history). The validity Netherlands, UK) have either universal
of diagnostic codes cannot be presumed with- government-funded health care or universal
out formal validation. Some EHR databases, health insurance. General practitioners (GPs),
including the Veterans Affairs (VA) and other and sometimes family pediatricians, act as
US EHRs, may not capture information from gatekeepers for medical care in many coun-
out-of-system care. Other EHR databases, tries. As a result, many European primary
Table 10.1 Overview of EHR databases.
Caserta
BIFAP SIDIAP LHU LPD Italy Pedianet CPRD IMRD IPCI IQVIA DA VA
Country Spain Spain Italy Italy Italy United United Netherlands Germany, France, United States
(region) (Catalonia) (Caserta) Kingdom Kingdom United Kingdom
Initiated 2001 2006 2000 1998 2000 1987 2002 1989 1992 1997
Patients, 7.9 M 5.6 M 0.9 M 1.6 M 0.4 M 59.5 M patients 19 M patients 2.4 M Germany: 34 M patients 14.5 M
follow-up patients; patients; patients, patients; pediatric (16.4 M active); (3.0 M active); patients; (including 17.2 M patients
time 49.7 M 5.7 M 10.2 M 19.2 M patients; >400 M >90 M >12.1 M German specialty (6.4 M active);
person- person- person- person- 1.8 M person-years person-years person-years patients); 54.5 M 168 M
years years years years person-years person-years person-years
France: 10.5 M patients;
6.0 M person-years
UK: 4.2 M patients;
17.9 M person-years
Ages All All All 15 16 All All All All All
included
(years)
Diagnostic ICD-9, ICD-10 ICD-9 ICD-9 ICD-9, Read, ICD-10, Read, ICPC ICD-10, Read (UK) ICD-9,
coding ICPC ICD-10 SNOMED SNOMED ICD-10, CPT
system
Drug ATC ATC, NDC ATC, NDC ATC ATC, NDC Gemscript Gemscript ATC ATC VA Drug
coding Classification
system System, NDC
Software Various, e-CAPTM Saniarp, Millewin Junior Bit Vision, EMIS Vision Various Various Various
used mainly Arianna
OMI-AP
ATC, Anatomical Therapeutic Chemical; CPT, Current Procedural Terminology; ICD, International Classification of Diseases; ICPC, Classification of Primary Care; M,
million; NDC, National Drug Code; NHS, National Health Service.
Table 10.2 Selected variables in EHR databases available for epidemiologic research.
BIFAP SIDIAP Caserta LHU LPD Italy Pedianet CPRD IMRD IPCI IQVIA DA VA
Health care Practice Age, sex, Physicians’ Practice Pediatricians’ Professional role Professional role N/A Physicians’ age, Professional role
professional location professional age, sex, years location age, sex, city of person entering of person entering sex, years in of person
demographics role, since of clinic; other data data practice entering
performance graduation information prescription
indicators (e.g. available by data
quality of care) request
Types of health Mainly GPs, Primary care GPs; claims GPs Family GPs GPs GPs Mainly GPs; Primary care
care other primary professionals: data from the pediatricians IQVIA DA physicians,
professionals care GPs, same Germany and physician
professionals pediatricians, catchment area France also specialists (e.g.
(e.g. dentists, include cardiologists),
pediatricians, nurses, specialists (e.g. other clinicians
nurses) midwives cardiologists, (e.g. nurse
dermatologists) practitioners,
pharmacists)
Practice and Practice Practice Practice Practice Practice Practice Practice Practice Practice Practice
patient Number of Location, Province Location Region, Region, practice Region, number of Number of Region, Region, facility,
demographics patients urban/rural, patients per size, practice-level
patients, employees community size, type of facility
Patient Patient
registered with number of DOB, sex, DOB, sex practice SES (Index of computerization available for patients per (clinics at
GP; number of patients, healthcare healthcare Patient Multiple date, Vision date, some practices practice, medical center
persons deprivation exemption exemption YOB, age, sex, Deprivation, Acceptable Patient number of vs. community-
registered in index (based on (based on region of Townsend scores, Mortality physicians, based), facility’s
DOB, sex
practice (MEDEA) salary, salary, residence, ~60%), date of lastReporting (see number of level of
available upon Patient: disability, disability) nationality, registration, text) employees, type complexity
request DOB, sex, chronic information Up-to-Standard Patient (e.g. GP vs Patient
Patient country of diseases) about parents date (see text) specialty)
YOB for adults, DOB, sex, race,
DOB, sex origin (e.g. Patient month/YOB for Patient ethnicity, zip
nationality, YOB for adults, children; Age, sex, health code
habits, blood month/YOB for patient-level, insurance status
group, children; sex, location-based SES (e.g. private,
mother’s ethnicity (~25% (Townsend statutory),
educational recorded; also deprivation scores, medical
level, SES) available via 95% recording), insurance
census data), region, ethnicity, company,
census-based SES; (20% recording), region, town
patient status patient status size (>100 000
(active, died, (active, died, vs. <100 000)
transferred out) transferred out)
Vital signs and Weight, BMI, BP, BMI, Height, weight, BMI, BP, Gestational Height, weight, Height, weight, BP, weight, BMI (~40%); BP, HR, height,
social history BP, smoking, smoking, BMI, smoking, smoking, age, birth BP, BMI BP, BMI BMI, smoking smoking and weight, SES,
alcohol use alcohol use, alcohol use alcohol use weight, birth (measurements (measurements (recorded alcohol education,
Framingham (25% of height, may be biased may be biased when recording marital status,
score. patients 65 neonatal toward patients toward patients considered unknown smoking (>90%)
Pediatric since 2013); BP jaundice; with clinical with clinical relevant)
screening data available (since growth indications); indications);
(height, 2016) measurements smoking smoking (86–94%),
weight, head (e.g. height, (83–93%), obesity obesity (73–83%),
circumference, weight); (61–79%), alcohol alcohol use
pubertal parental use (~80%) (75–85%)
development) smoking
Referrals, PCPs’ referrals Laboratory test Laboratory test Referral Apgar scores, Detailed Electronic Often N/A; test HbA1C, blood Specialist
procedures, to specialists results, results (~25%); data, laboratory/ information on referrals available; results may be glucose, referrals
results of and hospitals; diagnostic/ linkage to diagnostic imaging tests referrals, most outpatient available from cholesterol, available; all
investigations results from imaging hospital tests orders ordered and procedures, laboratory results letters from LDL, HDL laboratory
referrals may referrals; discharge data, reasons for laboratory tests hospitals or available; other results available
be recorded in spirometry; referral data, request; test (~75%, via linkage free text test results (require
coded fields or referrals for diagnostic tests results to HES) variably standardization)
as free text; therapeutic orders sometimes recorded, can be
self-referrals procedures; unavailable requested from
(less common) referrals to paper files
not available secondary/
tertiary care
(date, reason,
specialty
referred)
(Continued)
Table 10.2 (Continued)
Drug data types Drugs Drugs Drugs Drugs Drugs Drugs prescribed Drugs prescribed Drugs Drugs Drugs
prescribed and prescribed and prescribed in prescribed prescribed and in primary care; in primary care; prescribed in prescribed prescribed and
dispensed in dispensed in community in dispensed in some OTC drug vaccine data community dispensed in
community community setting; drug community community data available (see available setting; vaccine outpatient and
setting; vaccine setting (when dispensing by setting; setting; text); vaccine data data available; inpatient
data available covered by linkage to vaccine data inpatient drug available drug settings; vaccine
national claims available data available dispensing by data available
healthcare if reported to linkage to
system); pediatrician; Dutch
vaccine data non- PHARMO
available compulsory database
vaccine data
available,
remaining
vaccine data
identified via
linked claims
Available drug Drug name, ATC code, Drug ATC Drug name, Drug name, Drug name, route, Drug name, route, Drug name, Drug name, Drug name,
information active NDC, code, NDC route, dose, ATC code, strength, strength, quantity, route, dose, route, strength,
substance, indication, (with brand, frequency, indication, frequency, frequency, strength, dose frequency, dose, frequency,
number of profession of formulation, duration, Italian duration; duration; duration, cost quantity,
prescribed prescriber; units), cost MINSAN immunizations immunizations duration,
packages, prescribing indication code, NDC including batch; including batch; directions; cost
duration, data only: (with brand, cost upon request linkage available
prescribed start/end date, formulation, to cost
daily dose, drug units per units),
strength, day; dispensing number of
indication data only: prescribed
units per packages, dose
package, (not available
number of for 30%)
packages per
month, month
of drug
dispensation
Health care GP visits; PCP visits, GP visits; GP visits, Pediatrician GP visits, GP visits, GP visits; other GP visits, Outpatient
utilization referrals by GP referrals to hospital hospital visits, ED/ hospitalizations, hospitalizations data not hospitalizations, visits, ED visits,
to secondary secondary/ discharge discharge hospital consultant visits; entered by GP, sick available sick leave hospitalization
care/ED; tertiary care, letters, letters, admission if links to HES leave (if issued by unless hospital (including
hospital sick leave specialist specialist referred by provide detailed GP); links to HES discharge medical,
admissions (date, length, referrals, ED referrals pediatrician ward-level provides detailed letters sent to surgical,
available if reason), visits by resource ward-level GP intensive care
patients hospital linking with utilization resource units),
referred to GPs discharge claims (England only) utilization Community
after discharge (England only) Living Center
(VA nursing
home)
Identification ICD-9/ICPC Pregnancy, ICD-9 codes N/A May identify Pregnancy, Pregnancy, Some Pregnancy ICD-9/ICD-10
of pregnancy codes for pregnancy for pregnancy siblings pregnancy pregnancy birth-related variable, codes for
and families pregnancy; outcomes; or birth by outcomes; family/ outcomes; data available gynecologist pregnancy
cannot identify mother-baby linkage to household mother-baby link through records; family
families link available claims; cannot identification via family/ hospital data incomplete
identify number; household number discharge
families mother-baby and algorithm letters; cannot
linkage identify
families
Identification Date of death; Date of death Date of death Date of Date/cause of Date/cause of Death date, Date of death; Date and cause Date of death
of death and cause of death by linkage to death death death available sometimes cause cause of death of death seldom
cause of death not available claims via CPRD data of death; death available via recorded
consistently and linkage to certificates free-text
Office for available for fee
National Statistics
Additional Anonymized Hospital N/A N/A Free text by Hospital Hospital discharge Free text by N/A Additional data
data, e.g. free text notes discharge request discharge summaries, request (e.g. consult
consult records, available (30%); other summaries, consultant letters; records, free
free text, paper data by request consultant letters; free text by request text) by chart
files no free text review
available
(Continued)
Table 10.2 (Continued)
Questionnaires, Questionnaires Questionnaires Questionnaires N/A Participating Questionnaires Questionnaires for Questionnaires Questionnaires Chart review for
investigator- for GPs for GPs for GPs pediatricians for GPs and GPs and patients; for GPs upon request validation
initiated (sample) can interview patients; response response rates (response rates
outcome patients/ rates ~90% ~90% low)
validation families
Settings and Inpatient data, Inpatient data Inpatient data Inpatient Inpatient data Prescriptions in Prescriptions in Inpatient/ Secondary care Encounter/drug
types of most OTC N/A except (except via data, OTC N/A for 60%; secondary care, secondary care, specialist data, records, vaccine data from
missing data drugs admission/ discharge drugs, drug most OTC many OTC drugs, many OTC drugs, OTC drugs data, patient- healthcare
discharge data forms with dispensing, drug data; drug dispensing, drug dispensing, level linkage facilities outside
for hospitals of main pediatric adult health adherence adherence between VHA, including
the Catalan diagnoses), clinical and data primary care for local acute
Health laboratory prescribing and specialty care (e.g.
Institute; OTC results for 75%, data (any clinics stroke); some
drugs, drug OTC drugs, setting) inpatient
indications, vaccines medications
drugs not administered
covered by acutely
national health
system
ATC, Anatomical Therapeutic Chemical; BP, blood pressure; DOB, date of birth; ED, emergency department; GPs, general practitioners; HES, Hospital Episode Statistics;
ICD, International Classification of Diseases; ICPC, International Classification of Primary Care; N/A, not available; NDC, National Drug Code; OTC, over-the-counter;
PCP, primary care professional; SES, socioeconomic status; VHA, Veterans Health Administration; YOB, year of birth.
Descriptio 181
care-based EHR databases capture most of encompasses two datasets – CPRD Gold from
their patients’ health information, including practices across the UK, and CPRD Aurum
data from specialty and secondary care (e.g. predominantly from English practices – which
consultations, hospitalizations). European collectively contain data from about 25% of the
EHR databases may contain regional or nation- UK population and represent one of the largest
wide data, depending on the structure of the European EHR databases. IQVIA Medical
health care system and the database. Notably, Research Data (IMRD) (formerly The Health
EHR databases from countries where GPs have Improvement Network [THIN]) was later
less of a gatekeeper role (e.g. France, Germany) established as a collaboration between soft-
have less complete records of patients’ health ware and database companies (respectively,
information. Cegedim and Epic Database Research
US/VA: The VA’s Veterans Health Company Ltd., now part of IQVIA). The same
Administration (VHA) is one of the largest practices may contribute to both CPRD and
integrated health care systems in the United IMRD, but the proportion of overlap changes
States. Funded by the US government, the VA over time as new practices join or leave each
provides medical, surgical, and rehabilitative database. Information from both databases
care to military Veterans, active duty Reservists, may be combined to increase sample size and
and National Guard across nearly 150 hospi- improve statistical power and generalizability.
tals, >1200 outpatient clinics, and > 100 nurs- Merging data between CPRD and IMRD
ing homes organized within 18 regional requires identification and singular inclusion
integrated networks. In contrast to the general of practices contributing to both databases in a
US population, the VA population consists of given year.
predominantly older men (87% male, 47% over The IQVIA Disease Analyzer databases
age 65 as of 2017) who often have multiple (DA, previously known as Mediplus) contain
chronic medical or mental health conditions, anonymized patient records from primary
although the proportion of younger female care practices and office-based specialists
Veterans is rising. Most medications within the (e.g. cardiologists, dermatologists, gynecolo-
VHA are prescribed by VA clinicians and dis- gists, orthopedists) in France, Germany, and
pensed by VA pharmacies. Dual-care Veterans the UK. To preserve confidentiality, patients
with Medicare coverage (which covers virtu- who see both GPs and specialists have differ-
ally all US citizens aged 65 years and older) ent database identity codes, making it chal-
may receive medications through both the VA lenging to track patients across settings of
and the Medicare Part D Plan. Because the care.
VHA is not a closed medical system, Veterans Italy’s Health Search Longitudinal Patient
may receive out-of-network care, limiting one’s Database (LPD Italy from IQVIA) is the coun-
ability to study certain outcomes (See try’s largest EHR database. The Caserta Local
“Incompleteness of Clinical Data” below). Health Unit (LHU) database contains EHR
data from approximately 60% of inhabitants in
a province of southern Italy and represents the
Overview of Databases
only Italian database that systematically links
Europe: The UK was the setting of the first EHR and administrative claims data, including
European EHR database, now called the drug dispensing and hospital discharge data.
Clinical Practice Research Datalink (CPRD) Caserta LHU data may also be linked to com-
(formerly General Practice Research Database prehensive, multi-dimensional geriatric
[GPRD]), a research service of the UK assessments for almost three-quarters of the
Government that was established in 1987 for local elderly population, making it valuable for
conducting public health research. CPRD geriatric research. Another Italian database,
182 10 Electronic Health Record Databases
Pedianet, contains data on over 400 000 chil- Statistics [HES] data linked with CPRD and
dren throughout Italy. IMRD); and the International Classification of
In Spain, Base de Datos para la Investigación Primary Care (ICPC) (IPCI, most autonomous
Farmacoepidemiológica en Atención Primaria regions in BIFAP) (see Table 10.1).
(BIFAP) contains primary care data from 9 of The European EHR databases also vary in
17 autonomous communities representing how they record drug data. CPRD and IMRD
nearly 20% of the Spanish population. Sistema employ British National Formulary (BNF)
de Información para el Desarrollo de la codes through the Gemscript system. Non-UK
Investigación en Atención Primaria (SIDIAP) European databases record medications using
contains data from 85% of the population of a Anatomical Therapeutic Chemical (ATC) clas-
single autonomous community, Catalonia. sification codes; most countries also have
GPs across the Netherlands contribute data to national drug codes. All European EHR data-
Integrated Primary Care Information (IPCI, pre- bases discussed contain data on prescribed
viously Interdisciplinary Processing of Clinical medications. Caserta LHU, BIFAP, Pedianet,
Information), which contains data from approx- and SIDIAP also contain drug dispensing data.
imately 14% of the Dutch population. Additionally, Caserta LHU, BIFAP, Pedianet,
US/VA: The VA database contains demo- and (for roughly half of drugs) DA databases
graphic, clinical, and administrative data since specify drug indications.
1997 along with prescription data since 1999 and All the above databases are generally repre-
laboratory data starting in 2000. In 2017, the num- sentative of their underlying populations in
ber of Veterans in the VA database was about 2% terms of age/sex distribution and prevalence of
of the US population and 32% of US Veterans. most diseases and drugs used. However, cer-
tain diseases may vary in frequency across
databases depending on local or disease-
Data Collection and Structure
specific patterns of clinical care. BIFAP, CPRD,
Europe: Primary care practitioners use EHRs DA France and Germany, IPCI, Pedianet, LPD
to document clinical information about their Italy, and IMRD include most regions of their
patients, which can then be electronically respective countries.
extracted for research purposes, examined for Hospitalizations, referrals, and the resulting
completeness and accuracy, and anonymized. consultation letters are recorded to varying
Some EHR databases receive frequent data degrees in European EHR databases (see
updates (e.g. Pedianet, CPRD, IMRD) while Table 10.2). Data from social history, including
others receive data updates just one to two smoking and alcohol usage, is available to var-
times per year (e.g. BIFAP, IPCI, LPD Italy). ying extents in most EHR databases See “Data
Primary care EHR databases generally contain Quality: Accuracy and Completeness” below
a minimum common set of patient informa- and Table 10.2). Substance exposure informa-
tion, including demographics, medical diagno- tion is less consistently recorded in DA data-
ses, and drug prescriptions, but they differ bases and IPCI. Pedianet contains information
from one another in types of variables and data on parental smoking habits.
included (see Table 10.2). In most European EHR databases, data are
European EHR databases record diagnoses more commonly recorded using structured
using a variety of standardized coding systems: (coded) fields rather than unstructured free
Read codes (IMRD, CPRD); International text. BIFAP and IPCI also contain large vol-
Classification of Diseases, 9th edition (ICD-9) umes of unstructured data. These and certain
(Caserta LHU Caserta, one autonomous region other databases (e.g. Pedianet, SIDIAP) make
in BIFAP, LPD Italy, Pedianet); ICD-10 (IQVIA information from anonymized free text entries
DA, SIDIAP, Pedianet, Hospital Episode available to researchers for outcome validation
Descriptio 183
and supplemental data extraction. Several Data Quality: Accuracy

databases (e.g. BIFAP, CPRD, DA Germany, and Completeness
Pedianet, SIDIAP, IMRD) also allow research-
Europe: Data quality checks are performed
ers to administer questionnaires directly to cli-
periodically by European EHR databases on
nicians or patients for a fee.
three levels: (i) practitioner recording, (ii) data
US/VA: Local VA medical centers record
extraction; and (iii) database maintenance.
outpatient and inpatient clinical and adminis-
In the UK, national quality improvement
trative data within the Veterans Health
initiatives and advances in software have
Information Systems Technology and
increased overall capture and accuracy of data.
Architecture (VistA) system. Data from VistA
For example, the UK Quality and Outcomes
are available for research within the Corporate
Framework (QOF) used financial incentives to
Data Warehouse (CDW), which contains a
improve documentation for >100 quality indi-
wide range of data elements from outpatient
cators for 10 chronic diseases. UK GPs contrib-
and inpatient settings, including demograph-
uting to IMRD or CPRD receive software
ics, diagnoses (ICD-9 and ICD-10 codes), vital
training and regular evaluation of their data
signs, laboratory and radiology results, surgical
recording and prescribing behavior, including
procedures, free text notes, consults, and social
feedback reports with tips on improving
history (e.g. smoking). The VA Vital Status File
recording. Other database-directed quality
contains death data from multiple sources that
measures include audits of newly added prac-
are regularly cross-checked with the Social
tices and comparison of data to national data-
Security Administration Death Master File. In
bases (e.g. mortality, hospitalizations, cancer
addition, VA has several disease-specific regis-
and cardiovascular registries).
tries used for patient care and research (e.g.
CPRD only provides data from practices
cancer, diabetes, severe mental illness, amyo-
meeting quality standards (~90% of practices).
trophic lateral sclerosis, rheumatoid arthritis).
The Up-to-Standard date is a practice-based
The VA database contains information on
quality marker based on death rates and gaps in
both prescribing and dispensing of drugs in
recording, corresponding to when a practice in
both outpatient and inpatient settings. Drug
CPRD is considered to have continuous, com-
data may be found within several databases:
plete recording of data. For IMRD, IQVIA
the CDW; Pharmacy Benefit Management
employs a practice-specific quality measure
(PBM) database, which contains information
known as Acceptable Mortality Reporting
on nonprescription medication dispensing and
(AMR), denoting the first year that mortality
medical supplies; and the Bar Code Medication
reporting was deemed complete. Other
Administration (BCMA) database, which con-
European EHR databases have their own stand-
tains records of inpatient administrations of
ards for ensuring quality and completeness.
medications to patients. The VHA also main-
With regards to specific variables, data com-
tains its own adverse drug and vaccine event
pleteness varies among databases (See
reporting system (see Chapter 7).
“Incompleteness of Clinical Data” and Table
Investigators may also extract data for
10.2). Pregnancy, family structure, mortality,
research directly from the EHR through man-
and cause of death are heterogeneously
ual chart review or natural language process-
recorded and may be difficult to ascertain.
ing. Primary data collection from unstructured
CPRD offers a probabilistic mother–baby link
data fields, such as reports or free text notes,
algorithm and is linkable to mortality records
can facilitate outcome ascertainment or valida-
from the Office for National Statistics to
tion. In addition to EHR data, surveys of
improve death estimates and confirm cause of
Veterans or clinicians permit access to addi-
death. In IMRD, researchers may also use data
tional information.
algorithms to link family members or deter- ideal for nested case–control and cohort stud-
mine cause-specific mortality. ies. EHRs generally have a longer follow-up
US/VA: The CDW, a non-transformed mirror period compared to claims data. These charac-
of the medical record, is updated nightly. teristics, along with the large populations cov-
Updates of the Vital Status File occur monthly. ered by EHRs, make these data sources
The accuracy and completeness of data reflect particularly ideal to study rare diseases and
the EHR and beneficiary claims they source. rare outcomes.
Earlier years of demographic race and ethnic-
ity data can have up to 20% missing data.
Validity of Clinical Information
The PBM prescription database undergoes
daily quality assurance processes to ensure Epidemiologic studies in EHR databases
completeness and accuracy. As with medica- involve use of lists of codes, and sometimes
tions obtained through Medicaid (see coding algorithms, for identifying specific
Chapter 9), low or nil co-payments produce medical conditions, drugs/other exposures of
strong financial incentives for Veterans to interest, and covariates. The validity of such
obtain outpatient prescriptions through the VA. code lists and algorithms has often been stud-
ied in many of these databases. If a proposed
Data Access for Researchers outcome has not previously been validated,
researchers should strongly consider validat-
Europe: Research performed using European ing that outcome to ensure that diagnostic
EHR databases must generally first be reviewed codes or algorithms reflect patients’ true
by the local institutional review board (IRB) conditions.
and the respective database’s ethics board.
Given researchers’ inability to identify individ-
ual patients in anonymized databases, most Accuracy of Drug Information
studies meet criteria for IRB exemption. EHR databases contain information on name,
Several European databases are available for strength, and quantity of prescribed drugs,
licensing by investigators in government, aca- which can be used to estimate prescription
demia, and industry, while access to certain coverage. In the UK, unlike in other countries,
European databases (Caserta LHU, IPCI, LPD the prescription is the payment document.
Italy, Pedianet) requires collaboration with Refills can be accurately identified from pre-
affiliated researchers. scriptions in IMRD. The concordance between
US/VA: To use VA data, researchers must issued prescriptions and dispensed drugs in
first receive approval by the local or central VA IMRD is generally high, with some exceptions.
IRB. Access to VA data is limited to VA- Caserta LHU, BIFAP, Pedianet, SIDIAP, and
affiliated researchers and their collaborators. VA data contain information on outpatient
drug prescribing and dispensing. VA data also
capture inpatient drug dispensing and
Strengths
administration.
Population-Based Data, Sample
Size, and Length of Follow-up Ability to Access Original Health
Records
Population-based data from European EHRs
draw subjects from the general population, Some EHRs provide access to original
minimizing selection bias and improving anonymized healthcare records, such as free
validity and generalizability of pharmacoepi- text data (e.g. IMRD, IPCI), anonymized copies
demiologic studies. These data sources are of paper records (e.g. IMRD), or the entire
Limitation 185
EHR (e.g. VA), as well as access to clinicians or record laboratory or radiologic findings that are
patients via surveys (Table 10.2). Such access abnormal, while normal and some abnormal
can be useful for validation purposes. results may not be documented reliably.
Because European EHR databases are
designed to capture health information from
Linkage to External Patient-Level
primary care settings, they typically lack
Data
information from specialists. Researchers
Many EHR databases may be linked to other using IMRD or CPRD may access more exten-
health-related, patient-level information, thus sive and reliable data from other settings
extending the utility of EHR data. The data through linkage with HES data and other
source most commonly linked to CPRD and sources. Similarly, as mentioned above, EHR
IMRD is HES, which contain data on hospitali- data from Caserta LHU can be linked with
zations, accident and emergency episodes, and claims data.
specialized testing, including imaging. The The nature of illness can affect the pattern of
combination of data from primary care and data recording. Unlike in claims databases,
HES facilitates research on conditions man- codes for chronic diseases may be entered only
aged across multiple healthcare settings. once into some EHR databases. For this rea-
Linkage to official death records may improve son, episodes of care involving acute events
the accuracy of mortality studies and validate may be better recorded than chronic diseases.
mortality data from general practice. While many EHR databases capture infor-
Researchers can link EHR data with other data mation about race and ethnicity, smoking and
sources, including disease (e.g. cancer and alcohol use, BMI, socioeconomic status,
COVID-19) registries, mental health datasets, employment status, and occupation, these
and socioeconomic and deprivation indices. fields may be missing for many individuals
EHR data may also be linked to individual (see Table 10.2). UK EHR databases may miss
patient-generated data, including patient- many pediatric growth measurements that are
reported outcomes, environmental data, drug instead recorded locally on paper. Recording of
diaries, and biospecimens. EHR databases out- pediatric growth measurements is more com-
side the UK also permit linkages to other data prehensive in Pedianet.
sources: Caserta LHU EHR to claims data; VA Veterans in the VHA may receive health care
data to Medicare and Medicaid claims data and outside the VHA either by choice (especially
to genetic information from the Million older Veterans with Medicare coverage) or by
Veteran Program biobank. necessity (e.g. emergent care). As a result,
occurrences of acute conditions may be gener-
ally missed in inpatient data from VA hospi-
tals, potentially resulting in missing outcome
Limitations
data. The frequent omission of acute inpatient
outcomes is a major limitation of the VA data-
Incompleteness of Clinical Data
base; among veterans aged 65 and older, this
The accuracy and completeness of EHR data limitation can be overcome by linking VA data
rely heavily on the quality of the recording to Medicare claims data.
from physicians during routine care. Human
errors as well as systematic recording errors
Incompleteness of Drug Data
may occur, with the latter leading to bias. For
example, geriatric data in EHRs are likely to be Information on medication days’ supply and
selectively recorded for frailer patients. Another daily dosage may not be explicitly recorded in
example is that clinicians may be more likely to EHR data but can be imputed. Additionally,
algorithms have been developed to determine advances enable clinicians and researchers
daily dosage and other drug data (e.g. fre- to have greater access to increasing volumes
quency) from unstructured text. Only a few of data. As EHR systems evolve, systems
databases (BIFAP, Caserta LHU, Pedianet, and administrators, clinical informatics special-
to some extent DA) specifically link prescribed ists, and end users must address important
drugs to the specific indication of use. Without challenges, including missing data and vari-
this information, one can refer to diagnoses able recording. To optimize clinical care and
recorded in or around encounters that corre- facilitate high-quality research, health care
spond to prescribed drugs. systems must implement approaches to
Prescribing records do not indicate whether ensure consistent and complete clinical doc-
prescriptions were filled. Only BIFAP, Caserta umentation within the EHR. The vital need
LHU, SIDIAP, and the VA also contain drug to maintain individuals’ privacy and confi-
dispensing data. Data on over-the-counter dentiality must be balanced with the poten-
(OTC) drugs are frequently missing from EHR tial societal benefits of enhanced linkage
databases, but exceptions exist where health and sharing of data across disparate plat-
care systems pay for OTC drugs (e.g. long-term forms and data sources.
use of aspirin and nonsteroidal anti- The many advancements in EHR systems
inflammatory drugs in UK databases) and have important implications for the conduct of
where OTC drugs are prescribed (e.g. Caserta research. Technologic advances, such as natu-
LHU, Pedianet, VA). SIDIAP captures pre- ral language processing and machine learning,
scribing and dispensing of drugs irrespective enable researches to use complex EHR data in
of coverage by the national health system, novel ways. Large international research net-
leading to comprehensive recording of OTC works, such as TEDDY (Teddy European
medications. Network of Excellence for Paediatric Research)
In European EHR databases, data on medi- and OHDSI (Observational Health Data
cations restricted to specialist care, dispensed Sciences and Informatics), have demonstrated
from hospital pharmacies (e.g. biologics), and the capacity and power of international col-
given during hospitalization or upon hospital laborations to use EHR databases for large-
discharge may be missing. In the UK, patients scale research on drug use and outcomes,
generally receive a limited quantity of medica- including during the COVID-19 pandemic.
tions upon hospital discharge. In the VA, cer- Such collaborations lead to more generalizable
tain medications are recorded in the EHR but research with increased statistical power to
not the prescription databases, e.g. medica- study rare diseases, uncommon drugs, and rare
tions obtained from floor stock or adminis- outcomes. Linkage of EHRs with hospital data,
tered acutely in emergencies. Other drug administrative claims, and other data sources
administrations may occasionally be incom- (e.g. patient registries) helps maximize the
plete in VA prescription databases. advantages of each data source and minimize
their respective limitations. Furthermore, link-
age of EHR data to patient-generated data and
The Future biospecimens enhances discovery through
population-representative, patient-centered
Through interoperable platforms, health research and molecular pharmacoepidemiol-
information exchanges, patient portals with ogy. Researchers can also use EHR systems to
patient-generated data, and other techno- conduct large pragmatic trials. In addition,
logic advances, EHR databases continue to expansion of EHRs in low-and middle-income
evolve and expand in pursuit of delivering countries facilitates research in areas of
high-quality, high-value health care. Such great need.
Acknowledgmen 187
High-quality research conducted within (e.g. laboratory results, socioeconomic data,

EHR databases can favorably influence public dispensed drugs, mortality, free text); ability to
policy and public health. EHR databases also link to other data sources (e.g. claims data,
have an important regulatory role in post- inpatient data); and modes of access.
market evaluation, pharmacovigilance, and ●● With long durations of follow-up and often
risk minimization. With rapid, nearly real- large populations, EHR databases are suita-
time analyses of recent population-based ble settings for studying a wide variety of
data, EHR databases are useful settings to medical conditions, including rare diseases
conduct post-authorization safety studies as and rare outcomes, using numerous study
well as research addressing emergent health designs. Validation of key outcomes
crises, such as COVID-19. Through high- improves study validity.
quality clinical and translational research and ●● Incomplete information about certain types
pharmacovigilance, EHR systems of the of data (e.g. health-related behaviors, data
future will continue to serve as key platforms from specialists or hospitals) can lead to bias
for answering important questions and or other study limitations when using EHR
improving the health of patients, communi- databases. With some databases, investiga-
ties, and populations. tors may obtain additional information by
linking to other datasets (e.g. secondary
care, disease registries), reviewing free text,
ummary Points
S or sending questionnaires to physicians or
for Electronic Health Record patients.
Databases ●● Research using EHR databases can contribute
meaningfully to efforts related to public
●● European EHRs databases contain health, public policy, and pharmacovigilance.
anonymized, population-based data col-
lected in the context of patient care (e.g.
diagnoses, prescribed drugs, health-related Acknowledgment
behaviors, vital signs, referrals) predomi-
nantly from outpatient settings by primary The authors wish to thank Alexis Ogdie for her
care practitioners. contributions along with the following indi-
●● The United States Veterans Affairs databases viduals who provided information about spe-
contain a wide range of clinical data (e.g. cific databases: BIFAP: Miguel Gil García,
diagnoses, prescribed and dispensed drugs, Miguel Angel Maciá Martínez, Dolores
health-related behaviors, vital signs, labora- Montero Corominas; Caserta LHU: Valentina
tory data, mortality, free-text data) on US Ientile and Michele Tari; CPRD: Lucy Carty;
Veterans from outpatient and inpatient LPD Italy: Alessandro Pasqua, Iacopo Cricelli,
settings. Francesco Lapi; IPCI: Marcel de Wilde;
●● EHR databases vary widely in characteristics, Pedianet: Carlo Giaquinto, Luigi Cantarutti,
including data structure and coding systems; Anna Cantarutti; SIDIAP: Bonaventura
location, size, and proportion of population Bolíbar, Eduardo Hermosilla Pérez, Maria del
covered; population of focus (e.g. general, Mar García Gil, Talita Duarte Salles; VA: Lucy
children, Veterans); types of data included Pandey, Kwan Hur.
Case Example 10.1 Electronic Health Record Database in Spain (see de Abajo et al.,
Lancet, 2020)
Background (2019–2020 BIFAP data were not

●● Angiotensin-converting enzyme 2 (ACE2) available).
is the molecular receptor used by severe ●● Exposure was defined as a prescription for
acute respiratory syndrome coronavirus 2 RAAS inhibitors lasting until one month
(SARS- CoV-
2) to enter cells and cause before the index date (current use), com-
infection. pared with current use of other antihyper-
●● Renin–angiotensin–aldosterone system tensives (e.g. calcium channel blockers,
(RAAS) inhibitors (e.g. ACE inhibitors, beta blockers, diuretics).
angiotensin- receptor blockers [ARBs]) ●● Potential confounders were a history of
increase ACE2 expression in some animal diabetes, hyperlipidemia (defined as use of
models and may reduce angiotensin II- lipid-lowering drugs), atrial fibrillation,
associated lung injury, raising questions heart failure, ischemic heart disease,
about the role of RAAS inhibitors in thromboembolic disease, cerebrovascular
increasing or decreasing the risk of SARS- accident, asthma, cancer, chronic kidney
CoV-2 infection and severity of coronavirus disease, or chronic obstructive pulmonary
disease-2019 (COVID-19). disease; underlying cardiovascular disease
●● Early in the COVID- 19 pandemic, associa- (composite of comorbidities); cardiovascu-
tions between certain comorbidities (e.g. lar risk factors (composite of comorbidities
diabetes, hypertension) and severe including hypertension).
COVID-19 led to questions about the ●● Associations were estimated using multi-
safety of continuing RAAS inhibitors. variable conditional logistic regression,
adjusted for confounders. Potential effect
Issue modification by age, sex, and comorbidity
Using BIFAP and another local EHR database (e.g. hypertension, diabetes) was exam-
(HORUS), a population-based study was con- ined. Models were also stratified by
ducted to evaluate the association between COVID-19 severity.
use of RAAS inhibitors and severe ●● Secondary analyses considered associa-
COVID-19 in Madrid, Spain. tions for individual drug classes (e.g. ACE
inhibitors versus calcium channel
Approach blockers).
●● A case-
population design was used. ●● Sensitivity analyses considered impact of
●● Cases were adults consecutively admitted secular trends of RAAS use (given the two-
in March 2020 to hospitals in Madrid with year gap between case and control data)
a diagnosis of COVID- 19, stratified by and media alerts about the safety of RAAS
severity; patients with severe COVID- 19 inhibitors, among others.
required intensive care unit (ICU) admis-
sion or died. Results
●● Each case was matched by age, sex, and ●●Data were collected from 1139 cases and
index date (day and month of hospital 11 390 matched controls. Comorbidities
admission) to 10 random controls in and use of antihypertensives were more
Madrid using 2018 data from BIFAP prevalent among cases than controls.
●● Current use of RAAS inhibitors was 43.6% association between use of RAAS inhibi-
in cases and 33.6% in controls, compared tors and hospitalized/severe COVID-19.
with current use of other antihyperten-
sives in 13.6% of cases and 9.9% of
Limitations
controls.
●● Cases and controls were drawn from dif-
●● Adjusting for covariates, the odds ratio for
ferent databases (albeit from the same
the risk of COVID-19 requiring hospitaliza-
underlying population and source data)
tion among users of RAAS inhibitors was
and different years, but sensitivity analyses
0.94 (95% CI: 0.77, 1.15).
considering the influence of secular trends
●● Most secondary analyses based on drug
in use of RAAS inhibitors were similar.
class (ACE inhibitors, ARBs) and COVID-19
●● No adjustment for smoking, other lifestyle
severity were similar; larger associations
habits, or other potential unmeasured
were seen with aldosterone antagonists
confounders.
(adjusted OR 1.68, 95% CI 0.97, 2.91) and
●● Analyses considered drug prescription
short-term use of RAAS inhibitor (adjusted
data, not accounting for consumption or
OR 1.39, 95% CI 0.92, 2.10).
adherence.
●● Stratified analyses suggested potential
●● No consideration of dose effects of
effect modification by diabetes, with
medications.
adjusted OR of 0.53 (95% CI 0.34, 0.80) in
●● Certain secondary analyses suggested pos-
patients diagnosed with diabetes.
sible elevated risk with certain patterns of
●● Sensitivity analyses yielded similar
RAAS inhibitor use, of unclear significance.
results.
Strengths Key Points

●● Population-based data, active-comparator ●● Current use of RAAS inhibitors was not
design, and covariate adjustment reduced appreciably associated with altered risk of
bias from confounding and enhanced the hospitalized or severe COVID-19.
generalizability of the findings. ●● These findings did not support the prac-
●● Most secondary and sensitivity analyses tice of either stopping or starting RAAS
were consistent with the findings of a null inhibitors during the COVID-19 pandemic.
Further Readings
de Abajo, F.J., Rodríguez-Martín, S., Lerma, V. et al. pharmacoeconomic studies. Int. J. Clin. Pract.
(2020). Use of renin-angiotensin-aldosterone 47 (10): 617–626.
system inhibitors and risk of COVID-19 Cai, B., Xu, W., Bortnichak, E., and Watson, D.J.
requiring admission to hospital: a case- (2012). An algorithm to identify medical
population study. Lancet 395 (10238): practices common to both the general practice
1705–1714. research database and the health
Becher, H., Kostev, K., and Schroder-Bernhardi, improvement network database.
D. (2009). Validity and representativeness of Pharmacoepidemiol. Drug Saf. 21 (7): 770–774.
the disease analyzer patient database for use Carbonari, D.M., Saine, M.E., Newcomb, C.W.
in pharmacoepidemiological and et al. (2015). Use of demographic and
pharmacy data to identify patients included pharmacoepidemiology research.

within both the Clinical Practice Research Pharmacoepidemiol. Drug Saf. 16: 393–401.
Datalink (CPRD) and The Health Improvement Lewis, J., Bilker, W., Weinstein, R., and Strom, B.
Network (THIN). Pharmacoepidemiol. Drug (2005). The relationship between time
Saf. 24 (9): 999–1003. since registration and measured incidence
Dave, S. and Peterson, I. (2009). Creating rates in the general practice research
medical and drug code lists to identify cases in database. Pharmacoepidemiol. Drug Saf. 14:
primary care databases. Pharmacoepidemiol. 443–451.
Drug Saf. 18: 704–707. Lum, K.J., Newcomb, C.W., Roy, J.A. et al.
Evans, R.S. (2016). Electronic health records: (2017). Evaluation of methods to estimate
then, now, and in the future. Yearb. Med. missing days’ supply within pharmacy data of
Inform. 1: S48–S61. the Clinical Practice Research Datalink
Garvin, J.H., Kalsy, M., Brandt, C. et al. (2017). (CPRD) and The Health Improvement
An evolving ecosystem for natural language Network (THIN). Eur. J. Clin. Pharmacol. 73
processing in Department of Veterans Affairs. (1): 115–123.
J. Med. Syst. 41 (2): 32. Maguire, A., Blak, B., and Thompson, M.
Gellad, W.F. (2016). The veterans choice act and (2009). The importance of defining periods
dual health system use. J. Gen. Intern. Med. 31 of complete mortality reporting for
(2): 153–154. research using automated data from
Hall, G.C., Sauer, B., Bourke, A. et al. (2012). primary care. Pharmacoepidemiol. Drug
Guidelines for good database selection Saf. 18: 76–83.
and use in pharmacoepidemiology Marston, L., Carpenter, J.R., Walters, K.R. et al.
research. Pharmacoepidemiol. Drug Saf. 21 (1): (2014). Smoker, ex-smoker or non-smoker?
1–10. The validity of routinely recorded smoking
Haynes, K., Bilker, W.B., Tenhave, T.R. et al. status in UK primary care: a cross-sectional
(2011). Temporal and within practice study. BMJ Open 4 (4): e004958.
variability in The Health Improvement Oyinlola, J.O., Campbell, J., and Kousoulis, A.A.
Network. Pharmacoepidemiol. Drug Saf. 20 (2016). Is real world evidence influencing
(9): 948–955. practice? A systematic review of CPRD
Herrett, E., Thomas, S., Schoonen, S. et al. (2010). research in NICE guidances. BMC Health Serv.
Validation and validity of diagnoses in the Res. 16: 299.
general practice research database: a systematic Robb, M.A., Racoosin, J.A., Worrall, C. et al.
review. Br. J. Clin. Pharmacol. 69 (1): 4–14. (2012). Active surveillance of postmarket
Herrett, E., Gallagher, A.M., Bhaskaran, K. et al. medical product safety in the Federal
(2015). Data resource profile: Clinical Practice Partners’ collaboration. Med. Care 50 (11):
Research Datalink (CPRD). Int. J. Epidemiol. 948–953.
44 (3): 827–836. Smith, M.W. and Joseph, G.J. (2003). Pharmacy
Khan, N., Perera, R., Harper, S., and Rose, P. data in the VA health care system. Med. Care
(2010). Adaptation and validation of the Res. Rev. 60 (3 Suppl): 92S–123S.
Charlson index for read/OXMIS coded Sohn, M., Arnold, N., Maynard, C., and
databases. BMC Family Prac. 5 (11): 1. Hynes, D. (2006). Accuracy and completeness
Lester, H. (2008). The UK quality and outcomes of mortality data in the Department of
framework. BMJ 337: a2095. Veterans Affairs. Popul. Health Metrics 4:
Lewis, J.D., Schinnar, R., Bilker, W.B. et al. 2–8.
(2007). Validation studies of The Health Trifiro, G., Sultana, J., and Bate, A. (2018). From
Improvement Network (THIN) database for big data to smart data for pharmacovigilance:
the role of healthcare databases and evaluations of two exemplar trials. Health
other emerging sources. Drug Saf. 41 (2): Technol. Assess. 18 (43):
143–149. 1–146.
van Staa, T.P., Dyson, L., McCann, G. et al. Wolf, A., Dedman, D., Campbell, J. et al. (2019).
(2014). The opportunities and challenges of Data resource profile: Clinical Practice
pragmatic point-of-care randomised trials Research Datalink (CPRD) aurum. Int. J.
using routinely collected electronic records: Epidemiol. 48 (6): 1740–1740g.
192
11
Primary Data Collection for Pharmacoepidemiology

Priscilla Velentgas
Real World Solutions, IQVIA, Inc., Cambridge, MA, USA
Introduction “hybrid or enriched” studies may combine

existing data with primary data collection for
Primary data collection refers to data that are critical aspects of the patient or health care
collected specifically for a given research study provider (HCP) experience (see “Hybrid or
or program, while secondary data are data that enriched designs” below).
were collected to meet needs other than the Research questions that may require pri-
research for which they are being used. mary data include the following:
Primary data collection can be used in all types Designs involving randomization – pragmatic
of pharmacoepidemiologic study designs, both as well as explanatory randomized clinical trial
interventional and non-interventional, cohort, designs (see Chapter 17) necessitate at least
and case–control studies, as well as patient minimal site and/or patient contact for screen-
registries. Registries are conceptually a data ing to determine eligibility and consent to par-
collection structure for disease or product ticipate in the trial. While randomized studies
related studies, and in practice may closely frequently involve extensive primary data col-
resemble observational cohort study designs. lection to meet the trial objectives, these con-
Registry-based studies generally address effec- siderations overlap with those of observational
tiveness and safety of various medical treat- designs and will be addressed further in that
ments, and also are used to characterize context.
diseases including progression over time. Outcome assessment/adjudication – studies
may require collection of detailed primary data
for outcome assessment and/or for outcome
Research Questions that Require
adjudication. Collection of primary data for
Primary Data
outcome assessment becomes important when
The nature of the research question(s) and study outcomes are not consistently recorded
accompanying study design determine the in available health care data, or are not
need for collection of primary data from clini- recorded with the reliability, timing or fre-
cians, patients, and/or others to address a quency needed to meet study aims. Additional
study’s aims, and are generally weighed against primary data may also be collected to validate
the availability of the required information in or confirm outcomes collected through sec-
existing data sources (see Chapter 12). Some ondary data or patient self-report. Outcome
Introductio 193
validation can also be important for safety Commonly, prospective studies including
studies that use the patient as the primary planned analysis of laboratory data or imaging
reporter of potential adverse events (AEs), and studies over time may incorporate primary
clinical validation may be needed to confirm data collection to ensure complete collection
the endpoint of interest. For example, in the of assessments and that timing of assessments
European PROTECT (Pharmacoepidemio is aligned with the study follow-up period.
logical Research on Outcomes of Therapeutics) Additionally, use of a central lab to reduce vari-
Consortium, funded by the Innovative ability in laboratory measures may be consid-
Medicines Initiative, data were collected ered to further increase the validity of study
directly from pregnant women recruited on- results. The PROVALID study, (PROspective
line from the UK, Denmark, The Netherlands cohort study in patients with type 2 diabetes
and Poland. Researchers learned that women [T2D] mellitus for VALIDation of biomarkers),
could accurately report serious birth defects, launched in 5 EU countries (Austria, Hungary,
but there were many reports of potential Netherlands, Poland and Scotland), will obtain
abnormalities that were difficult to classify laboratory measurements on 4000 enrolled
without more clinical information. patients with type-2 diabetes treated in the pri-
Some observational as well as interventional mary care setting to examine the impact of
studies, especially those designed to meet medication and predict the clinical course of
post-marketing requirements with safety or disease, including renal and cardiovascular
effectiveness outcomes that require an addi- events. Sites that participate in PROVALID are
tional level of rigor, may include a full or mod- required to collect a minimum set of clinical
ified approach to clinical outcome review or data parameters, with the option to collect
adjudication by a central committee over and many additional laboratory and medical his-
above the reporting by individual study sites. tory characteristics on enrolled patients.
Reasons supporting the need for additional Characterization of patient-reported out-
outcome adjudication include concerns comes not captured in secondary data – It is
regarding investigator bias, if they hold strong often important to evaluate the burden of dis-
opinions as to the benefit or harm associated ease on patients and how that burden is
with a treatment under study, the need to affected by various treatments. Burden can
apply consistent standard definitions given include measures of disease related or general
variability in diagnostic criteria in usual prac- quality of life, ability to complete activities or
tice, and lack of detail or inconsistent use of attend work, assessment of pain or symptoms
standard coding practices in accessible sec- related to the underlying condition. Additional
ondary data sources. detail is provided in the section on “Patient-
Clinical assessments not consistently captured Generated Data” below.
in secondary data – Even as the collection of Studies of rare populations – when necessary
health care data from routine care is increas- to assemble as large and representative a sam-
ingly recorded and available from electronic as ple as possible from a rare population, one or
well as paper medical records, substantial vari- more existing data sources may not capture
ability in performing assessments on the part enough of the patient population of interest to
of HCPs according to individual and local address study aims. Registries of rare disease
practice, magnified by patients’ variability in and of pregnancy exposures commonly face
coming in for recommended routine visits, the challenge of a small number of patients
limits the extent to which clinical data from with the condition or exposure of interest, dis-
secondary sources can be used to address some tributed over many countries, for which no
research questions. See section on “Clinician single existing data source likely includes
or site-reported outcomes” below. enough patients to address research aims.
194 11 Primary Data Collection for Pharmacoepidemiology
Registries of rare diseases have been utilized Committee included the requirement to collect
in a number of ways to support the assessment data that would support the identification of
of potential treatments. These include provid- any significant change in frequency or severity
ing information on the natural history of dis- of reactogenicity in comparison with previous
ease to inform the design of studies and trials, years’ experience with the same vaccine com-
serving as historical comparators for single position. Reactogenicity AEs of interest
arm trials, and for identification of patients for include vaccination site reactions, headache,
studies of treatment effectiveness, including malaise, myalgia, shivering, rash, vomiting,
registry-based randomized trials. The validity nausea, arthralgia, decreased appetite, irrita-
of conclusions that may be drawn from analy- bility and crying (in pediatric vaccinees less
ses of registry data depend on the degree to than five years of age). Such symptoms, espe-
which subjects in the registry are not selec- cially when mild or moderate in severity, are
tively included because of their treatments and not commonly reported to HCPs and thus
outcomes, and the degree to which follow up medical records are not a useful source of data
and capture of necessary clinical data is suffi- to study these outcomes following vaccination.
ciently complete. Instead, studies and surveillance activities to
Pregnancy exposure registries are post- implement the requirements for enhanced
approval studies of the safety of drugs or vac- safety surveillance have incorporated patient
cines, which women may use during (and adult proxy for pediatric patients) self-
pregnancy. Depending on the product and report of occurrence of symptoms to obtain
indication, the number of exposed pregnant this information in a systematic manner, for
women may be very small and geographically example through distribution of Safety Report
dispersed. Recent draft FDA guidance for preg- Cards allowing patients to report these symp-
nancy registries states that “. . .pregnancy reg- toms by telephone or mail. Enrollment of vac-
istries remain an important tool for safety data cinated subjects at the point of vaccination
collection in the post-marketing setting allows for capture of the specific vaccine brand
because of the prospective design and the abil- and batch or lot number, either through inves-
ity to collect detailed patient level data.” Such tigators’ knowledge that only specific vaccines
data include detail of timing of exposures and are being distributed at the site, or through
clinical detail around pregnancy and offspring direct collection of this information from staff
outcomes, with requirements determined by administering the vaccines.
the research question and outcomes of interest Studies of medical devices – As with vac-
for each registry. cines, device studies may require information
Vaccine safety – When conducting studies on batch and manufacture location that are
of vaccine safety, detailed information on vac- not available in existing data sources (see
cine brand and formulation as well as batch or “Epidemiologic Studies of Implantable
lot numbers may not be readily available in Medical Devices” in Chapter 23). Additional
secondary data, and some AEs of interest may information on the “operator” or HCP implant-
not be routinely captured in existing data ing or administering the device may also help
sources (see “Special Methodological Issues in to provide a full characterization of product
Pharmacoepidemiologic Studies of Vaccine safety and effectiveness.
Safety” in Chapter 23). For example, data collected through the
Beginning in 2014, the European Medicines National Cardiovascular Data Registry (NCDR)
Agency (EMA) has required annual has been utilized for studies that compares the
enhanced safety surveillance for all seasonal effectiveness of different types of devices on
influenza vaccines. The interim guidance from cardiovascular outcomes, as well as allowing
the Pharmacovigilance Risk Assessment for examination of the role of manufacturing
Methods of Primary Data Collectio 195
site in device failure. In a 2017 report compar- an enriched study. The study objective was to
ing safety of the Mynx vascular closure device characterize and describe the global variation
to other similar devices using the NCDR in management of patients initiating second
CathPCI Registry, a safety signal was observed line therapy for T2D; data were collected from
in both the full identified analysis population patients at sites in 38 countries, with linkage of
and a subset of centers with greater experience electronic health records where feasible.
with the device.
Special requirements, controlled distribution
products – some products are approved by reg- ethods of Primary Data
M
ulatory authorities with special requirements Collection
for safety reporting. Often such requirements
necessitate data collection specific to the prod- Site-Based Data Collection
uct and the safety concern to ensure robust
monitoring of these concerns. Often primary data collection for pharmacoep-
Several active mandated safety registries and idemiologic studies begins with the identifica-
an example of a multi-sponsor pediatric safety tion of sites or HCPs who agree to participate
registry were described in a 2009 publication in the study and then for those HCPs to recruit
entitled, “Registries Rising: FDA Looking at patients following agreed upon inclusion and
TNF Inhibitors; AHRQ Updates Standards.” exclusion criteria as described in the protocol.
These examples included a pregnancy registry Traditionally, data collection was performed
for milnacipran; a pregnancy registry as part of on paper, but now most such data collection is
a restricted distribution program for eltrom- electronic through case report forms, which
bopag, a thrombopoeitin receptor agonist for are data collection forms designed specifically
treatment of idiopathic thrombocytopenia for the study purposes as part of an electronic
purpura; and a safety registry for teriparatide, data capture system. This model requires insti-
an anabolic treatment for osteoporosis and an tutional board review (ethical review) and site
expanded indication of glucocorticoid-induced contracts with each investigator. Generally,
osteoporosis. In a 2018 review of FDA post- investigators expect some payment for data
market requirements for new drugs and bio- collection, and such payments must be propor-
logics approved from 2009 to 2012, 97 out of tional to time spent.
110 new drugs or biologics had at least one
post-market requirement; of these nearly one- Clinician or Site-Reported
third were for prospective cohort studies, regis- Outcomes (ClinROs)
tries, or clinical trials.
Clinician-reported outcome measures, or clini-
cian or site-reported outcomes (ClinROs), are
Hybrid or Enriched Designs standardized, usually validated assessment
The terms “hybrid or enriched” are frequently tools, used to measure disease severity or pro-
used to describe study designs that draw upon gression, and often incorporated as endpoints
both primary and secondary data, with some for pharmacoepidemiologic studies. While
data collected de novo, specifically for the pur- some are recommended for use in routine clin-
poses of the study, and other study-specific ical practice, the consistency of uptake and
data collected via probabilistic or deterministic timing of administration is often not complete
linkage with other data sources, such as elec- enough to robustly address specific research
tronic health records, administrative claims questions of interest without additional
and billing data, vital records or existing regis- requirements for study-specific collection. For
tries. The DISCOVER study is an example of example, in rheumatoid arthritis, a number of
disease activity measures have been endorsed gov from 2000 to 2017, and showed a 34%
by the American College of Rheumatology, annual increase in use of these devices.
most requiring assessment by a physician of Continued quantification of the validity and
28 joints for tenderness and swelling. Although reliability of wearable sensors that collect
use of disease activity measures is strongly rec- information about physical activity and other
ommended to assess treatment effectiveness clinically useful data will encourage greater
and prevent or slow progression, regular use of use in longitudinal observational studies and
such measures in clinical practice remains low randomized trials.
because of barriers in implementing assess-
ments as part of the workflow in routine
Registries as Means of Data
visits.
Collection
As previously mentioned, registries may be
Patient-Generated Data
established to fill a need for data collection that
Patients are increasingly contacted directly to may support multiple research and/or public
provide a wide range of information for health surveillance objectives.
research studies including medical history, Population-based state, regional, and
exposures and outcomes, as well as being con- national cancer registries have played a major
tacted through devices that monitor their clini- role in cancer surveillance, by quantifying can-
cal status, behaviors such as medication cer incidence and mortality, and trends over
adherence, and physical activity. Studies may time throughout the world, and in pharma-
be designed more flexibly with options to col- coepidemiology, by providing data on prognos-
lect some data directly from patients either tic factors, treatment, and outcomes for
during an in-office study visit or electronically, analysis within single or across linked data-
outside of study visits. bases. In the United States, the Surveillance,
Multiple studies have established the valid- Epidemiology, and End Results (SEER) pro-
ity of patient-reported prescription medication gram of the National Cancer Institute works
use, laying a foundation for its reliability as from a network of 17 active cancer registries in
well as supporting the value of additional 14 states, that actively collect information on
information that can be obtained from patients all reported cancers diagnosed in their cover-
that is not typically available through second- age areas.
ary data, such as non-prescription medication Pharmacoepidemiologic studies using can-
use, recreational drug use, smoking and alco- cer registries have included case–control stud-
hol intake. There are a large number of vali- ies of hormonal contraception and breast,
dated patient reported outcome measures that ovarian, and endometrial cancer, and patterns
can provide important insights on the patient of care studies of the dissemination of
experience, including treatment satisfaction, advanced cancer treatment modalities
quality of life, burden of disease, ability to care throughout different population groups and
for oneself, work, etc. and new tools are con- into community practice. With approval,
stantly in development. researchers may be granted access to the SEER-
An expanding array of wearable devices and Medicare linked data files, which include
connected digital products may be used to Medicare claims prior to, during, and follow-
directly track physiologic and behavioral meas- ing cancer diagnosis and treatment. Topics
ures from patients without the involvement of studied include influences of treatment, facil-
a HCP or study investigator. Researchers have ity, and provider characteristics and interven-
reviewed mention of digital connected devices tions on survival and cost outcomes, as well as
in studies registered with http://clinicaltrials. disparities in care.
Limitation 197
Biobanks/Specimen Banks effectiveness are also applicable to studies that

use primary data collection and may help
Clinical data may be linked with biorepository
guide the actual study design. Extensive detail
data to guide researchers in the identification
and guidance regarding operational and scien-
of biomarkers that are predictive of clinical out-
tific considerations for primary data collection
comes and support the development of targeted
can be found in the Registries for Evaluating
therapies, e.g. by identifying patients whose
Patient Outcomes: a User’s Guide.
tumors harbor a genomic variant that can
potentially be targeted by a new drug. Some
biobanks have been established internationally, Strengths
such as the EuroBioBank network, developed
to support research on rare diseases. The UK A notable strength of primary data collection is
Biobank is another example of an international that it can address research objectives that
long-term registry accessible for research, require information that is not accessible or
which is following around 500 000 volunteers not consistently recorded in available second-
for at least 25 years to investigate the contribu- ary data sources including detail of vaccine
tions of genetic predisposition and environ- and medical device exposures, clinical out-
mental exposure (including nutrition and comes, and patient reported outcomes. This
lifestyle) in the disease development, and gain type of information can be particularly mean-
valuable insights to support advances in the ingful to clinicians, patients, regulators, payers
development of new medicines. An additional and those involved in drug development, and
area of data collection for the UK Biobank to can be more robust than inferences derived
support the inclusion of objective measures of from billing data and incomplete or inconsist-
physical activity in large scale observational ent electronic health data.
studies has been to obtain measures of physical Studies that use collect data directly from
activity from accelerometers from over 100 000 patients also provide the opportunity to follow
participants. Forty-five percent of those invited patients over long periods of time and to evalu-
to wear accelerometers for seven days accepted ate a range of outcomes including those of
the invitations; from these over 93% provided greatest priority to patients themselves. For
sufficient valid data for analysis. chronic diseases, patients may be followed for
years by their treating physicians, regardless of
whether a patient’s health insurance program
Guidelines on the Quality of Data changes – an important limitation of health
Collection insurance claims data. It is often the patient’s
relationship with the physician, and the physi-
There are a variety of guidelines that address
cian’s relationship with the research program
principles for pharmacoepidemiologic studies
that play an important role in long-term reten-
that use primary data collection, such as the
tion. Direct-to-patient data collection
Guidelines for Good Pharmacoepidemiologic
approaches also offer flexibility in study design
Practice developed by the International Society
and potential for further efficiencies in data
for Pharmacoepidemiology, the checklist for
collection and reduced study burden.
study protocols developed by the European
Network of Centers for Pharmacoepidemiology
and Pharmacovigilance, and principles of good Limitations
epidemiologic methods and practice. The
GRACE (Good Research for Comparative Primary data collection requires cooperation
Effectiveness) Principles for conducting and of data contributors, often over long periods of
evaluating observational studies of comparative time (follow-up). While it would seem that
altruism should be a sufficient motivation, surprisingly difficult and study retention was
experience has shown that successful primary low, speculating that although it was relatively
data collection requires an infrastructure sup- easy to send questionnaires frequently, partici-
porting patient and/or physician enrollment pants appeared to tire quickly of responding to
and retention, as well as an active program of the same questions over time.
data curation to assure that the data that are
collected are accurate and reliable. Further,
investigators need to consider the validity of Particular Applications
patient-centered endpoints, especially pertain-
ing to general and disease-specific quality of To provide further illustration of some of the
life assessments and to detailed information on applications of primary data collection in mod-
past exposures, such as in case–control studies ern pharmacoepidemiologic research, several
where the study outcome is known to the examples are described in further detail in this
patients at the time of the assessment. Like all section. These include a prospective compara-
data, the contribution to be made by patients tive effectiveness research study incorporating
in recall of past medical diagnoses and expo- collection of clinical endpoints and PROs, a
sures of interest to pharmacoepidemiologic novel hybrid study intended to provide data in
studies must be considered carefully in view of support of a label expansion with FDA, use of
their strengths and limitations. large registry data as framework for conduct of
Maintaining high subject retention in a multiple observational studies, and incorpora-
study over time can be difficult to achieve tion of measures of physical activity through
when using primary data collection. Retention accelerometry as part of the UK Biobank effort.
rates are often higher for studies that are both The Registry in Glaucoma Outcomes
(i) responsive to the needs of patients and phy- Research (RiGOR) study, funded by the US
sicians so they are motivated to continue par- Agency for Healthcare Research and Quality,
ticipating (a special concern for pregnancy was a prospective observational study that
registries and other vulnerable populations), used primary data collection to address
and (ii) parsimonious in their data collection. which treatment strategy for open-angle
Operational challenges relate to the need to glaucoma was associated with the greatest
deploy primary data collection systems that improvement in patient outcomes. The study
are easy to use, and that are simple enough to found that patients treated with incisional
encourage steady reporting but which do not surgery after failing at least one course of
result in reporting fatigue. Multiple methods medication were twice as likely as patients
of data entry such as internet, text messaging, treated with additional medication to
and/or mail, can be an advantage, considering achieve a 15% reduction in intraocular pres-
the demographics of the target population. sure (IOP) at 12 months, while patients
Many researchers believe that consistent per- treated with laser surgery had similar results
sonal interactions from study staff to clinical to those who were treated with additional
investigators and/or patients improves reten- medication. While IOP is routinely recorded
tion, and this approach is often used in preg- when glaucoma patients see their ophthal-
nancy registries. mologists, in order to ensure complete
While there is some optimism that newer assessment of IOP at around 6 and 12 months
technologies and the nearly universal adoption of follow-up, it was a required element in the
of smartphones would support the use of text study’s case report form, along with a vast
messaging and internet-based patient surveys, array of other detailed clinical information.
results from the PROTECT study raised a cau- The RiGOR study also included several vali-
tionary note. Researchers noted that internet- dated PROs assessments as secondary end-
based recruitment of pregnant women was points, which further required patients to
Key Point 199
complete these questionnaires at the time of that such data will ever contain all informa-
a study visit or at home through mail or elec- tion needed for every study purpose; thus,
tronic means. the need for primary data collection will
The Bioventus Exogen® device registry is an remain. Traditionally, HCPs have been the
example of a hybrid or enriched design involv- primary reporters/recorders of data for stud-
ing both primary and secondary data. This ies that use primary data collection, although
study was planned following extensive discus- data is increasingly collected directly from
sions with the US Food and Drug Administration patients as well. While methods for primary
Center for Drug Evaluation and Research (FDA data collection can and will change over
CDER) regarding this novel design for a label time, it is likely that researchers will always
expansion. The study utilizes a prospective need to invest time in data curation to assure
direct-to-patient product registry linked with a that the data are accurately represented and
propensity score matched comparator group to check for critical data that are systemati-
from a commercial claims database for a study cally missing. Primary data collection will
of a device used to treat bone fracture non- continue to be a mainstay of pharmacoepide-
union, currently used broadly outside of labeled miologic research, either as the sole method
indications to treat fracture. of data collection or as a key component of
Two European cancer registries provide research that uses multiple modes of data
examples of registry infrastructure created to collection or a mix of primary and secondary
address innumerable current and future data sources.
research questions pertaining to cancer inci-
dence and survival. The EUROCARE
(European Cancer Registry Based Study On Key Points
Survival And Care Of Cancer Patients) regis-
try is a very large collaborative research pro- ●● Primary data collection is needed for most
ject on cancer survival. The registry started pharmacoepidemiologic studies that require
in 1989, aiming to provide updated descrip- detailed clinical outcomes assessment,
tions of cancer survival time trends and dif- patient-reported outcomes, or physiological
ferences across European countries, to measures.
measure cancer prevalence, and to study pat- ●● Sources of primary data include assessments
terns of care of cancer patients. In its fifth and measures taken by an HCP or study site,
and current edition, EUROCARE-5 includes patient generated data, and connected digi-
data on more than 21 million cancer diagno- tal devices.
ses provided by 116 Cancer Registries in 30 ●● Primary data may be linked with secondary
European countries. At least 171 publica- data sources such as billing data, electronic
tions have been generated from health records, or existing registries utilizing
EUROCARE1-5, covering trends in survival “hybrid” or “enriched” approaches.
across a very broad range of cancer types as ●● Studies utilizing primary data must incorpo-
well as patterns of care, and predictors of rate data quality monitoring that reflects the
survival and other outcomes. nature of the data collection and potential
errors and the nature of the evidence need.
●● Long term follow-up of patients is both a
Conclusions potential strength of study designs that uti-
lize primary data collection in that there are
Despite the growing availability of large no limits on the follow-up period imposed by
amount of data on treatments and patients’ a database, and a challenge to maintain HCP
clinical experience as recorded in existing and patient engagement over long periods of
health records and billing data, it is unlikely time.
Case Example 11.1
European Cohort Study of Biomarkers in Type Results

2 Diabetes ●● The 4000 patients recruited for PROVALID
Source: Eder S et al. 2018, Heinzel A had a mean age of 63 years old and duration
et al. 2018. of diabetes of eight years. Mean values for
baseline renal function showed renal func-
Background
tion not indicative of CKD as measured by
The prevalence of T2D with and without
eGFR and albumin levels. Using the KDIGO
chronic kidney disease (CKD) varies substan-
grading system, 81% of patients were in
tially between European countries.
stages G1 or G2 for eGFR and 78% were in
Question stage A1 for albumin creatinine ratio at
The PROVALID study will use biomarkers and enrollment, indicating most had not yet
other detailed clinical and demographic data advanced to CKD or were in mild stages,
to predict the course of disease among which was expected given the population
patients and subgroups and to explain was identified from primary care. (Eder 2018).
regional differences in outcomes such as ●● From a longitudinal analysis of the rela-
rate of loss of estimated glomerular filtra- tionship of 17 plasma protein biomarkers

tion rate (eGFR) and renal and cardiovascular to a rate of progression of eGFR among
outcomes. 481 patients mostly at an early stage of
CKD, univariable analysis showed that nine
Approach
biomarkers differed significantly between
●● Prospective cohort study conducted in five
patients with stable eGFR and fast progres-
European countries (Austria, Hungary,
sion of eGFR decline, and 14 biomarkers
Netherlands, Poland and Scotland). Four
were significant predictors of rate of pro-
thousand T2D patients were enrolled from
gression of eGFR. However, results of multi-
primary care setting between 2011 and
variate analysis showed that no biomarkers
2015.
remained significant predictors of eGFR
●● Baseline clinical measures and laboratory
decline over time after adjustment for
measures corresponding to a minimal
baseline eGFR. (Heinzel 2018).
dataset as well as optional items were col-
lected for all enrollees. Strengths
●● Annual blood and urine samples are col- ●● Large cohort representing 5 EU countries.
lected to allow for validation of potential ●● Prospective collection of biomarkers and
biomarkers for renal disease diagnosis and annual blood and urine samples to be able
progression, with central laboratory analysis to investigate and validate future candi-
of lab parameters during the study period. date markers for prediction of renal and
●● Medication histories were obtained by cardiovascular disease outcomes.
interview of the patients or family
Limitations
members.
●● Potential for attrition bias over time.
●● Standard descriptive statistics with com-
●● Results will be most generalizable to T2D
parisons between groups using Chi2-tests
patients with no or early stage CKD.
or Kruskal-Wallis tests for baseline descrip-
tive analyses. Key points
●● Univariate and multivariate linear mixed The PROVALID cohort can make a potentially
models for analyses of relationship of clin- strong contribution to understanding of pre-
ical risk factors and protein biomarkers on dictors of progression of CKD and cardiovascu-
outcome of rate of change in eGFR (slope) lar disease among T2D patients and validation
over follow-up period. of potential biomarkers within this population.
Further Reading 201
Further Reading
Coleman, A., Lum, F., Velentgas, P. et al. (2016). Pharmacoepidemiological Research on

RiGOR: a prospective observational study Outcomes of Therapeutics by a European
comparing the effectiveness of treatment ConsorTium (PROTECT). http://www.
strategies for open-angle glaucoma. J. Comp. imi-protect.eu accessed 13 October 2020.
Eff. Res. 5 (1): 65–78. Jansen-van der Weide, M.C., Gaasterland,
Doherty, A., Jackson, D., Hammerla, N. et al. C.M.W., Roes, K.C.B. et al. (2018). Rare
(2017). Large scale population assessment of disease registries: potential applications
physical activity using wrist worn towards impact on development of new drug
accelerometers: the UK Biobank study. PLoS treatments. Orphanet J. Rare Dis. 13: 154.
One 12 (2): e0169649. https://doi.org/10.1371/ https://doi.org/10.1186/s13023-018-0836-0.
journal.pone.0169649. Mack C, Pavesio A, Kelly K, Wester T, Jones J,
Dreyer, N.A., Blackburn, S.C.F., Mt-Isa, S. et al. Maislin G, Irwin D, Brinkley E, Zura RD.
(2015 Jul-Dec). Direct-to-patient research: Breaking News: Study of Fracture Healing in
piloting a new approach to understanding Patients at Risk of Non-Union. Poster. 33rd
drug safety during pregnancy. JMIR Public ICPE: International Conference on
Health Surveill. 1 (2): e22. Pharmacoepidemiology & Therapeutic Risk
Eder, S., Leirer, J., Kershbaum, J. et al. (2018). A Management, Montreal, Canada, August
prospective cohort study in patients with type 23–27, 2017.
2 diabetes mellitus for validation of Marra, C., Chen, J.L., Coravos, A. et al. (2020).
biomarkers (PROVALID) – study design and Quantifying the use of connected digital
baseline characteristics. Kidney Blood Press. products in clinical research. NPJ Digit. Med.
Res. 43: 181–190. 3: 50. https://doi.org/10.1038/
EuroBioBank Network. http://www. s41746-020-0259-x.
eurobiobank.org National Cancer Institute, Surveillance,
Gliklich RE, Leavy MB, Dreyer NA (sr eds). Epidemiology, and End Results Program.
Registries for Evaluating Patient Outcomes: Overview of the SEER Program (web page).
A User’s Guide. 4th ed. (Prepared by L&M Available at http://seer.cancer.gov/about
Policy Research, LLC, under Contract No. accessed 13 October 2020.
290–2014-00004-C with partners OM1 and Pharmacovigilance Risk Assessment Committee
IQVIA) AHRQ Publication No. 19(20)- (PRAC) [Internet] Interim guidance on
EHC020. Rockville, MD: Agency for enhanced safety surveillance for seasonal
Healthcare Research and Quality; influenza vaccines in the EU. London:
September 2020. Posted final reports are European Medicines Agency (EMA) 2014.
located on the Effective Health Care Available at: http://www.ema.europa.eu/docs/
Program search page. DOI: 10.23970/ en_GB/document_library/Scientific_
AHRQEPCREGISTRIES4. guideline/2014/04/WC500165492.pdf
Heinzel, A., Kammer, M., Mayer, G. et al. (2018 accessed 13 October 2020.
Sep). Validation of plasma biomarker Resnic, F.S., Majithia, A., Marinac-Dabic, D. et
candidates for the prediction of eGFR decline al. (2017). Registry-based prospective, active
in patients with type 2 diabetes. Diabetes Care surveillance of medical-device safety. N. Engl.
41 (9): 1947–1954. https://doi.org/10.2337/ J. Med. 376: 526–535. https://doi.org/10.1056/
dc18-0532. NEJMoa1516333.
Hobbs MN. “Registries Rising: FDA Looking At Seltzer, J.H., Heise, T., Carson, P. et al. (2017).
TNF Inhibitors; AHRQ Updates Standards.” Use of endpoint adjudication to improve the
The Pink Sheet. August 2009;71(34):8–9. quality and validity of endpoint assessment
for medical device development and post Industry: Draft Guidance. May 2019. https://
marketing evaluation: rationale and best www.fda.gov/media/124746/download
practices. A report from the cardiac safety accessed 13 October 2020.
research consortium. Am. Heart J. 190: 76–85. Wallacj, J.D., Egilman, A.C., Dhruva, S.S. et al.
Singh, J.A., Saag, K.G., Bridges, S.L. Jr. et al. (2018). Postmarket studies required by the
(January 2016). 2015 American College of US Food and Drug Administration for
Rheumatology Guideline for the treatment of newdrugs and biologics approved between
rheumatoid arthritis. Arthritis Care Res. 68 (1): 2009 and 2012: cross sectional analysis. BMJ
1–26. https://doi.org/10.1002/art.39480. k2031: 361.
The DISCOVER study. http://www. World Health Organization. International
discoverdiabetes.com/DISCOVER-study.html Agency for Research on Cancer. Eurocare
Accessed October 13, 2020. study: Survival of cancer patients in
UK Biobank. www.ukbiobank.ac.uk accessed 13 Europe. http://www.eurocare.it/Home/
October 2020. tabid/36/Default.aspx accessed 13 October
US Food and Drug Administration. Postapproval 2020.
Pregnancy Safety Studies Guidance for
203
12
How Should One Perform Pharmacoepidemiologic Studies?

Choosing Among the Available Alternatives
Brian L. Strom
Introduction data resources described in the earlier chapters

of this book should be used. Although, to some
As discussed in the previous chapters, pharma- degree, the choice may too often be based upon
coepidemiologic studies apply the techniques a researcher’s familiarity with given data
of epidemiology to the content area of clinical resources and/or the investigators who have
pharmacology. Between 500 and 3000 individu- been using them, it is very important to tailor
als are usually studied prior to drug marketing. the choice of pharmacoepidemiologic resource
Most postmarketing pharmacoepidemiologic to the question to be addressed. One often may
studies need to include at least 10 000 subjects, want to use more than one data collection
or draw from an equivalent population for a strategy or resource, in parallel or in combina-
case–control study, in order to contribute suffi- tion. If no single resource is optimal for
cient new information to be worth their cost addressing a question, it can be useful to use a
and effort (see Chapter 3). This large sample number of approaches that complement each
size raises logistical challenges. Chapters 7 other. Indeed, this is probably the preferable
through 11 presented many of the different approach for addressing important questions.
data collection approaches and data resources Regardless, investigators are often left with a
that have been developed to perform pharma- difficult and complex choice.
coepidemiologic studies efficiently, meeting In order to explain how to choose among the
the need for these very large sample sizes. This available pharmacoepidemiologic data
chapter is intended to synthesize this material, resources, it is useful to synthesize the infor-
to assist the reader in choosing among the mation from the previous chapters on the rela-
available approaches. tive strengths and weaknesses of each of the
available pharmacoepidemiologic approaches,
examining the comparative characteristics of
hoosing Among the Available
C each (see Table 12.1). One can then examine
Approaches to Pharma the characteristics of the research question at
coepidemiologic Studies hand, in order to choose the pharmacoepide-
miologic approach best suited to addressing
Once one has decided to perform a pharma- that question (see Table 12.2). The assessment
coepidemiologic study, one needs to decide and weights provided in this discussion and in
which of the data collection approaches or the accompanying tables are arbitrary. They
Table 12.1 Comparative characteristics of pharmacoepidemiologic data resourcesa.
Relative Relative Relative Population- Cohort studies Case–control studies

Pharmacoepidemiologic approach size costa speed Representativeness based possible possible
Spontaneous reporting ++++ + ++++ ++ − − + (with external controls)

Health maintenance organizations/ ++ +++ +++ +++ ++ ++++ ++++
health plans
Commercial insurance databases ++ +++ +++ +++ ++ ++++ ++++
US Government claims databases +++ ++ ++ variable ++++ ++++ ++++
UK medical record databases ++ ++ +++ +++ +++ ++++ ++++
In-hospital databases + ++ +++ ++ − ++ ++
Canadian provincial databases ++ ++ +++ ++++ ++++ ++++ ++++
Pharmacy-based medical record linkage ++ ++ +++ ++++ ++++ ++++ ++++
systems
Ad hoc studies variable +++ + variable − − ++++
Case–control surveillance
Prescription-event monitoring +++ +++ + +++ ++ ++++ + (nested)
Registries variable +++ + variable variable +++ +++
Field studies as feasible +++ + as desired as desired − ++++
Ad hoc case–control studies
Ad hoc cohort studies as feasible ++++ − as desired as desired ++++ ++ (nested)
Randomized trials as feasible ++++ − − − ++++ ++ (nested)
Validity of Validity of Control of Inpatient drug Outpatient Loss to
Pharmacoepidemiologic approach exposure data outcome data confounding exposure data diagnosis data follow-up
Spontaneous reporting +++ ++ − +++ +++ N/A

Health maintenance organizations/ ++++ +++ ++ − ++ 3–15%/yr
health plans
Commercial insurance databases ++++ +++ ++ − ++ about 25%/yr
US Government claims databases ++++ +++ ++ − ++ variable
UK medical record databases +++ ++++ ++ − ++ Nil
In-hospital databases ++++ +++ ++ ++++ − Nil
Canadian provincial databases ++++ +++ ++ − ++ Nil
Pharmacy-based medical record ++++ + + − − Nil
linkage systems
Ad hoc studies ++ ++++ +++ − + N/A
Prescription-event monitoring +++ +++ ++ − +++ variable
Registries +++ +++ ++ + Variable N/A
Field studies ++ ++++ +++ ++ + N/A
Ad hoc cohort studies +++ +++ +++ ++ ++++ Variable
Randomized trials ++++ +++ ++++ ++ ++++ N/A
a
See the text of this chapter for descriptions of the column headings, and previous chapters for descriptions of the data resources.
Table 12.2 Characteristics of research questions and their impact on the choice of pharmacoepidemiologic data resourcesa.
Low Low
Pharmacoepidemiologic Hypothesis Hypothesis Study of benefits Incidence incidence prevalence
approach generating strengthening Hypothesis testingd (versus risk) rates desired outcome exposure
Spontaneous reporting ++++ + − − − ++++ ++++

Health maintenance organizations/ ++ ++++ +++ ++ +++ +++ +++
health plans
Commercial insurance databases ++ ++++ +++ ++ +++ +++ +++
US Government claims databases ++ ++++ +++ ++ +++ ++++ ++++
UK medical record databases ++ ++++ +++ ++ ++++ +++ +++
In-hospital databases + ++++ +++ ++ +++ + +
Canadian provincial databases ++ ++++ +++ ++ +++ +++ +++
Pharmacy-based medical record + ++ ++ ++ +++ +++ +++
linkage systems
Ad hoc Studies +++ +++ +++ +++ − ++++ +
Prescription-event monitoring ++ ++ +++ +++ +++ +++ +++
Registries + +++ +++ +++ +++ +++ +++
Field Studies + ++ +++ +++ + ++++ +
Ad hoc cohort studies + ++ +++ +++ ++++ ++ +++
Randomized trials + + ++++ ++++ ++++ + ++++
Pharmacoepidemiologic approach Important Drug use inpatient Outcome does not Outcome does not Outcome a Exposure a Urgent
confounders (versus outpatient) result in result in medical delayed effect new drug question
hospitalization attention
Spontaneous reporting − +++ ++++ + + ++++ ++++
Health maintenance organizations/ +++ − +++ − + ++ +++
health plans
Commercial insurance databases ++ − +++ − + +++ +++
US Government claims databases ++ − +++ − + to +++ ++ ++
UK medical record databases +++ − +++ − +++ +++ +++
In-hospital databases ++ ++++ − − − +++ +++
Canadian provincial databases ++ − +++ − +++ ++ +++
Pharmacy-based medical record + − − − ++ +++ +++
linkage systems
Ad hoc studies +++ + − − ++ + +
Prescription-event monitoring ++ + ++++ + + ++++ +
Registries ++ ++ + ++ ++ +++ +
Field studies +++ ++++ ++ − ++ + +
Ad hoc cohort studies +++ +++ ++++ +++ + ++++ +
Randomized trials ++++ +++ ++++ ++++ + ++++ +
a
See the text of this chapter for descriptions of the column headings, and previous chapters for descriptions of the data resources.
b
Hypothesis-generating studies are studies designed to raise new questions about possible unexpected drug effects, whether adverse or beneficial.
c
Hypothesis-strengthening studies are studies designed to provide support for, although not definitive evidence for, existing hypotheses.
d
Hypothesis-testing studies are studies designed to evaluate in detail hypotheses raised elsewhere.
208 12 How Should One Perform Pharmacoepidemiologic Studies? Choosing Among the Available Alternatives
are not being represented as a consensus of the maintenance organizations, depending on

pharmacoepidemiologic community, but rep- how many are included. The Canadian provin-
resent the judgment of this author alone, based cial databases again could be equivalently
on the material presented in earlier chapters of large, depending in part on how many are
this book. Nevertheless, I think that most included in a study. The other data resources
would agree with the general principles pre- are generally smaller. Case–control surveil-
sented, and even many of the relative ratings. lance, as formerly conducted by the Slone
My hope is that this synthesis of information, Epidemiology Unit, can cover a variable popu-
despite some of the arbitrary ratings inherent lation, depending on the number of hospitals
in it, will make it easier for the reader to syn- and metropolitan areas they include in their
thesize the large amount of information pre- network for a given study. The population base
sented in the prior chapters. of registry-based case–control studies depends
Note that there are a number of other data on the registries used for case finding. Ad hoc
sources not discussed here, some of which studies can be whatever size the researcher
have been, or in the future may be of impor- desires and can marshal resources for.
tance to pharmacoepidemiologic research. As to relative cost, studies that collect new
Examples include the old Boston Collaborative data are most expensive, especially rand-
Drug Surveillance data, MEMO, Pharmetrics, omized trials and cohort studies, for which
Aetna, Humana, and many others, many sample sizes generally need to be large and
reviewed in prior editions of this book. Given follow-up may need to be prolonged. In the
the wonderful proliferation of pharmacoepide- case of randomized trials, there are additional
miologic data resources, we are making no logistical complexities. Studies that use exist-
attempt to include them all. Instead, we will ing data are least expensive, although their
discuss them in categories of type of data, as cost increases when they gather primary medi-
we did in the chapters themselves. cal records for validation. Studies that use
existing data resources to identify subjects but
then collect new data about those subjects are
Comparative Characteristics
intermediate in cost.
of Pharmacoepidemiologic Data
With regard to the relative speed of study
Resources
completion, studies that collect new data take
Table 12.1 lists each of the different pharma- longer, especially randomized trials and cohort
coepidemiologic data resources that were studies. Studies that use existing data are able
described in earlier chapters, along with some to answer a question most quickly, although
of their characteristics. considerable additional time may be needed to
The relative size of the database refers to the obtain primary medical records for validation.
population it covers. Only spontaneous report- Studies that use existing data resources to
ing systems, US Medicare, some of the identify subjects but then collect new data
pharmacy-based medical record linkage sys- about those subjects are intermediate in speed.
tems, and Prescription Event Monitoring in Representativeness refers to how well the
the UK cover entire countries or large fractions subjects in the data resource represent the pop-
thereof. Of course, population databases differ ulation at large or a more specific population
considerably in size, based on the size of their of interest. US Medicare, Prescription Event
underlying populations. Aggregations of Monitoring in the UK, the provincial health
Medicaid databases are the next largest, with databases in Canada, and the pharmacy-based
the commercial databases approaching that. medical record linkage systems each include
The UK electronic health record databases entire countries, provinces, or states and, so,
would be next in size, as would the health are typical populations. Spontaneous reporting
Choosing Among the Available Approaches to Pharmacoepidemiologic Studie 209
systems are drawn from entire populations, that belong to the system and do not capture
but of course the selective nature of their all healthcare services a patient may receive.
reporting could lead to less certain representa- Thus, a patient diagnosed with and treated for
tiveness. Medicaid programs are limited to the venous thromboembolism in a participating
disadvantaged, and include a population that hospital could be readmitted to a different,
is least representative of a general population. non-participating hospital if the disease
Analogously, randomized trials include popu- recurred. This recurrence would not be
lations limited by the various selection criteria detected in a study using such a system. The
plus their willingness to volunteer for the data resources that are population-based are
study. The CPRD and THIN use a non-random those which use data from organized health
large subset of the total UK population, and care delivery or payment systems. Registry-
may be representative of the overall UK popu- based and ad hoc case–control studies can
lation. Health plans and commercial databases occasionally be conducted as population-based
are closer to representative populations than a studies, if all cases in a defined geographic area
Medicaid population would be, although they are recruited into the study, but this is unusual
include a largely working population and, so, (see also Chapter 11).
include few patients of low socioeconomic sta- Whether cohort studies are possible within a
tus and fewer than normal elderly. Some of the particular data resource would depend on
remaining data collection approaches or whether individuals can be identified by
resources are characterized in Table 12.1 as whether or not they were exposed to a drug of
“variable,” meaning their representativeness interest. This would be true in any of the
depends on which hospitals are recruited into population-based systems, as well as any of the
the study. Ad hoc studies are listed in Table 12.1 systems designed to perform cohort studies.
“as desired,” because they can be designed to Whether case–control studies are possible
be representative or not, as the investigator within a given data resource depends on
wishes. whether patients can be identified by whether
Whether a database is population-based or not they suffered from a disease of interest.
refers to whether there is an identifiable popu- This would be true in any of the population-
lation (which is not necessarily based in geog- based systems. Data from spontaneous report-
raphy), all of whose medical care would be ing systems can be used for case finding for
included in that database, regardless of the case–control studies, although this has been
provider. This allows one to measure incidence done infrequently.
rates of diseases, as well as being more certain The validity of the exposure data is most cer-
that one knows of all medical care that any tain in hospital-based settings, where one can
given patient receives. As an example, assum- be reasonably certain of both the identity of a
ing little or no out-of-plan care, the Kaiser prodrug and that the patient actually ingested it.
grams are population-based. One can use Exposure data in spontaneous reporting sys-
Kaiser data, therefore, to study medical care tems come mostly from health care providers
received in and out of the hospital, as well as and, so, are probably valid. However, one can-
diseases which may result in repeat hospitali- not be certain of patient adherence in sponta-
zations. For example, one could study the neous reporting data. Exposure data from
impact of the treatment initially received for claims data and from pharmacy-based medi-
venous thromboembolism on the risk of subse- cal record linkage systems are unbiased data
quent disease recurrence. In contrast, hospital- recorded by pharmacies, often for billing pur-
based case–control studies conducted outside poses, a process that is closely audited as it
a closed network like Kaiser are not population- impacts reimbursement. These data are likely
based: they include only the specific hospitals to be accurate with regard to medication
ossession, although, again, one cannot

p Control of confounding refers to the ability
assure adherence. Refill adherence though to control for confounding variables. Rando
has been found to correlate closely with mization is the most convincing way of
adherence measured using microchips controlling for unknown, unmeasured, or
embedded in medication bottles (see unmeasurable confounding variables.
Chapter 21). However, there are drugs that Approaches that collect sufficient information
may fall beneath a patient’s deductibles or co- to control for known and measurable variables
payments, or not be on formularies, so dis- are next most effective. These include health
pensed by the pharmacy but paid for in cash. plans, the UK medical record systems, case–
In claims databases, these scenarios may control surveillance, ad hoc case–control stud-
result in misclassification of a true medica- ies, and ad hoc cohort studies. Users of health
tion exposure as the patient would falsely databases in Canada, commercial databases,
appear unexposed. Also, since drug benefits and Medicaid (sometimes) can obtain primary
vary depending on the plan, pharmacy files medical records, but not all information neces-
may not capture all prescribed drugs if benefi- sary is always available in those records. They
ciaries reach the drug benefit limit or pay for are generally unable to contact patients directly
the prescription out-of-pocket. In the UK to obtain supplementary information that
medical record systems, drugs prescribed by might not be in a medical record. Finally, spon-
physicians other than the general practitioner taneous reporting systems do not provide
could be missed, although continued pre- enough systematically collected information
scribing by the general practitioner would be for control of confounding.
detected. Adhoc case–control studies gener- Relatively few of the data systems have data
ally rely on patient histories for exposure on inpatient drug use. The exceptions include
data. These may be very inaccurate, as spontaneous reporting systems, the in-hospital
patients often do not recall correctly the med- databases, and some ad hoc studies if designed
ications they are taking. However, this would to collect such.
be expected to vary, depending on the condi- Only a few of the data resources have suffi-
tion studied, type of drug taken, the question- cient data on outpatient diagnoses available
ing technique used, etc. (see Chapter 13). without special effort, to be able to study them
The validity of the outcome data is also most as outcome variables. Ad hoc studies can be
certain in hospital-based settings, in which the designed to be able to collect such information.
patient is subjected to intensive medical sur- In the case of ad hoc randomized clinical trials,
veillance. It is least certain in outpatient data this data collection effort could even include
from organized systems of medical care. There tailored laboratory and physical examination
are, however, methods of improving the accu- measurements. In some of the resources, the
racy of these data, such as using drugs, labora- outpatient outcome data are collected observa-
tory data, and procedures as markers of the tionally, but directly via the physician, and so
disease and obtaining primary medical records are more likely to be accurate. Included are
(see Chapter 13). The outcome data from auto- spontaneous reporting systems, the UK medi-
mated databases are listed as variable, there- cal record systems, HMOs, Prescription Event
fore, depending on exactly which data are Monitoring, and some ad hoc cohort studies.
being used, and how. The UK medical record Other outpatient data come via physician
systems analyze the actual medical record, claims for medical care, including Medicaid
rather than claims, and can access additional databases, commercial databases, and the pro-
questionnaire data from the general practi- vincial health databases in Canada. Finally,
tioner, as well. Thus, their outcome data may other data resources can access outpatient
be more accurate. diagnoses only via the patient, and so they are
less likely to be complete; although the diagno- generation as well, in addition to the tradi-
sis can often be validated using medical tional approach of using such data for hypoth-
records, it generally needs to be identified by esis testing. In the future, new approaches
the patient. These include most ad hoc case– using the internet (e.g. health websites with
control studies. consumer posting boards and other social
The degree of loss to follow-up differs sub- media) could potentially be used for hypothe-
stantially among the different resources. They sis generation of events, including those not
are specified in Table 12.1. coming to medical attention.
Hypothesis-strengthening studies are studies
designed to provide support for, although not
Characteristics of Research
definitive evidence for, existing hypotheses.
Questions and their Impact on the
The objective of these studies is to provide suf-
Choice of Pharmacoepidemiologic
ficient support for, or evidence against, a
Data Resources
hypothesis to permit a decision about whether
Once one is familiar with the characteristics of a subsequent, more definitive, study should be
the pharmacoepidemiologic resources availa- undertaken. As such, hypothesis-strengthening
ble, one must then examine more closely the studies need to be conducted rapidly and inex-
research question, to determine which pensively. Hypothesis-strengthening studies
resources can best be used to answer it (see can include crude analyses conducted using
Table 12.2). almost any dataset, evaluating a hypothesis
Pharmacoepidemiologic studies can be which arose elsewhere. Because not all poten-
undertaken to generate hypotheses about drug tially confounding variables would be con-
effects, to strengthen hypotheses, and/or to test trolled, the findings could not be considered
a priori hypotheses about drug effects. definitive. Examples would be the modular
Hypothesis-generating studies are studies studies conducted within Sentinel.
designed to raise new questions about possible Alternatively, hypothesis-strengthening stud-
unexpected drug effects, whether adverse or ies can be more detailed studies, controlling for
beneficial. Virtually all studies can and do raise confounding, conducted using the same data
such questions, through incidental findings in resource that raised the hypothesis. In this
studies performed for other reasons. In addi- case, because the study is not specifically
tion, virtually any case–control study could be undertaken to test an a priori hypothesis, the
used, in principle, to screen for possible drug hypothesis-testing type of study can only serve
causes of a disease under study, and virtually to strengthen, not test, the hypothesis.
any cohort study could be used to screen for Spontaneous reporting systems are useful for
unexpected outcomes from a drug exposure raising hypotheses, but are not very useful for
under study. In practice, however, the only set- providing additional support for those hypoth-
tings in which this has been attempted system- eses. Conversely, randomized trials can cer-
atically have been health plans, case–control tainly strengthen hypotheses, but are generally
surveillance, Prescription Event Monitoring, too costly and logistically too complex to be
and Medicaid databases. To date, the most pro- used for this purpose (Post-hoc analyses of ran-
ductive source of new hypotheses about drug domized trials can obviously be re-analyzed,
effects has been spontaneous reporting. for the purposes of generating or strengthen-
However, this is the goal of Sentinel, a ing hypotheses, but then they are really being
Congressionally mandated data system of over analyzed as cohort studies). Of the remain-
100 million US lives, initially built primarily ing approaches, those that can quickly access,
for hypothesis strengthening as “Mini-Sentinel,” in computerized form, both exposure data
although now being used for hypothesis and outcome data are most useful. Those
that can rapidly access only one of these data not in a way which usually is related to the out-
types, only exposure or only outcome data, are come under investigation in the study. It is
next most useful, while those that need to sometimes possible to address these questions
gather both data types are least useful, because using nonexperimental study designs.
of the time and expense that would be entailed. Generally, however, the randomized clinical
Hypothesis-testing studies are studies trial is far preferable, when feasible.
designed to evaluate in detail hypotheses In order to address questions about the inci-
raised elsewhere. Such studies must be able to dence of a disease in those exposed to a drug,
have simultaneous comparison groups and one must be able to quantify how many people
must be able to control for most known poten- received the drug. This information can be
tial confounding variables. For these reasons, obtained using any resource that can perform a
spontaneous reporting systems cannot be used cohort study. Techniques that need to gather
for this purpose, as they cannot be used to con- the outcome data de novo may miss some of
duct studies with simultaneous controls (with the outcomes if there is incomplete participa-
rare exception). The most powerful approach, tion and/or reporting of outcomes, such as
of course, is a randomized clinical trial, as it is with Prescription Event Monitoring, ad hoc
the only way to control for unknown or cohort studies, and outpatient pharmacy-based
unmeasurable confounding variables. cohort studies. On the other hand, ad hoc data
Instrumental variable analyses can approxi- collection is the only way of systematically col-
mate a randomized clinical trial, but only in lecting information about outcomes that need
the to-date limited circumstances that all the not come to medical attention (see below). The
underlying assumptions are met. (On the other only approaches that are free from either of
hand, studies of dose–response, duration- these problems are the hospital-based
response, drug–drug interactions, determi- approaches. Registry-based case–control stud-
nants of response, etc. are more readily done in ies and ad hoc case–control studies can occa-
non-randomized than randomized studies.) sionally be used to estimate incidence rates, if
Techniques which allow access to patients and one obtains a complete collection of cases from
their medical records are the next most power- a defined geographic area. The other
ful, as one can gather information on potential approaches listed cannot be used to calculate
confounders that might only be reliably incidence rates.
obtained from one of those sources or the To address a question about a low incidence
other. Techniques which allow access to pri- outcome, one needs to study a large population
mary records but not the patient are next most (see Chapter 3). This can best be done using
useful. spontaneous reporting, US Medicare,
The research implications of questions about Prescription Event Monitoring, or the
the beneficial effects of drugs are different, pharmacy-based medical record linkage sys-
depending upon whether the beneficial effects tems, which can or do cover entire countries.
of interest are expected or unexpected effects. Alternatively, one could use commercial data-
Studies of unexpected beneficial effects are bases, health plans, or aggregates of Medicaid
exactly analogous to studies of unexpected databases, which cover a large proportion of
adverse effects, in terms of their implications the United States, or the medical record sys-
to one’s choice of an approach; in both situa- tems in the UK. Canadian provincial databases
tions one is studying side effects. Studies of can also be fairly large, and one can perform a
expected beneficial effects, or drug efficacy, raise study in multiple such databases. Ad hoc
the special methodologic problem of con- cohort studies could potentially be expanded
founding by the indication: patients who to cover equivalent populations. Case–control
receive a drug are different from those who do studies, either ad hoc studies, studies using
registries, or studies using case–control sur- gather information on potential confounders

veillance, can also be expanded to cover large that might only be reliably obtained from one
populations, although not as large as the of those sources or the other. Techniques
previously-mentioned approaches. Because which allow access to primary records but not
case–control studies recruit study subjects on the patient are the next most useful.
the basis of the patients suffering from a dis- If the research question involves inpatient
ease, they are more efficient than attempting to drug use, then the data resource must obvi-
perform such studies using analogous cohort ously be capable of collecting data on inpatient
studies. Finally, randomized trials could, in drug exposures. The number of approaches
principle, be expanded to achieve very large that have this capability are limited, and
sample sizes, especially large simple trials (see include: spontaneous reporting systems and
Chapter 17), but this can be very difficult and inpatient database systems. Ad hoc studies
costly. could also, of course, be designed to collect
To address a question about a low prevalence such information in the hospital.
exposure, one also needs to study a large popu- When the outcome under study does not
lation (see Chapter 3). Again this can best be result in hospitalization, but does result in medi-
done using spontaneous reporting, US cal attention, the best approaches are rand-
Medicare, the pharmacy-based medical record omized trials and ad hoc studies which can be
linkage systems, or Prescription Event specifically designed to be sure this informa-
Monitoring, which cover entire countries. tion can be collected. Prescription Event
Alternatively, one could use commercial data- Monitoring and the UK medical record sys-
bases, large health plans, or aggregates of tems, which collect their data from general
Medicaid databases, which cover a large pro- practitioners, are excellent sources of data for
portion of the United States, or the medical this type of question. Reports of such outcomes
record databases in the UK. Ad hoc cohort are likely to come to spontaneous reporting
studies could also be used to recruit exposed systems, as well. Medicaid databases and com-
patients from a large population. Analogously, mercial databases can also be used, as they
randomized trials, which specify exposure, include outpatient data, although one must be
could assure an adequate number of exposed cautious about the validity of the diagnosis
individuals. Case–control studies, either ad information in outpatient claims. Canadian
hoc studies, studies using registries, or studies provincial databases are similar, as are health
using case–control surveillance, could theo- plans. Finally, registry-based case–control
retically be expanded to cover a large enough studies could theoretically be performed, if
population, but this would be difficult and they included outpatient cases of the disease
expensive. under study.
When there are important confounders that When the outcome under study does not
need to be taken into account in order to result in medical attention at all, the approaches
answer the question at hand, then one needs to available are much more limited. Only rand-
be certain that sufficient and accurate informa- omized trials and prospective cohort studies
tion is available on those confounders. can be specifically designed to be certain this
Spontaneous reporting systems cannot be used information is collected. Finally, occasionally
for this purpose. The most powerful approach one could collect information on such an out-
is a randomized trial, as it is the most convinc- come in a spontaneous reporting system, if the
ing way to control for unknown or unmeasur- report came from a patient or if the report
able confounding variables. Techniques which came from a health care provider who became
allow access to patients and their medical aware of the problem while the patient was vis-
records are the next most powerful, as one can iting for medical care for some other problem.
In the future, as noted above, new approaches studies or a randomized clinical trial are ideal
using the internet (e.g. health websites with for this, as they recruit patients into the study
consumer posting boards) could potentially be on the basis of their exposure. Spontaneous
used for hypothesis generation of events not reporting is similarly a good approach for this,
coming to medical attention. as new drugs are automatically and immedi-
When the outcome under study is a delayed ately covered, and in fact reports are much
drug effect, then one obviously needs more common in the first three years after a
approaches capable of tracking individuals drug is marketed. The major databases are next
over a long period of time. The best approach most useful, especially the commercial data-
for this are some of the provincial health databases, as their large population base will allow
bases in Canada. Drug data are available in one to accumulate a sufficient number of
some for more than 25 years, and there is little exposed individuals rapidly, so one can per-
turnover in the population covered. Thus, this form a study sooner. In some cases, there is a
is an ideal system within which to perform delay until the drug is available on the pro-
such long-term studies. Some health plans gram’s formulary; however, that especially can
have even longer follow-up time available. be an issue with HMOs. The US government
However, as health plans, they suffer from sub- claims databases (Medicare and Medicaid)
stantial turnover, albeit more modest after the have a delay in availability of their data, which
first few years of enrollment. Commercial makes them less useful for the newest drugs.
databases are similar. Any of the methods of Ad hoc case–control studies, by whatever
conducting case–control studies can address approach, must wait until sufficient drug expo-
such questions, although one would have to be sure has occurred that it can affect the out-
especially careful about the validity of the come variable being studied.
exposure information collected many years Finally, if one needs an answer to a question
after the exposure. Medicaid databases have urgently, potentially the fastest approach, if the
been available since 1973. However, the large needed data are included, is a spontaneous
turnover in Medicaid programs, due to changes reporting system; drugs are included in these
in eligibility with changes in family and systems immediately, and an extremely large
employment status, makes studies of long- population base is covered. Of course, one can-
term drug effects problematic. Similarly, one not rely on any adverse reaction being detected
could conceivably perform studies of long- in a spontaneous reporting system. The com-
term drug effects using Prescription Event puterized databases are also useful for these
Monitoring, the pharmacy-based medical purposes, depending on the speed with which
record linkage systems, ad hoc cohort studies, the exposures accumulate in that database; of
or randomized clinical trials, but these course, if the drug in question is not on the for-
approaches are not as well-suited to this type mulary in question, it cannot be studied.
of question as the previously-discussed tech- Modular analyses in Sentinel were designed
niques. Theoretically, one also could identify for exactly this purpose. The remaining
long-term drug effects in a spontaneous report- approaches are of limited use, as they take too
ing system. This is unlikely, however, as a phy- long to address a question. One exception to
sician is unlikely to link a current medical this is Prescription Event Monitoring, if the
event with a drug exposure long ago. drug in question happens to have been a sub-
When the exposure under study is a new drug, ject of one of its studies. The other, and more
then one is, of course, limited to data sources likely exception, is case–control surveillance if
that collect data on recent exposures, and pref- the disease under study is available in adequate
erably those that can collect a significant num- numbers in its database, either because it was
ber of such exposures quickly. Ad hoc cohort the topic of a prior study or because there were
Example 215
a sufficient number of individuals with the disappear in vital statistics data, except as a
ease collected to be included in control groups cause of hospitalization or death. Studying
for prior studies. death from upper gastrointestinal bleeding is
problematic, as it is a disease from which
patients usually do not die. Rather than
Examples studying determinants of upper gastrointes-
tinal bleeding, one would really be studying
As an example, one might want to explore determinants of complications from upper
whether nonsteroidal anti-inflammatory gastrointestinal bleeding, diseases for which
drugs (NSAIDs) cause upper gastrointestinal upper gastrointestinal bleeding is a compli-
bleeding and, if so, how often. One could cation, or determinants of physicians’ deci-
examine the manufacturer’s premarketing sions to withhold supportive transfusion
data from clinical trials, but the number of therapy from patients with upper gastroin-
patients included is not likely to be large testinal bleeding, for example: age, terminal
enough to study clinical bleeding, and the illnesses, etc.
setting is very artificial. Alternatively, one Alternatively, one might want to address a
could examine premarketing studies using similar question about nausea and vomiting
more sensitive outcome measures, such as caused by NSAIDs. Although this question is
endoscopy. However, these are even more very similar, one’s options in addressing it
artificial. Instead, one could use any of the would be much more limited, as nausea and
databases to address the question quickly, as vomiting often do not come to medical atten-
they have data on drug exposures that pre- tion. Other than a randomized clinical trial, for
ceded the hospital admission. Some data- a drug that is largely used on an outpatient basis
bases could only be used to investigate one is limited to systems which request infor-
gastrointestinal bleeding resulting in hospi- mation from patients, or ad hoc cohort studies.
talization (e.g. Kaiser Permanente, except As another example, one might want to fol-
via chart review). Others could be used to low-up on a signal generated by the spontane-
explore inpatient or outpatient bleeding (e.g. ous reporting system, designing a study to
Medicare, Medicaid, Canadian provincial investigate whether a drug, which has been on
databases). Because of confounding by ciga- the market for, say, five years, is a cause of a
rette smoking, alcohol, etc. which would not relatively rare condition, such as allergic
be well measured in these databases, one hypersensitivity reactions. Because of the
also might want to address this question infrequency of the disease, one would need to
using case–control or cohort studies, draw on a very large population. The best alter-
whether conducted ad hoc or using any of natives would be Medicare or Medicaid data-
the special approaches available, for exam- bases, health plans, commercial databases,
ple case–control surveillance or Prescription case–control studies, or Prescription Event
Event Monitoring. If one wanted to be able Monitoring. To expedite this hypothesis-testing
to calculate incidence rates, one would need study and limit costs, it would be desirable if it
to restrict these studies to cohort studies, could be performed using existing data.
rather than case–control studies. One would Prescription Event Monitoring and case–
be unlikely to be able to use registries, as control surveillance would be excellent ways
there are no registries, known to this author of addressing this, but only if the drug or dis-
at least, which record patients with upper ease in question, respectively, had been the
gastrointestinal bleeding. One would not be subject of a prior study. Other methods of con-
able to perform analyses of secular trends, as ducting case–control studies require gathering
upper gastrointestinal bleeding would not exposure data de novo.
As a last example, one might want to follow- the characteristics of the question to be
up on a signal generated by a spontaneous addressed, one should be able to choose those
reporting system, designing a study to investi- resources that are best suited to addressing the
gate whether a drug, which has been on the question at hand.
market for, say, three years, is a cause of an
extremely rare but serious illness, such as
aplastic anemia. One’s considerations would Key Points
be similar to those above, but even Medicare or
Medicaid databases would not be sufficiently ●● There are many different approaches to per-
large to include enough cases, given the delay forming pharmacoepidemiologic studies,
in the availability of their data. One would each of which has its advantages and
have to gather data de novo. Assuming the disadvantages.
drug in question is used mostly by outpatients, ●● The choice of pharmacoepidemiologic
one could consider using Prescription Event resource must be tailored to the question to
Monitoring or a case–control study. be addressed.
●● One may want to use more than one data
collection strategy or resource, in parallel or
Conclusion in combination.
●● By considering the characteristics of the
Once one has decided to perform a pharma- pharmacoepidemiologic resources available
coepidemiologic study, one needs to decide and the characteristics of the question to be
which of the resources described in the earlier addressed, one should be able to choose
chapters of this book should be used. By con- those resources that are best suited to address
sidering the characteristics of the pharma- the question at hand.
coepidemiologic resources available as well as
Further Reading
Anonymous (1986). Risks of agranulocytosis and Mitchell, A.A., Cottler, L.B., and Shapiro, S.
aplastic anemia. A first report of their relation (1986). Effect of questionnaire design on recall
to drug use with special reference to analgesics. of drug exposure in pregnancy. Am. J.
The international agranulocytosis and aplastic Epidemiol. 123: 670–676.
Anemia study. JAMA 256: 1749–1757. Paganini-Hill, A. and Ross, R.K. (1982).
Coulter, A., Vessey, M., and McPherson, K. Reliability of recall of drug usage and other
(1986). The ability of women to recall their health-related information. Am. J. Epidemiol.
oral contraceptive histories. Contraception 33: 116: 114–122.
127–139. Persson, I., Bergkvist, L., and Adami, H.O.
Glass, R., Johnson, B., and Vessey, M. (1974). (1987). Reliability of women’s histories of
Accuracy of recall of histories of oral climacteric oestrogen treatment assessed by
contraceptive use. Br. J. Prev. Soc. Med. 28: prescription forms. Int. J. Epidemiol. 16:
273–275. 222–228.
Klemetti, A. and Saxen, L. (1967). Prospective Rosenberg, M.J., Layde, P.M., Ory, H.W. et al.
versus retrospective approach in the search for (1983). Agreement between women’s histories
environmental causes of malformations. Am. of oral contraceptive use and physician
J. Public Health 57: 2071–2075. records. Int. J. Epidemiol. 12: 84–87.
Further Readin 217
Schwarz, A., Faber, U., Borner, K. et al. (1984). to drug use histories. Am. J. Epidemiol. 107:
Reliability of drug history in analgesic users. 226–235.
Lancet 2: 1163–1164. Strom, B.L., West, S.L., Sim, E., and Carson, J.L.
Stolley, P.D., Tonascia, J.A., Sartwell, P.E. et al. (1989). Epidemiology of the acute flank pain
(1978). Agreement rates between oral syndrome from suprofen. Clin. Pharmacol.
contraceptive users and prescribers in relation Ther. 46: 693–699.
219
Part III
Special Issues in Pharmacoepidemiology Methodology

221
13
Validity of Drug and Diagnosis Data in Pharmacoepidemiology

Mary Elizabeth Ritchey1,2, Suzanne L. West3, and George Maldonado4
1
Med Tech Epi, LLC, Philadelphia, PA, USA
2
Center for Pharmacoepidemiology and Treatment Science, Rutgers University, New Brunswick, NJ, USA
3
Gillings School of Global Public Health, University of North Carolina, Chapel Hill, NC, USA
4
Division of Environmental Health Sciences, School of Public Health, University of Minnesota, Minneapolis, MN, USA
Introduction linical Problems to be Addressed

C
by Pharmacoepidemiologic
Accurate pharmacoepidemiologic study con- Research
clusions require assessment of valid data,
regardless of whether the data originate from Of particular concern to the subject of this
questionnaires, administrative claims, or book is the validity of data on drug exposure
electronic health records (EHRs). This chap- and disease occurrence, because the typical
ter begins by discussing the validity of the focus of pharmacoepidemiologic research is
drug and diagnostic information used by cli- often the association between a medication
nicians in patients’ care. Next, we discuss and an adverse drug event. Further, many
measurement error, describing the different potential confounders of importance in phar-
types of error and error detection methods, macoepidemiologic research (although cer-
exploring how errors may affect the point tainly not all) are either drugs or diseases.
estimate, and describing current techniques Clinicians recognize that patients very often
for mitigation. In the remainder of the chap- do not know the names of the drugs they are
ter, we illustrate validity concerns when data taking. Thus, it is a given that patients have
from administrative claims, EHRs, or ques- difficulty recalling past drug use accurately, at
tionnaire responses are used, using as exam- least absent any aids to this recall. Superficially
ples studies of the associations between at least, patients cannot be considered relia-
nonsteroidal anti-inflammatory drugs ble sources of diagnosis information either; in
(NSAIDs) and myocardial infarction (MI), some instances, they may not even have been
and between NSAIDs and gastrointestinal told the correct diagnosis, let alone recall it.
(GI) bleeding. Yet, these data elements are crucial to
222 13 Validity of Drug and Diagnosis Data in Pharmacoepidemiology
harmacoepidemiology studies that ascertain

p validity, quantitative measurement of reliabil-
data using questionnaires. Special approaches ity, measurement error in pharmacoepidemio-
have been developed by pharmacoepidemiol- logic research, and adjusting measures of
ogists to obtain such data more accurately, association for measurement error.
from patients and other sources, but the suc-
cess of these approaches needs to be consid-
Indices of Measurement Error
ered in detail.
Relevant to Pharmacoepidemiologic
Besides self-reported data, pharmacoepide-
Research
miologists have been using administrative
claims data for more than 30 years to evaluate Two main comparisons may be drawn between
drug safety. We discuss validity issues with two (or more) methods of data collection or
using these data for research. However, the sources of information on exposure or out-
changing landscape of health care requires re- come: validity and reliability. Many different
assessing the validity of the data pharmacoepi- terms have been used to describe each, result-
demiologists are now using for their research ing in some confusion. Although the literature
and how these data impact clinical practice. uses the term validation or verification to
More and more, pharmacoepidemiologists describe the agreement between two sources
are turning to EHR data for their research. of information, concordance or agreement may
Whereas the increased granularity of EHR more appropriately indicate comparison
data is a benefit for their use in pharmacoepi- between data sources because validation
demiology, important limitations of these data requires a “gold standard.” Though, in recogni-
include their potential incompleteness and tion that a method or source can be superior to
lack of interoperability across health systems. another method or source without being per-
Unless EHR data arise from “closed” health fect, the term “alloyed gold standard” is used.
care systems where patients receive all their
outpatient and inpatient care, the EHR data
Quantitative Measurement
may represent only a portion of the patients’
of Validity
health problems and care received. If EHR
data from multiple health systems are used, For a binary exposure or outcome measure,
even if the health systems use the same EHR such as “ever” versus “never” using a particu-
vendor, the data may need to be restructured so lar drug, two measures of validity are used.
that they are consistent across all data arising Sensitivity measures the degree to which the
from all health systems. The clinician review- inferior source or method correctly identifies
ing evidence for patient care that arises from individuals who, according to the superior
studies using EHR data trusts that these data method or source, possess the characteristic of
have been curated sufficiently to produce interest (i.e. ever used the drug). Specificity
robust and valid study findings. measures the degree to which the inferior
source or method correctly identifies individu-
als who, according to the superior method or
Methodological Problems source, lack the characteristic of interest (i.e.
to be Solved by never used the drug). Figure 13.1 illustrates the
Pharmacoepidemiologic Research calculation of sensitivity and specificity.
Sensitivity and specificity are the two sides
There are five major methodologic problems of the validity “coin” for a dichotomous expo-
associated with validity of data for pharma- sure or outcome variable. In general, sources
coepidemiologic research: indices of measure- or methods with higher sensitivity tend to have
ment error, quantitative measurement of lower specificity, and methods with higher
Methodological Problems to be Solved by Pharmacoepidemiologic Researc 223
Gold standard
Exposed Not exposed
A B m1
Exposed
true positive false positive
Questionnaire
data
C D m2
Not exposed
false negative true negative
n1 n2 N
Sensitivity = A/A + C
Specificity = D/B + D
Figure 13.1 Formulas for calculating sensitivity and specificity.
specificity tend to have lower sensitivity. In c lassification method or information source,

these very common situations, neither of the not measures of validity. Predictive values
two sources or methods compared can be said depend not only on the sensitivity and specific-
to have superior overall validity. Depending on ity (i.e. on validity), but also on the true preva-
the study setting, either sensitivity or specific- lence of the exposure or outcome. Thus, if a
ity may be the more important validity meas- method or information source for classifying
ure. Moreover, absolute values of these persons with respect to outcome or exposure
measures can be deceiving. For example, if the has the same validity (i.e. the same sensitivity
true prevalence of ever using a drug is 5%, then and specificity) in two populations, but those
an exposure classification method or informa- populations differ in their outcome or expo-
tion source with 95% specificity (and perfect sure prevalence, the source or method will
sensitivity) will almost double the measured have different predictive values in the two
prevalence to about 10%. The ultimate crite- populations.
rion of importance of a given combination of In some validation studies, one method or
sensitivity and specificity is the degree of bias source may be used as a gold standard or as an
exerted on a measure of effect, such as an esti- alloyed gold standard to assess another method
mated relative risk due to measurement error. or source with respect to only one side of the
As measures of validity, sensitivity and spec- validity “coin.” Studies that focus on the com-
ificity have “truth” (i.e. the classification pleteness of one source, such as studies in
according to a gold standard or an alloyed gold which interview responses are compared with
standard) in their denominators. Investigators prescription dispensing records to identify
should take care not to confuse these measures drug exposures forgotten or otherwise unre-
with positive and negative predictive values ported by the respondents, may measure (more
(NPV), which include the inferior measure in or less accurately) the sensitivity of the inter-
their denominators. We distinguish here view data. However, such studies are silent on
between the persons who actually do or do not the specificity unless strong assumptions are
have an exposure or outcome, and those who made (e.g. that the respondent could not have
are classified as having it or not having it. The obtained the drug in a way that would not be
positive predictive value (PPV) is the proportion recorded in the prescription dispensing
of persons classified as having the exposure or records). Similarly, validation of cases in a
outcome who are correctly classified. The neg- case–control study using self-report or admin-
ative predictive value is the proportion of per- istrative claims data often provides only the
sons classified as lacking the exposure or PPV that the cases are true cases and does not
outcome who are correctly classified. Predictive evaluate the NPV that the controls are truly
values are measures of performance of a controls. Ideally, one would design a validation
study to calculate sensitivity and specificity, as equality, which is 45° from either axis. However,
well as positive and NPV and the key patient perfect correlation occurs when the points lie
characteristics and other variables on which along any straight line parallel to the line of
they depend. equality. It is difficult to tell from the value of a
For a drug exposure, a true gold standard is a correlation coefficient how much bias will be
list of all drugs the study participant has taken, produced by using an inaccurate measure of
including dose, duration, and dates of expo- disease or exposure.
sure. This drug list might be a diary of prescrip-
tions that the study participants kept or a
Quantitative Measurement
computerized database of filled prescriptions.
of Reliability
However, neither of these data sources is a
genuine gold standard. Prescription diaries When the same data collection method or
cannot be assumed to be kept in perfect accu- source of information is used more than once
racy. For example, participants may tend to for the same information on the same individ-
record drug use as more regular and complete ual, comparisons of the results measure the
than it actually was, or as more closely adher- reliability of the method or information source.
ing to the typically prescribed regimen. An example of a reliability study is a compari-
Similarly, substantial gaps may exist between son of responses in repeat interviews using the
the point at which a prescription is filled and same interview instrument. Reliability is not
the time that it is ingested, if it is ingested at all validity, although the term is sometimes used,
(see Chapter 21 for discussion of adherence). inaccurately, as such. The term reliability tends
Two methods are used to quantify the valid- to be used far too broadly to refer variously not
ity of continuously distributed variables, such only to reliability itself, but to agreement or
as duration of drug usage. The mean and validity as well. Researchers and others should
standard error of the differences between the take greater care with the way they use such
data in question and the valid reference meas- terms.
urement are typically used when the measure- When different data collection methods or
ment error is constant across the range of true different sources of information are compared
values (i.e. when measurement error is inde- (e.g. comparison of prescription dispensing
pendent of where an individual’s true expo- records with interview responses), and neither
sure falls on the exposure distribution in the of them can be considered distinctly superior
study population). With the caveat that it is to the other, the comparisons measure mere
generalizable only to populations with similar agreement. Agreement between two sources or
exposure distributions, the product–moment methods does not imply that either is valid.
correlation coefficient may also be used. To evaluate reliability or agreement for cate-
High correlation between two measures gorical variables, the percentage agreement
does not necessarily mean high agreement. For between two or more sources and the related
instance, the correlation coefficient could be (kappa, κ) coefficient are used. They are used
very high (i.e. close to 1), even though one of only when two imperfect classification
the variables systematically overestimates or schemes are being compared, not when one
underestimates values of the other variable. classification method may be considered a pri-
The high correlation means that the over- or ori superior to the other. The κ statistic is the
underestimation is systematic and very con- percentage agreement corrected for chance.
sistent. When the two measures being com- Agreement is conventionally considered poor
pared are plotted against each other and for a κ statistic less than zero, slight for κ
they have the same scale, full agreement between zero and 0.20, fair for a κ of 0.21–0.40,
occurs only when the points fall on the line of moderate for a κ of 0.41–0.60, substantial for a
κ of 0.61–0.80, and almost perfect for a κ of of these factors may incorrectly identify a risk
0.81–1.00. Figure 13.2 illustrates the percent- factor in the study that does not exist in the
age agreement and κ calculations for a reliabil- population or, conversely, may fail to detect a
ity assessment between questionnaire data and risk factor when one truly exists.
medical record information. For example, when questionnaire data are
The intraclass correlation coefficient is used used to study the association between drug A
to evaluate the reliability of continuous varia- and disease B, study participants who forgot
bles. It reflects both the average differences in their past exposure to drug A would be incor-
mean values and the correlation between meas- rectly classified as nonexposed. Similarly, if a
urements. The intraclass correlation coefficient provider uses a diagnosis code to document the
indicates the degree to which the total meas- process of testing and ruling out a disease and
urement variation is due to the differences then a researcher uses the diagnosis code as a
between the subjects being evaluated and to study outcome, then the person would be
differences in measurement for one individual. incorrectly classified as having the outcome.
When the data from two sets of measurements This misclassification is a measurement error.
are identical, the intraclass correlation coeffi- Although the measurement process often
cient equals 1.0. Under certain conditions, the involves some error, if this measurement error
intraclass correlation coefficient is exactly is of sufficient magnitude, the validity of the
equivalent to Cohen’s weighted κ. study’s findings is diminished.
It is impossible to translate values of meas- Surprisingly, measurement error is often
ures of agreement, such as κ, into expected ignored in epidemiologic studies. Jurek et al.
degrees of bias in exposure or disease reported the results of a random survey of
associations. studies published in three major epidemiology
journals; they concluded the following for
exposure-measurement error (EME): “Overall,
Measurement Error
the potential impact of EME on error in epide-
in Pharmacoepidemiologic
miologic study results appears to be ignored
Research
frequently in practice (page 871).”
Epidemiologic assessments of the effects of a Measurement error is a potentially serious
drug on disease incidence depend on an accu- cause for concern in epidemiologic studies,
rate assessment of the study exposure, disease and therefore, for several reasons, should not
occurrence, and variables to be adjusted in the be ignored when analyzing and interpreting
statistical analysis. Measurement error for any pharmacoepidemiologic study results. First,
Medical record
Exposed Not exposed
t
Exposed A B m1
Questionnaire
data
Not exposed C D m2
n1 n2 N
Accuracy = A + D/N
Chance agreement (expected)=((n1 × m1 ) + (n2 × m2))/N2
accuracy – chance agreement
κ=
1– chance agreement
Figure 13.2 Formulas for calculating the percent agreement and κ.

small amounts of measurement error can measuring the exposure and outcome are
cause large amounts of error in study results. not independent).
For example, consider a pharmacoepidemio- ●● Even when the above conditions beyond
logic study of NSAID A versus NSAID B on GI nondifferentiality are met, exact nondiffer-
bleed (Figure 13.3). In a study with a total entiality is required to guarantee bias is
number of study subjects equal to more than toward the null.
22 000, if only 10 subjects are misclassified ●● Also required to guarantee bias is toward the
with respect to their exposure or disease (five null is either (i) the absence of other study
with GI bleed who actually took NSAID B are biases (e.g. absence of confounding, absence
incorrectly classified as having taken NSAID of bias due to nonrandom subject selection/
A, and five users of NSAID A without GI bleed participation), or (ii) the combined effect of
are incorrectly classified as having GI bleed), all other biases is also toward the null.
the observed odds ratio (OR) would be ●● Bias is a statistical term that is defined as the
2.1 when the correct OR is in fact 1.0. difference between the true value and the
Second, measurement error can cause study expected value of an estimator (i.e. the aver-
results to overestimate or underestimate true age of study results over hypothetical repeti-
effect sizes, and there is no simple rule for pre- tions of the study). Bias is not the difference
dicting the direction of the error in real-life between the observed estimate for one repeti-
situations. We now understand that these old tion of the study and the true value. This
and often-cited heuristics are not necessarily important distinction was not appreciated in
true, except under special conditions that are earlier writings on this topic, and even today
not likely to occur in practice: (i) “nondifferen- we epidemiologists are not careful in our use
tial misclassification always produces bias of the term bias. Therefore, when bias is
toward the null,” and (ii) “bias toward the null toward the null, the expected value of the esti-
always produces an observed relative risk that mator is shifted toward the null, but an
is an underestimate of the true relative risk.” observed estimate can be an overestimate of
These heuristics are unlikely to be true in prac- the true relative risk due to the influence of
tice for the following reasons: random error. (Similarly, when there is no
bias of any kind, one observed estimate can
●● Conditions beyond nondifferentiality are
be an overestimate or an underestimate of the
required to guarantee bias is toward the
true relative risk simply due to random error.)
null (e.g. when the degree of exposure
measurement error systematically differs Third, error in measuring variables to be
across levels of a polychotomous or contin- adjusted in the analysis can result in only par-
uous exposure variable, or when errors in tial adjustment for the mismeasured variables.
Observed data (with 5 NSAID B cases misclassified as having taken

NSAID A, and 5 without GI bleed misclassified as GI bleed)
NSAID A NSAID B
GI Bleed 20 95
Not GI Bleed 1,995 20,000 Observed OR = 2.1
Data after correcting for measurement error

NSAID A NSAID B
GI Bleed 10 100
Not GI Bleed 2,000 20,000 True OR = 1.0
Figure 13.3 Small amount of measurement error can cause large error in study results.
Adjusting Measures of Association would have remained qualitative and unsub-

for Measurement Error stantiated. An important and well-known his-
torical example is the bias from nondifferential
One can use sensitivity analysis methods (also
misclassification of disease proposed by
known as uncertainty analysis and bias analy-
Horwitz and Feinstein to explain associations
sis) to adjust measures of association for
between early exogenous estrogen prepara-
measurement error as well as for other study
tions and endometrial cancer. When proper
biases. (As used in this context, the meaning
sensitivity analyses were conducted, only a
of the term sensitivity differs from its other epi-
negligible proportion of those associations
demiologic meaning as the counterpart to
were explained by bias.
specificity as a measure of classification valid-
Epidemiologic applications of quantitative
ity.) Sensitivity analysis is the last line of
methods with a long history in the decision sci-
defense against biases after every effort has
ences have become accessible for quantifying
been made to eliminate, reduce, or control
uncertainties about multiple sources of sys-
them in study design, data collection, and data
tematic error in a probabilistic manner. These
analysis. In a sensitivity analysis, one varies
methods permit incorporation of available
key assumptions or methods reasonably to see
validation data, expert judgment about meas-
how sensitive the results of a study are to
urement error, uncontrolled confounding (see
those variations.
Chapter 22), and selection bias, along with
One key assumption, usually implicit, in any
conventional sampling error and prior proba-
study that does not quantitatively account for
bility distributions for effect measures them-
the possibility of error in measuring the study
selves, to form uncertainty distributions. These
exposure or study outcome, is that the expo-
approaches have been used practically in phar-
sure and the outcome in a study have been
macoepidemiology studies such as in assessing
measured accurately. With estimates of sensi-
selection bias in a study of topical coal tar ther-
tivity and specificity from validation studies
apy and skin cancer among severe psoriasis
(from previous research or from a subsample
patients; exposure misclassification and selec-
within the study analyzed) or “guesstimates”
tion bias in a study of phenylpropanolamine
from expert experience and judgment, one can
use and stroke; selection bias, confounder mis-
modify this assumption and use sensitivity
classification, unmeasured confounding in a
analysis methods to “back calculate” what the
study of less than definitive therapy and breast
results might have looked like if more accurate
cancer mortality; and other clinical and non-
methods had been used to classify participants
clinical applications.
with respect to outcome, exposure, or both.
Sometimes biases can be shown to be of
For many years, a qualitative and informal
more concern and sometimes of less concern
version of this kind of assessment has been
than intuition or simple sensitivity analysis
conducted. However, the net result is contro-
might suggest. Using these methods, assess-
versy, with investigators judging the bias small
ment of systematic error can move from a
and critics judging it large. Further, in the
qualitative discussion of “study limitations,”
absence of a formal sensitivity analysis, intui-
beyond sensitivity analyses of one scenario at a
tive judgments, even those of the most highly
time for one source of error at a time, to a com-
trained and widely experienced investigators,
prehensive analysis of all sources of error
can be poorly calibrated in such matters.
simultaneously. The resulting uncertainty dis-
Formal sensitivity analysis makes the assess-
tributions can not only supplement but also
ment of residual bias transparent and quanti-
supplant conventional likelihood and p-value
tative, and forces the investigator (and other
functions, which reflect only random sampling
critics) to defend criticisms that in earlier times
error. As a result, much more realistic,
robabilistic assessments of total uncertainty

p studies indicate that people accurately remem-
attending to effect measure estimates are ber ever having used a medication and when
available. they first began using it, for some medications,
although they do not remember brand names
and duration of use as well. In general, inac-
Self-Reported Drug Data from
curacies correlated with more time elapsed
Ad hoc Survey Studies: Recall Accuracy
between occurrence of exposure and its subse-
The methodological literature on recall accu- quent reporting. Accuracy of self-reporting
racy discussed above indicates that study par- varies by medication. For example:
ticipants have difficulty remembering drug use
●● Chronically used medications (especially
from the distant past, which contributes to
those with more refills) being recalled more
misclassification of exposure in de novo stud-
often than acute exposures.
ies. Researchers are using best practices in
●● First and most recent brands in a class are
questionnaire design, including medication-
recalled more frequently than other medica-
specific and indication-specific questions,
tions in the class.
along with recall enhancements, which have
●● Multiple medications in one class are
been shown to produce better data. Calendars
recalled more frequently than single medica-
and photos of drugs augment recall to a greater
tion exposure.
degree than listing only the brand names of the
●● Salient exposures (those that prompted
drugs in question. These techniques – namely,
study initiation) are more accurately recalled
photos, calendars, and the two different types
than common and less disconcerting
of drug questions – have become the state of
medications.
the art for collecting self-reported drug data by
personal or telephone interview. For prescription drugs, recall between
The literature to date suggests that recall self-reported use and medical records was
accuracy of self-reported medication expo- moderately accurate, but over-the-counter
sures is sometimes, but not always, influenced medications and vitamin supplements had
by type of medication, drug-use patterns, poorer recall. Discrepancies were due to both
design of the data collection materials, and underreporting (e.g. respondent forgot that the
respondent characteristics. Given the current medication was taken) and overreporting (e.g.
state of the literature, epidemiologists who physician record patient’s use in chart from
plan to use questionnaire data to investigate previous visit, even though patient stopped
drug–disease associations will need to consider medication) and differed by therapeutic class.
which factors may influence recall accuracy in When self-reported data were compared with
the design of their research protocols. Case multiple sources, such as medical records and
example 13.1 summarizes the published litera- pharmacy dispensing, verification of self-
ture on studies that have validated drug expo- reported use was higher than for a single
sure information (NSAIDs) on health outcomes source.
(MI and GI bleeding) from questionnaire data.
The Influence of Questionnaire
The Influence of Medication Class Design
Several studies have compared self-reported As reported in a recent systematic review, sev-
recall accuracy for current or past medication eral factors affect the accuracy of medication
use with prospectively collected cohort data or exposure reported via questionnaire.
pharmacy, hospital, and outpatient medical Researchers can facilitate recall and reporting
record documentation. Overall, published of medication use by indication-specific ques-
tions, memory prompts (such as drug photo), a drugs, and 59% for three or more drugs,
list of drug names, or a calendar to record life although duration of use was not related to
events. Medication-specific or indication- recall. However, the questionnaire did not
specific questions can identify most medica- allow sufficient space to record all medica-
tions in current use, but a general medication tions used in the time period for this study.
question, such as “Have you taken any other Thus, if respondents were unable to record
medications?” failed to identify all medica- all medications due to space limitations, a
tions respondents were currently taking. misleading validation might have occurred:
Similarly, open-ended questions such as “Have it appeared that respondents were unable to
you ever used any medications?” yielded less recall all their medications dispensed,
than half of the affirmative responses for according to the database.
actual use of three different medications. Another methodological study evaluated
Using the filter question (“Did you use any whether question structure influenced the
medications in the three months before or dur- recall of currently used medications in 372
ing your pregnancy?”), van Gelder and col- subjects with hypertension who had at least
leagues noted that many women failed to 90 days of dispensing in the PHARMO data-
report medications that they had been dis- base. The questionnaire asked indication-
pensed for pain or infections. These findings specific questions first (e.g. medications used
could be attributed to poor recall, but they may for hypertension, medications used for diabe-
be also due to women having chosen not to tes), followed by an open-ended question that
take the dispensed medications. If researchers asked whether the participants used any other
choose to use open-ended medication ques- medications not already mentioned. For hyper-
tions, adding indication-specific questions that tension, the sensitivity was 91% for indication-
facilitate recall of medication exposures may specific questions and 16.7% for open-ended
be useful. Finally, 20–35% of respondents questions. About 20% of participants listed
reported drug exposure only when asked medications on the questionnaire that were
medication-name-specific questions. not in the database; a similar proportion failed
Response order may affect recall, as noted to list medications on the questionnaire that
with malaria medications when respondents were in use according to the pharmacy data-
had more than one episode of malaria. base. Based on these recall sensitivity results,
Medications listed earlier tended to be selected indication-specific questions invoke better
more frequently than those listed later – a find- recall accuracy. However, to adequately assess
ing that may be related to “satisficing,” which question structure, a questionnaire could be
occurs when respondents expend the least psy- designed to ask open-ended questions before
chological and emotional effort possible to pro- indication-specific questions. This sequencing
vide an acceptable answer to a survey question would allow a comparison of the number of
rather than an optimal answer. medications recalled by each question
A comparison of self-report for current structure.
and recent medication use (within the past
two years) with pharmacy records of dis-
The Influence of Patient
pensed prescriptions for multiple drug
Population
classes found that recall of the number of
drug dispensingswas highest for cardiovas- Few studies have evaluated whether demo-
cular medications (66%) and poorest for ali- graphic and behavioral characteristics influ-
mentary tract medications (48%). Recall was ence the recall of past medication use.
influenced by the number of regularly used Research suggests that education attainment
medications: 71% for one drug, 64% for two and race/ethnicity may affect recall accuracy.
Studies are inconsistent for age, socioeco- Both overreporting and underreporting was
nomic status, and smoking as predictors of noted for cardiovascular conditions, depend-
recall accuracy, and no study found that recall ing on the data source used for comparison. In
accuracy varies by gender. The inconsistencies most instances of recall error, respondents
on the effect of age on recall accuracy might who had incorrectly reported MIs and stroke
arise from differing study designs. The two had other conditions that they may have mis-
studies that reported an age effect were meth- takenly understood as coronary heart disease,
odological studies evaluating recall accuracy, MI, or stroke. Underreporting was the primary
whereas the two that reported no age effects reason for poor agreement comparing inter-
were etiologic studies that reported verifica- view data to clinical evaluation, although it is
tion of drug use as a measure of exposure mis- unclear whether this is due to the respondent’s
classification for the association under study. unwillingness to admit to mental illness or
Because of the paucity of information on pre- underdiagnosis of the conditions.
dictors of recall, further research in this area is
warranted.
The Influences of Timing
of Diagnosis and Its Emotional
Self-Reported Diagnosis Effects on the Patient
and Hospitalization Data from Factors influencing accuracy of past diagnoses
Ad hoc Studies: Recall Accuracy and hospitalizations also include the number
Just as recall accuracy of past medication use of physician services for that condition and the
varies by drug class, the recall accuracy of dis- recency of services. For reporting of diagnoses,
ease conditions varies by disease, particularly the longer the interval between the date of the
when it is chronic, like hypertension, or is last medical visit for the condition and the date
viewed as threatening, such as sexual transmit- of interview, the poorer the recall for that con-
ted infections. dition. These differences in recall may be
explained in part by recall interval, patient age,
a cohort (generational) effect, or some inter-
The Influences of Medical
twining of all three factors. Diagnoses consid-
Condition Type
ered sensitive by one generation may not be
Comparing patient self-report with data from a considered as such by subsequent generations.
provider questionnaire (gold standard) on pre- Further, terminology changes over time, with
vious history of cardiovascular disease or GI prior generations using different nomencla-
events, researchers found better agreement for ture compared with recent generations.
previous acute MI than for upper GI bleeding Conditions with substantial impact on a per-
(see Case Example 13.1). son’s life are more accurately reported than
●● The best reporting has been noted with con- those with little or no impact on lifestyle. More
ditions that are specific and familiar, such as patients with current restrictions on food or
diabetes mellitus, hypertension, asthma, and beverage due to medical problems reported
cancers such as breast, lung, large bowel, chronic conditions that were confirmed in
and prostate. However, assessing reporting medical records than did those without these
accuracy is more difficult for common, restrictions. Similarly, those who had restric-
symptom-based conditions such as sinusitis, tions on work or housework reported their
arthritis, low back pain, and migraine head- chronic conditions more often than those who
aches, which many people may have, or did not have these restrictions. The major deter-
believe they have, without having been diag- minant of recall for spontaneous abortions was
nosed by a clinician. the length of the pregnancy at the time the
Case Example 13.1 Fourrier-Reglat et al. (2010)
Background Results
●● Researchers may have to query the subject ●● Agreement between patients and prescrib-
to assess medication exposure and out- ers was substantial for MI (κ = 0.75, 95% CI:
come diagnoses. Accuracy of ad hoc ques- 0.71–0.80), and minimal for upper GI
tionnaire studies has been determined via bleeding (κ = 0.16, 95% CI: 0.11–0.22).
comparison with pharmacy, practitioner, ●● With prescriber data as the gold standard,
and hospital records. patient reports of MI provided moderately

●● In some cases, pharmacy or medical complete data (sensitivity: 77.7%; specific-
records may not be available or there may ity: 99.6%, PPV: 74.1%, NPV: 99.6%), but
be reason to question both the patient reports of upper GI bleeding by patients
and practitioner rather than conducting were less well documented in the pre-
medical record review. In addition, ques- scriber reports (sensitivity: 44.6%; speci-
tionnaires can provide concurrence ficity: 98.5%, PPV: 10.4%, NPV: 99.8%).
regarding patient history and indication ●● For the index NSAID indication, the pro-
for use, which are not available from portion of agreement ranged from 84.3%
claims data or easily found in many medi- to 99.4%, and concordance was almost
cal records. perfect (k = 0.81–1.00).
Question Strengths
When questionnaires are self-administrated, ●● Concurrent evaluation of patients and pre-
is there concordance of patient-derived and scribers allows corroboration of patient
physician-derived data on medical informa- history and indication for medication
tion such as previous medical history and usage, which are difficult to assess in
initial indication for NSAID prescriptions? claims and electronic health care records.
●● Self-
administered questionnaires provide
Approach
data on a list of potential confounders that
●● The Kappa statistic (κ) was used to
are often missing from electronic records.
measure concordance in self-
administered questionnaires completed Limitations
by 18 530 pairs of NSAID patients and ●● Study was carried out in an established
their prescribers for the French national cohort in a single country. Results may not
cohort study of NSAID and Cox-2 inhibi- be generalizable to other data in other
tor users. locations.
●● Both patients and prescribers were asked ●● Neither the patient nor the prescriber
about patients’ previous history of cardio- reports “truth”; accuracy of recall was not
vascular events, including MI, and GI events, assessed in this study.
including upper digestive hemorrhage.
Key Points
Patients and prescribers were asked to iden-
●● Prior history and indication for prescrip-
tify which of the following was the initial
tions are often missing from claims and
indication for NSAID use: rheumatoid arthri-
electronic health care record databases.
tis, psoriatic rheumatism, spondylarthritis,
Questionnaires from both patients and pre-
osteoarthritis, back pain, muscle pain/
scribers may collect data on these potential
sprain/tendonitis, migraine/headache, flu-
confounders, as well as determine the reli-
like symptoms, dysmenorrhea, or other
ability of the collected variables.
indication.
(Continued)
●● Patients and prescribers may have differ- over-the-counter medication use. However,
ing recall of patient history, especially as corroboration with another information
related to nonspecific diagnoses. source, such as data from the prescriber,
●● Relying on patient-reported data may be may provide an estimate of the reliability
necessary but inaccurate, especially for of patient-reported data.
event occurred: nearly all respondents who t horoughly evaluated, although the results are
experienced spontaneous abortions occurring conflicting. The most consistent finding is that
more than 13 weeks into the pregnancy remem- recall accuracy decreases with age, although
bered them, compared with just over half of this may be confounded by recall interval, or
respondents who experience such abortions cohort (generational) effects. Whether gender
occurring in the first six weeks of pregnancy. influences recall accuracy is uncertain. Men
Perhaps as a result of the emotional stress, have been found to report better than women,
lifestyle changes, and potential financial strain, independent of age, whereas conflicting evi-
hospitalizations tend to be reported accurately. dence found that women reported better than
Further, underreporting of hospitalizations men, especially in older age groups. Further
occurred in 9% of patients when surgery was studies indicate that gender and age differences
performed, compared with 16% of patients depended on the disease under investigation,
without a surgical procedure. Underreporting with women overreporting malignancies and
in those with only a one-day hospital stay was men overreporting stroke. No differences were
28% compared with 11% for two to four day found for reporting of hospitalizations by age or
stays and approximately 6% for stays lasting gender.
five or more days. Reporting of illnesses, procedures, and hos-
Researchers also agree that respondents pitalizations tends to differ by race/ethnicity,
remember the type of surgery accurately. but most studies had much larger proportions
Recall accuracy was very good for hysterec- of whites than nonwhites. Reporting by educa-
tomy and appendectomy, most likely because tional level was equivocal and was more com-
these surgeries are both salient and familiar to plete for self-respondents than for proxy
respondents. For induced abortions, marginal respondents. Those with a poor or fair current
agreement occurred, as noted by records from health status reported conditions more com-
a managed care organization: 19% of women pletely than those with good to excellent health
underreported their abortion history, 35% status.
overreported abortions, and 46% reported
accurately, according to their medical records.
The Influence of Questionnaire
Cholecystectomy and oophorectomy were not
Design
as well recalled and were subject to some over-
reporting. However, apparent overreporting Questionnaire design also influences the valid-
may have been due to possible incompleteness ity of disease and hospitalization data obtained
of the medical records used for comparison. by self report. Providing respondents with a
checklist of reasons for visiting the doctor
improves recall of all medical visits. Simpler
The Influence of Patient Population
questions yield better responses than more
The influence of demographic characteristics complex questions, presumably because com-
on reporting of chronic illness has been plex questions require the respondent to first
Currently Available Solution 233
comprehend what is being asked and then should attend to the cognitive processes
provide an answer. Inherent redundancy in involved in developing a response, especially
longer questions and greater time allowances those related to saliency for the respondent.
to develop an answer may increase recall; how- The typical rule of thumb for question sequenc-
ever, longer questions may tire the responding is to ask general questions before delving
ents, leading to satisficing, as well as increase into specific topics and to group questions
the cost of the research. according to topic. Response categories should
be unambiguous, nonoverlapping, and exhaus-
tive. When there is a possibility of biased
Currently Available response due to response ordering, it is best to
Solutions randomize the response options to minimize
the bias. Finally, satisficing is also possible
Following Best Practices when respondents are asked to identify the
for Questionnaire Design diagnoses they have been given previously.
With the increasing availability of broad-
Designing a questionnaire to collect epidemio-
band and the population’s access to the
logic data requires careful planning and pre-
Internet, more surveys are moving away from
testing. We suggest the following steps be
face-to-face and telephone interviewer admin-
considered during the design and initial analy-
istration to web-based surveys. This modality
sis stages of a study requiring data collection
requires the same considerations for question
via questionnaire:
design as described above, but because no
1) Use validated instruments or validated interviewer is available, usability and how effi-
questions whenever possible. ciently and effectively respondents can answer
2) Consider question banks if new questions the web-based questions, should be tested as
are required, such as World Bank’s Living well.
Standards Measurement Study and Q-Bank. To appreciate the accuracy of data derived
3) Use question assessment tools to determine by respondent recall, it is important to
the likelihood of response error. These tools understand the how people process, organ-
include the Question Appraisal System, the ize, and recall autobiographical information,
Survey Quality Predictor (SQP), and the which is key to the response process.
Question Understanding Aid (QUAID). Creating and retrieving information from
4) Strive for a fifth-grade literacy level if you autobiographical memories is a three-step
must develop new survey questions to be process. Information that comes in via sen-
used for a general population. sory or emotional input (e.g. visual, hearing,
5) Pretest questions using cognitive testing to semantic) is encoded into a construct that
assess respondent comprehension of new can be stored within the brain. The next step
questions. is storage, which refers to how the brain
retains the information, typically in either
short- or long-term memory. Retrieval or
Developing a De novo
recall of memories requires re-accessing
Questionnaire
information that was previously encoded
Although specific guidance on best practices and stored. The recall of encoded or cata-
for improving the ascertainment of diagnoses loged information from memory is thought
and hospitalizations is lacking, researchers to be facilitated by using important personal
developing questionnaires should be mindful milestones. Thus, when respondents are
of question wording and sequencing and asked to recall a visit to a doctor that may
response formats. Questionnaire designers have occurred at a particular point in time,
researchers believe that the respondents use The following example illustrates the
scripts (a generic mental representation of response process for recalling the date on
the event) to help retrieval. In general, which a respondent’s depression was diag-
underreporting of medical conditions and nosed (January 2015). The recall process
health visits is more widespread as the inter- begins with the respondent being uncertain
val since the event increases. Applying what whether the depression was diagnosed in 2014
we know about how autobiographical mem- or 2015. To work toward identifying the cor-
ory is organized and the recall process in rect year, the respondent recalls that the depres-
general helps us to understand survey sion occurred after he lost his job. The job loss
response. A respondent undergoes four key was particularly traumatic because he and
tasks when asked to answer a questionnaire: his wife just purchased their first home a few
months previously, and now, with the loss of
●● Question comprehension and interpretation.
his income, they were at risk of losing the
●● Search for and retrieval of information to
house. The home purchase was a landmark
construct an answer to the question.
event for this respondent, and he remembers
●● Judgment to discern the completeness and
that it occurred in mid-2014, just as their chil-
relevance of memory for formulating a
dren finished the school year. So, in 2014 he
response.
lost his job, near the end of the year because
●● development of the response based on
the holiday season was particularly grim. He
retrieved memories.
remembers that his depression was diagnosed
If survey instrument developers pay too little after the holidays, but was it January or
attention to the first two key tasks, their ques- February of 2015? It was January 2015
tions may be too vague or complex for respond- because he was already taking antidepres-
ents to marshal retrieval processes sants by Valentine’s Day, when he went out to
appropriately. For discussion of the theory of dinner with his wife and he could not drink
survey response and the cognitive process wine with his meal. This chronology is dia-
underlying retrieval, see Tourangeau et al. grammed in Figure 13.4. We describe below,
(2000). how to use the response process to design
When was your depression first diagnosed? The respondent knew it was in
either 2014 or 2015 but could not remember when. The depression occurred
because he lost his job.
2014 2015
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul
Kids finished Grim

school holidays Valentine’s
Day
Moved into Lost job
new home
Depression diagnosis
January 2015
Figure 13.4 Recall schematic for showing how date of depression diagnosis was determined.
questions to elicit the self-reported informa- Conducting Validation Studies

tion requested. to Assess Self-Reported Data
As illustrated in Figure 13.4, landmark events
Methodological studies that use alternative
probably serve as the primary organizational
data sources such as prospectively collected
units of autobiographical knowledge and
data or databases of dispensed drugs can meas-
anchor information retrieval. In particular, the
ure both sensitivity and specificity, if one
example shows how the respondent used land-
assumes that the prescription database is a
mark and other notable events, relationships
gold standard. In pharmacoepidemiology,
among datable events, and general knowledge
lower sensitivity is often more of a concern
(holiday period and children finishing the
than lower specificity. Questionnaires that
school year) to reconstruct when his major
underreport diseases or miss drug exposures
depression was first diagnosed. An important
because the medication was filled without
caveat is that this respondent described above
using the pharmacy plan (e.g. when the co-pay
was willing to expend considerable effort
is higher than the cost of the medication) –
searching his memory to determine when his
that is, data sources with low sensitivity – can-
depression was diagnosed, which may not be
not be used to rigorously evaluate drug-disease
true for all respondents.
associations. Alternatively, low specificity is
The process of satisficing occurs when
often less of a problem in pharmacoepidemiol-
respondents expend the least psychological
ogy unless the characteristic with low specific-
and emotional effort possible to provide an
ity also has very low prevalence in the
acceptable answer to a survey question rather
population being studied. For example,
than an optimal answer. To minimize satisfic-
because the incidence of Stevens–Johnson
ing, questionnaire developers should consider
syndrome is rare, a small degree of misclassifi-
the length of the instrument and the number
cation when using administrative claims data
of response categories. When faced with a long
in which the case definition uses the ICD-
list of choices and depending on the mode of
9-CM code 695.1 will include several skin
questionnaire administration (i.e. telephone
problems other than Stevens–Johnson (i.e. the
versus self-administered), respondents may
false-positive rate will be high).
choose answers at the top or bottom of the list
Besides the need for completeness on the
to minimize effort. Respondents with lower
individual level, the comparator database must
cognitive skills and less education, when chal-
have information for all persons whose infor-
lenged with discerning the best possible
mation is to be assessed for accuracy.
response, are more apt to settle for a satisfac-
Systematic omissions of specific population
tory rather than an optimal response. Because
groups, such as certain ethnic or racial groups,
accuracy of response is critical for pharma-
diminish the quality of the database.
coepidemiologic research, questionnaire
developers must consider methods to mini-
mize response burden leading to satisficing.
Considering the Influence
In addition to survey design and respond-
of Comparator Selection
ent motivation, measurement error can be
on Validation Studies
attributed to improper training of interview-
ers and poor data entry quality. The degree to Regardless of whether the medical record is
which one understands the measurement paper or electronic, one needs to understand
error associated with key variables is critical its availability, completeness, and accuracy to
to the analysis can be assessed using several determine whether it is adequate for evaluat-
different modeling approaches, which ing the accuracy of self-reported information.
Biemer, 2009 discusses in more detail. Retrieval of medical records depends not only
on a person’s ability to remember and report databases are available for pharmacoepidemio-
who prescribed the drug or diagnosed the con- logic research, the structure, strengths, and
dition in question, but on whether the health limitations of which were reviewed in
care provider recorded the information (and Chapters 8–10. One major advantage of using
recorded it accurately) and on the availability such databases for pharmacoepidemiologic
of the medical record for review. If the medical research is the comparative validity of the drug
record cannot be retrieved because the health data in lieu of questionnaire data, where recall
care provider could not be identified, the pro- bias is of concern, as previously described.
vider had retired, or their record was destroyed The drawbacks and limitations of these data
or lost, the events cannot be verified. systems are important. Their most critical limi-
In the US, health care is fragmented. Patients tation for pharmacoepidemiologic research is
see multiple providers, are treated in several dif- the manner in which health insurance is cur-
ferent health settings (e.g. chiropractors, podia- rently covered in the United States, typically
trists), and may become inpatients at several through the employer. If the employer changes
different hospitals. Thus, accessing patients’ plans, which may occur annually, the employee
outpatient and inpatient medical records does changes among the plans offered by the
not guarantee that a researcher will have all employer, or the employee changes jobs, then
medical care provided and drugs prescribed to the plan no longer covers that employee or his
the patient. For example, if a researcher is able or her family. Thus, the continual enrollment
to access only the patients’ primary care records, and disenrollment of plan members hinders
it is possible that the results of cardiology tests the opportunity for extended longitudinal
to confirm a diagnosis or medications for that analyses in both administrative claims and
diagnosis are not available. However, within EHRs.
integrated delivery systems that include pri- Diagnoses, procedures, medications, and
mary care, multiple specialties, and inpatient other therapeutics are included in administra-
care, there is a greater likelihood that the EHR tive claims and EHR data through structured
will contain most of the care provided and med- coding systems. Each coding system has its
ications prescribed to the patient. own ontology and is separated into specific
In addition, exposure information about codes, based on an established hierarchy.
medications or important confounders (e.g. Further, the coding systems are updated peri-
smoking) may be incomplete if clinicians do odically to reflect changes in the practice of
not ascertain this information and correctly health care as well as to incorporate new thera-
enter it into the EHR. Another problem intro- pies and processes. Both codes and the general
duced by EHRs is the potential for errors inher- structure and hierarchy differ between coding
ent to electronic data entry, such as copying systems. In many cases, a single code is insuf-
and pasting of incorrect data from other parts ficient to define a variable and an algorithm,
of the record, of expired or irrelevant clinical with required timing of codes or a sequential
information, or of incorrect or unverified med- process for determining the level(s) of the vari-
ication lists. able, is needed. Algorithms developed in one
coding system likely require translation to be
useful in another coding system.
Validation of Pharmacoepidemiologic
In addition to the structured data, many
Drug and Diagnosis Data from
aspects of health care are captured within cli-
Electronic Encounter Databases
nician notes, images and descriptions of proce-
In addition to conducting de novo studies to dure results, and other unstructured data.
evaluate drug–disease associations, a variety of Performance of an algorithm can be enhanced
computerized, administrative claims and EHR through use of this unstructured information,
by converting it into structured information medication to be adherent throughout a

(e.g. manually) for a specific project, or to 12 month period; this finding may be consist-
modify and improve algorithm performance as ent with intermittent or “as needed”
cases are identified over time (e.g. machine utilization.
learning). Previously developed algorithms are In summary, drug and medical intervention
sometimes used for comparison, however data are often considered to be correct when
patient charts (electronic or paper) are still using administrative claims data and EHRs for
often used as the reference standard for assess- research. Although this is generally the case,
ing validation. researchers should be aware of whether and
Completeness and validity of data are the how prescribing, dispensing, and administra-
most critical elements in the selection of a tion of drugs are captured within each data-
database for research. Completeness is defined base they are contemplating using. We will
as the proportion of all exposures, events of likely see greater emphasis on data linkage and
interest, or both that occurred in the popula- incorporation of more unstructured data from
tion covered by the database that appear in the clinical notes into pharmacoepidemiologic
computerized data. Missing subjects, expo- research, which may lead to increased need for
sures, or events could introduce bias in the validation of drug and medical intervention
study results. For example, completeness of exposures in the future.
the drug data may vary by income level if per-
sons with higher incomes and drug copay-
Special Considerations with
ments choose to obtain their medications at
Diagnosis and Hospitalization Data
pharmacies not participating in a prescription
plan, which is how pharmacy data are col- Unlike drug data where many researchers are
lected. Similarly, bias may be introduced in the comfortable with accuracy and completeness,
association between a drug and a serious inpatient and outpatient diagnoses in these
adverse drug reaction if hospitalizations for databases raise considerable concern for
that adverse reaction are missing, e.g. if the investigators. The accuracy of outpatient diag-
researcher only has access to outpatient clinic noses is more uncertain than inpatient diag-
data in the EHR. noses for several reasons. Hospitals employ
experienced people to code diagnoses for
reimbursement, which may not occur in phy-
Special Considerations with Drug
sicians’ offices where outpatient diagnoses are
Data
determined. Moreover, hospital personnel
A handful of studies to date have assessed dis- scrutinize inpatient diagnoses for errors, mon-
pensing associated with prescriptions via itoring which does not typically occur in the
linked administrative claims and EHR data. outpatient setting.
These studies indicated that 70–77% of initial Systematic errors as a result of diagnostic
prescriptions are dispensed. Prescribed analge- coding may influence the validity of both inpa-
sics (i.e. pain medications, including NSAIDS) tient and outpatient diagnostic data. For exam-
and lifestyle drugs (e.g. phosphodiesterase type ple, diseases listed in administrative claims
5 inhibitors) are least likely to be dispensed, databases are often coded using the
while antimicrobials are most likely to be dis- International Classification of Disease (ICD)
pensed for an initial prescription. Substantial system. Poorly defined diseases are difficult to
variation in dispensing was seen across medi- code using the ICD system, and no way exists
cations within a class. In addition, results from to indicate that an ICD code is coded for “rule-
Rowan suggests that <20% of patients taking out” purposes. How health care plans address
analgesics and NSAIDS possessed adequate “rule-out” diagnoses is unclear; they likely do
Case Example 13.2 Bustamante et al. (2019)
Background ○○ Identified common terms from progress

●● A variety of electronic administrative data- notes, included character distance
bases are available for pharmacoepidemi- between “aspirin” and dose indicators –
ologic research. One major advantage of identified terms consistent with aspirin
using such databases for pharmacoepide- use and non-use
miologic research is the comparative ○○ Combined structured and unstructured
validity of the drug data in lieu of ques- data

tionnaire data, for which recall bias is ●● Assessed sensitivity, specificity PPV, and
always a concern. NPV of charts compared to manual chart
●● Algorithms using structured and unstruc- review of free- text progress notes and
tured data may have differing validity. pharmacy data
Unstructured data may provide useful
Results
information for identifying an over- the-
Table 13.1 presents the sensitivity, specific-
counter medication.
ity, PPV, and NPV for the three algorithms.
Question The sensitivity dropped (minimally) as a
Within the Veteran’s Affairs (VA) administra- wider array of data was incorporated into the
tive data, what is the reliability and validity of algorithm, while the specificity, PPV, and NPV
various measures of dispensing for aspirin? increased. Using both structured and
unstructured data led to the highest speci-
Approach
ficity, PPV, and NPV.
●● A retrospective cohort database study was
conducted of 1 869 439 veterans who Strengths

underwent colonoscopy within the VA in ●● Evaluation of sensitivity, specificity, PPV,
1999–2014 to develop and evaluate three and NPV provides the full picture of relia-
algorithms for identifying which patients bility and validity for algorithms assessing
received aspirin (over-the-counter or pre- both prescription and over- the-
counter
scription). Charts reviewed to develop medication use.
algorithms (number not stated); tested ●● Multiple predictive algorithms were
algorithms in 100 charts categorized as assessed, showing the improvement in
having aspirin use and 100 charts catego- validity with the use of unstructured data
rized with non-use. (i.e. progress notes).
○○ Structured data from VA Meds: use of ●● Data were from routine clinical and billing
aspirin documented as medicine taken processes, so no additional burden of data

and allergy to aspirin assumed to con- collection for research purposes was
firm non-use of aspirin necessary.
Table 13.1 Sensitivity, specificity, PPV and NPV for three algorithms.
Sensitivity (%) Specificity (%) PPV (%) NPV (%)
Structured data 98 61 94 80
Unstructured data 98 95 95 98
Both structured and unstructured data 96 100 99 98
Limitations Key Points

●● The study was conducted in one database ●● Validation of the specific codes or algo-
(VA) only; validation results for the algorithm used to assess a medication or treat-
rithm may not be applicable to different ment can improve capture of
databases or patient populations. over-the-counter medication use.
●● Algorithms did not include other measures ●● Inclusion of unstructured data may sub-
of patient condition, severity or treatment. stantially increase specificity, but it does

Additional factors (e.g. age, diagnoses) may so at the expense of sensitivity – this may
improve the reliability and validity of be warranted with specificity increases
algorithms. substantially with only a small decrease in
sensitivity, especially when PPV and NPV
also increase.
become part of administrative claims data. In Special Considerations

addition, reimbursement standards and with Distributed Data Systems
patient insurance coverage limitations may
Multiple health data sources may be included
influence the selection of ICD codes for billing
within a single study or for ongoing surveil-
purposes. The potential for abuse of diagnostic
lance. Simultaneous assessment of multiple
codes, especially outpatient codes, may occur
data sources allows for better understanding of
when physicians apply to either an insurance
a larger population while also observing a
carrier or the government for reimbursement
diverse set of patients. These multidatabase
and may be less likely in group model health
studies or distributed data systems may have
maintenance organizations (HMOs), such as
differences in information collected, coding
Kaiser Permanente.
systems, language (e.g. across different coun-
Continuing with the NSAID example, we
tries), and even the underlying practice of
conducted a literature scan of published
medicine and overarching system of health
studies validating, MI or GI bleeding out-
care. Thus, even in the situation where distrib-
comes with use of NSAIDs in administrative
uted data systems use a common data model,
claims databases. Administrative claims data
careful consideration is warranted regarding
are often compared with medical records in a
how to assess validity of drugs, other therapeu-
validation study. Most of these studies pro-
tics, diagnoses, procedures, and health-related
vide only a PPV that indicates whether the
events within each administrative claims data
coding scheme is accurately classifying
source contributing to the distributed data sys-
observed measures as compared with
tem. Whenever EHR data are utilized, differ-
another source. Case Example 13.3 summa-
ences across sites warrant assessment of
rizes the findings from a study using the data
validity within each health system to improve
from the Veteran’s Administration to iden-
overall accuracy.
tify upper GI events with an algorithm con-
sisting of ICD-9 and Current Procedural
Terminology (CPT) codes. The measurement
Best Practices
characteristics of the diagnosis codes
selected for the algorithm depended on For the data in an administrative database to
which codes were used. be considered valid, subjects who appear in the
Case Example 13.3 Abraham et al. (2006)
Background ●● Multivariable logistic regression analy-

●● Databases differ in terms of demographic, sis was used to derive a predictive algo-
clinical, diagnosis, and procedure data rithm using ICD-9 codes, CPT codes,
capture. source of code (outpatient/inpatient),
●● New algorithms for identifying an out- and patient age. The algorithm was
come may have different reliability and developed and tested in separate cohorts
validity than previously used algorithms, from VA data.
even in the same database. It is important Results
to validate claims within each new data- Table 13.2 presents the sensitivity, specific-
base and to validate new algorithms upon ity, PPV, and NPV for the three claims-based
development. diagnoses of UGIE evaluated and the algo-
Question rithm PPV among NSAID users. The sensitiv-
Within the Veteran’s Affairs (VA) administra- ity dropped as diagnosis included broader
tive data, what is the reliability and validity parameters, while the PPV increased. Using
of various measures of diagnosis for NSAID– both ICD-9 and CPT codes led to the highest
related upper gastrointestinal events (UGIE)? specificity and NPV.
Approach Strengths
●● A retrospective cohort database and medi- ●● Evaluation of sensitivity, specificity, PPV,
cal record abstraction study was conducted and NPV provides the full picture of relia-
of 906 veterans to determine the reliabil- bility and validity of claims for both
ity and validity of ICD-9 and CPT codes to patients diagnosed with UGIE and those
determine UGIE and to develop an algo- without the diagnosis.
rithm to predict events among NSAID ●● Multiple diagnoses codes and a predictive
users. algorithm were assessed, providing a

●● The ICD- 9 codes used for UGIE were robust evaluation for this data set, one
531.x-534.x, and 578.x. CPT procedural that is suitable for future study and clinical
codes were 432xx, 443xx, 435xx, 436xx, decision-making.
440xx, 446xx, 44 120, 78 278, 7424x, ●● Data were from routine clinical and billing
7425x, and 74 260. processes, so no additional burden of data
Table 13.2 Sensitivity, specificity, PPV and NPV for claims-based diagnoses of UGIE.
Claims-based diagnosis Positive predictive Negative predictive

of UGIE Sensitivity (%) Specificity (%) value (%) value (%)
Only ICD-9 codes for 100 96 27 100

UGIE
ICD-9 and CPT codes 82 100 51 99
for UGIE
ICD-9 and CPT 66 88 67 88
algorithm for UGIE
Algorithm in only NA NA 80 NA
NSAID users
Source: Adapted from Abraham et al. (2006) with permission from John Wiley and Sons.
NA = not applicable. UGIE = upper gastrointestinal events.
collection for research purposes was diagnoses captured. For instance, use of
necessary. both ICD-9 and CPT codes for UGIE sub-
stantially decreased the sensitivity and
Limitations
increased the PPV for UGIE within claims
●● The study was conducted in one database
data. Thus, validation of the specific codes
(VA) only; validation results for the algo-
or algorithm used to assess a diagnosis is
rithm may not be applicable to different
key to understanding the findings from
databases or patient populations.
pharmacoepidemiologic research.
●● Multivariable logistic regression did not
●● Inclusion of a broader base of diagnoses,
include multiple demographic or clinical
procedures, or other factors forming an
factors that may affect occurrence of out-
algorithm to provide a diagnosis may
come. Additional factors may improve the
increase specificity, but it does so at the
reliability and validity of algorithm.
expense of sensitivity. Sensitivity, speci-
Key Points ficity, PPV, and NPV are all necessary to
●● Within the same database, different algo- provide a complete picture of the agree-
rithms or codes used to assess a diagnosis ment between claims data and medical
may substantially affect the number of records.
computerized files as having a drug exposure The investigator must be aware of the limi-
or disease should truly have that attribute, and tations of both the administrative database
those without the exposure or disease should and the chosen comparison data set. The cho-
truly not have the attribute. Validating the case sen comparator should provide sufficient data
definition for observational studies by compar- to validate both the exposure and outcome
ing administrative databases with original doc- used for the study. A variable that provides
uments, such as inpatient or outpatient linkage between the files in a data source,
medical records, is an important step to such as a medical record number, should be
enhance research quality and credibility. available so that accuracy can be evaluated
Although many studies have reviewed original within a subset of known study patients. For
documents to validate the diagnoses under example, if a single claim contains six diagno-
study or have referenced validation studies, a sis codes and six months of claims were used
need still exists for validation of drug expo- to determine outcomes in patients, then all six
sures and disease diagnoses in databases in diagnosis codes for all claims across the six-
which no previous validation has been per- month study time must be available in a com-
formed. As medical practice changes, further parison data set to establish the validity of the
validation of previously validated claims will outcome. A validation assessment should
be warranted. include evaluation of patients with and with-
Evaluating the completeness of the data- out the exposure or outcome. PPV, NPV, sensi-
bases is much more difficult because it requires tivity, and specificity combined provide a
an external data source known to be complete. complete understanding of the agreement
Although administrative claims and EHR between the two data sources.
databases have greatly expanded our ability to The following is a broad overview of how to
do pharmacoepidemiologic research, we need conduct a validation study in administrative
to ensure that tools, including the databases claims or EHR data. First, choose a meaningful
used for analyses, are complete and of the number of patients for validation. This sample
highest quality. size should be statistically grounded; however,
considerations of data availability, cost, and will be warranted for use of data from weara-
labor are understandable. Next, extract the bles and prior to integration of new data from
variables needed to determine cohort selec- biobanks, mobile apps, social media, or other
tion, exposure, outcome, and other variables sources into a rigorous research framework.
for validation. Calculate measures of agree- As part of the standardization process, data
ment and error rates (e.g. standard deviations) holders will have to document that their data
between the two data sets. Finally, consider are valid for conducting research and surveil-
strengths and limitations of the two data sets lance activities. This will require investigators
to ascertain validity and completeness of the to apply their knowledge and practices from
data source to answer the study question. use of administrative claims and EHR data to
linked data from these novel sources. Both
medication exposure and outcome diagnosis
The Future data from these novel sources do not carry the
same level of comfort regarding validity as
Methods for conducting pharmacoepidemio- claims and EHR data. As these data are consid-
logic studies have shifted over the past several ered for research, we hope and expect to see
decades from reliance on studies requiring de studies validating their use.
novo data collection from individuals, to exten-
sive use of electronic data from either adminis-
trative claims or EHRs to linked data sources
and distributed data networks. Yet, de novo Key Points
data collection will continue to be required to
ascertain information on quality of life, patient ●● The validity of self-reported diagnosis and
reported outcomes, and medications either not drug use data is a function of two properties:
included in pharmacy dispensing files or not how accurately persons who have medical
reliably entered into EHRs, such as herbal and conditions or use drugs of interest are ascer-
over-the-counter medications. In fact, with the tained (sensitivity), and the accuracy with
advent of wearables and the Internet of Things, which those who do not have the conditions
we anticipate that de novo collection of health or do not use the drugs are identified
data may increase in the coming years. (specificity).
The improved computer technology that ●● Misclassification of drug and diagnosis
resulted in faster processor speeds and information obtained from study partici-
increased storage capacity facilitated storage of pants by questionnaires or interviews
health care data in an electronic format, i.e. depends on factors such as the training and
EHRs, and allowed development of distributed experience of interviewers, the elapsed time
data networks, using data from multiple health since the events of interest took place, and
plans. The availability of these data for research characteristics of the participants, such as
has improved researchers’ ability to conduct their medical status and age.
studies, and increasing uptake of EHRs is lead- ●● Misclassification can lead to over- or under-
ing to increased availability of more granular estimation of the true association between
clinical data for pharmacoepidemiologic the drug exposure and the outcome of
research (e.g. lab results and clinical notes). interest.
Initial evaluation of EHR data suggest great ●● The medical record is typically used as the
promise, and increased data quality and stand- gold standard for verifying drug and diagno-
ardization of terminology and codes will be sis information, but it may be incomplete
required to make these data, collected for clini- and, with the increasing focus on privacy,
cal care, useful for research. Similar processes may be difficult to obtain.
Further Reading 243
Further Reading
Abraham, N.S., Cohen, D.C., Rivers, B., and analysis of 195,930 electronic prescriptions.
Richardson, P. (2006). Validation of Journal of General Internal Medicine 25 (4):
administrative data used for the diagnosis of 284–290.
upper gastrointestinal events following Flegal, K.M., Keyl, P.M., and Nieto, F.J. (1991).
nonsteroidal anti-inflammatory drug Differential misclassification arising from
prescription. Alimentary Pharmacology & nondifferential errors in exposure
Therapeutics 24 (2): 299–306. measurement. American Journal of
Bank, T.W. (2018). Living Standards Epidemiology 134 (10): 1233–1244.
Measurement Survey -Questionnaires. [cited Fourrier-Reglat, A., Cuong, H.M., Lassalle, R.
27 January 2018] Available from: http://go. et al. (2010). Concordance between prescriber-
worldbank.org/HZVSQMD5S0. and patient-reported previous medical history
Biemer, P. Measurement errors in sample and NSAID indication in the CADEUS cohort.
surveys. In: Handbook of Statistics – Pharmacoepidemiology and Drug Safety 19 (5):
Sample Surveys: Design, Methods and 474–481.
Applications (eds. D. Pfeffermann and C.R. Gama, H., Correia, S., and Lunet, N. (2009).
Rao), 281–316. The Netherlands: Questionnaire design and the recall of
North-Holland. pharmacological treatments: a systematic
Bland, J.M. and Altman, D.G. (1986). Statistical review. Pharmacoepidemiology and Drug
methods for assessing agreement between two Safety 18 (3): 175–187.
methods of clinical measurement. Lancet 1 Geisen, E. and Bergstrom, J.R. (2017).
(8476): 307–310. Respondent-survey interaction. In: Usability
Bustamante, R., Earles, A., Murphy, J.D. et al. Testing for Survey Research (ed. M.
(2019). Ascertainment of aspirin exposure Kaufmann), 21–49. Cambridge, MA:
using structured and unstructured large-scale Elsevier.
electronic health record data. Medical Care 57 Graesser, A.C., Wiemer-hastings, K., Kreuz, R.
(10): e60–e64. et al. (2000). QUAID: a questionnaire
Copeland, K.T., Checkoway, H., McMichael, A.J., evaluation aid for survey methodologists.
and Holbrook, R.H. (1977). Bias due to Behavior Research Methods, Instruments, and
misclassification in the estimation of relative Computers 32: 254–262.
risk. American Journal of Epidemiology 105 Greenland, S. (2005). Multiple-bias modelling
(5): 488–495. for analysis of observational data [with
Dean, E., Caspar, R., McAvinchey, G. et al. discussion]. Journal of the Royal Statistical
(2007). Developing a low-cost technique for Society: Series A (Statistics in Society) 168 (2):
parallel cross-cultural instrument 267–306.
development: the Question Appraisal System Greenland, S., Lash, T.L., and Rothman, K.J.
(QAS-04). International Journal of Social (2008). Bias analysis. In: Modern Epidemiology
Research Methodology 10 (3): 227–241. (eds. K.J. Rothman, S. Greenland and T.L.
Dosemeci, M., Wacholder, S., and Lubin, J.H. Lash), 345–380. Philadelphia, PA: Wolters
(1990). Does nondifferential misclassification Kluwer Health/Lippincott Williams & Wilkins.
of exposure always bias a true effect toward Greenland, S. (2001). Sensitivity analysis, Monte
the null value? American Journal of Carlo risk analysis, and Bayesian uncertainty
Epidemiology 132 (4): 746–748. assessment. Risk Analysis: An Official
Fischer, M.A., Stedman, M.R., Lii, J. et al. Publication of the Society for Risk Analysis 21
(2010). Primary medication non-adherence: (4): 579–583.
Horwitz, R.I. and Feinstein, A.R. (1978). derived from medical records. Vital and health
Alternative analytic methods for case–control statistics. Series 2. Data Evaluation and
studies of estrogens and endometrial cancer. Methods Research 23: 1–84.
The New England Journal of Medicine 299 Maldonado, G. (2008). Adjusting a relative-risk
(20): 1089–1094. estimate for study imperfections. Journal of
Jurek, A.M., Maldonado, G., Greenland, S., and Epidemiology and Community Health 62 (7):
Church, T.R. (2006). Exposure-measurement 655–663.
error is frequently ignored when interpreting Maldonado, G., Greenland, S., and Phillips, C.
epidemiologic study results. European Journal (2000). Approximately nondifferential
of Epidemiology 21 (12): 871–876. exposure misclassification does not ensure
Jurek, A.M., Greenland, S., Maldonado, G., and bias toward the null. American Journal of
Church, T.R. (2005). Proper interpretation of Epidemiology 151: S39.
non-differential misclassification effects: Mitchell, A.A., Cottler, L.B., and Shapiro, S.
expectations vs observations. International (1986). Effect of questionnaire design on recall
Journal of Epidemiology 34 (3): 680–687. of drug exposure in pregnancy. American
Jurek, A.M., Maldonado, G., and Greenland, S. Journal of Epidemiology 123 (4): 670–676.
(2013). Adjusting for outcome Naleway, A.L., Belongia, E.A., Greenlee, R.T.
misclassification: the importance of et al. (2003). Eczematous skin disease and
accounting for case–control sampling and recall of past diagnoses: implications for
other forms of outcome-related selection. smallpox vaccination. Annals of Internal
Annals of Epidemiology 23 (3): 129–135. Medicine 139 (1): 1–7.
Krosnick, J. (2000). The threat of satisficing in Newton, K.M., Peissig, P.L., Kho, A.N. et al.
surveys: the shortcuts respondents take in (2013). Validation of electronic medical
answering questions. Survey Methods record-based phenotyping algorithms: results
Newsletter 20: 4–8. and lessons learned from the eMERGE
Lash, T.L., Fox, M.P., and Fink, A.K. (2009). network. Journal of the American Medical
Applying Quantitative Bias Analysis to Informmatics Association 20 (e1): e147–e154.
Epidemiologic Data. New York: Springer. Nichols, E., Olmsted-Hawala, E., Holland, T.,
Lash, T.L., Fox, M.P., MacLehose, R.F. et al. and Riemer, A.A. (2016). Best Practices of
(2014). Good practices for quantitative bias Usability Testing Online Questionnaires at the
analysis. International Journal of Census Bureau: How rigorous and repeatable
Epidemiology 43 (6): 1969–1985. testing can improve online questionnaire
Lash, T.L. and Fink, A.K. (2003). Semi- design. In: International Conference on
automated sensitivity analysis to assess Questionnaire Design, Development,
systematic errors in observational data. Evaluation, and Testing (QDET2). Miami, FL.
Epidemiology (Cambridge, Massachusetts) 14 Phillips, C.V. (2003). Quantifying and reporting
(4): 451–458. uncertainty from systematic errors.
Lessler, J.T. and Harris, B.S.H. (1984). Medicaid Epidemiology (Cambridge, Massachusetts) 14
Data as a Source for Postmarketing (4): 459–466.
Surveillance Information. Research Triangle Phillips, C.V. and LaPole, L.M. (2003).
Park, NC: Research Triangle Institute. Quantifying errors without random sampling.
Maclure, M. and Willett, W.C. (1987). BMC Medical Research Methodology 3: 9.
Misinterpretation and misuse of the kappa Poole, C. (1985). Exceptions to the rule about
statistic. American Journal of Epidemiology nondifferential misclassification. American
126 (2): 161–169. Journal of Epidemiology 122: 508.
Madow, W.G. (1967). Interview data on chronic C.f.D.C.a. Q-Bank. Prevention [cited 27 January
conditions compared with information 2018]. https://wwwn.cdc.gov/qbank
Further Reading 245
Ray, W. (2003). Evaluating medication effects solutions. The Annals of Family Medicine 7
outside of clinical trials: new-user designs. (2): 100–103.
American Journal of Epidemiology 158: Tourangeau, R., Rips, L.J., and Rasinski, K.
915–920. (2000). The Psychology of Survey Response.
Ritchey, M.E., West, S.L., and Maldonado, G. Cambridge, MA: Cambridge University
(2019). Validity of pharmacoepidemiology drug Press.
and diagnosis data. In: Pharmacoepidemiology, Trifiro, G., Coloma, P.M., Rijnbeek, P.R. et al.
6e (eds. B.L. Strom, S. Hennessy and S.E. (2014). Combining multiple healthcare
Kimmel), 948–990. Sussex: Wiley. databases for postmarketing drug and vaccine
Rodgers, A. and MacMahon, S. (1995). safety surveillance: why and how? Journal of
Systematic underestimation of treatment Internal Medicine 275 (6): 551–561.
effects as a result of diagnostic test inaccuracy: Tsou, A.Y., Lehmann, C.U., Michel, J. et al.
implications for the interpretation and design (2017). Safe practices for copy and paste in the
of thromboprophylaxis trials. Thrombosis and EHR. Systematic review, recommendations,
Haemostasis 73 (2): 167–171. and novel model for health IT collaboration.
Rothman, K.J., Greenland, S., and Lash, T.L. Applied Clinical Informatics 8 (1): 12–34.
(2008). Precision and statistics in Upadhyaya, S.G., Murphree, D.H., Ngufor, C.G.
epidemiologic studies. In: Modern et al. (2017). Automated diabetes case
Epidemiology, 3e (eds. K.J. Rothman, S. identification using electronic health record
Greenland and T.L. Lash), 148–167. data at a tertiary care facility. Mayo Clinic
Philadelphia, PA: Lippincott, Williams & Proceedings. Innovations, Quality & Outcomes
Wilkins. 1 (1): 100–110.
Rothman, K.J., Greenland, S., and Lash, T.L. van Gelder, M.M., van Rooij, I.A., de Walle,
(2008). Validity in epidemiologic studies. In: H.E. et al. (2013). Maternal recall of
Modern Epidemiology (eds. K.J. Rothman, S. prescription medication use during
Greenland and T.L. Lash), 128–147. pregnancy using a paper-based
Philadelphia, PA: Wolters Kluwer Health/ questionnaire: a validation study in the
Lippincott Williams & Wilkins. Netherlands. Drug Safety 36 (1): 43–54.
Rowan, C.G., Flory, J., Gerhard, T. et al. (2017). Wacholder, S., Armstrong, B., and Hartge, P.
Agreement and validity of electronic health (1993). Validation studies using an alloyed
record prescribing data relative to pharmacy gold standard. American Journal of
claims data: a validation study from a US Epidemiology 137 (11): 1251–1258.
electronic health record database. Watson, D.L. (1965). Health interview responses
Pharmacoepidemiology and Drug Safety 26 (8): compared with medical records. Vital and
963–972. Health Statistics 1 (46): 1–74.
Smith, T.W. (2016). Optimizing Questionnaire West, S.L., Blake, C., Zhiwen, L. et al. (2009).
Design in Cross-National and Cross-Cultural Reflections on the use of electronic health
Surveys. In: International Conference on record data for clinical research. Health
Questionnaire Design, Development, Informatics Journal 15 (2): 108–121.
Evaluation, and Testing (QDET2). Miami, FL. West, S.L., Savitz, D.A., Koch, G. et al. (1995).
Survey Quality Predictor (SQP) 2.1. [cited Recall accuracy for prescription medications:
27 January 2018]. http://sqp.upf.edu self-report compared with database
Stange, K.C. (2009). The problem of information. American Journal of
fragmentation and the need for integrative Epidemiology 142 (10): 1103–1112.
246
14
Assessing Causality from Case Reports

Bernard Bégaud1 and Judith K. Jones†
1
Clinical Pharmacology and Pharmacoepidemiology, Medical School, University of Bordeaux, Bordeaux, France
†
Formerly Principal Consultant PharmaLex, Inc. 9302 Lee Highway. Suite 700, Fairfax, VA 220131, and Adjunct Faculty: The University
of Michigan School of Public Health Summer Program, 1415 Washington Heights, 1700 SPH I, Ann Arbor, MI 48109-2029, USA
Introduction Moreover, adverse reactions to drugs can be

acute, subacute, or chronic, can be reversible
An important component in clinical and epi- or not (e.g. birth defects and death), and can
demiological research is the assessment of the be quite rare or common. They can be patho-
degree to which an observed or reported event logically unique or mimic almost the entire
is causally associated with a drug treatment. range of human pathology. Thus, defining gen-
Several approaches to assessing this probabil- eral data parameters and criteria for assessing
ity of a causal connection have been devel- causality that will apply to most types of sus-
oped. This chapter reviews the basic principles, pected adverse reactions is tricky.
current approaches, and discusses their regula- Finally, in some instances, the assessment
tory context and application. will have little public health or economic
impact. Conversely, if continuation of a
development program depends upon this
assessment, the rigor of the approach used
linical Problems to be
C
becomes critical.
Addressed by
Pharmacoepidemiologic
Research The Two Paradigms
of Causality Assessment
The basic clinical problem is that adverse
events may be associated with multiple causal For a clinical pharmacologist, causal infer-
factors. The task is to make a differential ence, i.e. stating that a drug can cause a given
diagnosis and evaluate the degree to which the event, can only ensue from an experimental
occurrence of an event is linked to a particular design such as a randomized clinical trial. In
suspected causal agent: a drug or other this case, owing to the random allocation of
pathophysiological mechanism. exposure, the compared groups can be consid-
In case reports of suspected adverse reactions ered as identical with respect to all variables
to a medicinal product, data are often incom- that could influence the probability of occur-
plete so causal assessment is challenging. rence of the considered event. Therefore, a
When is Assessing Causation from Cases Reports Usefu 247
s tatistically significant excess of cases in the of massive under-reporting, worsened by a

exposed group, absent any other biases or the probable subjective selection bias, this sample
play of chance, signifies drug causation. has a good chance to differ, in terms of its
Pharmacoepidemiology deals with the real characteristics, from the whole population of
world where exposure is not allocated by exposed cases, the number of which remains
chance but by human behavior. For this reason, unknown.
compared groups never can be considered as Answering the “Did it” question requires
identical in risk, and an excess risk observed in working only from what we know about this
the exposed can also be explained by other case, and then extracting what can be
factors or characteristics over- or under- considered as relevant for assessing causation.
represented in the treated population. Drug This assessment will be based on common
causation for a significant risk increase among sense, medical knowledge, and/or probability
the exposed can be established only after these theory.
confounders and putative other biases have
been properly accounted for.
For both the clinical and epidemiological When is Assessing
approaches, statistical inference relies on the Causation from Cases
same 2 × 2 contingency table: Reports Useful?
Event No event Here are some examples where case-by-case

causal assessment can offer real added value:
Exposed a b
Unexposed c d
Spontaneous Reporting
n
Historically, assessing causation at the indi-
If we leave the populational level, a quite dif- vidual level was the trademark of pharma-
ferent question is: Did this drug cause this event covigilance. For the most relevant adverse drug
in this particular person? This is the scope of reactions reported by health professionals or
this chapter. patients, e.g. those which were serious or not
Assessing drug causation from individual previously reported, it was crucial to know
case reports may seem incongruous for a statis- whether the drug-event association had a good
tician. Indeed, here we are dealing only with chance of being causal or simply coincidental.
exposed cases, without the help of figures and During the last 30 years of the twentieth cen-
data from the non-exposed or from exposed tury, dozens of drugs were withdrawn from the
but non-diseased persons: market owing to a safety profile judged “unac-
ceptable” on the basis of case reports assessed
Event No event as causal by the manufacturer or a regulatory
agency. From a methodological point of view,
Exposed a ? this indisputably was the golden era of case-
Unexposed ? ? by-case causality assessment and more than 30
? different methods and approaches were pro-
posed during this fruitful period.
The situation is even worse since in many In several countries, assessment methods
instances, such as passive safety surveillance and algorithms were or still are widely used to
schemes like spontaneous reporting, a is only discard spurious drug-event associations in
the reported fraction of the cases which have order to decrease the background noise and
occurred in the exposed population. Because the generation of false-positives.
248 14 Assessing Causality from Case Reports
As big health databases have been made treatment is vital and the help provided by a
accessible and allow large pharmacoepidemio- reliable diagnostic method is particularly
logic studies to be designed, regulatory deci- valuable.
sions have become more and more based on a
causation assessed at the populational level.
Reports of Adverse Drug Reactions
However, the case-by-case approach remains
to Medical Journals
essential for associations difficult to appraise by
epidemiological approaches, such as serious Cases published in the literature are one of the
cases reported during the early phase of mar- most respected and influential sources of data
keting or of a suspected very rare occurrence, for drug safety surveillance. Owing to the
when the constitution of an appropriate control impact of such publications, arguments for
group is difficult if not impossible or when a and against drug causation should be
conservative decision could be made on the extensively listed and discussed. Sadly, many
basis of a series of serious case reports. reports do not provide information on
confounding therapies, medical conditions or
data deemed essential for considering causal-
Clinical Trials
ity. Here too, operational guidelines or algo-
and Pharmacoepidemiology
rithms can have some added value, and some
It is obvious that the clinical development of a journals like Therapies or Annals of
novel drug may be hampered by the occurrence Pharmacotherapy have made the use of an
of a single case of an adverse reaction if this operational algorithm mandatory.
event is serious and drug causation is the more
probable option. For example, for one case of
death identified during clinical development Hypothesis Generation
totaling 900 exposed subjects, the risk estimate and Research
derived from Poisson probabilities would range The scope of causal assessment goes far
from 3.3/100 000 to 6.2/1000. This emphasizes beyond the issue of adverse drug reactions
the strategic importance of a precise and and pharmacovigilance. Its basic principles
reliable assessment of the causal nature of the and structured approaches can apply to the
link in this particular case. broader question of the link between a factor,
A more common practice is to classify trait or exposure and an event, whatever its
adverse events observed in the study groups nature.
according to the likelihood of drug causation,
e.g. not related, possible, probable, etc. The
same applies to pharmacoepidemiologic
approaches like large post-marketing cohort
ethodological Problems
M
studies, mainly when a dedicated control to be Addressed by
group is not available. Pharmacoepidemiologic
Research
Clinical Practice and Prescription
Adverse drug reactions are a major cause of
Appraising whether a drug treatment is mortality and morbidity, both in developed
responsible or not for the occurrence of an and developing countries. Decisions to mini-
adverse event or of the worsening of a patient’s mize their public health and economic impact
state is a daily challenge in medical practice. In must be based on reliable conclusions as to the
some instances, (e.g. oncology, cardiology), the nature – causal or not – of the association
decision of pursuing or not pursuing the between a drug exposure and a harmful event.
Approaches for Assessing Causation from Individual Case 249
Two divergent philosophies have devel- mation on the case, refer to current medical
oped. Some discount the value of causality knowledge, and judge the likelihood that
assessment of individual reactions, deferring the adverse event resulted from drug expo-
judgment to the results of formal epidemio- sure. However, if not structured, global
logic studies or clinical trials. In contrast, introspection is known to suffer severe limi-
others contend that the information in single tations, notably judgment errors and poor
reports can be evaluated to determine some reproducibility.
degree of association, and that this can be It has been repeatedly shown that several
useful – sometimes critical – when consider- experts working separately on the same set of
ing discontinuation of a clinical trial or case reports may express marked disagreements
development or market withdrawal of a drug. in their judgments. Moreover, the process is
The latter view has spurred the evolution of uncalibrated; for example, one assessor’s
causal assessment from expert consensual “possible” might be another’s “probable” or
opinion (global introspection) to structured “doubtful.” Worse, as for any subjective
algorithms and elaborate probabilistic judgment, the same expert repeating the
approaches. process after some weeks or months has a good
chance of not duplicating his/her previous
opinions.
Despite these weaknesses, global
pproaches for Assessing
A introspection continues to be used, notably by
regulatory agencies and manufacturers, for
Causation from Individual Cases
assessing spontaneous reports whose serious
events may call for a decision to be made. The
For the sake of simplicity, three approaches for
same is true for severe or serious cases
assessing causation from individual cases can
identified during the clinical development of a
be contrasted:
novel drug in the framework of “safety advisory
●● Expert judgment/Global introspection. boards.”
●● Structured guidelines and algorithms. Here, we are touching on the paradox of
●● Probabilistic approaches. global introspection, which is not a reliable
instrument per se, but can be considered as the
In practice, the differences among these
gold standard in some circumstances. Such
approaches is not totally watertight: expert
may be the case when a multidisciplinary
judgment can be used as a gold standard for
group of senior experts is called upon in a sys-
weighting the decision criteria of an
tematic, structured, and interactive process
operational algorithm and probability concepts
such as the Delphi Method. For example, drug
may be used when designing an algorithm.
causation may be assessed by several experts
working separately. They are required to
express their final judgment for each case and
Expert Judgment/Global
to provide detailed arguments to substantiate
Introspection
it. Next, all disagreements are listed and then
Since case-by-case causality assessment interactively discussed until a consensus is
tends to reproduce the medical diagnosis reached on each case. Even if time-consuming,
process, e.g. to determine the most probable and therefore not applicable to large case
cause of a disease, the most natural and series, this approach can be considered as a
common approach for assessing drug causa- sort of gold standard and has been used as
tion is indisputably global introspection. such to calibrate various algorithmic or proba-
One or more experts review available infor- bilistic methods.
Structured Guidelines Lasagna, proposed the first structured

and Algorithms algorithm for assessing case reports. The two
approaches shared similar basic data elements:
Because expert judgment/global introspection,
which is often plagued by subjectivity, haspoor ●● Timing relative to drug exposure,
reliability and reproducibility, an obvious ●● “Dechallenge”,
improvement has been to force experts to ●● “Rechallenge”,
follow a structured common thread, e.g. by ●● Presence/absence of other potential causal
obliging them to make their final judgment on factors or alternative causes,
the successive assessment of relevant criteria. ●● Other supportive data, e.g. if previous cases
These criteria can make reference to formal were reported.
logic or common sense, to medical knowledge,
During the following two decades, no less
or to literature data.
than 30 methods, algorithms and others were
A good example of formal logic is the
designed and, for the most part, published.
temporal association between the drug treat-
Although their functional architecture can be
ment and the event. The “cause” should always
quite diverse, they generally consist in
precede the first symptoms of the event (“chal-
successively answering a certain number of
lenge”); the suppression of the cause, i.e. stop-
questions in order to end up with an assessment
ping the treatment is expected to improve or to
of the probability of drug causation that is
cure the situation (“dechallenge”), while the
generally expressed in the form of qualifiers
re-introduction of the cause, if any, is expected
(e.g. “Not-related,”“doubtful,”“possible,”“prob
to make the event reappear (“rechallenge”).
able,”“certain”) or, for some algorithms, of
This should be interpreted in the light of the
figures (e.g. 1, 2, 3, 4) or of a probability scale.
characteristics of the molecule (pharmacoki-
Although, the number (from less than 10 up to
netic parameters) and of the event. For
84) and the wording of questions may differ
instance, for some of them, such as with death
greatly, they roughly include five basic criteria:
or irreversible lesions, no relapse can be
timing, dechallenge, rechallenge, confounding,
expected. Similarly, as for the Bradford Hill cri-
and prior history of the event. Information
teria, a clear relationship between the dose and
relevant to each criterion is elicited by
the intensity/severity of the event can be sup-
questions. The answers are restricted to “yes/
portive of drug causation.
no” (and for some methods “don’t know”) and
For criteria calling upon medical knowledge,
then aggregated to derive the final causality
there are arguments that allow alternative
score. This aggregation process can be discrete
causes for the event to be ruled out or the
(the various questions being interdependent)
biological or pharmacological plausibility:
relying on an algorithm, a decisiontree or a
does the pharmacological properties of this
combination table, or continuous by summing
drug allow such a manifestation to be expected?
notes or scores as in Figure 14.1 below.
For literature data, one is obviously more
The core issue that remains is the weight
inclined to draw a conclusion of drug causation
attributed to each question in the final score.
if this type of reaction has been repeatedly
On this point, each method has its own “per-
reported with the suspected drug.
sonality” and it comes as no surprise that
The first attempt at structuring the
when used on the same set of case reports,
assessment of drug causation in individual
several methods can provide discrepant esti-
cases was proposed in 1976 by Nelson Irey, a
mates of the probability of drug causation. In
pathologist at the US Armed Forces Institute of
fact, the two dozen methods proposed so far
Pathology. One year later, in 1977, two clinical
do not purport to have the same goals. Some
pharmacologists, Fred Karch and Louis
are very basic algorithms aiming at roughly
Approaches for Assessing Causation from Individual Case 251
CAUSALITY ASSESSMENT
NARANJO SCORED ALGORITHM
QUESTION ANSWER SCORE
Yes No Unk
Previous reports? +1 0 0
Event after drug? +2 –1 0
Event abate on drug removal? +1 0 0
+ Rechallenge? +2 –1 0
Alternative causes? –1 +2 0
Reaction with placebo? –1 +1 0
Drug blood level toxic? +1 0 0
Reaction dose-related? +1 0 0
Past history of similar event? +1 0 0
ADR confirmed objectively? +1 0 0
Total Score
Figure 14.1 Simplified presentation of method proposed by Naranjo et al. Likelihood of drug causation is
qualified by score (Total Score) obtained by summing individual scores from each of 10 questions.
classifying case reports in order to reject those markedly increase the reproducibility of
for which drug causation is quite improbable. estimates, and overall, provide a structured
In other cases, the method was designed for a approach that makes the observer aware of the
specific application like signal detection in questions to be explored before concluding in
routine safety surveillance where sensitivity terms of drug causation.
is preferred, often at the cost of specificity, for
a specific type of event such as liver injury, or
Probabilistic Approaches
for a given type of drug (vaccines, anticancer
drugs, etc.). The most ambitious ones aim to Bayesian Approaches
provide “exact” appreciations of the probabil- In essence, the issue of drug causation belongs
ity of drug causation by means of meticu- to the domain of probability, the aim being to
lously tuning the respective weight of each quantify (from 0 to 1 or 0% to 100%) the
question. probability that drug A caused the adverse
Fewer than one third of these methods have event presented by a given person. In the
actually been used in routine practice: some absence of any relevant information for
because their use was recommended or made judging the nature of the link, the probability
mandatory by some regulatory agencies (as it of drug causation is 50% (i.e. in the middle of a
was the case for Australia and the Food and probability scale ranging from 0 to 1). In this
Drug Administration and still is in practice for case, the chances for drug causation and non-
France); others because they were developed drug causation are equal (i.e. 50% each). Any
by a pharmaceutical company for their own evidence in favor of drug causation will move
pharmacovigilance, and others because they the estimate to the right (probability >0.5),
were adopted by medical journals for the while an argument against it will shift it left
reporting of adverse drug reactions. (probability <0.5). These basic probability
Despite their obvious weaknesses, algo- rules have led some researchers to propose the
rithms offer a compromise between simplicity Bayes’ Theorem as the most satisfactory
and added value for whomever intends to approach to the issue of drug causation.
deal with drug causation. They dramatically Indeed, conditional probabilities allow the
decrease discrepancies between experts, calculation of the probability that an event will
occur or that an affirmation will be correct what we learn about this particular case in
under certain conditions, e.g. that a test is order to obtain a final estimate called posterior
positive or negative or a symptom is present or probability.
absent. The approach developed by Thomas For the sake of simplicity, it is easier to use
Bayes (1702–1761) was to start from ana priori odds rather than probabilities, the original
estimate of this probability (prior probability) Bayes’ formulae being rather troublesome to
that is then altered (increased or decreased) by use. In this case, the formulation is:
Posterior Odds Prior Odds LR1 LR 2 LR 3 LR j
The odds (Posterior or Prior) are simply the s tarting point and has a notable influence on
probability that the drug was the cause of the the final result. In the absence of any pre-
event divided by the probability that it was not. analysis information, the prior odds can be
The odds obviously being 1 when these proba- set arbitrarily to 1, i.e. one assumes that the
bilities are equal (50% or 0.5 each). Indeed: chances that the drug under study was or
Odds Probability / 1 Probability was not the cause of the event are equal. In
more favorable cases, it is possible to set the
LRs are the Likelihood Ratios corresponding value of the prior odds more satisfactorily.
to each information component judged to be This is for example the case when the drug is
relevant for assessing the probability of drug already known or suspected to cause the dis-
causation in this particular case. For example, ease and a literature search provides reliable
they may refer to the delay between the start of data on the strength of association between
treatment and the occurrence of the event, to a exposure to this drug and the probability of
particular clinical or biological sign, to the past occurrence of the event in question. For
medical history of the person, or to whatever example, if a pharmacoepidemiologic study
appears relevant in this particular context. The has shown that such a drug treatment
value of each LR (which is simply an odds) is increased the probability of the disease by a
obtained by dividing the probability of the sign factor (hazard ratio) of 4.5, one can deduce
being present if the drug was the cause by the that the etiologic fraction of the risk among
probability of the sign being present if it was the exposed is:
not the cause. For example, if we learned from AFR E HR 1 / HR or 4.5 1 / 4.5 0.78
literature that, on average, 67% of persons pre-
senting with the considered adverse event are In other words, on average, the probability
female while the proportion of females among for drug causation is 78% for each exposed
users is 42% in the source population, the LR individual in this study. Therefore, the prior
for gender would be: 0.67/0.42 = 1.6 if the per- odds (probability of drug causation divided by
son was female and 0.33/0.58 = 0.57 if this per- the probability of non-drug causation) is: 0.78/
son was male. (1 – 0.78) = 3.5. Or, more simply:
David Lane and colleagues have made this
Prior Odds HR 1
approach more operational by superposing the
various LRs to be computed with the criteria Starting from the two examples above, the
universally used in algorithmic approaches: probability estimate would be (if the case was a
challenge, dechallenge, rechallenge, etc. An male):
automated computerized version of the
Posterior Odds 3.5 0.57 1.99
method has even been developed.
One can see from the Bayes’ formula that At this stage, and before pursuing the com-
the value used for the prior odds is the putation with other, and probably more
Choosing the Appropriate Approac 253
r elevant LRs (timing, clinical and/or biological ble than it is not. To that end, odds can easily
feature, medical history, alternative causes, be, if wished, converted into a probability for
etc.), the drug is twice as likely to be responsi- drug causation:
Probability Odds / Odds 1 or, for our interim result : 1.99 / 2.99 66%.
Although rather complex to handle, a major or was not the cause of an event. For exam-
advantage of the Bayesian approach is that it ple, this might involve a well-documented
relies on healthy probabilistic concepts, reaction with this drug that occurs after an
respects the basic rules of probability and pro- expected delay, which relapses after treat-
vides a scientifically valid estimate of ment cessation, reappears after its re-
probability. introduction and for which no other cause
has been found. However, following such a
Other Probabilistic Approaches caricatural logic gives no guarantee that the
The logistic function was proposed by Arimone method will provide consistent results in the
et al. to convert the individual estimates of situations where uncertainty is the main rea-
seven criteria (roughly those commonly used son for using it such as in the vast majority of
by algorithmic and Bayesian approaches) into cases.
a probability of drug causation varying from 0 ●● A more sensible approach would be to use,
to 1: when available, the data from a cohort study
1 or a large-scale trial using a non-exposed
p control group. The value of the hazard ratio
7
1 exp iXi
makes it possible to calculate the etiologic
i 1 fraction of the risk in exposed and, therefore,
the proportion and the number of cases of
α is similar to the Bayesian prior odds, X1 to the event attributable to exposure. In an
X7 are the results of the individual assessments ideal world, this proportion and number
of the seven criteria, and ß1 toß7 are the rela- should be confirmed by the individual
tive weights attributed to each criterion after assessment of all exposed cases with the
the validation against a gold standard. method under study.
Assessments were made simple owing to an
automated computerized process.
hoosing the Appropriate

C
Calibration Approach
When designing a method for assessing drug To date, it is obvious that there is no unique
causation, the main challenge is to make sure method for assessing drug causation that can
that the estimates it provides are valid. As no be adopted by all stakeholders in the domain
indisputable gold standard exists, several of drug safety. Moreover, if used on the same
workarounds have been proposed: set of cases, one could expect the results to be
discrepant. Nevertheless, one should bear in
●● The most satisfactory ones may arguably, be mind the following:
Bayesian approaches and multidisciplinary
experts’judgment optimized by the Delphi ●● A structured approach is always preferable
process. to a non-structured assessment, which
●● It would be tempting to use cases for which intrinsically has poor reliability and poor
one can be almost certain that the drug was reproducibility.
●● Beyond the Holy Grail quest for the “univer- given the fact that there is no reliable standard
sal” method, the choice should first be terminology.
guided by the context: All of these factors suggest the need for fur-
–– How will the evaluation be used and who ther work in several areas:
will perform the evaluation?
–– The importance of the accuracy of the judg- ●● Determining the applications of causality
ment. If the evaluation will determine assessment, i.e. the “output” of the process,
either the continuation of a clinical trial to better define the desired rigor, accuracy,
or the continued marketing of a drug, and usability of the methods. There will
then its accuracy may be critical, so a probably always be the need for simpler and
probabilistic approach would be more rougher methods, as well as more complete
appropriate. Conversely, if little hinges and rigorous methods when the determina-
upon the judgment, unrefined estimates tion has considerable impact.
and methods may suffice. ●● Further defining the critical elements needed
–– The number of causality evaluations to be for evaluation of causality for different types
made must also be weighed against the of adverse reactions (e.g. hepatic, hemato-
time required to make judgments on large logical, skin, etc.) so that this information
numbers of reports, a dilemma for regula- may be collected at the time of reporting or
tory agencies and sponsors. Here, the publishing a spontaneous event.
need for accurate judgments is pitted ●● Gathering of data on critical elements of the
against the volume of evaluations to be specific adverse events in the course of both
considered. clinical trials and epidemiologic studies.
Risk factor, history, timing, characteristics,
and resolution patterns of adverse events
The Future should be described in these studies and
incorporated into general data resources on
The field of adverse reaction causality assess- the characteristics of medical events and
ment has many unresolved issues, both meth- diseases.
odological and practical. Originally there was ●● Further work on automation of the causality
hope a consensus method would be found, but evaluation process. Global introspection is
the current state of the field suggests that this still widely used because of the cumbersome
is unlikely for several reasons. nature of many of the more complete meth-
First, some individuals and institutions ods. Convenient access to the proper ques-
have adopted one or a few methods and have tions, set out in logical order, as well as
committed to their use, often through their background data meeting quality criteria on
choice of data collecting systems or software. the state of information to date, has poten-
Second, practical aspects in using them appear tial for considerably improving the state of
to play a very real role. Although discussed adverse reaction causality evaluation.
with excitement as the possible “gold stand- ●● Consideration of new and different
ard” for adverse reaction causality, the approaches. Although it is likely that further
Bayesian method has not been widely work will be based on one or more of the
embraced, partly because it is difficult to use many available methods, other interesting
without automation. With the lifting of that approaches have emerged. For example, as
barrier, and with further practical applica- part of work on patient safety in the US,
tions, its potential may be realized. Third, the “root cause analysis” is used to identify
misuse of judgment terms or scores within the important contributors to adverse events in
legal arena has generated concern, particularly clinical settings. This approach creates
Further Readin 255
f unctional maps of possible contributing fac- does not further the development of the
tors to identify not only a cause but also hypothesis raised in the report of the event.
preventative measures. Another approach is ●● The choice of a method is based upon the
the N-of-1 trial, which can evaluate the cau- use of judgment; if pivotal to the continued
sality of adverse events in individuals, par- development of a drug, the most rigorous
ticularly those who have experienced methods such as the Bayesian approach may
multiple reactions to drugs. Other poten- help; if used to sort out well-documented
tially promising approaches include expert cases that may probably be associated with a
systems, machine learning, neural networks drug, then simple algorithms or scoring
and other sophisticated approaches. algorithms usually suffice.
●● The components of causality assessment
methods can help structure data collection
on individual and groups of cases; ulti-
Key Points mately, these aggregate data can improve the
description of the event of interest, and pos-
Applications of a structured causality
sibly its relationship to a drug, or the disease
●●
method can standardize and help reduce

of indication.
biases in assessing the possible cause-effect
●● The detailed probabilistic and explicit
relationship of an event to a drug exposure.
approach in the Bayesian method can, if
The use of a clinical non-structured
data are available, provide a basis for devel-
●●
approach (“global introspection”) to assess

oping more precise statements of the hypoth-
adverse events believed to be associated with
esis that is posed in a spontaneous report of
a drug has been shown to yield inconsistent
a suspected adverse drug reaction.
results between raters; its lack of structure
Further Reading
Agbabiaka, T.B., Savović, J., and Ernst, E. (2008). Collet, J.-P., MacDonald, N., Cashman, N., and
Methods for causality assessment of adverse Pless, R. (2000). The advisory committee on
drug reactions: a systematic review. Drug Saf. causality assessment. Monitoring signals for
31 (1): 21–37. vaccine safety: the assessment of individual
Arimone, Y., Bégaud, B., Miremont-Salamé, G. adverse event reports by an expert advisory
et al. (2006). A new method for assessing drug committee. Bull. World Health Organ. 78:
causation provided agreement with experts’ 178–185.
judgment. J. Clin. Epidemiol. 59 (3): 308–314. Drug Information Association (1986).
Benahmed, S., Picot, M.C., Hillaire-Buys, D. Proceedings of the Drug Information
et al. (2005). Comparison of Association Workshop, Arlington, Virginia,
pharmacovigilance algorithms in drug February. Drug Inf. J. 20: 383–533.
hypersensitivity reactions. Eur. J. Clin. Irey, N.S. (1976). Adverse drug reactions and
Pharmacol. 61 (7): 537–541. death: a review of 827 cases. JAMA 236:
Bénichou, C. and Danan, G. (1994). A new 575–578.
method for drug causality assessment: Jones, J.K. (2020). Assessing causation from case
RUCAM. In: Adverse Drug Reactions. A reports. In: Pharmacoepidemiology, 6e (eds.
Practical Guide to Diagnosis and Management, B.L. Strom, S.E. Kimmel and S. Hennessy),
277–284. New York: Wiley. 725–745. Wiley.
Karch, F.E. and Lasagna, L. (1977). Toward the Naranjo, C.A., Busto, U., Selers, E.M. et al.
operational identification of adverse drug (1981). A method for estimating the
reactions. Clin. Pharmacol. Ther. 21: probability of adverse drug reactions. Clin.
247–254. Pharmacol. Ther. 30: 239–245.
Kelly, W.N., Arellano, F.M., Barnes, J. et al. Péré, J.C., Bégaud, B., Harambaru, F., and Albin,
(2007). Guidelines for submitting adverse H. (1986). Computerized comparison of six
event reports for publication. adverse drug reaction assessment procedures.
Pharmacoepidemiol. Drug Saf. 16: 581–587. Clin. Pharmacol. Ther. 40: 451–461.
Kramer, M.S., Leventhal, J.M., Hutchinson, T.A., Stephens, M.D. (1987). The diagnosis of adverse
and Feinstein, A.R. (1979). An algorithm for medical events associated with drug
the operational assessment of adverse drug treatment. Adverse Drug React. Acute
reactions. I background, description, and Poisoning Rev. 6: 1–35.
instructions for use. JAMA 242: 623–632. Venulet, J., Ciucci, A.G., and Berneker, G.C.
Lane, D.A., Kramer, M.S., Hutchinson, T.A. et al. (1986). Updating of a method for causality
(1987). The causality assessment of adverse assessment of adverse drug reactions. Int. J.
drug reactions using a Bayesian approach. Clin. Pharmacol. Ther. Toxicol. 24: 559–568.
Pharm. Med. 2: 265–283. Yerushalmy, J. and Palmer, C.E. (1959). On the
Macedo, A.F., Marques, F.B., and Ribeiro, C.F. methodology of investigations of etiologic
(2006). Can decisional algorithms replace factors in chronic diseases. J. Chronic Dis. 10
global introspection in the individual causality (1): 27–40.
assessment of spontaneously reported ADRs? Zapater, P., Such, J., Perez-Mateo, M., and
Drug Saf. 29 (8): 697–702. Horga, J.F. (2002). A new Poisson and
Miremont-Salamé, G., Théophile, H., Bayesian-based method to assign risk and
Haramburu, F., and Bégaud, B. (2016). causality in patients with suspected hepatic
Causality assessment in pharmacovigilance: adverse drug reactions: a report of two new
the French method and its successive updates. cases of Ticlopidine-induced hepatotoxicity.
Therapie 71: 179–186. Drug Saf. 25: 735–750.
257
15
Molecular Pharmacoepidemiology
Christine Y. Lu1 and Stephen E. Kimmel2
1
Harvard Medical School & Harvard Pilgrim Health Care Institute, Boston, MA, USA
2
University of Florida, Gainesville, FL, USA
Introduction ispensing, adherence with prescribed drug

d
regimen (see Chapter 21), and last, but not
Precision medicine has been defined by the least, the genetic profile of the patient.
National Institutes of Health (NIH) in the The effects of genes and other biomarkers on
United States as an “approach to disease pre- drug response can be studied using molecular
vention and treatment based on people’s indi- pharmacoepidemiology. Molecular pharma-
vidual differences in environment, genes and coepidemiology is the study of the manner in
lifestyle.” Genomic technologies are increas- which molecular biomarkers alter the clinical
ingly available, and their use in clinical care effects of medications in populations. Just as
has grown substantially over the last decade. the basic science of pharmacoepidemiology is
There are different types and applications of epidemiology, applied to the content area of
genomic technologies, including disease clinical pharmacology, the basic science of
screening, diagnosis, prognosis, risk assess- molecular pharmacoepidemiology is epidemi-
ment or susceptibility, informed reproductive ology in general and molecular epidemiology
choices, and pharmacogenetics. specifically, also applied to the content area of
One of the most challenging aspects of clini- clinical pharmacology. Thus, many of the
cal pharmacology and pharmacoepidemiology methods and techniques of epidemiology
is to understand why individuals and groups of apply to molecular pharmacoepidemiologic
individuals respond differently to a specific studies. However, there are features of molecu-
drug therapy, both in terms of beneficial and lar pharmacoepidemiology that are unique, as
adverse effects. Reidenberg observed that, discussed later in this chapter. Most of the dis-
while the prescriber has basically two deci- cussion will focus on studies related to genes,
sions to make while treating patients (i.e. but the methodological considerations apply
choosing the right drug and choosing the right equally to studies of proteins (e.g. proteomics)
dose), interpreting the inter-individual varia- and other biomarkers, such as the microbiome
bility in outcomes of drug therapy includes a (the genes within the microbial cells, primarily
wider spectrum of variables, including the bacteria in the gut, harbored within each per-
patient’s health profile, prognosis, disease son) and mRNA (messenger RNA that results
severity, quality of drug prescribing and from DNA transcription.
258 15 Molecular Pharmacoepidemiology
On average for each medication, it has been about genomic regions that do not encode
estimated that about one out of three treated RNA or protein, but play important roles in
patients experience beneficial effects, one out gene expression and regulation such as epige-
of three do not show the intended beneficial netics (changes in DNA expression that occur
effects, 10% experience only side effects, and but are not related to the base order, such as
the rest of the patient population is nonadher- DNA-methylation). In addition, changes in the
ent so that the response to the drug is difficult DNA of microbial cells (the microbiome) can
to assess. This highlights the challenge of indi- influence human response to medications.
vidualizing therapy to produce a maximal ben- Thanks to numerous human genome initia-
eficial response and minimize adverse effects. tives, we also have substantial information
Although many factors can influence medica- about inter-individual variability in the human
tion efficacy and adverse effects, including age, genome. The most common form of genomic
drug interactions, and medication adherence variability is a single nucleotide polymorphism
(see Chapter 21), genetics is an important con- (SNP), which represents a substitution of one
tributor in the response of an individual to a nucleotide (i.e. the basic building block of
medication. Genetic variability can account for DNA, also referred to as a “base”) for another,
a large proportion (e.g. some estimates range which is present in at least 1% of the popula-
from 20% to 95%) of variability in drug disposition. Each person has inherited two copies of
tion and medication effects. each allele (one from the paternal chromo-
In addition to altering dosing requirements, some and one from the maternal chromo-
genetics can influence response to therapy by some). The term allele refers to the specific
altering drug targets or the pathophysiology of nucleotide at one point in the genome inher-
the disease states that drugs are used to treat. ited either from the father or mother, and the
combination of alleles in an individual is
denoted a genotype. When the two alleles are
identical (i.e. the same nucleotide sequence on
Definitions and Concepts
both chromosomes), the genotype is referred
to as “homozygous;” when the two alleles are
Genetic Variability
different (i.e. different nucleotide sequences
Building on the success of the various human on each chromosome), the genotype is referred
genome initiatives, it is now estimated that to as “heterozygous.” Approximately 10 mil-
there are approximately 25 000 regions of the lion SNPs are thought to exist in the human
human genome that are recognized as genes genome, with an estimated two common mis-
because they contain deoxyribonucleic acid sense (i.e. amino acid changing) variants per
(DNA) sequence elements including exons gene.
(sequences that encode proteins), introns However, SNPs are not the only form of
(sequences between exons that do not directly genetic variation that may be relevant to
encode amino acids), and regulatory regions human traits and diseases. For example, copy
(sequences that determine gene expression by number variants (CNV), sections of the
regulating the transcription of DNA to RNA, genome that have repeats of base pairs, have
and then the translation of RNA to protein). also been recently identified as another com-
Some of these sequences have the ability to mon form of genomic variation that may have
encode RNA (ribonucleic acid, the encoded a role in disease etiology. DNA methylation,
messenger of a DNA sequence that mediates where methyl groups are added to DNA, thus
protein translation) and proteins (the amino changing the activity of DNA (which itself is
acid sequence produced by the translation of regulated by genetics), and variability in the
RNA). In addition, we are learning a great deal gut microbiome can also alter drug response.
The Interface of Pharmacogenetics and Pharmacogenomics with Molecular Pharmacoepidemiolog 259
Finally, we also recognize that the genome is easurements of these surrogates in small
m
not simply a linear nucleotide sequence, but groups of patients in highly controlled set-
that population genomic structure exists in tings. Molecular pharmacoepidemiology
which regions as large as 100 kilobases (a kilo- focuses on the effects of genetics on clinical
base being a thousand nucleotides, or bases) in outcomes and uses larger observational and
length define units that remain intact over evo- experimental methods to evaluate the effec-
lutionary time. These regions define genomic tiveness and safety of drug treatment in the
block structure that may define haplotypes, population. Molecular pharmacoepidemiol-
which are sets of genetic variants that are ogy uses similar methods as pharmacoepide-
transmitted as a unit across generations. miology to answer questions related to the
Thus, the complexity of genome structure effects of genes on drug response. Thus,
and genetic variability that influences response molecular pharmacoepidemiology answers
to medications provides unique challenges to questions related to:
molecular pharmacoepidemiology.
1) The population prevalence of SNPs and
other genetic variants,
Pharmacogenetics 2) Evaluating how these genetic variants alter
and Pharmacogenomics disease outcomes,
3) Assessing the impact of gene–drug and
While the term pharmacogenetics is predomi-
gene– gene interactions on drug response
nantly applied to the study of how genetic vari-
and disease risk, and
ability is responsible for differences in patients’
4) Evaluating the usefulness and impact of
responses to drug exposure, the term pharma-
genetic tests in populations exposed, or to
cogenomics also encompasses approaches
be exposed, to drugs.
simultaneously considering data about thou-
sands of genotypes, as well as responses in There are, however, some aspects of molecu-
gene expression to existing medications. lar pharmacoepidemiology that differ from the
Although the term “pharmacogenetics” is rest of pharmacoepidemiology. These include
sometimes used synonymously with pharma- the need to understand the complex relation-
cogenomics, the former usually refers to a ship between medication response and the vast
candidate-gene approach as opposed to a number of potential molecular and genetic
genome-wide approach in pharmacogenomics influences on this response; a focus on interac-
(both discussed later in this chapter). tions among these factors and interactions
between genes and environment (including
other medications) that raise issues of sample
size and has led to interest in novel designs;
The Interface
and the need to parse out the most likely asso-
of Pharmacogenetics ciations between genes and drug response
and Pharmacogenomics from among the massive number of potentially
with Molecular important genes identified through bioinfor-
Pharmacoepidemiology matics (the science of developing and utilizing
computer databases and algorithms to acceler-
Pharmacogenetic and pharmacogenomic ate and enhance biological research). As stated
studies are usually designed to examine previously, the basic science of epidemiology
intermediate endpoints between drugs and underlies molecular pharmacoepidemiology
outcomes (such as drug levels, pharmacody- just as it underlies all pharmacoepidemiology.
namic properties, or surrogate markers of What is different is the need for approaches
drug effects) and often rely on detailed that can deal with the vast number of potential
genetic influences on outcomes; the possibility Pharmacokinetic Gene–Drug Interactions

that “putative” genes associated with drug Genes may influence the pharmacokinetics of
response may not be the actual causal genes, a drug by altering its metabolism, absorption,
but rather a gene near or otherwise associated or distribution. Metabolism of medications
with the causal gene on the chromosome in can either inactivate their effect or convert an
the population studied (and that may not be inactive prodrug into a pharmacologically
similarly linked in other populations); the active compound. The genes that are responsi-
potential that multiple genes, each with a rela- ble for variable metabolism of medications are
tively small effect, work together to alter drug those that code for various enzyme systems,
response; and the focus on complex interac- especially the cytochrome P450 enzymes.
tions between and among genes, drugs, and The gene encoding CYP2D6 represents a
environment. By discussing the potential good example of the various ways in which
approaches to these challenges in this chapter, polymorphisms can alter drug response. Many
it is hoped that both the similarities and differ- drug-CYP2D6 genetic variant interactions
ences between pharmacoepidemiology and have been reported based on experimental or
molecular pharmacoepidemiology will be epidemiologic associations. CYP2D6 is one of
made clear. the most common pharmacogenomic markers
included in drug labeling by the US Food and
Drug Administration (FDA) and the European
Medicines Agency (EMA). Some of the genetic
C
variants lead to low or no activity of the
Addressed by CYP2D6 enzyme whereas some individuals
Pharmacoepidemiologic have multiple copies of the gene, leading to
Research increased metabolism of drugs. Thus, patients
using CYP2D6-dependent antipsychotic drugs
It is useful to conceptualize clinical problems (e.g. haloperidol) who are poor metabolizers
in molecular pharmacoepidemiology by think- (low CYP2D6 activity) are more than four
ing about the mechanism by which genes can times more likely to need anti-Parkinsonian
affect drug response. medication to treat side effects of the antipsy-
chotic drugs than high metabolizers. The
decreased metabolic activity of CYP2D6 may
Three Ways That Genes Can Affect
also lead to lower drug efficacy, as illustrated
Drug Response
for codeine, which is a prodrug that is metabo-
The effect that a medication has on an indi- lized to the active metabolite, morphine, by
vidual can be affected at many points along CYP2D6. It has been estimated that approxi-
the pathway of drug distribution and action. mately 6–10% of Caucasians have variants that
This includes absorption and distribution of result in CYP2D6 genotypes that encode dys-
medications to the site of action, interaction functional or inactive CYP2D6 enzyme, in
of the medication with its targets, metabo- whom codeine is an ineffective analgesic.
lism of the drug, and drug excretion (see There also is important interethnic variability
Chapter 4). These mechanisms can be cate- of CYP2D6 alleles and phenotypes. An analy-
gorized into three general routes by which sis of CYP2D6 allele-frequency data from
genes can affect a drug response: pharma- >60 000 individuals suggests that diplotype
cokinetic, pharmacodynamic, and gene– frequencies predicting poor metabolism are
drug interactions in the causal pathway of highest among Europeans and the Ashkenazi
disease. These will be discussed in turn Jewish population (about 5–6%) and lowest
below. among East and South Central Asians,
Clinical Problems to be Addressed by Pharmacoepidemiologic Researc 261
Oceanians, and Middle Easterns (<1% in each methods that can identify the true causal
of these populations). In contrast, diplotype variant(s) that may reside in an LD block.
frequencies predicting ultrarapid metabolism
were highest in Oceanian (21.2%), followed by Pharmacodynamic Gene–Drug
Ashkenazi Jews and Middle Easterns (about Interactions
11% in each of these populations) and lowest Once a drug is absorbed and transported to its
in East Asians (1.4%). target site, its effect may be altered by differ-
However, predicting clinical outcomes in ences in the response of drug targets. Therefore,
daily practice based on such CYP2D6 genetic polymorphisms in genes that code for drug tar-
data in a valid fashion remains complex. Drug– gets may alter the response of an individual to
gene associations shown in one study cannot a medication.
always be replicated in another one. Obviously, For example, polymorphisms of the β(2)-
variance in drug response has many determi- adrenergic receptor (β(2)-AR) might affect
nants and singling out only one genetic factor response to β-agonists (e.g. albuterol) in
fails to account for the co-occurrence, inter- asthma patients. In particular, the coding vari-
play, and interactions of several other factors ants at position 16 within the β(2)-AR gene
(e.g. disease severity, exposure variability over (β(2)-AR-16) have been suggested to determine
time, physiological feedback mechanisms, and patient response to albuterol treatment (see
testing bias), all also important for molecular Case Example 15.1).
pharmacoepidemiology. Pharmacodynamic gene–drug interactions
In addition to metabolism, genes that alter may also affect the risk of adverse reactions.
the absorption and distribution of medications One example is a polymorphism in the gene
may also alter drug levels at tissue targets. coding for the bradykinin B2 receptor that has
These include, for example, genes that code for been associated with an increased risk of angi-
transporter proteins such as the ATP-binding otensin converting enzyme (ACE) inhibitor-
cassette transporter proteins (ABCB, also induced cough. Cough is one of the most
known as the multidrug-resistance [MDR]-1 frequently seen adverse drug reactions (ADRs)
gene), which has polymorphisms that have in ACE therapy and very often a reason for dis-
been associated with, for example, resistance continuation of therapy. The TT genotype and
to antiepileptic drugs. Patients with drug- T allele of the human bradykinin B(2) receptor
resistant epilepsy (approximately one of three gene were found to be significantly higher in
patients with epilepsy is a nonresponder) are patients with cough. However, similar to many
more likely to have the CC polymorphism of other studies, replication of these findings has
ABCB1, which is associated with increased been limited. Further research using genome-
expression of this transporter drug-efflux pro- wide association studies (GWAS) has sug-
tein. Of note, the ABCB1 polymorphism falls gested that other SNPs are related to intolerance
within an extensive block of linkage disequi- to ACE inhibitors, but again requires
librium (LD). LD is defined by a region in replication.
which multiple genetic variants (e.g. SNPs) are
correlated with one another due to population Gene–Drug Interactions and the
and evolutionary genetic history. As a result, a Causal Pathway of Disease
SNP may be statistically associated with dis- Along with altering the pharmacokinetic and
ease risk, but is also in LD with the true causa- pharmacodynamic properties of medications,
tive SNP. Therefore, the SNP under study may genetic polymorphisms may also alter the dis-
not itself be causal but simply linked to a true ease state that is the target of drug therapy. For
causal variant. One of the major challenges in example, antihypertensive medications that
genetics research at this time is developing work by a particular mechanism, such as the
Case Example 15.1
Background ●● Candidate genes enhances biological

Regular use of inhaled B-agonists for asthma plausibility.
may produce adverse effects and be no more
effective than as-needed use of these drugs. Limitations
●● Multiple polymorphisms tested on multi-
Question ple outcomes leads to concern of false
Can genetic polymorphisms in the B2- positives.
agonist receptor alter responsiveness to ●● LD: polymorphisms identified could be
inhaled B-agonists? “innocent bystanders” by being linked to
Approach the true causative mutations.
●● Perform genetic analysis within a rand-
omized clinical trial of regular versus as- Key Points

needed use of inhaled B-agonists, and ●● Genetic polymorphisms of drug targets
●● Compare the effects of multiple genetic may alter drug response.
polymorphisms on drug-response. ●● Because of the concern of false positives
and/or LD, replication studies and mecha-

Results
nistic studies remain critical to identifying
●● Regular use of inhaled B-agonist is associ-
true putative mutations that alter drug
ated with decline in efficacy among those
response.
with B(2)-AR-16 variants but not among
●● Effects of a gene may vary by the pattern
those with other variants tested.
of drug use, making it important to
●● No effect of genotype in those using inhaled
c
onsider all aspects of drug use (dose,
B-agonists in an as-needed manner.
duration, frequency, regularity, etc.)
Strengths in molecular pharmacoepidemiology
●●Randomized trial design eliminates con- studies.
founding by indication for frequency of
medication use.
increasing sodium excretion of some antihy- (Herceptin®), which is used for the treatment
pertensive medications, may have different of metastatic breast cancer patients with over-
effects depending on the susceptibility of the expression of the HER2 oncogene. The HER2
patient to the effects of the drug. Patients with protein is thought to be a unique target for tras-
a polymorphism in the α-adducin gene may tuzumab therapy in patients with this geneti-
have greater sensitivity to changes in sodium cally associated overexpression, occurring in
balance. A case–control study has suggested 10–34% of females with breast cancer. The case
that those with the α-adducin polymorphism of trastuzumab, together with another anti-
may be more likely to benefit from diuretic cancer drug, imatinib, which is especially
treatment than those without the effective in patients with Philadelphia
polymorphism. chromosome-positive leukemias, has pio-
Genetic variability in disease states also can neered successful genetically targeted therapy.
be critical for tailoring drug therapy to patients The association of somatic mutations to drug
with a specific genotype related both to the dis- response has received substantial interest.
ease and drug response. One example is the There are many targeted therapies now availa-
humanized monoclonal antibody trastuzumab ble that block the growth and spread of cancer
by interfering with specific molecules that are The Interplay of Various

involved in the growth, progression, and Mechanisms
spread of cancer.
It is useful to conceptualize how the effects of
Genetic polymorphisms that alter disease
genetic polymorphisms at different stages of
states can also play a role in drug safety. For
drug disposition and response might influence
example, factor V Leiden mutation, present in
an individual’s response to a medication. As an
about 1 out of 20 Caucasians, is considered an
example, an individual may have a genotype
important genetic risk factor for deep vein
that alters the metabolism of the drug, the
thrombosis and embolism. A relative risk of
receptor for the drug, or both. Depending on
about 30 in factor V carriers and users of oral
the combination of these genotypes, the indi-
contraceptives compared to noncarriers and
vidual might have a different response in terms
non-oral-contraceptive users has been
of both efficacy and toxicity (see Table 15.1). In
reported. This gene–drug interaction has also
the simplified example in Table 15.1, there is
been linked to the differential thrombotic risk
one genetic variant that alters drug metabo-
associated with third-generation oral contra-
lism and one genetic variant that alters recep-
ceptives compared with second-generation
tor response to a medication of interest. In this
oral contraceptives. Despite this strong associ-
example, among those who are homozygous
ation, Vandenbroucke et al. have calculated
for the alleles that encode normal drug metab-
that mass screening for factor V would result in
olism and normal receptor response, there is
denial of oral contraceptives for about
relatively high efficacy and low toxicity.
20 000 women positive for this mutation in
However, among those who have a variant that
order to prevent 1 death. Therefore, these
reduces drug metabolism, efficacy at a stand-
authors concluded that reviewing personal
ard dose could actually be greater (assuming a
and family thrombosis history, and only if suit-
linear dose–response relationship within the
able, factor V testing before prescribing oral
possible drug levels of the medication) but tox-
contraceptives, is the recommended approach
icity could be increased (if dose-related).
to avoid this adverse gene–drug interaction.
Among those who have a variant that reduces
This highlights another important role of
receptor response, drug efficacy will be reduced
molecular pharmacoepidemiology: determin-
while toxicity may not be different from those
ing the utility and cost-effectiveness (see also
who carry genotypes that are not associated
Chapter 18) of genetic screening to guide drug
with impaired receptor response (assuming
therapy.
Table 15.1 Hypothetical response to medications by genetic variants in metabolism and receptor genes.
Drug response
Gene affecting metabolisma Gene affecting receptor responsea Efficacy Toxicity
Wild-type Wild-type 70% 2%

Variant Wild-type 85% 20%
Wild-type Variant 20% 2%
Variant Variant 35% 20%
Source: Data from Evans and McLeod (2003).

a
Wild-type associated with normal metabolism or receptor response and variants associated with reduced
metabolism or receptor response.
that toxicity is not related to the receptor characteristic regardless of the type or subtype
responsible for efficacy). Among those who of cancer. For instance, in 2017 the US FDA
have variants for both genes, efficacy could be approved pembrolizumab for treatment of
reduced because of the receptor variant (per- unresectable or metastatic solid tumors that
haps not as substantially as those with an iso- have a biomarker, microsatellite instability-high
lated variant of the receptor gene because of (MSI-H) or mismatch repair deficient (dMMR),
the higher effective dose resulting from the in adult and pediatric patients. This is the first
metabolism gene variant), while toxicity could drug approval based on a tumor’s biomarker
be increased because of the metabolism without regard to the tumor’s original location.
variant.
Methodological Problems
The Progression to be Addressed by
and Clinical Application Pharmacoepidemiologic
of Molecular Research
The same methodological problems of phar-
Medications with a narrow therapeutic index macoepidemiology must be addressed in
are good targets for the use of molecular phar- molecular pharmacoepidemiology. These
macoepidemiology to improve the use and problems include those of chance and statisti-
application of medications. One example is cal power, confounding, bias, and generaliza-
warfarin. This example illustrates both the log- bility (see Chapters 2, 3, and 22).
ical progression of pharmacogenetics through However, the complex relationship between
molecular pharmacoepidemiology and the medication response and molecular and
complexity of moving pharmacogenetic data genetic factors generates some unique chal-
into practice. The enzyme primarily responsi- lenges in molecular pharmacoepidemiology.
ble for the metabolism of warfarin to its inac- These challenges derive from the large number
tive form is the cytochrome P450 2C9 variant of potential genetic variants that can modify
(CYP2C9). Case Example 15.2 illustrates both the response to a single drug, the possibility
the logical progression of pharmacogenetics that there is a small individual effect of any
through molecular pharmacoepidemiology one of these genes, the low prevalence of many
and the complexity of moving pharmacoge- genetic variants, and the possibility that a pre-
netic data into practice. sumptive gene–drug response relationship
Another pertinent example is in oncology. may be confounded by the racial and ethnic
Markers predicting response to anti-cancer mixture of the population studied. Thus, the
drugs are mostly related to the fact that drug methodological challenges of molecular phar-
efficacy can be greatly influenced by alterations macoepidemiology are closely related to issues
in drug targets and in related proteins present of statistical interactions, type I and type II
in tumor cells. Therefore, cancer targeted ther- errors, and confounding. First and foremost,
apies, directed to a specific cancer alteration, however, molecular pharmacoepidemiologic
may only be indicated for the subgroup of studies rely on proper identification of putative
patients with tumors carrying that molecular genotypes. In addition, in all research of this
target. Examples include trastuzumab and type, use of appropriate laboratory methods,
imatinib mentioned earlier in the chapter. such as high-throughput genotyping technolo-
It is important to note that novel treatment gies, is necessary. Similarly, appropriate qual-
strategies can be based on a specific genetic ity control procedures must be considered to
Methodological Problems to be Addressed by Pharmacoepidemiologic Researc 265
Case Example 15.2

The complexity of the progression and clini- dosing in African Americans. The EU-PACT
cal application of molecular pharmacoep - UK trial demonstrated improvement in
idemiology anticoagulation control. A third trial, the
GIFT trial, examined pharmacogenetic dos-
Background
ing in orthopedic patients, compared with
Warfarin is a narrow therapeutic index drug.
a clinical algorithm dosing arm. This trial
Underdosing or overdosing, even to a mini-
demonstrated benefit from pharmacoge-
mal degree, can lead to significant morbidity
netic dosing, but could not address the
(thromboembolism and/or bleeding).
question of the effects of pharmacogenet-
Question ics in African Americans due to the enrol-
Can genetic variants be identified and used ment of very few African American patients.
to alter the dosing of warfarin and thus Together, these trials demonstrate the
improve safety and effectiveness? need for, and benefit of, pharmacogenetic
clinical trials testing different strategies in
Approach and Results
different patient populations.
●● Multiple study designs have been used to
address this question. Strengths
●● Pharmacogenetic studies identified the A logical series of studies, each with its own
effect of CYP2C9 polymorphisms on war- strengths and limitations, has improved our
farin metabolism. Additional research understanding of genetic variability in
clearly demonstrated that the vitamin K response to warfarin.
epoxide reductase complex 1 (VKORC- 1)
Limitations
gene is responsible for coding the target
●● No randomized trials have yet shown that
for warfarin.
one can reduce adverse events and
●● Numerous observational studies then
enhance effectiveness of warfarin by
demonstrated differential response to
knowing a patient’s genetic make-up.
warfarin in patients with CYP2C9 and
●● Only about 50% of variability in warfarin
VKORC1 variants, particularly for the out-
response can be explained by existing
come of the required maintenance dose of
algorithms–other polymorphisms or clini-
warfarin.
cal factors (e.g. adherence) may be
●● The development of algorithms to predict a
important.
maintenance warfarin dose that combines
clinical and genetic data then suggested Key Points
that improvements may be made by incor- ●● The process of fully understanding the
porating genetic data into dosing algo- effects of polymorphisms on drug response
rithms. However, small clinical trials did not requires multiple studies and often sub-
demonstrate the utility of genotyping. stantial resources.
●● Large- scale trials were performed to ●● Our understanding of genetic variants has
answer the important question of clinical progressed so rapidly that new questions
utility. The Clarification of Optimal are often raised that have implications for
Anticoagulation through Genetics (COAG) clinical practice even as old ones are
trial demonstrated no benefit of answered.
pharmacogenetic dosing on anticoagula- ●● Before using genetic data to alter drug
tion control overall, and worsening of anti- prescribing, prospective evaluation is

coagulation control with pharmacogenetic needed.
obtain meaningful data for research and clini- epending on the presence or absence of the
d
cal applications. Recent next-generation genetic variant. This difference can be either
sequencing (NGS) techniques have only high- on the multiplicative or additive scale. On the
lighted further the need for, and complexity of, multiplicative scale, interaction is present if
obtaining valid genotyping results. This sec- the effect of the combination of the genotype
tion will focus on the methodological chal- and medication exposure relative to neither is
lenges of studying interactions, minimizing greater than the product of the measure of
type I and type II errors, and accounting for effect of each (genotype alone or medication
confounding, particularly by population alone) relative to neither. On the additive scale,
admixture (defined below). interaction is present if the effect of the combi-
nation of the genotype and medication expo-
sure is greater than the sum of the measures of
Interactions
effect of each alone, again all relative to
Along with examining the direct effect of genes neither.
and other biomarkers on outcomes, molecular For studies examining a dichotomous medi-
pharmacoepidemiologic studies must often be cation exposure (e.g. medication use versus
designed to examine effect modification nonuse), a dichotomous genetic exposure (e.g.
between medication use and the genes or bio- presence versus absence of a genetic variant),
markers of interest. That is, the primary meas- and a dichotomous outcome (e.g. myocardial
ure of interest is often the role of biomarker infarction occurrence versus none), there are
information on the effect of a medication. For two ways to consider presenting and analyzing
purposes of simplicity, this discussion will use interactions. The first is as a stratified analysis,
genetic variability as the measure of interest. comparing the effect of medication exposure
Effect modification is present if there is a versus non-exposure on the outcome in two
difference in the effect of the medication strata: those with the genetic variant and those
Table 15.2 Two ways to present effect modification in molecular pharmacoepidemiologic studies using
case–control study as a model.
Stratified analysis
Genotype Medication Cases Controls Odds ratio Information provided
+ + a b ad/bc Effect of medication vs. no medication

among those with the genotype
− c d
− + e f eh/fg Effect of medication vs. no medication

among those without the genotype
− g h
2 × 4 Table
Genotype Medication Cases Controls Odds ratio Information provided

+ + a b ah/bg = A Joint genotype and medication vs. neither
+ − c d ch/dg = B Genotype alone vs. neither
− + e f eh/fg = C Medication alone vs. neither
− − g h Reference Reference Group
without (e.g. see Table 15.2). The second is to exposures (2 × 4 table) as well as the effect of
present a 2 × 4 table (also shown in Table 15.2). the medication in the presence or absence of
In the first example (stratified analysis), one the genotypic variant (stratified table).
compares the effect of the medication among
those with the genetic variant to the effect of
Type I Error
the medication among those without the
genetic variant. In the second example (the The chance of type I error (concluding there is
2 × 4 table), the effect of each combination of an association when in fact one does not exist)
exposure (i.e. with both genetic variant and increases with the number of statistical tests
medication; with genetic variant but without performed on any one data set (see also
medication; with medication but without Chapter 3). It is easy to appreciate the potential
genetic variant) is determined relative to the for type I error in a molecular pharmacoepide-
lack of exposure to either. The advantage of miologic study that examines, simultaneously,
the 2 × 4 table is that it presents separately the the effects of multiple genetic factors, the
effect of the drug, the gene, and both relative to effects of multiple nongenetic factors, and the
those without the genetic variant and without interaction between and among these factors.
medication exposure. In addition, presentation One of the reasons cited for nonreplication of
of the data as a 2 × 4 table allows one to directly study findings in molecular pharmacoepide-
compute both multiplicative and additive miology is type I error. Limiting the number of
interactions. In the example given in Table 15.2, associations examined to those of specific can-
multiplicative interaction would be assessed didate genetic variants that are suspected of
by comparing the odds ratio for the combina- being associated with the outcome is one
tion of genotype and medication exposure to method to limit type I error in pharmacoepide-
the product of the odds ratios for medication miology. However, with increasing emphasis
alone and genotype alone. Multiplicative inter- in molecular pharmacoepidemiologic studies
action would be considered present if the odds on identifying all variants within a gene (and
ratios for the combination of medication and all variants within the genome) and examining
genotype (A in Table 15.2) was greater than the multiple interactions, this method of limiting
product of the odds ratios for either alone type I error is often not tenable. Some other
(B × C). Additive interaction would be consid- currently available solutions are discussed in
ered present if the odds ratio for the combina- the next section.
tion of genotype and medication use (A) was
greater than the sum of the odds ratios for
Type II Error
medication use alone and genotype alone
(B + C). The 2 × 4 table also allows the direct Because it has been hypothesized that much of
assessment of the number of subjects in each the genetic variability leading to phenotypic
group along with the respective confidence expression of complex diseases results from
interval for the measured effect in each of the the relatively small effects of many relatively
groups, making it possible to directly observe low prevalence genetic variants, the ability to
the precision of the estimates in each of the detect a gene–response relationship is likely to
groups and therefore better understand the require relatively large sample sizes to avoid
power of the study. Furthermore, attributable type II error (concluding there is no associa-
fractions can be computed separately for each tion when in fact one does exist). The sample
of the exposures alone and for the combination size requirements for studies that examine the
of exposures. In general, presenting the data in direct effect of genes on medication response
both manners is optimal because it allows the will be the same as the requirements for exam-
reader to understand the effect of each of the ining direct effects of individual risk factors on
outcomes. With relatively low prevalences of among those with the genetic variant, only 385
polymorphisms and often low incidence of participants contributed to the analyses.
outcomes (particularly in studies of ADRs), In order to minimize false negative findings,
large sample sizes are typically required to further efforts must be made to ensure ade-
detect even modest associations. For such quate sample sizes for molecular pharmacoep-
studies, the case–control design (see Chapter 2) idemiologic studies. Because of the complex
has become a particularly favored approach for nature of medication response, and the likeli-
molecular pharmacoepidemiologic studies hood that at least several genes are responsible
because of its ability to select participants for the variability in drug response, studies
based on the outcome of interest, and its ability designed to test for multiple gene–gene and
to study the effects of multiple potential geno- gene–environment interactions (including
types in the same study. other medications, environmental factors,
Studies that are designed to examine the adherence to medications, and clinical factors)
interaction between a genetic polymorphism will, similarly, require large sample sizes.
and a medication will require even larger sam-
ple sizes. This is because such studies need to
Confounding by Population
be powered to compare those with both the
Admixture
genetic polymorphism and the medication
exposure with those who have neither. As an When there is evidence that baseline disease
example, the previously mentioned case–con- risks and genotype frequencies differ among
trol study of the α-adducin gene and diuretic ethnicities, the conditions for population strat-
therapy in patients with treated hypertension ification (i.e. population admixture or con-
examined the effects of the genetic polymor- founding by ethnicity) may be met. Population
phism, the diuretic therapy, and both in combi- admixture is simply a manifestation of con-
nation. There were 1038 participants in the founding by ethnicity, which can occur if both
study. When comparing the effect of diuretic baseline disease risk and genotype frequency
use with no use and comparing the effect of vary across ethnicity. For example, the African–
the genetic variant with the nonvariant allele, American population represent admixture of
all 1038 participants were available for com- at least three major continental ancestries
parison (Table 15.3). However, when examin- (African, European, and Native American).
ing the effect of diuretic therapy versus nonuse The larger the number of ethnicities involved
Table 15.3 Gene-exposure interaction analysis in a case–control study.
Odds Ratio (OR) for Stroke

Diuretic Use Adducin Variant Cases Controls Myocardial Infarction
0 0 A00 B00 1.0

103 248
0 1 A01 B01 1.56
85 131
1 0 A10 B10 1.09
94 208
1 1 A11 B11 0.77
41 128
Source: Data from Psaty et al. (2002).

in an admixed population, the less likely that heritability” of complex disease suggests that
population stratification can be the explana- other classes of genetic variation may explain
tion for biased associations. Empirical data much of the genetic contribution to common
show that carefully matched, moderate-sized disease.
case–control samples in African–American There are two primary approaches for gene
populations are unlikely to contain levels of discovery: candidate gene association studies
population admixture that would result in sig- and genome-wide studies. In the former, genes
nificantly inflated numbers of false-positive are selected for study on the basis of their
associations. There is the potential for popula- plausible biological relevance to drug
tion structure to exist in African–American response. While this allows for identification
populations, but this structure can be elimi- of variants with a priori biological plausibility,
nated by removing recent African or Caribbean it is limited by our partial knowledge of which
immigrants, and limiting study samples to genetic variants may actually be responsible
resident African–Americans. Based on the lit- for variable drug effects. In the latter, DNA
erature that has evaluated the effects of consequences are examined for associations with
founding by ethnicity overall, and specifically outcomes, initially irrespective of biological
in African–Americans, there is little empirical plausibility. The benefit of this approach is
evidence that population stratification is a that it does not rely on our limited knowledge
likely explanation for bias in point estimates or of genetics; the disadvantage is that the bio-
incorrect inferences. Nonetheless, population logical plausibility of the findings may then
admixture must be considered in designing need to be confirmed.
and analyzing molecular pharmacoepidemio- One example of the genome-wide approach
logic studies to ensure that adequate adjust- are GWAS. GWAS rely on LD, defined above as
ment can be made for this potential confounder. the correlation between alleles at two loci. The
It is important to note that poor study design GWAS approach uses DNA sequence variation
may be more important than population strati- (e.g. SNPs) found throughout the genome, and
fication in conferring bias to association does not rely on a priori functional knowledge
studies. of gene function. A number of factors influ-
ence the success of these studies. Appropriate
epidemiologic study designs and adequate sta-
urrently Available
C tistical power remain essential. Thorough
characterization of LD is essential for replica-
Solutions
tion of GWAS: the haplotype mapping
(HapMap) consortium and other groups have
Identifying Additional Genetic
shown that the extent of LD varies by ethnicity,
Contributions to Drug Response
which may affect the ability to replicate find-
A great concern of the identification of low ings in subsequent studies. Particularly
penetrance alleles (those in which few of the informative SNPs that best characterize a
individuals who have the allele exhibit the genomic region can be used to limit the amount
clinical symptoms associated with that allele) of laboratory and analytical work in haplotype-
is that they have not yet been able to explain based studies. It has been hypothesized that
the majority of the estimated genetic contribu- studies that consider LD involving multiple
tion to disease etiology. Based on studies of SNPs in a genomic region (i.e. a haplotype) can
families or phenotypic variability, most loci increase power to detect associations by
have been found to explain less than half (and 15–50% compared with analyses involving
at times as little as 1%) of the predicted herita- only individual SNPs. Finally, even if genome-
bility of many common traits. This “missing wide scans may identify markers associated
with the trait of interest, a challenge will be to often a major methodological and logistical
identify the causative SNPs. challenge in case–control studies. One limita-
Newer, sequencing technologies have made tion of the case-only design is that it relies on
it possible to study rarer genetic variants. the assumption of independence between
While Sanger sequencing is still considered the exposure (medication use) and genotype.
gold standard in clinical testing, its limitations Although this assumption may be valid (in the
include low throughput and high cost. Broadly, absence of knowing genotype clinically, it may
NGS describes technologies that utilize clon- be reasonable to assume that the use of the
ally amplified or single-molecular templates medication is not related to patients’ geno-
that are then sequenced in a massively parallel types), it is certainly possible that, within
fashion. The advance of NGS technologies has observational studies, the genotype, by altering
been enabled by innovation in sequencing response to medications targeted at a specific
chemistries, better imaging, microfabrication, disease, could affect the medications being
and information technology. In addition, bio- prescribed to patients. Another method is to
informatics tools for data analysis and man- perform the case-only study within a rand-
agement and sample preparation methods omized trial, where drug use is randomly
have rapidly evolved along with the sequenc- assigned.
ing technologies, translating to reductions in
the amount of input materials required. In
Type I Error and Replication
2013, the US FDA approved marketing for the
first time for a next-generation sequencer, Given concerns of type I error (along with
Illumina’s MiSeqDx, which allows the devel- other methodologic concerns such as uncon-
opment and use of innumerable new genome- trolled confounding, publication bias, and
based tests. LD), a key issue in molecular epidemiology is
Clearly, candidate gene and genome-wide the ability to replicate association study find-
approaches are not mutually exclusive. Both ings. Replication of association studies is
have the potential to identify important vari- required not only to identify biologically plau-
ants that may be clinically useful. sible causative associations, but also to con-
clude that a candidate gene has a meaningful
etiological effect.
Interactions
The lack of replication can be explained by
Along with traditional case–control and cohort false positive reports (e.g. spurious associa-
studies, the case-only study can be used for tions), by false negative reports (e.g. studies
molecular pharmacoepidemiologic studies that are insufficiently powerful to identify the
designed to examine interactions between association), or by actual population differ-
genes and medications. In this design, cases, ences (e.g. the true associations are different
representing those with the outcome or pheno- because of differences in genetic background,
type of interest, are selected for study, and the exposures, etc.).
association between genetic variants and med- In order to achieve believable, replicable
ication use is determined among these cases. association results, investigators must consider
Assuming that the use of the medication is factors that influence the design, analysis, and
unrelated to the genotype, the case-only study interpretation of these studies. These include
provides a valid measure of the interaction of adequate sample size, proper study design, and
the genotype and the medication on the risk of characterization of the study population, par-
the outcome. ticularly when replication studies themselves
One strength of the case-only study design is are not comparable in terms of participant
that it eliminates the need to identify controls, characteristics or other confounding factors.
The Futur 271
Adequate reporting of genetic association as interactions) are combined in order to

studies is important to allow assessment of improve the ability to detect such associations
their strengths and weaknesses. The STREGA (see Chapter 20).
statement (Strengthening the Reporting of
Genetic Association studies) is an extension of
Confounding by Population
the STROBE statement (Strengthening the
Admixture
Reporting of Observational Studies in
Epidemiology) that provides a checklist to help Although population stratification is unlikely
researchers and journals. to be a significant source of bias in epidemio-
Data analytical methods can complement logic association studies, this assumes ade-
replication studies to address multiple testing quate adjustment for population genetic
and type I error. Bonferroni correction is the structure. A number of analytical approaches
most basic approach for adjusting multiple exist to either circumvent problems imposed
testing. However, this method is considered by population genetic structure or that use this
too conservative for tightly linked SNPs and it structure in gene identification. The “struc-
may wipe out many small effects that one may tured association” approach identifies a set of
actually expect (i.e. increase risk of type II individuals who are drawing their alleles from
errors). The false discovery rate (FDR) different background populations or ethnici-
approach is less conservative for controlling ties. This approach uses information about
for multiple analyses of the data. The FDR genotypes at loci that lie in regions other than
method estimates the expected proportion of the location of the gene of interest (i.e.
false-positives among associations that are “unlinked markers”) to infer their ancestry
declared significant, which is expressed as a (often referred to as ancestry informative
q-value. Under a Bayesian approach there is no markers) and learn about population struc-
penalty for multiple testing because the prior ture. It further uses the data derived from these
probability of an association should not be unlinked markers to adjust the association test
affected by the tests that the investigator statistic. By adjusting for these ancestry
chooses to conduct. However, without strict informative markers, one can adjust for differ-
standards, one may choose an exaggerated ences in ancestry.
prior plausibility of a model that is supported a
posteriori.
The Future
Type II Error
Scientific and clinical developments in biology
Reducing type II error (concluding that there is and molecular biology, particularly in the field
no association when one does exist in fact) of genomics and other biomarkers, have and
essentially involves a logistical need to ensure will continue to affect the field of pharmacoep-
adequate sample size (see also Chapter 3). One idemiology in a significant way. Translating
approach to increasing the sample size of biomarkers from the lab and experimental
molecular pharmacoepidemiologic studies is studies to clinical practice has been a difficult
to perform large, multicenter collaborative path. Often, initial promising findings on drug-
studies. Another is to combine multiple, sepa- gene interactions to predict clinical drug
rately performed cohorts. responses could not be replicated in subse-
Another potential solution to minimizing quent studies. For sure, the ability of genes and
type II error is through meta-analysis, whereby other biomarkers to improve patient care and
smaller studies, which are, individually, not outcomes will need to be tested in properly
powered to detect specific associations (such controlled studies, including randomized
c ontrolled trials in some circumstances. The nificant amount of information will be novel
positive and negative predictive value of carry- and/or of unknown clinical importance.
ing a genetic variant will be important deter- A major area that requires further develop-
minants of the ability of the variant to improve ment is in establishing the clinical utility of the
outcomes. Those genetic variants with good identified markers/strategies for patients and
test characteristics may still need to be evalu- healthcare systems. The level of evidence
ated in properly controlled trials. Such studies required to establish that a marker is clinically
could examine several ways to incorporate useful and should be introduced for routine
genetic testing into clinical practice, including use has been discussed extensively but consen-
the use of genetic variants in dosing algo- sus has not been reached. Genetic and molecu-
rithms, in selection of a specific therapeutic lar studies are increasingly being incorporated
class of drug to treat a disease, and in avoid- in large clinical trials, which can lead to the
ance of using specific medications in those at identification of subgroups of patients with
high risk for ADRs. These scientific advances clear benefit from drugs, accelerating the dis-
are also finding their way into drug discovery covery of effective therapies for selected popu-
and development in order to rationalize drug lations. Another challenge to the
innovation and to identify good and poor implementation of genetic testing is the fact
responders, both in terms of efficacy and that pharmacogenetics knowledge is con-
safety, of drug therapy in an earlier phase. The stantly being updated. Clinicians need to inter-
cost-effectiveness of such approaches is also of pret the results of these tests in accordance
great interest because the addition of genetic with current understanding of the association
testing adds cost to clinical care (see also between pharmacogenetic variation and drug
Chapter 18). Research will be needed to deter- effects.
mine the cost-effectiveness of new biomarker What this all means for the future of phar-
and genetic tests as they are developed. macoepidemiology is a challenging question.
NGS will also require the development of Genotype data will increasingly become avail-
novel approaches to data analyses. There are able and will enrich pharmacoepidemiologic
three levels of analysis that are conducted by analysis. New methods (e.g. sequencing) will
NGS technologies: (i) targeted gene panels provide new opportunities but also new chal-
focus on a limited set of genes allowing for lenges to analyzing pharmacoepidemiologic
greater depth of coverage. The advantages data. Further, although it is useful to charac-
include higher analytical sensitivity and speci- terize the three different pathways of how
ficity, and improved ability to interpret the drug-gene interactions may occur as was done
results in a clinical context because only genes in this chapter, this stratification is most likely
with an established role in the disease are an oversimplification of the large plethora of
sequenced, (ii) exome sequencing tests all cod- possible mechanisms of how drugs, genes, and
ing regions of the human genome, and (iii) patient outcomes are interrelated. All these
whole-genome sequencing analyzes the entire may have consequences for how molecular
three billion bases of the genome. The targeted pharmacoepidemiologic studies are designed,
approach to genome sequencing is the more conducted, and analyzed. In addition, the
widespread clinical implementation of NGS more that genotype testing is applied in clini-
technologies. This is because only some of the cal practice, the more drug exposure will be
enormous amount of genetic information gen- influenced by such tests, making genotype and
erated by exome or whole-genome sequencing drug exposure non-independent factors.
can be interpreted and is actionable. Along Finally, just as for all research, the ethical,
with the bioinformatics challenges of manag- legal, and social implications of genetic testing
ing and validating such large data sets, a sig- must be considered and addressed (see also
Key Point 273
Chapter 16). Pharmacogenetic testing raises genome-wide approach; these approaches

issues of privacy concerns, access to health are complementary, not mutually
care services, and informed consent. For exam- exclusive.
ple, concern has been raised that the use of ●● The methodological challenges of molecular
genetic testing could lead to targeting of thera- pharmacoepidemiology are closely related to
pies to only specific groups (ethnic or racial) of issues of statistical interactions, type I and
patients, ignoring others, and to loss of insur- type II errors, and confounding.
ance coverage for certain groups of individu- ●● Case-only studies can be used to measure
als. There is also a concern that medicines will the interaction between genetic variants
be developed only for the most common, com- and medications and eliminate the diffi-
mercially attractive genotypes, leading to culty and inefficiency of including a con-
“orphan genotypes.” Equally importantly, as trol group. However, they rely on the
more and more genetic data are collected on assumption of independence between
individuals as part of routine clinical care, the medication use and genetic variants among
requirements and methods for returning unan- those without disease, an assumption that
ticipated genetic results must be carefully may not be met.
determined and implemented. ●● Given concerns of type I error (along with
All of these issues are challenges to over- other methodological concerns such as
come as we continue to reap the benefits of the uncontrolled confounding and LD), a key
tremendous strides made in determining the issue in molecular epidemiology is the
molecular basis of disease and drug response. ability to replicate association study
findings.
●● Because genetic variability leading to pheno-
Key Points typic expression of complex diseases results
from the relatively small effects of many rel-
●● Genes can affect a drug response via: altera- atively low prevalence genetic variants, the
tion of drug pharmacokinetics, pharmaco- ability to detect a gene–response relation-
dynamic effects on drug targets, and gene–drug ship is likely to require relatively large sam-
interactions in the causal pathway of ple sizes to avoid type II error. Methods to
disease. ensure adequate sample sizes include the
●● Molecular pharmacoepidemiology is the use of large, multicenter collaborative stud-
study of the manner in which molecular bio- ies; assembly and genotyping of large, rela-
markers (often, but not exclusively, genes) tively homogenous populations for multiple
alter the clinical effects of medications in studies; and meta-analysis.
populations. ●● Population stratification can distort the
●● Molecular pharmacoepidemiology answers gene-medication response association.
questions related to: the population preva- Although unlikely to be a significant source
lence of SNPs and other genetic variants; of bias in well-controlled epidemiological
evaluating how these SNPs alter disease out- association studies, a number of analytical
comes; assessing the impact of gene–drug approaches exist to either circumvent prob-
and gene–gene interactions on disease risk; lems imposed by population genetic struc-
and evaluating the usefulness and impact of ture or that use this structure in gene
genetic tests in populations exposed, or to be identification.
exposed, to drugs. ●● The ability of genes and other biomarkers to
●● Identifying genes that alter drug response improve patient care and outcomes needs to
for molecular pharmacoepidemiology stud- be tested in properly controlled studies,
ies can use a candidate gene approach or a including randomized controlled trials in
some cases. Similarly, the cost-effectiveness ●● The ethical, legal, and social implications of
of such approaches must be justifiable given genetic testing must be considered and
the additional costs of genetic testing in clin- addressed, just as they must be considered
ical care. for all research.
Further Reading
Botto, L.D. and Khoury, M.J. (2004). Facing the Klein, T.E., Altman, R.B., Eriksson, N. et al.
challenge of complex genotypes and gene– (2009). Estimation of a warfarin dose with
environment interaction: the basic clinical and pharmacogenetic data.
epidemiologic units in case–control and International Warfarin Pharmacogenetics
case-only designs. In: Human Genome Consortium. N. Engl. J. Med. 360: 753–764.
Epidemiology (eds. M.J. Khoury, J. Little and Lohmueller, K.E., Pearce, C.L., Pike, M. et al.
W. Burke), 111–126. New York: Oxford (2003). Meta-analysis of genetic association
University Press. studies supports a contribution of common
Bournissen, F.G., Moretti, M.E., Juurlink, variants to susceptibility to common disease.
D.N. et al. (2009 Apr). Polymorphism of Nat. Genet. 33: 177–182.
the MDR1/ABCB1 C3435T drug-transporter Mallal, S., Phillips, E., Carosi, G. et al. (2008).
and resistance to anticonvulsant drugs: HLA-B*5701 screening for hypersensitivity to
a meta-analysis. Epilepsia 50 (4): abacavir. N. Engl. J. Med. 358: 568–579.
898–903. Manolio, T.A., Collins, F.S., Cox, N.J.
Caraco, Y., Blotnick, S., and Muszkat, M. (2008). et al. (2009). Finding the missing heritability
CYP2C9 genotype-guided warfarin of complex diseases. Nature 461:
prescribing enhances the efficacy and safety 747–753.
of anticoagulation: a prospective randomized Psaty, B.M., Smith, N.L., Heckbert, S.R. et al.
controlled study. Clin. Pharmacol. Ther. 83: (2002). Diuretic therapy, the alpha-adducin
460–470. gene variant, and the risk of myocardial
Evans, W.E. and McLeod, L.J. (2003). infarction or stroke in persons with treated
Pharmacogenomics–drug disposition, drug hypertension. JAMA 287:
targets, and side effects. N. Engl. J. Med. 348: 1680–1689.
528–549. Roden, D.M., McLeod, H.L., Relling, M.V. et al.
Fleeman N, Dundar Y, Dickson R, et al. (2019 Aug 10). Pharmacogenomics. Lancet
Cytochrome P450 testing for prescribing 394 (10197): 521–532.
antipsychotics in adults with schizophrenia: Roses, A.D. (2008). Pharmacogenetics in drug
systematic review and meta-analyses. Pharm. discovery and development: a translational
J. 2011 Feb;11(1):1–14. perspective. Nat. Rev. Drug Discov. 7 (10):
Gaedigk, A., Sangkuhl, K., Whirl-Carrillo, M. 807–817.
et al. (2017). Prediction of CYP2D6 phenotype Thomas, D. (2010 Apr). Gene-environment-
from genotype across world populations. wide association studies: emerging
Genet. Med. 19: 69–76. https://doi. approaches. Nat. Rev. Genet. 11 (4):
org/10.1038/gim.2016.80. Epub 2016 Jul 7. 259–272.
Kimmel, S.E., French, B., Kasner, S.E. et al. Tse, S.M., Tantisira, K., and Weiss, S.T. (2011).
(2013). A pharmacogenetic versus a clinical The pharmacogenetics and
algorithm for warfarin dosing. N. Engl. J. Med. pharmacogenomics of asthma therapy.
369: 2283–2293. Pharm. J. 11: 383–392.
Further Reading 275
Veenstra, D.L. (2004). The interface between stratification in epidemiologic studies of

epidemiology and pharmacogenomics. In: gene–gene or gene–environment
Human Genome Epidemiology (eds. M.J. interactions. Cancer Epidemiol. Biomark.
Khoury, J. Little and W. Burke), 234–246. Prev. 15 (1): 124–132. https://doi.
New York: Oxford University Press. org/10.1158/1055-9965.EPI-05-0304. PMID:
Wang, Y., Localio, R., and Rebbeck, T.R. (2006). 16434597.
Evaluating bias due to population
276
16
Bioethical Issues in Pharmacoepidemiologic Research

Laura E. Bothwell1, Annika Richterich2, and Jeremy Greene3
1
Yale School of Public Health, New Haven, Connecticut, USA
2
Maastricht University, Maastricht, The Netherlands
3
Johns Hopkins University, Baltimore, Maryland, USA
Introduction universal normative cornerstones of pharma-

coepidemiology ethics. The same goes for the
Because the bioethical issues involved in phar- injunction that objectives and results of phar-
macoepidemiologic research are closely related macoepidemiologic research should be inde-
to changing patterns of drug usage and chang- pendent from economic and promotional
ing technologies of surveillance and data anal- interests of pharmaceutical companies or
ysis, it is impossible to understand them device manufacturers. Yet these principles are
without attention to historical and sociological not simple to implement systematically at an
perspectives. The field of pharmacoepidemiol- international level. In this chapter, we explore
ogy emerged as a result of broader recent the emergence and conduct of pharmacoepi-
developments in medical therapeutics, con- demiologic research in three major global set-
comitant to the expansion and refinement of tings in which the field developed (North
the field of bioethics. Some key bioethical prin- America, Europe, and East Asia) and some of
ciples relevant to pharmacoepidemiologic the key challenges, tensions, and trends in his-
research have remained significant over time, toric and current international ethical policies
others have only gained attention in recent toward pharmacoepidemiology.
years. This chapter briefly introduces historical
and sociological dimensions of pharmacoepi-
demiology from an international perspective,
linical Problems to be Addressed
C
with an eye toward commonalities and differ-
ences in national variations in ethical
approaches to the field. Research
It is widely believed that pharmacoepidemi-
ologic studies should create data that benefit The Emergence, Changing Methods,
public health, improve drug safety, and ensure and Moral Stakes of Pharmacoepidemiology
efficacy. The protection of research subjects’ in Twentieth Century North America
rights and safety, their wellbeing, dignity, In 1962 a series of epidemiological reports initi-
autonomy and privacy, as well as the reliability ated by the German physician Widukind Lenz
and robustness of generated data are relatively connected a recent increase in phocomelia, a
birth defect which resulted in grossly visible fraction of all cases and the total number of
limb deformities, with maternal use of the pop- patients receiving a drug was unknown.”
ular new anti-nausea medicine Contergan (tha- The 1962 Kefauver-Harris Amendments –
lidomide). Images of thalidomide children passed largely on the strength of popular moral
became an international symbol of the ethical outrage over thalidomide – demanded phar-
failure of the medical profession and the regu- maceutical manufacturers establish records
latory state to protect vulnerable populations and make reports to the FDA of “data relating
from the harmful effects of widely marketed to clinical experience and other data or infor-
new drugs. Contergan had been extensively mation, received or otherwise obtained” for all
marketed to physicians and consumers alike, new drugs. By 1967 the agency had developed
and its premarket testing and post-market pro- a protocol requiring manufacturers to seek and
motion had emphasized its remarkably non report any reported or published case reports
toxic safety profile by available standards of related to putative side effects of their mar-
clinical pharmacology. As Lenz’s work was keted products. Any novel or unexpected
read internationally, his careful use of the cor- adverse effect was to be reported to the agency
relative techniques of infectious disease epide- within 15 days; other information “pertinent to
miology within the terrain of prescription drug the safety or effectiveness of the drug” was to
use documented not only the unseen dangers be reported quarterly for the first year after
of newly marketed drugs but also the need for a approval, twice in the second year, and annu-
new discipline of pharmaceutical epidemiol- ally thereafter. Yet this kind of information
ogy to scour observational data for therapeutic could become actionable only after years of
effects and adverse reactions that could clearly case reports, and then only if one of the rela-
be associated with drug use in clinical practice. tively few FDA staffers took active interest in
The recognition that the risks of new drugs pursuit of a specific question of drug harm.
could be better understood when they were The hospital became the center of early pro-
consumed by broad numbers of patients had grams of pharmacoepidemiologic surveillance.
been evident long before Lenz’s epidemiology By 1964, the FDA and AMA had built a surveil-
of thalidomide-associated phocomelia. Indeed, lance program involving more than 600 hospi-
the history of federal drug regulation in the tals, which became the focus of early
United States can be recounted as a succession pharmacoepidemiologic research by Johns
of measures taken in response to dangers of Hopkins University’s Leighton Cluff, Harvard
drugs that became apparent after widespread University’s Thomas Chalmers, and Tufts
consumption by the general public. However, University’s Hershel Jick. Yet the data were
until the 1960s the Food and Drug still only as good as the reporting physicians’
Administration (FDA) had very limited author- records. As Leighton Cluff noted, an early vali-
ity in the post-market regulation of drugs. The dation system of reporting efforts at the Johns
agency had neither direct means to control Hopkins Hospital “proved completely unsatis-
physician prescriptions nor resources to gather factory for detecting drug reactions. . .[d]uring
data on prescribing of newly marketed drugs. recent daily intensive surveillance of one hos-
While the Committee on Pharmacy and pital service, four times as many reactions
Chemistry of the American Medical were detected than had been reported on the
Association (AMA) nominally maintained cards from the entire hospital.” Would-be epi-
more influence in both arenas, it depended demiologists of adverse drug effects needed a
entirely upon voluntary physician reports, and way to circumvent the physician as reporting
Committee members complained loudly that device – and the digitization of data provided
the system itself was doomed to failure; as one an appealing solution. Cluff’s attempts at com-
report noted, “physicians reported only a small puterized drug monitoring involved the
278 16 Bioethical Issues in Pharmacoepidemiologic Research
c reation of three linked datasets for every drug than had previously been anticipated. More
received by every patient in a dedicated hospi- than one-third of patients on the Shattuck
tal ward. D. J. Finney, another early theorist of wards experienced at least one drug-associated
computerized drug monitoring, expressed adverse reaction during the first year of study.
these data sets as a linked “P-D-E system,” in By 1967 the Boston group had established a
which P(atient) population data would be sys- numerator/denominator approach for com-
tematically gathered within a set geographic or paring drug usage between long-term and
hospital catchment area, the D(rug) data acute hospitals through a network of five hos-
would include records of all relevant prescrip- pitals in Boston. By 1968, over 2500 patients
tions, and E(vent) collection would record all had been entered and discharged from the sur-
untoward reactions potentially attributable to veillance system, with over 26 000 monitored
the drugs prescribed. drug exposures, representing more than
Proponents of drug monitoring imagined a 700 individual drugs. Commonly-prescribed
linked system of inpatient surveillance wards drugs, such as digoxin and heparin, could be
circling the globe, which could act as pharma- reported in detail, yielding novel information
covigilance sensors, detecting early signals of related to their clinical pharmacology and
possible drug harms and providing descriptive their interactions with other drugs. The system
data regarding their frequency, severity, and enabled the observation of not only obvious
relative strength of association. Finney pre- drug reactions (such as a rash) but also other
dicted that surveillance would change phar- clinical events (such as heart attacks or kidney
macoepidemiology from a reactive into a failure) that could only be associated with
proactive field. Allowing that “much is due to drugs by careful epidemiologic surveillance.
Lenz for his discovery in 1961 [that thalido- As the Boston Collaborative Drug
mide was associated with phocomelia],” he Surveillance Program escalated its activities
also boasted that “a monitor could have sig- and exported its methods to other sites, these
naled a warning 1½–2 years earlier.” new data provoked a series of drug scandals
Automated inpatient surveillance systems lib- that emphasized both the utility and the limi-
erated pharmacoepidemiology from the “weak tations of the new forms of pharmacovigilance.
link” of the reporting physician. With public Clioquinol, an anti-infective that had been in
and private support from the United States use since the 1930s, was found to be associated
Public Health Service and the Pharmaceutical with subacute myelo-optic neuropathy in 1970,
Manufacturers Association, Dennis Slone, over three decades after its initial introduction.
Hershel Jick, and Ivan Borda demonstrated the An association between the synthetic estrogen
feasibility of implementing an automated diethylstilbestrol (DES) and a rare form of cer-
hospital-based drug monitor system in 1966. vical clear cell adenoma was reported in 1971,
Based at the Lemuel Shattuck Hospital, the with evidence of a 20-year latency period
Boston Collaborative Drug Surveillance between the use of the drug and detection of
Program bypassed the physician by hiring a the cancer. The beta-blocker practolol became
drug surveillance nurse “whose primary role is the focus of a broad scandal after it was associ-
the acquisition of accurate data.” The Boston ated with a potentially fatal inflammation of
team became a model for an automated drug the skin and soft tissues (oculomucocutaneous
surveillance program that functioned “largely syndrome) some five years after its broad
independent of clinical judgment in establish- release on the British market. These examples
ing a connection between a drug and an simultaneously elucidated the scientific and
adverse event.” ethical necessity for drug surveillance units
Early results showed that drug-related and underscored the impossibility of inpatient
events were both more frequent and less severe surveillance systems to capture drug-disease
associations in which three decades or more had provided irrefutable evidence that pre-
might pass between drug exposure and adverse scription drugs were ill-used in American soci-
events. As Jick warned, in a systematic pro- ety. Kennedy called for Congress to work with
posal for the theory and design of the emerging the medical profession and the pharmaceutical
field of pharmacoepidemiology, the ability to industry to sponsor a public-private body of
study “drug-illness relations” required distinct expertise whose explicit purpose would be to
methods depending on the time course and establish a post-market surveillance system for
prevalence of prescription-related adverse prescription drugs. As the Commission would
events. High-frequency events in high- note in its final report, the purpose of system-
prevalence diseases could be detected swiftly atic prescription surveillance was “not merely
by case report, low-frequency events in high- to learn ‘something’ about a drug but to glean
prevalence diseases required careful active information that is useful in improving the
ongoing surveillance, and low-frequency rational use of drugs.”
events in low-prevalence diseases might sim- Conceived as a public-private venture, the
ply never be adequately described. Commission ran from 1976 until 1979 and
Many early pharmacoepidemiologic research- issued its final report in the first month of 1980.
ers viewed scientific quality and ethics as com- The Commission worked to integrate the social,
plementary: more rigorous data collection of epidemiological, marketing, and policy inter-
drug-related events carried ethical benefits by ests in prescriptions as a source of data. Initially,
enhancing medical practitioners’ capacity to the prospects for a harmonization of these four
“do no harm” to patients. As early pharma- perspectives seemed auspicious. At the first
coepidemiologic work also coincided with the meeting, Howard L. Binkley, Vice President for
development of bioethics as a field, critical Research and Planning of the Pharmaceutical
principles of informed consent, external review Manufacturers Association, provided a descrip-
of research protocols, and protection of patient tion and critique of presently available sources
privacy began to influence pharmacoepidemio- of data on trends in the prescribing and dis-
logic investigators’ thinking in the US and pensing of prescription drugs, with an empha-
internationally, as described below. sis on how market research data could be linked
To address the growing problems of drug to broader systems of private and public claims
safety, prescription surveillance needed to and outcomes data. Yet as the Commission
extend outwards: spatially, from the monitored assessed its findings by 1979, it became clear
wards of the hospital to the messier universe of that although several data sets existed, no indi-
outpatient care; temporally, from links visible vidual data set contained enough information
in days or weeks of measurable hospital time to deliver sufficient granularity to allow the full
to the longer stretches of months and years assessment of drug use in outpatient practice.
required to understand the impacts of chronic The Commission began to interview hybrid
medication use; and thematically, from the iso- data sources that illustrated new links between
lated connection of drug and disease to the the public and private nature of prescriber data
study of all steps of diagnosis, prescription, sets. Fledgling health maintenance organiza-
adherence, consumption, and presentation tions such as Kaiser Permanente and the
that might extend in between. In the United Group Health Cooperative of Puget Sound
States, this project would find its boldest form developed in-house proprietary databases that
in the Joint Commission on Prescription Drug linked both prescription claims and outcomes
Use, formed in response to a press conference data in the same place. Exploratory work by
held by Senator Edward Kennedy on Hershel Jick following the use of the
30 November 1976, at which he announced blockbuster anti-ulcer drug Tagamet (cimeti-
that the new science of drug utilization studies dine) in Puget Sound pharmacies suggested
that this approach could be quite promising systems by the second half of the twentieth
indeed. Another hybrid form was introduced century. Scandinavian countries in particular
by Noel Munson, a spokesman from had long histories of centrally organized phar-
Prescription Card Services (PCS), a private pre- macy records and more tightly controlled
scription data company that acted as a “fiscal national formularies of allowable drugs.
intermediary” for public payment groups like Moreover, the World Health Organization had
Medicare and Medicaid and other groups that set up a regional European Drug Utilization
paid for prescription drugs. But these individ- Group in Oslo which held a prominent confer-
ual companies (e.g. PCS) appeared to code ence on the overprescribing of prescription
their data according to their own proprietary drugs in 1969 and then proceeded to develop
software. Even within the Medicaid system, methods of comparing utilization across drug
the promise of effortless data linkage remained classes and across national pharmacy stand-
a dream in the late 1970s, complicated by wide ards. Ironically, even in countries such as
state-by-state discrepancies in patterns of cod- Sweden, much of the prescription data came
ing, storing, and retrieving prescription data. from the private sector. Still, pharmacoepide-
If the 1980 publication of the Joint miologic research in Europe continued to
Commission report represented a high point of receive substantial public support throughout
collaboration between market researchers, epi- the 1970s, 1980s, and 1990s.
demiologists, policy reformers, and sociolo- The founding of the European Medicines
gists in imagining an early “big data” universe Agency (EMA) in 1995 was a crucial step
for therapeutic surveillance, it also represented toward a pan-European supervision of medi-
a dream of collaborative work that would soon cines. The decentralized agency is critical to
dissipate. Like many other grand designs for the European Medicines Regulatory Network
federally-sponsored health programs con- (EMRN), partnering with the European
ceived in the later 1970s and proposed in the Commission (EC) and national authorities of
early 1980s, its speculative structures would European Economic Area (EEA) member
never materialize, its measures would be left states (the Heads of Medicines Agencies
unfunded, and subsequent calls for a center for [HMA] network). The EMRN’s main objective
post-marketing surveillance would be is to achieve a consistent approach to medi-
repeated, and unfunded, every few years for cines regulation across the European Union
several decades. Only in the past decade, with (EU). In collaboration with the network part-
the passage of the Food and Drug ners, the EMA oversees the scientific evalua-
Administration Amendments Act of 2007 tion, safety and efficacy monitoring of human
(FDAAA), would a substantial US public (and veterinary) medicines in the EU. For most
investment be made in the construction of a innovative medicines, including those for rare
linked public prescription database for phar- diseases, a central assessment and marketing
macoepidemiologic research with the creation authorization coordinated by the EMA is com-
of the FDA’s new automated pharmacovigi- pulsory. In cases of human medicines, the
lance program, the Sentinel Initiative, which EMA’s Committee for Medicinal Products for
officially launched in 2016. Human Use (CHMP) carries out a scientific
assessment, based on which the EC decides
whether to grant marketing authorization.
European Pharmacoepidemiologic
Once granted, such a centralized marketing
Trends and Ethics
authorization is valid across the EU.
In Europe, several nations with centralized Predominantly though, medicines in the EU
national health systems like England and are authorized by member states’ national
Sweden created prescription surveillance authorities.
Shared, key ethical requirements in In response to the benfluorex scandal, the

European pharmacoepidemiologic research legislation was amended in 2012. Servier phar-
came to include beneficence, transparency, sci- maceuticals’ Mediator (benfluorex), marketed
entific independence, and integrity. Yet, incon- as an add-on for diabetes and hyperlipidemia,
sistent application and authorization was under pharmacovigilance investigation in
procedures for clinical studies in EU and EEA France since 1998. It was found in 2003 that
member states have long been criticized. This the drug caused cardiovascular complications.
also applies to pharmacoepidemiology and In response, Servier did not re-apply for mar-
pharmacovigilance. Especially for multi- keting authorization in Spain and Italy, effec-
national, non-interventional studies (NIS), it tively withdrawing the product from the
has been lamented that “[. . .] a patchwork of market in those countries. However, benfluo-
regulations and codes of conduct have to be rex continued to be available and approved for
followed.” diabetes treatment in France and other coun-
Partly in response to some of these issues, tries until 2009. Only then was the benfluorex
since the early 2000s new EU regulations, authorization fully revoked; its efficacy was
directives and guidelines have been intro- found to be limited and it risked causing car-
duced. These aim to facilitate ethical, effective diac valvulopathy. Subsequently, EU regula-
pharmacoepidemiologic practices in and tion No. 1027/2012 and Directive 2012/26/EC
across different member states. Currently, cru- were published, amending the 2010 EU phar-
cial regulatory changes are underway that will macovigilance legislation. The amendments
affect pharmacoepidemiology and pharma- especially addressed the issue that safety meas-
covigilance in the EU. ures for medicinal products need to be imple-
The EU pharmacovigilance legislation aims mented consistently and in a timely fashion in
to minimize risks and harms posed by adverse all member states where respective products
drug reactions (ADRs). Its implementation is were authorized.
overseen by the EMA, EU member state The benfluorex scandal points to broader
authorities, and the EC. Key legal documents challenges regarding pharmacovigilance and
for the pharmacovigilance legislation and pharmacoepidemiologic research in the EU:
pharmacoepidemiologic studies are EU regu- regulations and guidelines need to be applied
lation No. 1235/2010 and directive 2010/84/ across multiple states and to different actors,
EC. In effect since 2012, the regulation out- including national marketing authorization
lines measures for safeguarding patients’ holders and applicants. While the legislation
safety and rights and asserts the crucial role outlines fairly broad objectives, responsibili-
of healthcare professionals in reporting ties, and issues, these are specified in concrete
ADRs. It moreover acknowledges the neces- deliverables. One of these deliverables was the
sity to develop EU/EEA wide “[. . .] harmo- founding of the EMA Pharmacovigilance Risk
nized guiding principles for, and regulatory Assessment Committee (PRAC) which moni-
supervision of, post-authorization safety stud- tors and assesses drug safety in the EU.
ies that are requested by competent authori- Moreover, it initiated the development of the
ties and that are non-interventional, that are EMA’s Good Pharmacovigilance Practices
initiated, managed or financed by the market- (GVP) guideline, (described below).
ing authorization holder.” Among other deliv- The European Network of Centres for
erables, the regulation established the Pharmacoepidemiology and Pharmacovigilance
EudraVigilance database as a main platform (ENCePP) was established in 2006 and is coor-
for the obligatory reporting of ADRs by mar- dinated by the EMA. It is an expertise and
keting authorization holders and respective resource network focused on pharmacoepide-
national authorities. miology and pharmacovigilance in Europe. It
consists of partners that are public and not-for- in non-profit organizations who must apply
profit organizations, including research and and undergo ethical review.
pharmacovigilance centers, university hospi- The Korea Food and Drug Administration
tals, healthcare database hosts, and electronic (KFDA) launched an ADR reporting system in
registry sponsors. For-profit organizations, e.g. 1988, although the reporting rate was initially
contract research institutions, may only partici- very low. In 2004, the KFDA mandated that
pate if they conduct pharmacoepidemiologic pharmacists and pharmaceutical companies
and/or pharmacovigilance studies commis- report ADRs. The KFDA also established
sioned by third parties. While pharmaceutical regional pharmacovigilance centers in univer-
companies are not eligible for becoming sity hospitals that now provide nearly com-
ENCePP partners, the network provides rele- plete coverage of the country. The KFDA
vant resources and allows for these companies funded a pharmacovigilance research network
to be involved in public document reviews. (PVNet) among these centers, and researchers
The ENCePP offers crucial guideline docu- in the network use their data for studying
ments for pharmacoepidemiology and phar- adverse events. The Korean national health
macovigilance: a Code of Conduct; the insurance (NHI) database also contains all
ENCePP Checklist for Study Protocols; and information on insurance claims made and
the ENCePP Guide on Methodological prescriptions for approximately 50 million
Standards in Pharmacoepidemiology. The Koreans, and this has been used for
Code lays down rules and principles aimed at pharmacovigilance.
ensuring transparency and scientific inde- In Japan, drug manufacturers are required to
pendence. While adherence to the Code is report ADRs to the Pharmaceuticals and
voluntary, it is required to receive the ENCePP Medical Devices Agency (PMDA). A partial
Seal. Conditions for receiving the Seal are, ADR dataset is available to researchers through
among others, that a study is entered in the the PMDA website. Healthcare professionals
EU post-authorization study (PAS) Register report adverse drug events to the Ministry of
and that it is of scientific and public health Health, Labor and Welfare. Japan made its
relevance, rather than mainly pursuing national insurance claims database available
results which may promote certain medicinal for drug safety researchers in 2011. The data-
products. The Checklist is meant to ensure base covers the entire population of 128 mil-
that studies adhere to epidemiological princi- lion and includes basic patient characteristics,
ples, while also considering methodological drug prescription and dispensing, medical pro-
transparency and the need for public cedures, hospital admission, and annual health
outreach. check data (for some patients). To protect
patient privacy, Japan’s national database is
usually not available for purchase and may
East Asian Pharmacoepidemiologic
only be shared in some cooperative research
Trends and Ethics
projects. The Japanese government has also
East Asian investigators have made major con- created the Medical Information for Risk
tributions to the field of pharmacoepidemiol- Assessment Initiative (MIHARI) to access data
ogy. Researchers in South Korea, Japan, and from different sources and create a central
Taiwan have linked into comprehensive data database with a common data format.
systems on insurance claims created through Taiwan requires mandatory reporting of
universal insurance coverage of these entire serious adverse reactions by medical institu-
national populations. To help protect patient tions, pharmacies, and drug and device com-
privacy, these databases have been made avail- panies, as well as obligatory safety reports for
able for drug safety research only to researchers newly marketed drugs over a five-year
Methodologic Problems to be Solved by Pharmacoepidemiologic Researc 283
s urveillance period. In Taiwan, the National rights of the liberal subject that characterize
Adverse Drug Reactions Reporting System has much of Western bioethics. The family is often
been the primary source for post-marketing depicted as responsible for taking care of mem-
surveillance of adverse drug events. Taiwan’s bers who become sick, and medical decision-
single-payer NHI program was created in 1995 making has often been family-based. Some
and covers more than 99% of the population. have also noted a plurality of ethical perspec-
The NHI Research Database is thus a highly tives within East Asia, contending that a sim-
comprehensive data set including basic patient ple Eastern and Western bioethical dichotomy
data, care record and expenditure claims, and of communitarian versus individualistic val-
pharmaceutical reimbursements. There are ues would be overly simplistic. Others have
also subject datasets available to researchers viewed bioethics as a Western entity, promot-
on topics such as traditional Chinese medi- ing the development of Asian bioethics based
cine, cancer, diabetes, dental, catastrophic ill- more on the traditions, philosophies, religions,
ness, or psychiatric care. Patients and medical and perspectives of the region’s cultures.
facilities are de-identified for pharmacoepide- Future policies should consider these issues as
miologic research use of the NHI Research core principles for pharmacoepidemiologic
Database. To protect patient privacy, research- research ethics are discussed.
ers using Taiwan’s NHI Research Database also
receive data for 10% or less of the population.
Ethical policies for data privacy stipulate that Methodologic Problems
no individual-level data can be shared with to be Solved by
researchers from other countries. Pharmacoepidemiologic
China and other East Asian countries also Research
have been creating national healthcare claims
databases. In China, the Shanghai Center for More work remains to establish international
Adverse Drug Reaction Monitoring has oper- ethical policy harmonization while also
ated a drug surveillance and evaluation system promoting practices that support cultural vari-
since 2001 that works with patient information ation in ethical values. Yet, as pharmacoepide-
from 10 Shanghai hospitals. The Asian miological practices developed in different
Pharmacoepidemiology Network (AsPEN) national contexts that have been incorporated
was recently established as a multi-national into increasingly globalized flows of pharma-
research network for pharmacoepidemiologi- ceuticals and pharmaceutical-related infor-
cal research that promotes international com- mation, a number of ethical principles and
munication among academia, government, practices have been adopted widely across
industry, and consumers. The network func- international settings in an effort to establish
tions to promptly identify drug safety issues. consistent pharmacoepidemiologic methodol-
Pharmacoepidemiology ethics in East Asia ogy. The expansion of the field of pharma-
are similar in many ways to those of Western coepidemiology has coincided with the
countries, including features such as institu- establishment and institutionalization of the
tional ethical review and guiding principles discipline of bioethics. Numerous critical ethi-
such as beneficence, justice, autonomy, and cal concepts took hold early in pharmacoepi-
data privacy. However, experts on East Asian demiology and have remained significant over
bioethics have also recognized some distinc- time. For example, privacy of medical data is a
tions. For example, scholars have contended historically consistent value, guiding the eth-
that much East Asian bioethical thinking ics of global pharmacoepidemiologic research.
reflects value systems that emphasize the fam- Pharmacoepidemiologic research protocols
ily and public interest ahead of the individual and/or database designs have also often been
subjected to evaluation by institutional review conduct between countries and increases the
boards as external review has become increas- cost of assembling multinational data. This
ingly widespread for biomedical research poses challenges for conducting large interna-
since the second half of the twentieth century, tional studies capable of detecting rare events.
although there is variation in the nature of Additionally, requirements of explicit individ-
this review (for example, some pharmacoepi- ual informed consent are problematic in that
demiologic research has been reviewed by they can corrupt data by preventing a post-
institutional or national ethics boards, as well marketing pharmacoepidemiologic study from
as by privacy boards). Some countries also do detecting fatal or serious events since people
not require ethical review for de-identified who have died are unable to provide informed
data sets. consent. Thus, it is unsurprising that ethicists
weighing risks and benefits have tended to
contend that individual consent is not essen-
Informed Consent
tial for the use of patient records in pharma-
Informed consent became increasingly valued coepidemiologic research.
as a critical standard of international research However, over time it has become normative
ethics following its establishment as a corner- that pharmacoepidemiologists must also meet
stone of the 1964 Declaration of Helsinki, a certain requirements when conducting
groundbreaking statement of international research in which participant consent is
human experimentation ethics. However, the waived. These requirements often include that
role of informed consent has been perceived the use of protected health information
differently in interventional versus non- involves no more than minimal risk to patients,
interventional research studies. Many ethicists the research could not be effectively conducted
of international human subject research have without access to the protected health infor-
argued that since pharmacoepidemiologic mation and/or the waiver of individual con-
research involves relatively low risks to partici- sent, the privacy risks to individuals are
pants, patient consent is necessary only for reasonable in relation to any value to the indi-
studies that involve contact with patients/ viduals of the knowledge expected to result
research subjects, such as for direct interven- from the study, there is a sound plan to protect
tion or prospective gathering of information. patients from the improper use or disclosure of
There has been a broad acceptance among their information, there is a plan to destroy
ethicists allowing researcher access to identifi- identifiers at the earliest opportunity consist-
able medical records for pharmacoepidemio- ent with the research, and the data will not be
logic research without explicit individual shared with external parties to the research.
subject authorization. Research has also found Recent attention has been given to the waiver
that public opinion has echoed the views of of patient informed consent to use data on sub-
professional ethicists that pharmacoepidemi- stances of abuse or drugs that carry social
ologists should be permitted to use identifiable stigma. Patient privacy is essential in these
patient records, without patient consent, to areas of research; however, requiring informed
study drug safety as long as existing ethical consent for each patient or allowing retraction
guidelines and relevant laws are followed. of sensitive drug information from patient
A number of nations, however, require records leads to partial data sets that impede
explicit informed consent from each study par- the ability of researchers to study the impact of
ticipant, and there are also international varia- these substances on patient health outcomes.
tions in requirements for electronic consent The negative consequences of failing to collect
versus hard copy written consent. Ethical regu- sound pharmacoepidemiologic data on the
latory disharmony causes differences in study health effects of these substances are likely
Methodologic Problems to be Solved by Pharmacoepidemiologic Researc 285
worse than the relatively minimal risk associ- some common stances toward ethical review
ated with waiver of patient consent. However, of drug surveillance. Certain pharmacoepide-
in such circumstances, the highest precautions miologic research tends to qualify as exempt
should be taken to protect patient privacy, such from ethics board review or qualifies for expe-
as de-identifying data through secure codes or dited review by an ethics board chair or a des-
potentially having extra ethics training require- ignated member. For studies in which it is not
ments for all researchers using data on stigma- possible for investigators to identify the indi-
tized or abused substances. vidual patients, ethics board review is often not
required. For example, the US 45 Code of
Federal Regulations 46.101 exempts from insti-
Ethics of Surveillance
tutional review “research involving the collec-
Surveillance has long evoked public concern tion or study of existing data, documents,
regarding privacy, confidentiality, and auton- records, pathological specimens, or diagnostic
omy. This is relevant to post-marketing surveil- specimens, if these sources are publicly availa-
lance, since health information is seen as highly ble or if the information is recorded by the
sensitive and personal. Thus, pharmacoepide- investigator in such a manner that subjects
miologic researchers need to balance possible cannot be identified, directly or through iden-
risks to a larger population against the harms tifiers linked to the subjects.” In many coun-
concerning individuals, such as a possible tries, research is also often eligible for expedited
infringement of privacy. While privacy is highly review if it poses no more than minimal risk to
important to the ethics of pharmacoepidemio- patients and involves a retrospective analysis
logic research, privacy is not an absolute value, of existing records. Still, ethics review policies
nor does it seem to have been perceived as such vary internationally and by institutional prac-
in public health surveillance history. Rather, tice, depending inter alia on respective
privacy is one of multiple values that are bal- national/state regulations, posing challenges
anced in public health surveillance. It has been for global collaborative studies. This may lead
argued that ensuring privacy is part of the to inconsistent risk–benefit assessments and
broader value of protecting autonomy. Yet variations in balancing subjects’ protection
other key principles to be balanced in pharma- (e.g. regarding safety and privacy) against pub-
coepidemiologic research include beneficence lic health interests.
to promote research that adds to the existing
knowledge base of medicine to improve patient
Ethical Benefits
health and prevent mortality; non-maleficence,
of Pharmacoepidemiologic
or the prevention of patient harm; and justice,
Research for Data Integrity
which manifests as the fair distribution of
research burdens and benefits among people. From a broader ethical perspective, it is
Risks of surveillance can be minimized increasingly clear that the expansion of phar-
through confidentiality and data anonymiza- macoepidemiological research can provide
tion. Such strategies are ethically imperative, added benefits to drug research by detecting
since they safeguard individuals’ rights, pri- groups at risk for adverse events. Thus, the
vacy, autonomy, and dignity. Applying the field can play an important role in reducing
highest ethical standards and communicating drug safety data inequalities. For example,
with the public about potential criticism is also expanding drug outcomes data for groups such
important for a positive public perception of as minorities or small/rare genetic subpopula-
pharmacoepidemiology. tions who may have treatment outcome varia-
While there have been some disagreements, tions that can only be identified and/or
international ethics policies have developed adequately quantified and measured through
large post-marketing pharmacoepidemiologic advancement depend on obtaining external

studies may provide substantial benefits for funding also can exacerbate the ethical prob-
members of these populations. There is also lems resulting from direct relationships
limited data on the efficacy and safety of between drug companies and the pharmacoep-
drugs in children due to the fact that histori- idemiologists evaluating their products.
cally, children have often not been included Investigators in such environments are under
in randomized controlled trials (RCTs). professional pressure to secure funding, and in
Pharmacoepidemiologic research helps to fill a climate of heightened competition for public
these research gaps. However, it would be ethi- funding sources, an academician who estab-
cally problematic for pharmacoepidemiology lishes a positive working relationship with a
to be relied on solely to provide missing data pharmaceutical research sponsor may increase
on children, minorities, or other subgroups in his/her chances of obtaining future funding
lieu of RCTs, particularly in cases when RCTs from that sponsor. This can create an incen-
could produce more robust data. tive, whether subconscious or acknowledged,
Further, pharmacoepidemiologic studies are for researchers to conduct studies that spon-
usually conducted after drug approval, and soring drug companies will find favorable.
there is high variability in the frequency and Indeed, studies have shown a trend toward
design of post-marketing pharmacoepidemio- more favorable efficacy results and conclu-
logic research. Such studies are not necessarily sions for industry-sponsored drug research
required, and so are not a consistently reliable than research sponsored by other sources,
source of information on drug outcomes finding a bias in industry-funded research that
among diverse demographic groups. Clinical cannot be otherwise explained by standard
trials are usually required for drug approval assessments of risk of bias. There are a number
and are thus a mechanism for ensuring broader of feasible solutions to address the ethical con-
implementation of policies requiring the inclu- flicts of interest in industry-funded research,
sion of diverse research subjects. Ultimately, as described below.
consistent with recurring concerns over ethical
practices in pharmacoepidemiologic research
in general, ethicists have noted that pharma- urrently Available
C
coepidemiology related to subpopulations
Solutions
would benefit from a more explicit legal ethical
framework, particularly to clarify ethical
Good Pharmacoepidemiology
requirements for data sharing.
and Pharmacovigilance Practices
The International Society for Pharma
Problems of Conflicts of Interest
coepidemiology (ISPE) has created Guidelines
for Drug Industry Research
for Good Pharmacoepidemiology Practice
Academia-industry collaborations have (GPP), which provide a model for key pharma-
become a critical area of concern for the ethics coepidemiologic research ethics policies. The
of pharmacoepidemiologic research, particu- guidelines recommend that researchers
larly in recent decades as pharmaceutical prof- include a description of quality control proce-
its have soared and the stakes have been raised dures; plans for protecting human subjects;
for the outcomes of research on drug safety confidentiality provisions; ethical conditions
and efficacy. There is an inherent conflict of under which a study would terminate; the use
interest in research that is funded by drug com- of Data Safety Monitoring Boards where
panies to assess their own products. Academic appropriate; institutional review board and
settings in which researcher success and informed consent considerations in accordance
with local laws; research study registration; by the ENCePP. As the ethics review procedure
and plans for disseminating study results. and requirements for respective committees
However, ISPE GPP policies are nonbinding depend on national legislation, information on
and therefore do not resolve concerns regard- individual application procedures is not
ing national variations in ethical oversight and included in the GVP. While there is no EU reg-
requirements by regulatory agencies for post- ulation or directive for NIS, interventional
marketing pharmacoepidemiologic work. studies are covered in the Clinical Trials
EU policies provide a useful example of Regulation (CTR).
transnational efforts at regulatory standardiza- In the EU, methodological, ethical, and legal
tion of GVPs. EU documents concerning bio- requirements for pharmacoepidemiologic
medical research in general and research hinge significantly on whether a
pharmacoepidemiologic research in particular study is categorized as a “clinical trial” or as
commonly speak of two types of clinical stud- “non-interventional/non-experimental.” Both
ies, broadly speaking: interventional, i.e. categories are defined as “clinical studies”
experimental, and non-interventional, some- aimed at discovering or confirming the
times called observational research. On the (adverse) effects of medicinal products. For
one hand, pharmacoepidemiologic research pharmacoepidemiologic studies involving
relies on non-interventional study designs clinical trials, the introduction of the EU CTR
such as case–control or cohort studies. On the No. 536/2014 will soon be decisive. The CTR
other hand, interventional, randomized clini- was adopted on 16 April 2014 and entered into
cal trials (RCTs) are an important element of force on 16 June 2014. For the regulation to
post-marketing pharmacoepidemiology stud- become applicable, an EU-wide clinical trials
ies (see Chapter 17). portal and database is required. Both need to
The EMA defines GVPs as “a set of measures undergo an independent audit. According to
drawn up to facilitate the performance of the the EMA, the Regulation was supposed to
safety monitoring of medicines in the European come into application in late 2019, starting a
Union.” It includes chapters on pharmacovigi- transition period of three years. However, due
lance processes as well as product- and to technical difficulties concerning the plat-
population-specific considerations. For EU form and database, this has been postponed
pharmacoepidemiologic post-authorization and EMA’s Management Board agreed to pro-
safety studies (PASS), module VIII is particu- ceed with the audit in December 2020 (see
larly relevant. PASS may be interventional or European Commission n.d.). The CTR is
non-interventional. Although the module meant to harmonize research practices and to
touches upon interventional studies too, ensure the highest methodological and ethical
emphasis is put on non-interventional PASS. standards across all EU as well as EEA EFTA
In accordance with the EU pharmacovigi- (European Economic Area, European Free
lance legislation, the GVP stipulates that the Trade Association) member states. To what
EMA needs to ensure that protocols and extent it will deliver on these promises is under
abstracts of PASS results are published. While discussion. The regulation replaces the
the primary/lead investigator is responsible for “Clinical Trials Directive” 2001/20/EC which
the information provided, the registration may is said to have “[. . .] failed to achieve its goal of
be made by, for example, research center staff simplifying the scientific and ethical review of
or representatives of pharmaceutical compa- clinical trials in the EU.”
nies funding a study. Where possible, this Moreover, the ENCePP had problematized
should be done before the study commences. the NIS definition given in the 2001 directive.
Practically, registration and publication are The ENCePP raised the issue that the defini-
processed through the EU PAS register, hosted tion was not sufficiently specific and created
uncertainty as to what counts as NIS or RCT. also has an interest in maintaining public trust
Pharmacoepidemiologic prospective case- in product integrity, as well as an interest in
control studies − like the International Primary compliance with regulatory ethical and meth-
Pulmonary Hypertension Study investigation odological requirements to obtain drug
of primary pulmonary hypertension (PPH) approval. Thus, there is some incentive for
occurrence in association with anorectic industry to address concerns about conflicts of
agents − would classify as a clinical trial interest. The Board of Directors of the ISPE has
according to the 2001 directive. Its ambiguous published a set of principles for academia-
NIS definition was thus criticized for impeding industry collaboration that can be helpful in
the conduct of pharmacoepidemiologic managing industry conflicts of interest. It
studies. includes the importance of transparent
The ENCePP Guide on Methodological research agreements, open and complete dis-
Standards in Pharmacoepidemiology (Revision closure of conflicts of interest, registration of
6, July 2017) lays down rules and principles for research protocols in public sites such as the
transparency and scientific independence. ENCePP registry or http://ClinicalTrials.gov,
Chapter 9 of the Guide deals with ethical compliance with local laws, clarity on confi-
aspects of pharmacoepidemiology, focusing on dentiality of proprietary information while
patient and data protection and scientific also ensuring reporting of all relevant and
integrity and ethical conduct. It identifies key important information to regulators, the
values based on documents such as the potential value of having a steering committee
ADVANCE Code of Conduct for Collaborative and/or an independent advisory committee to
Vaccine Studies, the GPPs of the ISPE, and the the research, and an obligation to disseminate
Good Epidemiology Practice (GEP) guidelines and publish research findings of potential sci-
of the International Epidemiological entific or public health importance irrespec-
Association. The Guide highlights that “[p]rin- tive of results.
ciples of scientific integrity and ethical con- While all of these principles are helpful in
duct are paramount in any medical research” managing financial conflicts of interest, they
and points out that the abovementioned do not eliminate the inherent problem of drug
ENCePP code of conduct “[. . .] offers stand- companies having a stake in the outcomes of
ards for scientific independence and transpar- research that they sponsor or the ethical con-
ency of research in pharmacoepidemiology cerns associated with the power dynamics of
and pharmacovigilance.” In addition, the industry directly funding investigators as
Guide highlights core values, such as best sci- described above. To eliminate these underlying
ence, strengthening public health, and improv- ethical problems, the direct relationships in
ing transparency, as stressed by the ADVANCE which companies fund individual investigators
Code of Conduct. It moreover emphasizes the to assess specific products would need to be
need for ensuring scientific autonomy, benefi- severed. Alternative models that eliminate
cence, non-maleficence and justice, according these ethical conflicts can be easily envisaged.
to the four general ethical principles defined in For example, the British Drug Safety Research
the GEP guidelines. Unit (DSRU), an independent charity sup-
ported by the National Health Service, con-
ducts publicly funded pharmacoepidemiologic
Protections against Conflicts
research. Still, the organization conducts a
of Interest for Drug Industry-
large amount of research funded by uncondi-
Sponsored Research
tional donations from pharmaceutical compa-
While industry-sponsored research creates real nies. Yet, the companies have no control on the
challenges for conflicts of interest, industry conduct or the publication of the studies
The Futur 289
c onducted by the DSRU. This helps to mitigate ethical pharmacoepidemiologic research that
the pressure of inherent conflicts of interest in effectively identifies and reports ADRs, thus
industry-funded research. Given that industry allowing for timely responses. New policies
funding may lead to biased study results, a must also be more thoroughly transnational
comprehensive solution could build from the and attentive to global variations in ethical
DSRU model, for example by requiring spon- beliefs. A main challenge is and will be to trans-
sors of new drugs to contribute an uncondi- late inevitably general documents into practical
tional fee to drug regulators that would fund instructions and relevant local practices.
pharmacoepidemiologic research. By making In the future, national regulatory authori-
such contributions mandatory rather than vol- ties, universities, and research centers will
untary, investigators could conduct studies continue working to align requirements
without concern as to whether results may toward coherent pharmacoepidemiologic
influence future industry donation decisions. research ethics. It is to be expected that further
In the US, for example, the expansion of the regulatory efforts will be invested in streamlin-
FDA’s Prescription Drug User Fee could easily ing requirements for ethics review boards and
establish a fund for pharmacoepidemiologic ethical guidelines for NIS, especially across the
research. EU. Although recent regulations and directives
in the EU hope to address several, pressing
issues, many of these are complicated anew by
The Future the United Kingdom’s announced withdrawal
from the EU. This has already triggered practi-
The ethical conduct of pharmacoepidemio- cal changes, such as the relocation of the EMA
logic studies is of crucial importance for sub- from London to Amsterdam in March 2019.
jects’ safety, health and wellbeing. Moreover, it Moreover, legal uncertainties are underway, as
is decisive for the public perception of pharma- it has been disclosed by the UK Department for
coepidemiology. Research in this field is rooted Exiting the European Union that the post-Brexit
in the moral obligation to preempt or at least guidelines for clinical studies in the UK may
minimize medicine-related harms and health deviate from EU legislation.
hazards. Implementing highest ethical stand- Transparency has been stressed as a key ele-
ards helps to avoid potential damage to the ment for ensuring ethical pharmacoepidemio-
public image of the field and public trust in logic practices. Moreover, data sharing is
claims of pharmacoepidemiological research pivotal for effective pharmacoepidemiology
as a disinterested form of expert knowledge. and pharmacovigilance. At the same time,
Such damage may be related to research prac- researchers are required to safeguard the sub-
tices compromised by economic interests or jects’ privacy and dignity. Developments such
misconduct of the pharmaceutical industry. as the open data movement on the one hand
Thus, scientific integrity, independence, and and regulations aimed at protecting individu-
transparency will continue to be crucial for the als’ privacy on the other hand put researchers
ethics of pharmacoepidemiologic research. in a difficult position. At an increasing rate,
Even in the recent past, regulatory amend- there is a tendency to require public accessibil-
ments relevant to pharmacoepidemiology and ity of scientific results and even data.
pharmacovigilance were often triggered by Simultaneously, privacy concerns and poten-
scandals, although a dream to make pharma- tial regulations may pose challenges for data
coepidemiology a proactive, rather than a reac- (re-)use in pharmacoepidemiologic studies.
tive, field can be traced back to the 1960s if not Heightened attention has already been paid
earlier. Adjusted, new, and emerging to the environmental, polluting effects of phar-
regulations and guidelines aim at promoting maceutical residues. Regulatory documents,
such as the EU pharmacovigilance legislation, and leaks have led to public skepticism con-
acknowledge that “[t]he pollution of waters cerning the use of big data. Not only such neg-
and soils with pharmaceutical residues is an ative connotations, but also scientific concerns
emerging environmental problem.” Research regarding users’ consent, autonomy, and pri-
examining the adverse effects of pharmaceuti- vacy raise ethical questions about big data
cals on the environment has been labeled as research. Pharmacoepidemiologic research
pharmacoenvironmentology. With its focus on involving big data requires careful ethical con-
the environmental impact of drugs given at siderations for the individuals’ generating such
therapeutic doses, it is considered part of phar- data, for example users of social networking
macovigilance. Assuming that environmental sites. Moreover, pharmacoepidemiologists
issues will continue to be high on the political need to consider the biases inherent to digital
and scientific agenda, pharmacoepidemiologic data sources: such bias can be caused by big
expertise will be increasingly needed to assess data retrieved from populations that do not
medicines as pollutants. In this context, phar- allow for generalizations. For instance, since
macoepidemiologists will need to employ and individuals included in a digital data sample
expand their methodological repertoire for may represent only those using an expensive/
studies investigating medicines’ adverse effects innovative technical device or service, these
on the environment. This development might users could be on average younger or above
also imply an amplified need for novel, inter- average in access to health-promoting
disciplinary research collaboration involving resources. In addition, the quality of such data
pharmacoepidemiologists. Such collaboration may differ from other sources of data (e.g.
is also characteristic for another emerging medical records).
intersection, between pharmacoepidemiology, Research involving these alternative sources
computer, and data science. of data is subject to different laws and regula-
Research at the intersection of digital ser- tory frameworks when conducted in different
vices, big data, and public health is a poten- global settings. For the United States, access to
tially promising, but precarious field. It has health relevant information via social net-
been demonstrated that emerging, digital data working sites such as Facebook is at present
sources like social networking sites can func- legally possible, due to the lack of protection
tion as complementary resources for pharma- for health-relevant data retrieved outside of
coepidemiology. The use of such data sources, the traditional health care and research sys-
often referred to as a type of “big data,” is atypi- tem. With regards to medical privacy, the
cal for pharmacoepidemiologic studies, but Electronic Frontier Foundation (EFF) points
may become more common in the future. out that social networking sites and other
Research drawing on “big data” may take place online services compromise US citizens’ con-
outside of medical departments or hospitals, trol over their health data: “The baseline law
e.g. conducted by data scientists. Big data and for health information is the Health Insurance
emerging data science approaches have cre- Portability and Accountability Act (HIPAA).
ated new possibilities for pharmacoepidemio- HIPAA offers some rights to patients, but it is
logic research. For example, Freifeld et al. used severely limited because it only applies to an
data from the social networking site and entity if it is what the law considers to be either
microblogging service Twitter to monitor a ‘covered entity’ − namely: a health care pro-
ADRs. vider, health plan, or health care clearinghouse
The term “big data” has become associated − or a relevant business associate (BA).” This
with various leaks and scandals. The United also implies that content such as Facebook or
Kingdom Science and Technology Committee Twitter data, despite their actual use as health
concluded in a 2015 report that data misuses indicators, are currently not protected under
Key Point 291
HIPAA. Yet, although arguably unlikely, this that occurred as a result of the release of
may change in the future. In addition, scien- poorly understood and inadequately tested
tists should not conflate legal with ethical drugs onto the marketplace. More rigorous
requirements. pharmacoepidemiologic data on drug out-
With regard to biomedical research, it has comes enhanced medical practitioners’
been pointed out that the ethical implications capacity to “do no harm” to patients.
of big data research are, at least partly, ●● Pharmacoepidemiology has developed in
uncharted territory. Additional ethical consid- varied international social contexts in recent
erations for pharmacoepidemiologic research history. Global researchers share common
involving big data are thus needed. This applies experiences and ethical challenges such as
to the autonomy of data subjects, but also to concerns over informed consent, external
new corporate stakeholders and public-private review of research protocols, protection of
partnerships. The latter may not merely involve patient privacy, and questions of conflicts of
pharmaceutical companies or device manufac- interest. However, pharmacoepidemiologic
turers. Internet and technology corporations work within different national or private
may also play a role and require ethical as well data systems has created specific areas of
as legal oversight, since they control access to ethical demands and emphases in different
digital data that could further complement settings. This history is reflected in the var-
pharmacoepidemiology in the future. ied global landscape of ethical discussions
and policies.
●● There is more work to be done to consider
Acknowledgement how pharmacoepidemiology ethics can be
inclusive of diverse cultural approaches to
The authors would like to thank Philip Phan bioethics while also streamlining global ethi-
for assistance with literature searches. cal regulatory standards where appropriate.
●● In the future, as the field of pharmacoepide-
miology continues to rely on changing tech-
Key Points nological and data platforms and evolves to
address new areas of pharmacovigilance
●● From an ethical perspective, pharmacoepi- such as pharmacoenvironmentology, ethical
demiologic research has helped to address reflections and policies should continue to
the ethical problem of drug safety scandals evolve and develop, as well.
Case Example 16.1
Evaluation of Facebook and Twitter monitor- t erritory. Also, their methodological effec-
ing to detect safety signals for medical prod- tiveness, that is whether they can indeed
ucts: an analysis of recent FDA safety alerts complement pharmacoepidemiologic
(Pierce et al. 2017) approaches, and their reliability are still
under investigation.
Background
The use of “big data” retrieved from social Question
networking sites is rather uncommon for The authors aimed to examine whether,
pharmacoepidemiologic studies, at least for post-approval, adverse effects were reported
now. Due to the methodological novelty of sooner on social media than via the US Food
approaches involving such data, ethical and Drug Administration’s Adverse Event
issues are to some extent still uncharted Reporting System (FAERS).
(Continued)
Approach privacy (Sharon 2016). Studies such as the

In a retrospective approach, the authors one conducted by Pierce et al. need to be
retrieved data from the social networking carefully examined with regard to power
site Facebook and the microblogging service asymmetries in access to data and control
Twitter. These were scanned for signals cor- over technology platforms. For example,
relating with 10 post-marketing safety sig- while some of the authors were employed by
nals reported in FAERS and flagged by the the US FDA, others indicated that they were
FDA. For example, the authors examined employees of a technology company intend-
whether tweets indicated symptoms of vas- ing to commercialize the software platform
culitis caused by Dronedarone or angi- used in the research – a conflict of interest.
oedema caused by Pradaxa. Potentially Such public–private partnerships also lead
relevant posts were then compared with to dependencies between partners and put
those reported in FAERS, assessing where involved companies in a powerful position:
symptoms were reported first. this concerns the corporate ownership of
public health data on the one hand and
Results
researchers’ reliance on commercial tech-
The study did not yield a clear outcome.
nologies on the other. Ethical issues aside,
While some relevant social media data indi-
one should also not forget that the effective-
cated adverse effects for Multaq
ness of such novel approaches and their
(Dronaderone) before these were reported
place in the methodological repertoire of
by the FDA, this was not the case for others.
pharmacoepidemiology are still under
Strengths scrutiny.
Provided that approaches like these were to
Key Points
become effective and reliable, harvesting
●● Data, in this case user-generated content,
and analyzing data from social media could
retrieved from social networking sites
complement existing pharmacoepidemio-
such as Facebook and Twitter are being
logic methods. Data from social media might
explored as complementary sources for
particularly serve as early warning signals.
pharmacovigilance.
Limitations ●● The methodological effectiveness and
In a letter to the editor, Wiwanitkit (2017) reliability of pharmacoepidemiologic stud-
raised concerns over this study, questioning ies involving data from online sources
the validity and reliability of social media such as social media have not been
data. The commentator criticized that “[. . .] confirmed.
the basic issue with these social media plat- ●● Ethical issues concerning such studies go
forms is of confidentiality and privacy.” This beyond matters of privacy and confidenti-
raises the issue of whether it is ethical to use ality: they raise issues of power asym-
social media data in the first place. In this metries, dependencies, and conflicts of
context, one should also consider the com- interest.
plexity and challenge of fully anonymizing ●● Ethics boards and comparable oversight
big data. Moreover, others have pointed out committees must gain further expertise to
that when entering public-private collabora- review pharmacoepidemiologic research
tions, researchers need to go beyond ques- projects and approaches involving user-
tions concerning the quality of research or generated big data.
Further Reading 293
Further Reading
Abou-El-Enein, M. and Schneider, C. (2016). Pratt, B., Van, C., Cong, Y. et al. (2014).
Deciphering the EU clinical trials regulation. Perspectives from South and East Asia on
Nature Biotechnology 34: 231–233. clinical and research ethics. Journal of
Carpenter, D.P. (2010). Reputation and Power. Empirical Research on Human Research Ethics
Princeton, NJ: Princeton University Press. 9 (2): 52–67.
European Commission (n.d.). ‘Clinical trials - Ramirez, I. (2015). Navigating the maze of
Regulation EU No. 536/2014’. Retrieved from: requirements for obtaining approval of
https://ec.europa.eu/health/human-use/ non-interventional studies (NIS) in the
clinical-trials/regulation_en (accessed on 26 European Union. GMS German Medical
October 2020). Science 13 https://doi.org/10.3205/000225.
Freifeld, C.C., Brownstein, J.S., Menone, C.M. Sethi, N. (2014). The promotion of data sharing
et al. (2014). Digital drug safety surveillance: in pharmacoepidemiology. European Journal
monitoring pharmaceutical products in of Health Law 21 (3): 271–296.
twitter. Drug Safety 37 (5): 343–350. Sharon, T. (2016). The Googlization of health
Hedgecoe, A. (2016). Scandals, ethics, and research: from disruptive innovation to
regulatory change in biomedical research. disruptive ethics. Personalized Medicine 13 (6):
Science, Technology, & Human Values 42 (4): 563–574.
577–599. Urushihara, H., Parmenter, L., Tashiro, S. et al.
International Society of Pharmacoepidemiology (2017). Bridge the gap: the need for
Public Policy Committee (2016). Guidelines harmonized regulatory and ethical standards
for Good Pharmacoepidemiology Practice for postmarketing observational studies.
(GPP). Pharmacoepidemiology and Drug Safety Pharmacoepidemiology and Drug Safety 26:
25: 188–191. 1299–1306.
Kimura, T., Matsushita, Y., Kao Yang, Y.-H. et al. Vayena, E., Salathé, M., Madoff, L.C., and
(2011). Pharmacovigilance systems and Brownstein, J.S. (2015). Ethical challenges of
databases in Korea, Japan, and Taiwan. Pharma big data in public health. PLoS Computational
coepidemiology and Drug Safety 20: 1237–1245. Biology 11 (2): e1003904.
Merz, J.F. (2001). Introduction: a survey of Wiwanitkit, V. (2017). Comment on: “Evaluation
international ethics practices in pharma of Facebook and Twitter monitoring to detect
coepidemiology and drug safety. Pharma safety signals for medical products: an
coepidemiology and Drug Safety 10 (7): 579–581. analysis of recent FDA safety alerts”. Drug
Pierce, C.E., Bouri, K., Pamer, C. et al. (2017). Safety 40 (8): 755–755.
Evaluation of Facebook and Twitter Wettermark, B. (2013). The intriguing future
monitoring to detect safety signals for medical of pharmacoepidemiology. European
products: an analysis of recent FDA safety Journal of Clinical Pharmacology 69 (1):
alerts. Drug Safety 40 (4): 317–331. 43–51.
294
17
The Use of Randomized Controlled Trials

for Pharmacoepidemiology
Samuel M. Lesko1, Allen A. Mitchell2, and Robert F. Reynolds3
1
Northeast Regional Cancer Institute, and Geisinger Commonwealth School of Medicine, Scranton, PA, USA
2
Slone Epidemiology Center at Boston University, Boston, MA, USA
3
Epidemiology, Research and Development, GlaxoSmithKline, New York, NY, USA
Introduction world clinical practice led to selecting a large

simple trial design to assess the risk of serious
When properly conducted, randomized con- but rare adverse reactions. The resulting expe-
trolled trials (RCTs) are considered the gold rience of studying risks associated with pediat-
standard for demonstrating the efficacy and ric ibuprofen and atypical antipsychotic use
safety of a medicine for regulatory approval. serves as the basis for this chapter (see Case
During the premarketing phases of drug devel- Example 17.1 and Further Reading) and may
opment, RCTs typically involve highly selected prompt others to consider randomized trials,
subjects and in the aggregate include at most a including large simple and pragmatic trials, for
few thousand patients. These studies are the postmarketing assessment of drug safety.
designed to be sufficiently large to provide evi-
dence of a beneficial clinical effect and to
exclude large increases in risk of common linical Problems to be
C
adverse clinical events. However, premarket- Addressed by
ing trials are rarely large enough to detect rela- Pharmacoepidemiologic
tively small differences in the risk of common Research
adverse events or to estimate reliably the risk
of rare events. Identification and quantifica- Pharmacoepidemiologic methods are used to
tion of these potentially important risks require quantify risks and benefits of medications that
large studies, which typically are conducted could not be adequately evaluated in studies
after a drug has been marketed. Because of performed during the premarketing phase of
design complexity and costs, large controlled drug testing. While this chapter considers only
trials are not generally conducted in the post- the assessment of the risks of medications, the
marketing setting. Rather, observational principles involved also apply to the postmar-
designs are commonly used to evaluate the keting evaluation of the benefits of
safety of medicines post-approval. The authors’ medications.
search for the best method to minimize the As noted in Chapters 1 and 3, premarketing
potential for bias yet remain relevant to real studies are typically too small to detect modest
Case Example 17.1 (See Lesko et al.)
Background failure, anaphylaxis and Reye syndrome) was

●● The use of nonsteroidal anti- not significantly greater among children
inflammatory drugs is associated with an treated with ibuprofen compared to those
increased risk of GI bleeding and renal treated with acetaminophen.
failure in adults.
Strengths
●● In 1989, ibuprofen suspension (an NSAID)
●● The large sample size allowed evaluation
was approved for use in children by pre-
of rare events.
scription only.
●● Randomization effectively controlled for
●● The risk of rare but serious adverse events
confounding, including confounding by
among children treated with ibuprofen
indication.
suspension must be documented before
this medication can be considered for a Limitations
switch from prescription to over- the- ●● The use of an active control treatment
counter use in children. (acetaminophen) precludes using these

●● Confounding by indication is likely in data to compare the risk of ibuprofen to
observational studies of prescription ibu- that of placebo in febrile children.
profen use in children. ●● Because medication exposure was limited
to the duration of an acute illness, this

Question
study cannot be used to assess the risk of
Is the use of ibuprofen suspension in chil-
long-term ibuprofen use in children.
dren associated with an increased risk of rare
but serious adverse events? Key Points
●● When confounding by indication is likely, a
Approach
RCT may be the only study design that will
●● Conduct a large simple randomized trial of
provide a valid estimate of a medication’s
ibuprofen use in children.
effect.
●● A randomized trial involving nearly 84 000
●● Large, simple, RCTs can be successfully
children 12 years of age and younger with
conducted to evaluate medication safety.
a febrile illness was conducted.
●● By keeping the study simple, it is possible
Results to conduct a large, practice-based study

The risk of rare but serious adverse events and collect data that reflects current
(hospitalization for GI bleeding, acute renal ambulatory medical practice.
differences in the incidence rates (e.g. relative satisfy the sample sizes needed to identify (or
risks of 2.0 or less) for common adverse events rule out) the relevant risks. However, findings
or even large differences in the incidence rates from observational studies are often contested
for rare events, such as those that affect 1/1000 as the basis for regulatory and clinical deci-
treated patients. Modest differences in risk of sions. Epidemiologic studies of medication
non-life-threatening adverse events can be of exposures and their effects have difficulty
substantial public health importance, particu- measuring and controlling for confounding in
larly if the medication is likely to be used by general and confounding by indication for
large numbers of patients. If there are post- drug use (and/or severity of disease) specifi-
licensing questions about the safety of a drug, cally. Uncontrolled or incompletely controlled
large observational studies are typically used to confounding can easily account for modest
296 17 The Use of Randomized Controlled Trials for Pharmacoepidemiology
associations between a drug and an adverse the underlying illness (or its severity) and not
clinical event (see Chapter 2 and Case to any effect of the drug (see also Chapter 22).
Example 17.1). As with any other form of confounding, one
In observational studies, weak associations can, in theory, control for its effects if one can
deserve attention with respect to uncontrolled reliably measure and control for the underly-
confounding. Although there are important ing risk of illness. In practice, however, this
exceptions, the general view is that the stronger often is not easily done.
the association, the more likely the observed When confronted with the task of assessing
relationship is causal. This is not to say that a the safety of a marketed drug product, the
weak association (e.g. a relative risk 1.5) can pharmacoepidemiologist must evaluate the spe-
never be causal; rather, it is more difficult to cific hypothesis to be tested, estimate the mag-
infer causality because such an association, nitude of the hypothesized association, and
even if statistically significant, can more likely determine whether confounding by indication
be an artifact of confounding. As an example, is possible. If incomplete control of confound-
consider an analysis where socioeconomic sta- ing is likely, it is important to recognize the
tus is a potential confounder and education is limitations of observational research designs
used as a surrogate for this factor. Because the and consider conducting a randomized study
relation between years of education completed design. There is nothing inherent about a ran-
(the surrogate) and socioeconomic status (the domized design that precludes a pharmacoepi-
potential confounder) is, at best, imperfect, demiologist from designing and carrying it out.
analyses controlling for years of education can To the contrary, the special skills of a pharma-
only partially control for confounding. Thus, it coepidemiologist can be very useful in per-
is advisable to use extreme caution in making forming large scale randomized trials that use
causal inferences from small relative risks epidemiologic follow-up methods.
derived from observational studies. When
there is concern about residual confounding
Overview of Classic RCTs
prior to embarking on an observational study,
one may wish to consider using a non- As noted above, RCTs are most commonly
observational study design. used during the premarketing phases of drug
development to demonstrate a drug’s efficacy
(and to gather general information concerning
Methodological Problems safety). By randomization, one hopes to equal-
to be Solved by ize the distributions of confounding factors,
Pharmacoepidemiologic whether they are known or unknown.
Research Therefore, the assigned treatment is the most
likely explanation for any observed difference
Confounding by indication (also referred to as between treatment groups in the clinical out-
indication bias, channeling, confounding by comes (improvement in the illness or the
severity, or contraindication bias) may be a occurrence of adverse clinical events). By defi-
particular problem for postmarketing drug nition, participants in observational studies
studies. According to Slone et al., confounding are not assigned treatment at random. In clini-
by indication exists when “patients who cal practice, the choice of treatment may be
receive different treatments differ in their risk determined by the stage or severity of the ill-
of adverse outcomes, independent of the treat- ness, by the underlying risk of developing the
ment received.” In general, confounding by outcome of interest, or by the patient’s poor
indication occurs when an observed associa- response to or adverse experience with alterna-
tion between a drug and an outcome is due to tive therapies, any of which can introduce bias.
Sample Size nethical in some circumstances. Further, it

u
In homogeneous populations, balanced treat- may be difficult to maintain masking in
ment groups in RCTs can be achieved with placebo-controlled studies, as noted above.
relatively small study sizes. In heterogeneous Studies employing an active control typically
populations (e.g. children less than 12 years of utilize usual drug treatments, which frequently
age), a large sample size may be required to represent the standard of care. Although often
insure the equal distribution of uncommon considered more ethical and easier to keep
confounders among study groups (e.g. infants concealed because the illness and symptoms
versus toddlers versus school-age children). are not left untreated, these studies do not per-
Study size is determined by the need to assure mit comparison with the natural history of the
balance between treatment groups and the illness.
magnitude of the effect to be detected (see
Chapter 3). Large randomized studies mini- Data Collection
mize the chance that the treatment groups are Data collection in a premarketing clinical trial
different with respect to potential confounders is generally resource intensive. Detailed
and permit the detection of small differences descriptive and clinical data are collected at
in common clinical outcomes or large differ- enrollment, and extensive clinical and labora-
ences in uncommon ones (see Chapter 3). tory data are collected at regular and often fre-
quent intervals during follow-up. In addition
Masking Treatment Assignment to the data needed to test the hypothesis of a
Concealed (or masking) treatment assignment clinical benefit, premarketing trials of medica-
is used to minimize detection bias and is par- tions must also assess general safety and there-
ticularly important where the outcome is sub- fore must collect extensive data on symptoms,
jective. Reporting of subjective symptoms by physical signs, and laboratory evaluations.
study participants and the detection of even
objectively defined outcome events may be Data Analysis
influenced by knowledge of the medications In observational studies, data analyses may be
used. For example, if a patient complains of quite complex because of the need to adjust for
abdominal pain, a physician may be more potential confounders. In contrast, analysis of
likely to perform a test for occult blood in the the primary hypothesis in many clinical trials
stool if that patient was being treated with an is straightforward and involves a comparison
NSAID rather than acetaminophen. Thus, fol- of the outcome event in different groups.
low-up data collection will only be unbiased if Analyses involving repeated measures, sub-
both parties (patient and investigator) are una- groups of study subjects, or adjustment to con-
ware of the treatment assigned. Concealing trol for incomplete or ineffective randomization
may be difficult to achieve and maintain, par- may be performed, albeit adding complexity.
ticularly if either the study or control medica-
tion produces specific symptoms (i.e. side Generalizability of Results
effects) or easily observable physiologic effects The usual clinical trial conducted during the
(e.g. nausea or change in pulse rate). pre-marketing evaluation of a drug almost
always involves highly selected patients; as a
Choice of Control Treatment consequence, the results of the trial may not be
The hypothesis being tested determines the generalizable to the large numbers of patients
choice of control treatment. Placebo controls who may use the medication after licensing.
are most useful for making comparisons with Observational studies offer an advantage in
untreated disease but may not represent stand- that they can reflect the real-world experience
ard of care and have been challenged as of medication use and clinical outcomes, and
because their modest costs permit studies Salk vaccine trial of the 1950s is an early exam-
involving large numbers of patients. ple of a very large trial. More recently, large
randomized trials have been used to test the
efficacy of therapeutic interventions, espe-
Limitations of RCTs
cially in cardiology, or to evaluate dietary sup-
Methodological strengths notwithstanding, plements or pharmaceuticals for primary
there are several features of the classic RCT prevention of cardiovascular disease and can-
that limit its use as a postmarketing study cer. This approach has also been used success-
design. First, it may be unethical to conduct a fully to evaluate the risk of adverse drug effects
study in which patients are randomly assigned when the more common observational designs
a potentially harmful treatment. For example, have been judged inadequate. LSTs are just
an RCT to test the hypothesis that cigarette very large randomized trials made simple by
smoking increases the risk of heart disease reducing data collection to the minimum
would not be acceptable. Second, the complex- needed to test only a single hypothesis (or at
ity and cost of traditional premarket RCTs, most a few hypotheses). Randomization of
with their detailed observations and resource- treatment assignment is the key feature of the
intensive follow-up, make very large studies of design, which controls for confounding by
this type generally infeasible. However, if the both known and unknown factors, and the
study can be simplified and use the epidemi- large study size provides the power needed to
ologist’s tools to track patients and collect fol- evaluate small risks of common events as well
low-up data, it may be possible to both control as large risks of rare events.
costs and make a large study feasible. Third, It is useful to note that while LSTs may
RCTs, by design, do not study the safety of a appear to be similar to pragmatic trials, they
medicine as it is actually prescribed by physi- can differ in important ways. Both are rand-
cians and used by patients once on the market. omized studies; pragmatic trials are intended
A simplified design, such as the large, simple to provide widely generalizable results and
trial (LST), merges the ideal characteristics of while they may be large, that is not always the
the RCT (randomization) with those of an case; by definition, LSTs are large to assess
observational epidemiology study (follow-up rare events and small relative risks but may
with minimal intervention). The adoption of not be generalizable. The two designs share
as many pragmatic elements as possible to similarities when LSTs may have so few exclu-
mimic usual care practice while protecting sion criteria as to make them “pragmatic”
study validity in theory yields results that are (generalizable).
more informative for regulatory decisions and
public health policy. How Simple Is Simple?
Yusuf et al. (1984) suggest that very large
randomized studies of treatment-related
Currently Available mortality collect only the participants’ vital
status at the conclusion of the study. Because
Solutions
the question of drug safety frequently con-
cerns outcomes less severe than mortality,
Large Simple Trials
these ultra simple trials may not be sufficient.
LSTs may be the best solution when it is not Hasford has suggested an alternative in
possible to completely control confounding by which “large trials with lean protocols”
means other than randomization, and the vol- include only relevant baseline, follow-up, and
ume and complexity of data collection can be outcome data. Collecting far fewer data than
kept to a minimum (see Table 17.1). The US is common in the usual RCT is the key feature
Table 17.1 Typical design characteristics of a large simple trial (LST) compared to those of a randomized
controlled trial (RCT).
Design Characteristic LST RCT
Randomization Yes Yes

Medicine Assignment Concealed if feasible, Concealed
assignment may be known, dose
adjustment permitted
Sample size Larger (thousands) Smaller (hundreds)
Inclusion criteria Broad (e.g. per approved Narrow (e.g. excludes patients with
medicine label) co-morbid conditions, using multiple
medications, pregnant women, elderly)
Questionnaire/Case Report Minimal, brief Complex, lengthy
Form (CRF)/Interview
Endpoints Hard endpoints (mortality, Virtually Any
hospitalization or life-
threatening events)
Required patient visits and Few, if any; Follows normal Yes, frequent; Visits and tests far
procedures practice schedule and greater than expected in clinical
assessments practice
Primary source of investigators Primary care provider/ Clinical research/Academic centers
or enrolling physicians Community-based
Site monitoring Minimal Frequent
Followed after randomized Yes No, or for limited duration post-
treatment discontinued discontinuation (e.g. 30 days)
Primary analytic method Intent to treat (ITT) ITT
of both approaches. With simplified proto- long latency periods may be best studied with
cols that take advantage of epidemiologic long duration follow-up. However, power is
follow-up methods, very large trials can be not the only factor to consider. For example,
conducted to test hypotheses of interest to while an elderly population may be at high risk
pharmacoepidemiologists. for gastrointestinal bleeding or cardiovascular
events, a study limited to this group may lack
Power/Sample Size generalizability and would not provide infor-
Study power is a function of the number of mation on the risk of these events in younger
events observed during the course of the study, adults or children.
which in turn is determined by the incidence
rate for the event, the sample size, and the Data Elements
duration of observation or follow-up (see The data collection process can be kept simple
Chapter 3). Power requirements can be satis- by examining primary endpoints that are objec-
fied by studying a population at high risk, tive, easily identified, and verifiable. Because
enrolling a large sample size, or conducting confounding is controlled by randomization,
follow-up for a prolonged period. The appro- data on potential confounders need not be col-
priate approach will be determined by the goal lected. Rather, a few basic demographic varia-
of the study and the hypothesis to be tested. bles can be collected at enrollment in order to
Allergic or idiosyncratic events may require a characterize the population and to confirm that
very large study population, and events with randomization was achieved.
Data Collection nausea may not be trivial for the individual

The data collection process itself can be simpli- patient but may not warrant the expense of a
fied; follow-up data can be collected by mail or large study. On the other hand, if the question
web-based questionnaires, or telephone inter- involves the risk of premature death, perma-
views conducted with the study participants. nent disability, hospitalization, or other seri-
Because the study will involve clear and objec- ous events, the cost may well be justified.
tive outcomes (see below), which can be con-
firmed by medical record review or other Uncertainty Must Exist
means, self-report by the study participants An additional condition has been referred to as
can be an appropriate source of follow-up data. the “uncertainty principle.” Originally
Other sources of follow-up data could include described by Gray et al. as a simple criterion to
electronic medical records (e.g. for LSTs con- assess subject eligibility in LSTs, it states that
ducted among subscribers of a large health “both patient and doctor should be substan-
insurance plan) or vital status records for fatal tially uncertain about the appropriateness, for
outcomes (e.g. the US National Death Index). this particular patient, of each of the trial treat-
The primary advantage of simplicity is that it ments. If the patient and doctor are reasonably
allows very large groups of study participants certain that one or other treatment is inappro-
to be followed at reasonable cost. However, a priate then it would not be ethical for the
simple trial cannot answer all possible ques- patient’s treatment to be chosen at random.”
tions about the safety of a drug but must be This principle (also known as equipoise) should
limited to testing, at most, a few related be applied to determine the appropriateness of
hypotheses. performing all RCTs, including an LST of an
adverse clinical event. Very large randomized
When Is a Large Simple Randomized trials are justified only when there is true
Trial Appropriate? uncertainty about the risk of the treatment in
LSTs are appropriate when the conditions in the population. Apart from considerations of
Table 17.2 apply. benefit, it would not be ethical to subject large
numbers of patients to a treatment that was
Important Research Question reasonably believed to place them at increased
Although a simple trial will cost less per sub- risk, however small, of a potentially serious or
ject than a traditional clinical trial, the total permanent adverse clinical event. The concept
cost of a large study (in money and human of uncertainty can thus be extended to include
resources) will still be substantial. The cost will a global assessment of the combined risks and
usually be justified only when there is a clear benefits of the treatments being compared. One
need for a reliable answer to a question con- treatment may be known to provide superior
cerning the risk of a serious outcome. A minor therapeutic benefits, but it may be unknown
medication side effect such as headache or whether the risks of side effects outweigh this
advantage. For example, the antiestrogen
tamoxifen may improve breast cancer survival,
Table 17.2 Conditions appropriate for the conduct but may do so only at the cost of an increased
of a large simply randomized trial.
risk of endometrial cancer. Appropriately, a
randomized trial was undertaken to resolve
The research question is important.
uncertainty in this situation.
Genuine uncertainty exists about the likely results.
Confounding by indication is likely. Power and Confounding
The absolute risk is small, or the relative risk is LSTs will only be needed if: (i) the absolute
small, regardless of the absolute risk.
risk of the study outcome is small and there
are concerns about confounding by indica- Table 17.3 Conditions which make a large, simple
tion, or (ii) the relative risk is small (in which randomized trial feasible.
case, there are inherent concerns about
The study question can be expressed as a simple
residual confounding from any source). By
testable hypothesis.
contrast, LSTs would not be necessary if the
The treatment to be tested is simple
absolute risk were large, because premarket (uncomplicated).
or other conventional RCTs should be ade-
The outcome is objectively defined (e.g.
quate, or where uncontrollable confounding hospitalization, death).
is not an issue, because observational studies Epidemiologic follow-up methods are appropriate.
would suffice. Also, if the relative risk were
The patient and physician population are
large (and confounding by indication and motivated to participate by a meaningful research
other potential biases inherent in observa- question.
tional studies are not concerns), observa-
tional studies would be appropriate.
Simple Hypothesis
LSTs are best suited to answer focused and
No Interaction Between Treatment
relatively uncomplicated questions. For exam-
and Outcome
ple, an LST can be designed to test the hypoth-
An additional requirement for LSTs is that
esis that the risk of hospitalization for any
important interactions between the treatment
reason, or for acute gastrointestinal bleeding,
and patient characteristics (effect modifica-
is increased in children treated with ibuprofen.
tion) are unlikely. In other words, the available
However, it may not be possible for a single
evidence should suggest that the association
LST to answer the much more general ques-
will be qualitatively similar in all patient sub-
tion, “Is ibuprofen safe with respect to all pos-
groups. Variation in the strength of the associ-
sible outcomes in children?”
ation is acceptable among subgroups, but there
should be no suggestion that the effect would Simple Treatments
be completely reversed in any subgroup. Simple therapies (e.g. a single drug at a fixed
Because of the limited data available in a truly dose for a short duration) are most amenable
simple trial, it may not be possible to test to study with LSTs. They are likely to be com-
whether an interaction has occurred, and the monly used, so that it will be feasible to enroll
data collected may not be sufficient to identify large numbers of patients, and the results will
relevant subgroups. Because randomization be applicable to a sizeable segment of the pop-
only controls confounding for comparisons ulation. Complex therapeutic protocols are dif-
made between the groups that were rand- ficult to manage, can reduce patient adherence,
omized, subsets of these groups may not be and by their very nature may not be compatible
strictly comparable with respect to one or more with the simple trial design.
confounding factors. Thus, if clinically impor-
tant interaction is considered likely, additional Objective and Easily Measured
steps must be taken to permit the appropriate Outcomes
analyses (e.g. stratified randomization). This The outcomes to be studied should be objective
added complexity may result in a study that is and easy to define, identify, and recall. An
no longer a truly simple trial. example might include hospitalization for
acute gastrointestinal bleeding. Study partici-
pants may not correctly recall the details of a
When is an LST Feasible?
hospital admission, or even the specific reason
LSTs are feasible when the conditions in for admission, but they likely will recall the
Table 17.3 are met. fact that they were admitted, the name of the
hospital, and at least the approximate date of emails and mobile phone text messages) may
admission. Medical records can be obtained to be appropriate for some studies. Because suc-
document the details of the clinical events that cess depends on the cooperation of multiple
occurred. Events of this type can be reliably health care providers and a large number of
recorded using epidemiologic follow-up meth- patients, it is best to limit the demands placed
ods (e.g. questionnaires, telephone interviews, on each practitioner (or his/her clinical prac-
or linkage with public vital status records). On tice). One approach is to have the practitioner
the other hand, clinical outcomes that can be identify eligible subjects, obtain permission to
reliably detected only by detailed in-person pass their names to a central study staff, and
interviews, physical examinations, or exten- leave to the study staff the task of explaining
sive physiologic testing may not be amenable study details, enrollment, and obtaining
for study in simple trials. informed consent. Another approach is to
provide comprehensive training prior to site
Cooperative Population initiation followed by support to local
A study population motivated by the research administrative staff throughout the course of
question will greatly increase the probability of the study, particularly for research-naïve and
success. Striking examples are the large popu- inexperienced sites. Obtaining informed con-
lations in the Physicians’ and Women’s Health sent, baseline data, and the medicine assign-
Studies; the success of these studies is at least ment are best handled during a single visit.
partly due to the willingness of large numbers To facilitate patient recruitment and to maxi-
of knowledgeable health professionals to par- mize generalizability of the results, minimal
ticipate. Because of the participants’ knowl- restrictions should be placed on patient eligi-
edge of medical conditions and symptoms and bility. Patients with a medical contraindication
participation in the US health care system, or known sensitivity to either the study or con-
relatively sophisticated information was trol drug should not, of course, be enrolled, but
obtained using mailed questionnaires, and other restrictions should be kept to a minimum
even biologic samples were collected. and should ideally reflect only restrictions that
would apply in a typical clinical setting.
Substantial bias can be introduced if either
Logistics of Conducting an LST
physician or patient can choose not to partici-
An LST may be appropriate and feasible, but it pate after learning (or guessing) which treat-
will only succeed if all logistical aspects of the ment the patient has been assigned. Therefore,
study are kept simple as well. In general, LSTs patients should be randomized only after eligi-
will involve an oversight body, sometimes bility has been confirmed and the enrollment
organized as a Scientific Steering Committee process completed.
comprised of epidemiologic, statistical and
clinical experts who are responsible for the sci- Importance of Complete Follow-up
entific conduct of the study, as well as a central Because dropouts and losses to follow-up may
data coordinating facility, and a network of not be random but may be related to adverse
enrollment sites (e.g. offices of collaborating treatment effects, it is important to make every
physicians or other health care providers). effort to obtain follow-up data on all subjects.
Health care professionals (e.g. physicians, A study with follow-up data on even tens of
nurse practitioners, and pharmacists in private thousands of patients may not be able to pro-
practice or members of large health care vide a valid answer to the primary study ques-
organizations) can participate by recruiting eli- tion if this number represents only a modest
gible patients. Alternative methods to identify proportion of those randomized. The duration
and enroll eligible subjects (e.g. direct mailings of the follow-up period can affect the com-
to professional groups, print or online ads, pleteness of follow-up data collection. If it is
Analysi 303
too short, important outcomes may be missed although as previously noted this can make the
(i.e. some conditions may not be diagnosed trial far more complex. In addition, a search of
until after the end of the follow-up period). On public records (e.g. the National Death Index
the other hand, as the length of the follow-up in the US) can identify study subjects who have
period increases, the number lost to follow-up died during follow-up.
or exposed to the alternate treatment (contam-
inated exposure) increases. In the extreme, a
randomized trial becomes an observational Analysis
cohort study because of selective dropouts in
either or both of the treatment arms. Beyond Primary Analysis
choosing a motivated and interested study
Analyses of the primary outcomes are usually
population, the investigators can minimize
straightforward and involve a simple compari-
losses to follow-up by maintaining contact
son of incidence rates between the treatment
with all study participants. Regular mailings of
and control groups. Under the assumption
supplies of medication, a study newsletter, or
that confounding has been controlled by the
email reminders can be helpful, and memory
randomization procedure, complex multivari-
aids such as medication calendar packs or
ate analyses are not necessary (and may not be
other devices may help maintain adherence
possible because only limited data on poten-
with chronic treatment schedules.
tial confounders are available). Descriptive
data collected at enrollment should be ana-
lyzed by treatment group to test the randomi-
Follow-up Data Collection
zation procedure; any material differences
An important element of a successful LST is
between treatment groups suggest an imbal-
that the burden to health care providers for
ance despite randomization. As noted above,
follow-up data collection is minimized. Busy
it is assumed that there is no material interac-
health care providers cannot be expected to
tion between patient characteristics and medi-
commit the time required to obtain systemati-
cation effects, thus eliminating the need for
cally even minimal follow-up data from large
complex statistical analyses to test for effect
numbers of subjects. However, the clinician
modification.
who originally enrolled the subject may be able
to provide limited follow-up data (e.g. vital sta-
tus) or a current address or telephone number
Subgroup Analyses
for the occasional patient who would other-
wise be lost to follow-up. A mailed or elec- It is important to remember that confounding
tronic questionnaire, supplemented by factors will be distributed evenly only among
telephone follow-up when needed, is effective. groups that were randomized; subgroups which
The response rate will likely be greatest if the are not random samples of the original rand-
questions are both simple and direct and mini- omization groups may not have similar distri-
mal time is required to complete the question- butions of confounding factors. For example,
naire. Medical records can be reviewed to participants who have remained in the study
verify important outcomes, such as rare (i.e. have not dropped out or been lost to follow-
adverse events, and the work needed to obtain up) may not be fully representative of the origi-
and abstract the relevant records should be nal randomization groups and may not be
manageable. If there is a need to confirm a comparable with respect to confounders among
diagnosis or evaluate symptoms, a limited the different groups. Despite all efforts, com-
number of participants can be referred to their plete follow-up is rarely achieved, and because
enrolling health care provider for examination only the original randomization groups can be
or to have blood or other studies performed, assumed to be free of confounding, the primary
analysis should include all enrolled study sub- be a growing need for LSTs to evaluate larger
jects regardless of whether or not they adhered relative risks. Because of few restrictions on
to taking the assigned therapy (i.e. an intention- participant eligibility, LSTs are more likely
to-treat analysis) should be performed. Also, than classical randomized clinical trials to
unless a stratified randomization scheme was reflect the true benefits and risks of medica-
used, one cannot be certain that unmeasured tions when used in actual clinical practice. The
confounding variables will be evenly distrib- generalizability of the results of LSTs and other
uted in subgroups of participants, and the pragmatic clinical trials makes these studies
smaller the subgroup, the greater the potential particularly attractive to regulators and policy-
for imbalance. Therefore, subgroup analyses makers and may lead to increased use of these
will be subject to the same limitations as obser- studies.
vational studies (i.e. the potential for uncon- One possible approach that may improve
trolled confounding). efficiency in large studies is to conduct trials
involving patients who receive care from very
large health delivery systems (see Chapters 8
Data Monitoring/Interim Analyses
and 10). These data arise from electronic
Because of the substantial commitment of health records (EHRs) with information
resources and large number of patients poten- recorded by clinical staff at the point of care
tially at risk for adverse outcomes, it is appro- (e.g. in hospitals or outpatient clinics), admin-
priate to monitor the accumulating data during istrative claims data (e.g. the Veterans Affairs
the study. A study may be ended prematurely if database in the US), national/regional regis-
participants experience unacceptable risks, if tries (e.g. population-wide databases in Sweden
the hypothesis can be satisfactorily tested ear- and Denmark), and patient disease/condition/
lier than anticipated, or if it becomes clear that drug registries (e.g. CorEvitas rheumatoid
a statistically significant result cannot be arthritis registry). EHRs and registries have
achieved, even if the study were to be com- been used in conventional RCTs (e.g. the West
pleted as planned. A data monitoring commit- of Scotland Coronary Prevention Study, the
tee, independent of the study investigators, Scandinavian Simvastatin Survival Study and
should conduct periodic reviews of the data the EHR4CR project) to identify potential trial
using an appropriate analysis procedure. participants and collect long-term follow-up
data. The most recent development is the use
of registries for identification, recruitment and
The Future as follow-up for trials, providing an efficient
and reusable infrastructure for LSTs. While
With accelerated approval of new medications uncommon, recent simplified trials conducted
and rapid increases in their use, we may see a in the UK Clinical Practice Research Data, the
greater need for large, randomized postmar- US Veterans Affairs Healthcare System and
keting studies to assess small differences in PCORnet, the National Patient Centered
risk. This is particularly the case for medica- Clinical Research Network, illustrate the
tions considered for over-the-counter switch, approach. Research questions where the study
because the risks of rare and unknown events results are directly relevant to patient care offer
that would be acceptable under prescription the most promise for a successful trial in large
status might be unacceptable when the drug is health delivery systems. With more experi-
self-administered by much larger and more ence, we anticipate more systems will recog-
diverse populations. In the absence of tech- nize the potential benefits and efficiencies of
niques that reliably control for confounding by this approach, developing the capability to
indication in observational studies, there will conduct embedded point-of-care LSTs.
Further Reading 305
It is clear that very large simple controlled Key Points

trials of drug safety can be successfully carried
out. It is less clear, however, how frequently ●● Randomization usually controls for con-
the factors that indicate the need for a very founding, including confounding by
large trial (Table 17.1) will converge with those indication.
that permit such a trial to be carried out ●● A large study allows assessment of small to
(Table 17.2). As a discipline, pharmacoepide- modest associations of common events and
miology is well suited to conduct LSTs and to large associations with rare events and
develop more efficient methods of subject assures that randomization produces bal-
recruitment and follow-up data collection that anced treatment groups.
can make these studies a more common option ●● Large RCTs are feasible if data collection is
for the evaluation of small but important risks kept simple and outcome events are objec-
of medication use. tive and verifiable.
Further Reading
Choudhry, N.K. (2017). Randomized, controlled Hennekens, C.H. and Buring, J.E. (1989).
trials in health insurance systems. N. Engl. J. Methodologic considerations in the design
Med. 377: 957–964. and conduct of randomized trials: the U.S.
Connolly, S.J., Ezekowitz, M.D., Yusuf, S. et al. Physicians’ Health Study. Control. Clin. Trials
(2009). Dabigatran versus warfarin in patients 10: 142S–150S.
with atrial fibrillation. N. Engl. J. Med. 361 Hennekens, C.H., Buring, J.E., Manson, J.E.
(12): 1139–1151. et al. (1996). Lack of effect of long-term
D’Avolio, L., Ferguson, R., Goryachev, S. et al. supplementation with beta carotene on the
(2012). Implementation of the department of incidence of malignant neoplasms and
Veterans Affairs’ first point-of-care clinical cardiovascular disease. N. Engl. J. Med. 334:
trial. J. Am. Med. Inform. Assoc. 19 (e1): 1145–1149.
e170–e176. Lee, I.M., Cook, N.R., Manson, J.E. et al. (1999).
DeMets, D.L. (1998). Data and safety monitoring Beta-carotene supplementation and incidence
boards. In: Encyclopedia of Biostatistics (eds. P. of cancer and cardiovascular disease: the
Armitage and T. Colton), 1067–1071. Women’s Health Study. J. Natl. Cancer Inst. 91
Chichester: Wiley. (24): 2102–2106.
Fisher, B., Costantion, J.P., Wickerham, D.L. Lesko, S.M. and Mitchell, A.A. (1995). An
et al. (1998). Tamoxifen for prevention of assessment of the safety of pediatric
breast cancer: report of the National ibuprofen: a practitioner-based randomized
Surgical Adjuvant Breast and Bowel Project clinical trial. JAMA 273: 929–933.
P-1 Study. J. Natl. Cancer Inst. 90: Mitchell, A.A. and Lesko, S.M. (1995). When a
1371–1388. randomized controlled trial is needed to assess
Francis, T. Jr., Korns, R., Voight, R. et al. (1955). drug safety: the case of pediatric ibuprofen.
An evaluation of the 1954 poliomyelitis Drug Saf. 13: 15–24.
vaccine trials: summary report. Am. J. Public ONTARGET Investigators, Yusuf, S., Teo,
Health 45 (suppl): 1–50. K.K. et al. (2008). Telmisartan, ramipril, or
Hasford, J. (1994). Drug risk assessment: a case both in patients at high risk for vascular
for large trials with lean protocols. events. N. Engl. J. Med. 358 (15):
Pharmacoepidemiol. Drug Saf. 3: 321–327. 1547–1559.
Peto, R., Collins, R., and Gray, R. (1995). during 10-year follow-up of the Scandinavian
Large-scale randomized evidence: large, Simvastatin Survival Study (4S). Lancet 364
simple trials and overviews of trials. J. Clin. (9436): 771–777.
Epidemiol. 48 (1): 23–40. Strom, B.L., Faich, G.A., Reynolds, R.F. et al.
Reynolds, R.F., Lem, J.A., Gatto, N.M., and Eng, (2008). The ziprasidone observational study of
S.M. (2011). Is the large simple trial design used cardiac outcomes (ZODIAC): design and
for comparative, post-approval safety research? baseline characteristics. J. Clin. Psychiatry 69:
A review of a clinical trials registry and the 114–121.
published literature. Drug Saf. 34: 799–820. Strom, B.L., Eng, S.M., Faich, G. et al. (2010).
Rothman, K.J. and Michels, K.B. (1994). The Comparative mortality associated with
continuing unethical use of placebo controls. ziprasidone and olanzapine in real-world use
N. Engl. J. Med. 331: 394–398. among 18,154 patients with schizophrenia:
Slone, D., Shapiro, S., Miettinen, O.S. et al. (1979). the ziprasidone observational study of cardiac
Drug evaluation after marketing. A policy outcomes (ZODIAC). Am. J. Psychiatry 168:
perspective. Ann. Intern. Med. 90: 257–261. 193–201.
van Staa, T.P., Dyson, L., McCann, G. et al. Yusuf, S., Collins, R., and Peto, R. (1984). Why
(2014). The opportunities and challenges of do we need some large, simple randomized
pragmatic point-of-care randomised trials trials? Stat. Med. 3: 409–420.
using routinely collected electronic records: Yusuf, S. (1993). Reduced mortality and
evaluations of two exemplar trials. Health morbidity with the use of angiotensin-
Technol. Assess. (Winch. Eng.) 18 (43): converting enzyme inhibitors in patients
1–146. with left ventricular dysfunction and
Strandberg, T.E., Pyörälä, K., Cook, T.J. et al. congestive heart failure. Herz 18 (suppl):
(2004). Mortality and incidence of cancer 444–448.
307
18
Pharmacoeconomics
Economic Evaluation of Pharmaceuticals
Kevin A. Schulman
Clinical Excellence Research Center, Stanford University, Stanford, CA, USA
Introduction The Economics of Drug

Development
The science of drug development and assess- Drug development starts with an investment
ment has been well described throughout this in science. Historically, public grant funds
book. The economics of this process brings through the National Institutes of Health or
together concepts of finance, health economics, the National Science Foundation, or private
and behavioral economics in a manner that is funds through programs like the Howard
truly unique. Hughes Medical Institute, would support
fundamental science that might be years or
even generations away from translation into
C medical products. Academic researchers and
Addressed by the pharmaceutical industry would use the
Pharmacoeconomic insights from this work to begin an effort at
Research translation, moving from fundamental science
to specific interventions for specific diseases.
At its core, drug development is about innova- This work could still be publicly funded, but
tion, and bringing the benefits of the science of might be more likely to be funded through
medicine to patients. Drug development is private resources such as pharmaceutical
time consuming, expensive, and risky, so we firms, or even applied science efforts such as
need to ensure there are adequate financial the Bill and Melinda Gates Foundation. These
incentives to bring these new innovations to efforts serve to translate biology into drug
market, especially for areas of unmet clinical targets, identifying potential pathways to alter
need. At the same time, we need to ensure the target through the identification of small
access to medications, including their afforda- molecules or biologics that have (hopefully)
bility to public or private payers. unique effects on the target. The discovery of a
Pharmacoeconomics provides a pathway to candidate drug would lead to the filing of a
understand the conflicts between all of these patent, an opportunity for the inventor to own
interests and perspectives. the rights to the discovery and to preclude
308 18 Pharmacoeconomics
others from patenting the invention (under a Switzerland, and $3930 in the United States.
provision called the Bayh–Dole Act, universi- Not only are drug prices high, but cancer ther-
ties own the patent rights to discoveries even if apies have experienced significant price
the work was funded using federal grants). growth. In one analysis, the monthly cost of
The patent rights are critical in the next stage oncology products has increased from approxi-
of drug development. This is when drug mately $100 as recently as 1980 to $10 000 by
development moves out of the laboratory and 2010.
into human testing. This is the most expensive
step in drug development, and for the most
Health Economics and Health Care
part the work is privately funded. Investors
Financing
justify their willingness to invest in a molecule
given the opportunity for financial returns Health Insurance
resulting from their ownership of the molecule Insurance is a mechanism for sharing risk
through the patent. This transition from public across individuals. Generally, insurance works
to private support is challenging for many best when the occurrence being insured is
discoveries. The “Valley of Death” is the gap infrequent, can be catastrophic to the
between science that is funded by public grants individual, and is not influenced by the
and the ability to attract private investment to individual or organization being insured. In
the development of a molecule. insurance markets characterized by these
Clinical testing of pharmaceuticals is conditions, insurance can be a relatively
carefully regulated by the Food and Drug inexpensive proposition. Health insurance has
Administration (FDA) and other global different characteristics. We use healthcare
regulatory bodies. For most products, services frequently, and with medications even
regulatory authorities require proof of safety more frequently. While some healthcare costs
and efficacy of products before they are can be catastrophic, not all are. While paying
approved for sale. Clinical testing can require for a monthly medication is not enjoyable, for
up to a decade to complete and can require most people it is not financially catastrophic.
more than $1 billion in direct outlays. Finally, consumption of healthcare services is
At the end of drug development, with inherently influenced not only by individuals
product approval by the FDA, the manufacturer (do you want to go to the clinician for that bad
can set a price and market the product. Prices cold or sprained ankle?), but also by
set by manufacturers reflect their significant pharmaceutical manufacturers and healthcare
investment in clinical development, and the systems encouraging you in prime time televi-
inherent risk they were required to assume, sion advertisements to consume more health-
but also considerations of market access or care resources. As a result, health insurance is
barriers to full reimbursement for patients. often a very expensive insurance product. In
The price can reflect the marginal cost of truth, most health insurance combines the
producing a product, but often this is a idea of prepayment for usual healthcare ser-
relatively minor consideration. The prices of vices with a catastrophic medical benefit (to
specialty pharmaceutical products can be some extent, high-deductible health plans try
extraordinary, reaching $475 000 per patient to separate out these two different elements of
for Novartis’s CAR-T therapy. Prices also vary healthcare financing).
across markets, with a 30-day supply of A lot of attention has been focused on the
Janssen’s Xarelto® priced at $48 in South Africa, impact of health insurance on the prices of
$102 in Switzerland, and $292 in the United pharmaceutical products. Historically, pre-
States, or 400 mg of Genentech’s Avastin® scription drugs were relatively affordable,
priced at $956 in South Africa, $1752 in and so were paid for by patients. As
Clinical Problems to be Addressed by Pharmacoeconomic Researc 309
edications became more effective, the con-

m have our own tastes, preferences, and needs
cept of prescription drug insurance began to which form our assessment of value.
develop. In 1960 in the United States, 96% of Third-party payment alters this fundamental
prescription drug spending was out of pocket calculus. Consider going out to dinner with a
by individuals. By 1980, out-of-pocket spend- group of friends. After the menu is passed
ing was still 71% of total spending. By 1990, around, you notice items of lower and higher
it was down to 57%, 2000 to 28%, 2010 to 18%, price, say salad and steak. You can approach
and 2015 to 14%. Prescription drug coverage payment in one of two ways: individual checks
has led to a transformation of the pharma- or splitting the check. If you all decide on indi-
ceutical market. Over this same period, the vidual checks before you order, you may decide
prescription drug market in the US has to purchase the lower-cost salad since you are
grown from $2.7 billion in 1960 to $479 bil- on a budget. However, what happens if you
lion in gross sales in 2018 (resulting in net decide to split the check before you order? You
pharmaceutical sales to manufacturers of may be worried that everyone else at the table
$344 billion). is likely to order the higher-priced steak, and
you will have to pay your share of their higher-
Moral Hazard priced meals. Since you are paying for their
Health economists have long been worried steak, why not order your own steak so at least
about the economic impact of health insurance you get the benefit of the higher price you will
on the patterns of consumption of healthcare pay for dinner? In this simple illustration, your
due to a concept called “moral hazard.” Moral behavior changes between self-payment and
hazard describes the change in individual third-party payment models.
behavior between conditions of self-pay and Health insurance is one form of third-party
conditions of third-party payment. Kenneth payment. Under health insurance, rather than
Arrow was awarded the Nobel Prize in paying the full cost of medical products, you
economics for developing this framework, and pay only a co-payment (fixed amount), or co-
Mark Pauly further developed the theory to insurance (a percentage payment) for medical
focus on demand. products. As illustrated in Figure 18.1, products
The basic framework is easy to understand. 1 through 3 offer at least $1.00 of value for
We all make purchases based on our concept of $1.00 of cost. In a self-payment model, you
value. We generally make purchases of goods would be expected to purchase only products 1
or products for $1.00 when we perceive that through 3 since only these products have a
they offer $1.00 worth of value. This concept of value of $1.00. In an insurance model, however,
value is an individual determination: we all you only pay the co-payment of $0.20. Now,
Product Product Product Product Product Product Product

1 2 3 4 5 6 7
Value 1.50 1.25 1.0 .75 .50 .25 .10

Cost
$1.00 $1.00 $1.00 $1.00 $1.00 $1.00 $1.00
(No Insurance)
Cost (Insurance) $0.20 $0.20 $0.20 $0.20 $0.20 $0.20 $0.20
Figure 18.1 Moral Hazard and Product Choice. Note: Cost (No insurance)-assumes only cash payments for
the product. Cost (insurance)-assumes the product is covered by an insurance policy with a 20% coinsurance
requirement.
products 1 through 6 offer value equal to or circumstances when suppliers have power to
greater than the $0.20 copayment, so using the influence prices, especially in healthcare.
same rule (only buying products that offer Suppliers can have market power when they
value greater than or equal to the price you have a barrier to market entry such as a patent
pay), you would purchase products 1 through awarded to a pharmaceutical manufacturer or
6. Again, behavior changes under conditions a product developed for a niche category, such
of third-party payment. While many econo- as an orphan drug, which is too small to attract
mists have argued that health insurance competition. In these cases, suppliers can
increases the overall cost of healthcare due to increase the price of product 1 based on value.
these changes in demand, there is also the con- If they decide to price at the total value of the
cept of good moral hazard where people can product, they could raise the price from $1.00
purchase goods or products through insurance to $1.50 to capture the full value to patients.
that would otherwise be unaffordable. It is pos- Under our conceptual model, this pricing
sible to develop a direct estimate of the increase strategy would be attractive to patients even in
in prescription drug spending between those a cash pay market. However, under conditions
with and without insurance. of third-party payment, suppliers can consider
To this point, the discussion has focused on an even more aggressive pricing strategy by
the impact of moral hazard on the demand for considering that patients measure value
healthcare products. However, the impact of against their co-insurance, not the full cost of
moral hazard also extends to the supply side of the product. Under these conditions, suppliers
healthcare. While much of the literature exam- can raise the price to $7.50 while consumers
ines the impact of moral hazard on the provi- would have a cost-share of $1.50, or an amount
sion of services, there is also an impact on the equal to the value they expect to receive from
price of products. Given insurance, the suppli- the therapy. As a result of supply-side moral
ers of high-value products can realize that hazard, the cost increased from $1.00 to
products are perceived as being significantly $7.50 in this simple example. Co-payment
underpriced since insured patients only con- coupons or patient assistance programs can
sider the out-of-pocket costs and not the full exacerbate this effect by artificially decreasing
cost of therapy. Applying a value framework to the amount individuals have to pay. This
pricing can lead manufacturers to raise their “discount” on out-of-pocket payments can
prices to meet the value threshold rather than allow suppliers even more latitude to raise
simply developing a price to meet their inter- prices under this framework.
nal financial expectations. This supply-side
moral hazard effect on the price of pharmaceu- Behavioral Economics
tical products have been much less discussed This concept of patients being risk-averse is
in the literature. consistent with the idea of buying health
Again, going back to the basic example of insurance in the first place. Buying health
product 1 in Figure 18.1, this product provides insurance is seen as a risk-averse financial
great value to patients under conditions of self- decision. People pay some money annually for
payment and even more under conditions of health insurance to avoid the potential
third-party payment. Sophisticated suppliers financial consequences associated with the
will notice these conditions. In a competitive rare risk of becoming severely ill. Consumers
market, suppliers will have little ability to may even buy certain policies with limits on
influence the welfare surplus enjoyed by things that are not important to them when
patients in this example since the price is they are healthy – narrow networks of
determined by the market and is driven by providers, for example, or limits on the drug
entry and exit of firms. However, there are formulary for specialty pharmaceutical
Clinical Problems to be Addressed by Pharmacoeconomic Researc 311
products. However, buying health insurance is patients. This study of the psychology of deci-
not the same as buying health care. Whether sion making in real-world setting has been
the risk-averse decision-making approach to called behavioral economics.
buying insurance carry-over to making treat- More recently, Kahneman and others have
ment decisions for health care products or ser- focused on the role of emotion in decision
vices is an open question. making. They have developed a framework
Let’s consider a clinical scenario. Assume an which considers two different ways of making
otherwise healthy patient comes into a physi- decisions, System 1 and System 2. System 1
cian’s office. They feel great, have a full social decision processes are autonomous decision-
and work life, exercise regularly, and have a lot making efforts that represent our “gut” or
to look forward to. Given a history of smoking emotional response to an uncertain situation.
in the past, the physician had ordered a chest System 1 processes easily incorporate societal
X-ray. Unfortunately, the chest X-ray showed attitudes and is subject to many systematic
that the patient had a spot on their lung. After flaws. System 2 decision processes are more
further work-up, it was found to be lung cancer data driven and analytical, but they have a
that had spread. This otherwise healthy person high cognitive burden. In the normal course of
now has a life-threatening condition. events, we make most decisions using the
Obviously, this is a significant loss in life expec- System 1 framework, despite its limitations, so
tancy for the patient. How do they react to the that we minimize our cognitive burden in
shock of their diagnosis? They seek treatment making simple choices or completing simple
for their condition. In this case, the patient will job tasks. However, we have System 2 processes
accept a treatment which has any chance of available for more complex decision making.
restoring their health, irrespective of the side Importantly, in a heightened emotional state,
effects of the therapy. They definitely do not we generally rely on System 1 processes for
ask about the cost of treatment. Under condi- decision making. This can be critically
tions of loss, the way that patients make deci- important in understanding medical decision
sions changes from how they felt about future making, where patients (or their loved ones)
potential treatment choices when they bought can experience significant anxiety from the
their health insurance policy. care process, the diagnosis itself, or can be in a
This idea that people make different heightened emotional state from the experience
decisions under conditions of gains and losses of the symptoms of the illness, especially when
earned Daniel Kahneman the Nobel Prize in suffering from a disease with an acute
Economics in 2002 (he collaborated with Amos presentation.
Tversky in developing prospect theory, but While the role of loss can make patients
Amos passed away before the prize was appear to be risk seeking in making treatment
awarded). Under conditions of gains, we are choices, we have suggested that the role of
risk-averse, and under conditions of loss we emotion can also lead to the same type of
are risk-seeking. When a 70-year-old patient decision making by patients.
refuses a flu shot because of her concerns that
she does not want to get sick from the shot she Economic Evaluation
is under a condition of gain (full health) and is of Pharmaceuticals
being risk-averse. The unfortunate patient Across the globe, technology assessment
with lung cancer is an example of decision agencies have been established to help provide
making under conditions of loss. The applica- or evaluate economic data as part of the
tion of this framework to treatment choices by reimbursement process. In light of all of the
patients with life threatening diseases helps to challenges of third-party payment, the charge
explain the apparently risk-seeking behavior of to these agencies is to understand the value of
new medical therapies (value is the cost in Figure 18.2 with which readers should become
relationship to the benefit), and to make a rec- familiar. Along the X-axis are three types of
ommendation about whether they should be economic analysis – cost identification, cost-
included in a benefit package for patients. In effectiveness, and cost–benefit. Along the
the UK, the National Institute for Health and Y-axis are four points of view, or perspectives,
Care Excellence (NICE) provides guidance to that one may take in carrying out an analysis.
the National Health Service. In Germany, the One may take the point of view of society in
Institute for Quality and Efficiency in Health assessing the costs and benefits of a new medi-
Care (IQWiG) evaluates the effectiveness of cal therapy. Alternatively, one may take the
drugs. In the US, the Institute for Clinical and point of view of the patient, the payer, or the
Economic Review (ICER) is a private organiza- provider. Along the third axis, the Z-axis, are
tion publishing independent economic analy- the types of costs and benefits that can be
ses of new pharmaceutical products. included in economic analysis of medical care.
These costs and benefits, which will be defined
below, include direct costs and benefits, pro-
M ductivity costs and benefits, and intangible
to be Addressed by costs and benefits.
Pharmacoeconomic
Research Types of Analysis
Cost–Benefit Analysis
In considering economic analysis of medical Cost–benefit analysis of medical care com-
care, there are three dimensions of analysis, pares the cost of a medical intervention to its
represented by the three axes of the cube in benefit. Both costs and benefits are measured
s
Intangible
fit
ne
be
Productivity
d
an
Direct non-medical
s
st
Co
Direct medical
Society
Patient
Perspective
Payer
Provider
n ss it
tio e nef
ifica en be
nt ctiv st-
id
e ffe Co
st
st-e
Co Co
Analysis
Figure 18.2 The three dimensions of economic evaluation of clinical care. Reproduced from Bombardier
and Eisenberg (1985) with permission from The Journal of Rheumatology.
Methodological Problems to be Addressed by Pharmacoeconomic Researc 313
Case Example 18.1 Economic Evaluation of High-Dose Chemotherapy plus Autologous Stem
Cell Transplantation for Metastatic Breast Cancer
Background greater morbidity and economic costs with

A clinical trial of high-dose chemotherapy no improvement in survival.
plus autologous hematopoietic stem cell ●● Results of the sensitivity analyses sug-
transplantation versus conventional dose gested that the findings were robust even
chemotherapy in women with metastatic when important cost assumptions varied.
breast cancer found no significant differ-
Strengths
ences in survival between the two treatment
●● By studying resource use and estimating
groups. Thus, the economic evaluation would
costs, the authors were able to quantify
provide decision makers with important
the economic burden associated with the
additional information about the two
two treatments.
therapies.
●● The study allowed the investigators to
Question provide novel information about the “clini-

What were the differences between the two cal trajectories” of patients with metastatic
treatment groups with regard to course of breast cancer.
treatment and resources consumed? ●● The economic evaluation did not place any
additional data collection burden on

Approach
investigators, but yielded important sec-
●● The researchers abstracted the clinical
ondary findings.
trial records and oncology department
●● Economic evaluation allowed the research-
flow sheets retrospectively to document
ers to quantify the economic burden asso-
resource use.
ciated with interventions.
●● Each patient’s course of treatment and
resource use was analyzed in four phases. Limitations

Based on these clinical phases, patients ●● Collection of resource use data from the
were grouped into one of three clinical clinical trial records may have resulted in
“trajectories.” underestimation of treatment costs.
●● Costs were estimated using the Medicare ●● Resource costs were estimated rather than
Fee Schedule for inpatient costs and aver- directly observed.

age wholesale prices for medications.
Key Points
●● Sensitivity analyses examined changes in
●● By studying resource use and estimating
the discount rate, hospital costs, and the
costs, the authors were able to quantify
number of cycles of the chemotherapeutic
the economic burden associated with the
drugs paclitaxel and docetaxel.
two treatments and to provide information
Results about the “clinical trajectories” of patients
●● Patients undergoing transplantation used with metastatic breast cancer.
more resources, mostly due to inpatient ●● Sensitivity analyses are crucial in studies
care. that rely on numerous estimates and

●● The investigators also found differences assumptions.
by clinical trajectory, and these differences ●● Economic analysis can provide important
were not consistent between treatment additional information to decision makers

groups. in cases where no differences were
●● High-dose chemotherapy plus stem-cell observed between treatment groups on
transplantation was associated with the primary clinical endpoint.
Case Example 18.2 Economic Analysis of a Novel SARS-CoV-2 Vaccine
Background a cost-effectiveness analysis, or a cost–

SARS-CoV-2 emerged in late 2019 and benefit analysis? What about a benefit
spawned a global pandemic in 2020. Since pool analysis?
this was a novel virus, there was no native ●● How does choice of perspective influence
immunity in infected populations. There was your assessment of the question?
a global rush to develop a vaccine. You are ●● How does societal value relate to the
now a member of a national panel looking to determination of a market price? What
determine the appropriate market price for other considerations might be in play?
the vaccine.
●● Would your analysis change if you were
Question and Issue risk-averse (someone who survived an
●● What are the economic benefits of the asymptomatic infection with SARS-
vaccine? CoV-2), or risk-seeking (high-risk individ-
●● How would you assess the value of the vac- ual who has not been infected with
cine based on estimates of the population SARS-CoV-2)?
that have yet to be exposed to the virus? ●● How would consideration of the role of
●● How should these economic benefits public vs private investment in vaccine
relate to the price you would agree is a fair development impact your assessment?
market price for the vaccine? For devel-
oped countries? For emerging market
Key Points
economies?
This analysis of a fair market price for a
Approach SARS-CoV-2 vaccine highlights many of the
●●How would your answer differ if you were challenges in economic analysis of pharma-
asked to consider a budget impact analysis, ceutical products.
in the same (usually monetary) units (e.g. dol- utcomes as part of the evaluation. While cost
o
lars). These measurements are used to deter- generally is still calculated only in monetary
mine either the ratio of dollars spent to dollars terms (e.g. dollars spent), effectiveness is deter-
saved or the net saving (if benefits are greater mined independently and may be measured
than costs) or net cost. All else being equal, an only in clinical terms, using any meaningful
investment should be undertaken when its clinical unit. For example, one might measure
benefits exceed its costs. One potential diffi- clinical outcomes in terms of number of lives
culty of cost–benefit analysis is that it requires saved, complications prevented, or diseases
researchers to express an intervention’s costs cured. Alternatively, health outcomes can be
and outcomes in the same units. Thus, mone- reported in terms of a change in an intermedi-
tary values must be associated with years of life ate clinical outcome, such as cost per percent
lost and morbidity due to disease and with change in blood cholesterol level. These results
years of life gained and morbidity avoided due generally are reported as a ratio of costs to clin-
to intervention. ical benefits, with costs measured in monetary
terms but with benefits measured in the units
Cost-Effectiveness Analysis of the relevant outcome measure (for example,
Cost-effectiveness analysis provides an alter- dollars per year of life saved).
native approach that avoids the dilemma of When several outcomes result from a medi-
assessing the monetary value of health cal intervention (e.g. the prevention of both
death and disability), cost-effectiveness analy- is a set of procedures in which the results of a
sis may consider these two outcomes together study are recalculated using alternate values
only if a common measure of outcome can be for some of the study’s variables in order to test
developed. Frequently, analysts combine dif- the sensitivity of the conclusions to these
ferent categories of clinical outcomes accord- altered specifications. Such an analysis can
ing to their desirability, assigning a weighted yield several important results by demonstrat-
utility, or value, to the overall treatment out- ing the independence or dependence of a
come. A utility weight is a measure of the result on particular assumptions, establishing
patient’s preferences for his/her health state or the minimum or maximum values of a variable
for the outcome of an intervention. The that would be required to affect a recommen-
comparison of costs and utilities sometimes is dation to adopt or reject a program, and identi-
referred to as cost–utility analysis, with the fying clinical or economic uncertainties that
denominator expressed as quality-adjusted life require additional research. In general, sensi-
years (QALYs). tivity analyses are performed on variables that
have a significant effect on the study’s conclu-
Cost Identification Analysis sions but for which values are uncertain.
An even less complex approach than cost–ben-
efit or cost-effectiveness analysis would be
Types of Costs
simply to enumerate the costs involved in med-
ical care and to ignore the outcomes that result Another dimension of economic analysis of
from that care. This approach is known as cost clinical practice illustrated by Figure 18.2 is
identification analysis, by which the researcher the evaluation of costs of a therapy. Economists
can determine alternative ways of providing a consider three types of costs: direct, productiv-
service. The analysis might be expressed in ity, and intangible.
terms of the cost per unit of service provided.
For example, a cost identification study might Direct Medical Costs
measure the cost of a course of antibiotic treat- The direct medical costs of care usually are
ment, but it would not calculate the clinical associated with monetary transactions and
outcomes (cost-effectiveness analysis) or the represent costs incurred during the provision
value of the outcomes in units of currency of care. Examples of direct medical costs
(cost–benefit analysis). Cost identification include payments for purchasing a pharma-
studies, which include comparisons among ceutical product, payments for physicians’
different treatments based upon their costs fees, salaries of allied health professionals, or
alone, are appropriate only if treatment out- purchases of diagnostic tests. Because the
comes or benefits are equivalent among the charge for medical care may not accurately
therapies being evaluated. reflect the resources consumed, accounting or
statistical techniques may be needed to deter-
Sensitivity Analysis mine direct costs.
Most cost–benefit and cost-effectiveness stud-
ies require large amounts of data that may vary Direct Nonmedical Costs
in reliability and validity, and could affect the Monetary transactions undertaken as a result
overall results of the study. This is especially of illness or healthcare to detect, prevent, or
the case when models are developed for the treat disease are not limited to direct medical
economic analysis using secondary data costs. There is another type of cost that is often
sources, when data collection is performed ret- overlooked: direct nonmedical costs. These
rospectively, or when critical data elements are costs are incurred because of illness or the
unmeasured or unknown. Sensitivity analysis need to seek medical care. They include the
cost of transportation to the hospital or physi- and benefits can be calculated with respect to
cian’s office, the cost of special clothing needed society’s, the patient’s, the payer’s, and the pro-
because of the illness, the cost of hotel stays for vider’s points of view. A study’s perspective
receiving medical treatment at a distant medi- determines how costs and benefits are meas-
cal facility, and the cost of special housing (e.g. ured, and the economist’s strict definition of
modification of a home to accommodate an ill costs (the consumption of a resource that
individual). Direct nonmedical costs, which could otherwise be used for another purpose)
are generally paid out of pocket by patients and no longer may be appropriate when perspec-
their families, are just as much direct costs as tives different from that of society as a whole
are expenses that are more usually covered by are used. For example, a hospital’s cost of pro-
third-party insurance plans. viding a service may be less than its charge.
From the hospital’s perspective, then, the
Productivity Costs charge could be an overstatement of the
In contrast to direct costs, productivity costs, resources consumed for some services.
sometimes referred to as indirect costs, do not However, if the patient has to pay the full
stem from transactions for goods or services. charge, it is an accurate reflection of the cost of
Instead, they represent the cost of morbidity the service to the patient. Alternatively, if the
(e.g. time lost from work) or mortality (e.g. pre- hospital decreases its costs by discharging
mature death leading to removal from the patients early, the hospital’s costs may decrease
workforce). They are costs because they repre- but patients’ costs may increase because of the
sent the loss of opportunities to use a valuable need for increased outpatient expenses that are
resource, a life, in alternative ways. A variety of not covered by their health insurance plan.
techniques are used to estimate productivity Because costs will differ depending on the
costs of illness or healthcare. Sometimes, as perspective, the economic impact of an inter-
with varicella vaccination, the productivity vention will be different from different per-
costs of an illness are substantially greater spectives. To make comparisons of the
than the direct costs of the illness. economic impact across different interven-
tions, it is important for all economic analyses
Intangible Costs to adopt a similar perspective. It has been rec-
Intangible costs are those of pain, suffering, ommended that, as a base case, all analyses
and grief. These costs result from medical ill- adopt a societal perspective.
ness itself and from the services used to treat In summary, economic analysis of medical
the illness. They are difficult to measure as part technology or medical care evaluates a medical
of a pharmacoeconomic study, though they are service by comparing its dollar cost with its
clearly considered by clinicians and patients in dollar benefit (cost–benefit), by measuring its
considering potential alternative treatments. dollar cost in relation to its outcomes (cost-
Although investigators are developing ways to effectiveness), or simply by tabulating the costs
measure intangible costs – such as willingness- involved (cost identification). Direct costs are
to-pay analysis whereby patients are asked to generated as services are provided. In addition,
place monetary values on intangible costs – at productivity costs should be considered, espe-
present these costs are often omitted in clinical cially in determining the benefit of a service
economics research. that decreases morbidity or mortality. Finally,
the perspective of the study determines the
costs and benefits that will be quantified in
Perspective of Analysis
the analysis, and sensitivity analyses test the
The third axis in Figure 18.2 is the perspective effects of changes in variable specifications for
of an economic analysis of medical care. Costs estimated measures on the results of the study.
Using Economic Data product on the shelf tastes better but costs
more. Should we buy this product? The first
Health economics helps to understand both
question is a budget question – do we have the
the supply and demand for pharmaceutical
money to make this purchase? We have not said
products, while pharmacoeconomics provides
anything about prices but obviously, this is an
insight into the value of products to patient,
important part of our consideration. If the new
payers, and the marketplace. With all of these
product is enormously more expensive (it is
elegant data, the next challenge for analysts
hand crafted in small batches and infused with
and policy makers is to relate the results of the
gold and sold in a crystal decanter), we may not
economic analysis to purchase decisions for
have the money to afford the new product and
pharmaceutical products. Should a patient,
so the question becomes moot. This is a budget
hospital, or payer (public or private) make a
constraint. This constraint does not have to be
decision to approve payments for a therapy (for
so extreme; we can make a budget of $100 for a
example, by adding the product to an approved
grocery list, and hold ourselves to meeting our
drug list or formulary)?
budget in our shopping trip. With this budget,
Generally, if a product is thought to add clin-
even a modestly priced new item may not meet
ical benefit and save money, it is an easy deci-
our budget constraint.
sion to add the therapy. These types of products
are described as cost-saving or dominant (see Of course, we may decide we have some
Figure 18.3). For example, vaccines sometimes room in our budget to increase spending at the
fall into this category, as do generic drugs (in supermarket. For simplicity, let’s assume we
comparison with brand-name products). If a have two choices to consider for our increased
product worsens clinical outcomes but raises spending: products A and B. Product A is the
cost, this is also usually an easy decision to not tastier version of one of our shopping staples,
add the therapy. These types of products are while product B is a new item that a friend rec-
described as dominated but are much less ommended. Product A costs $5.00 more than
common. The major challenges in making for- our usual item, and product B costs $5.00, so
mulary decisions generally relate to therapies our budget would now be $105 if either is added
that add clinical benefit at additional cost. This to the shopping list (a 5% increase in cost). How
analysis requires further discussion on how would you choose between these two? This is a
this decision can be approached. cost-effectiveness question. Given the increase
Outside healthcare environments, we make in cost, which product would provide more
these types of decisions frequently. The new value to you as a consumer - enhanced flavor
Cost
Increase Decrease
Cost-effective
Improve (amount of benefit added Cost-saving or dominant
per additional cost)
Clinical
outcomes
Cost-effective
Worsen Dominated (amount of money saved per
reduction in benefit)
Figure 18.3 Cost and clinical outcomes.

from A or the novelty of B? We make these remiums is interesting. We have called this
p
types of decisions all the time, and the answer approach the “benefit pool” perspective. It sug-
greatly depends on individual taste preferences. gests that we can calculate how much addi-
So, in this simple case, cost is transparent and tional premium we would all have to contribute
value is based on individual preferences. for us to have an increase in our pharmaceuti-
Now, back to the pharmaceutical market. cal budget. The budget model looks at the issue
Budget constraints are built into healthcare from the insurer perspective, while the benefit
spending. National health insurance programs pool perspective looks at the same issue from
often have a fixed allocation from government the perspective of everyone buying insurance
for annual spending. In private health insur- (or paying taxes). With an increase in pre-
ance markets, health insurers estimate premi- mium, we would still have a budget constraint,
ums for the coming year before selling policies but one that allows for growth in pharmacy
during open-enrollment season, or as much as spending.
18 months in advance of actual spending. So, With the additional resources, we would
the introduction of a new product that adds need to develop a process for increasing our
cost can face real budget constraints depend- pharmacy spending. In our cash payment
ing on the potential magnitude of the spending shopping model, the willingness to allocate
increase. Sofosbuvir, a drug used to treat hepa- additional resources to our budget was based
titis C, had a list price of $84 000 in the US and on consumers’ individual perceptions of value.
a potential market size of 3.2 million people In healthcare, we do not pay for our medicines
when it was first launched. This would require directly, so the organization administering the
a staggering budget of $269 billion to treat eve- benefit pool (which could be a public or private
ryone with the infection. So, while the clinical payer) needs to make this determination. Here,
potential of this therapy was tremendous, the cost-effectiveness analysis can be used to assess
budget constraint resulted in policies to limit the relative value of additional investments in
its adoption. different therapies. From here, one can
As with the shopping example, budget con- consider the relative value across products and
straints can be absolute. For example, the fund the product that is the most economically
Medicaid program is jointly funded by federal attractive (the lowest cost-effectiveness ratio)
and state governments, and states are not first. In this way, we will ensure that incremen-
allowed to run budget deficits by law. Thus, if tal spending is for the product delivering the
increased spending on a new therapy would most value. You can continue adding therapies
require an increase in outlay by the Medicaid in this way until all of your resources are
program, states may be forced to not offer the allocated.
therapy, or to cut back in other areas of spend- Another way to use a cost-effectiveness ratio
ing to stay within their budget. Going forward, is to set a criterion of what represents good
if states wanted to add to their Medicaid spend- value for money, or what is “economically
ing, they would need to raise revenues (taxes) attractive.” In the US, dialysis care has long
to support this increased spending, or find been used to provide a benchmark of good
other parts of the budget to cut (education, for value for money. Dialysis was added to the
example). Medicare program in 1972, after consideration
Again, as with the shopping example, budget of the cost of care for patients with end-stage
constraints may not be absolute. Imagine that renal disease. As a result, we have an example
a health insurance company calculated its pre- of a clinical program where Congress made an
miums to incorporate new spending on drugs explicit decision to add a benefit to Medicare,
to be introduced in the coming year. Thinking one that added cost but that also extended life
about the budget from the perspective of expectancy for beneficiaries who need the
s ervice. Since patients on dialysis can cost lower cost-effectiveness ratio more likely to be
$50 000 per annum, the benchmark for value recommended by NICE.
as reflected in Congressional approval of this Returning to the benefit pool perspective can
service was seen to be $50 000 per year of life be another way of looking at this question.
gained. Since dialysis requires treatment three This analysis looks at the impact on the health
times a week for several hours at a time, the insurance premium resulting from the addi-
benchmark was thought to be even higher tion of a new product to the formulary. For
when considering quality-adjusted survival. example, the addition of PCSK-9 inhibitors
Does that mean that we can add therapies to (used to lower cholesterol), which was initially
the formulary that offer good value? The priced at more than $14 000 annually per
answer is, “it depends.” Again, while cost- patient, was calculated to add $140 to the pre-
effectiveness analysis does a good job assessing mium for everyone in the insurance pool under
the relative value of different therapies, the modest adoption assumptions. This perspec-
ratio itself is not tied to a budget impact or pre- tive can be generalized to the consideration of
mium. In other words, a product that is not specialty pharmaceutical products more
good value for money for a rare condition broadly (Figure 18.4). Here, we can see that
would have a relatively modest budget impact, health insurance premiums increase $250 for
while a drug that is good value for money for a every 0.25% of the population that receives a
common condition could have a significant $100 000 drug for any indication. The relation-
impact on budgets. The value of a product can ship of access to innovation and affordability is
also change by indication. For example, since an area of ongoing debate.
patients with known heart disease have higher Finally, we have the quadrant where the
risk for cardiovascular events than patients products save money but at the expense of
without heart disease, secondary prevention worse clinical outcomes. Actually, the use of
can provide more value for money than pri- therapies that meet this criterion is relatively
mary prevention. To date, efforts such as pric- common, as long as the amount of money
ing by indication have been challenging to saved is large relative to the loss of health ben-
implement. efits (a cost-effectiveness ratio of savings
Rather than providing an absolute recom- related to benefits lost where a higher number
mendation, the UK has an implicit relative is the most economically attractive). For exam-
framework for value, with products that have a ple, amoxicillin is recommended for first-line
5,000 150
Premium increase ($) Premium increase (%)

4,000 120
Premium increase (%)
Premium increase ($)
3,000 90
2,000 60
1,000 30
0 0
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00
Prevalence rate
Figure 18.4 Specialty Pharmaceuticals and Premiums. Source: Adapted from Hirsch et al. (2014).
therapy for otitis media, despite the high level and fellow author in previous editions of this
of resistance to this antibiotic. This is because work. The author also acknowledges Henry
of the low cost of the therapy, and the low like- Glick PhD, Harris Koffer Pharm D, Dan Polsky
lihood of significant complications of failure PhD, and Shelby Reed PhD for contributions to
of this initial treatment. previous versions of this chapter.
The Future Key Points
The cost of pharmaceutical products is an ●● The economics of pharmaceuticals are

important challenge for everyone involved in driven by the value of products, the health
the healthcare value chain, including patients, condition, and the type of payment model
physicians, payers, and government. While we for these products.
all desire innovation in healthcare, it is not ●● In general, programs that cost more and are
clear if it is possible to afford innovation at any more effective (and perhaps even some pro-
price. We have outlined an expanded version grams that cost less and have reduced clini-
of the study of the economics of pharmaceutical outcomes) are more likely be adopted if
cals to begin to flesh out a fuller discussion of both their incremental cost-effectiveness
this fascinating and complex subject. ratios and budget impact from a benefit pool
perspective fall within an acceptable range.
●● Results of an economic analysis are just one
Acknowledgements component of the decision-making process
regarding the adoption of an intervention;
The author dedicates this chapter to John M. social, legal, political, and ethical issues,
Eisenberg, MD (1946–2002), mentor, friend, among others, are also important.
Further Reading
Arrow, K.J. (1963). Uncertainty and the welfare Hirsch, B.R., Balu, S., and Schulman, K.A.
economics of medical care. AER 53 (5): 141–149. (2014). The impact of specialty
Bombardier, C. and Eisenberg, J. (1985). Looking pharmaceuticals as drivers of health care
into the crystal ball: can we estimate the costs. Health Affairs 33 (10): 1714–1720.
lifetime cost of rheumatoid arthritis? The Hlatky, M.A., Owens, D.K., and Sanders, G.D.
Journal of Rheumatology 12 (2): 201–204. (2006). Cost-effectiveness as an outcome in
DiMasi, J.A., Grabowski, H.G., and Hansen, randomized clinical trials. Clinical Trials 3 (6):
R.W. (2016). Innovation in the pharmaceutical 543–551.
industry: new estimates of R&D costs. Journal Kahneman, D. (2013). Thinking Fast and Slow.
of Health Economics 47: 20–33. New York: Farrar, Straus and Giroux.
Drummond, M.F., Sculpher, M.J., Torrance, G.W. Lakdawalla, D. and Sood, N. (2009). Innovation
et al. (2005). Methods for the Evaluation of and the welfare effects of public drug
Health Care Programs, 3e. New York, insurance. Journal of Public Economics 93
New York: Oxford Medical Publications. (3–4): 541–548.
Eisenberg, J.M. (1989). Clinical economics: a Outlook for Global Medicines Through 2021:
guide to the economic analysis of clinical Balancing Cost and Value. Quintiles IMS
practices. JAMA 262 (20): 2879–2886. Institute. (2016). http://www.imshealth.com/
Further Reading 321
en_US/thought-leadership/quintilesims- Reed, S.D., Anstrom, K.J., Bakhai, A. et al.

institute/reports/outlook_for_global_ (2005). Conducting economic evaluations
medicines_through_2021 (accessed 6 June 2017. alongside multinational clinical trials: toward
Pauly, M.V. (1968). The economics of moral a research consensus. American Heart Journal
Hazard: comment. AER 58 (3): 531–537. 149 (3): 434–443.
The Staffs Of Ranking Member Ron Wyden And Reed, S.D., Anstrom, K.J., Ludmer, J.A. et al.
Committee Member Charles E. Grassley. (2004). Cost-effectiveness of imatinib versus
(2015). The Price Of Sovaldi And Its Impact On interferon-alpha plus low-dose cytarabine for
The U.S. Health Care System, DECEMBER patients with newly diagnosed chronic-phase
(2015). US Senate, Committee of Finance. chronic myeloid leukemia. Cancer 101 (11):
https://www.finance.senate.gov/imo/media/ 2574–2583.
doc/1%20The%20Price%20of%20Sovaldi%20 Reed, S.D., Friedman, J.Y., Velazquez, E.J. et al.
and%20Its%20Impact%20on%20the%20 (2004). Multinational economic evaluation of
U.S.%20Health%20Care%20System%20 valsartan in patients with chronic heart
(Full%20Report).pdf (accessed 8 July 2017). failure: results from the Valsartan Heart
Ramsey, S., Willke, R., Briggs, A. et al. (2005). Failure Trial (Val-HeFT). American Heart
Good research practices for cost-effectiveness Journal 148 (1): 122–128.
analysis alongside clinical trials: the ISPOR Reed SD, Radeva JI, Weinfurt KP, McMurray JJV,
RCT-CEA Task Force report. Value in Health 8 Pfeffer MA, Velazquez EJ, et al. Resource use,
(5): 521–533. costs, and quality of life among patients in the
Rasiel, E.B., Weinfurt, K.P., and Schulman, multinational Valsartan in Acute Myocardial
K.A. (2005). Can prospect theory explain Infarction Trial.
risk-seeking behavior by terminally ill Weinstein, M.C. and Stason, W.B. (1982).
patients? Medical Decision Making 25 (6): Cost-effectiveness of coronary artery bypass
609–613. surgery. Circulation 66 (5): 56–68.
322
19
Patient Engagement and Patient Reported Outcomes

Esi M. Morgan1,2 and Adam C. Carle3,4,5
1
Seattle Children’s Hospital, Seattle, WA, USA
2
University of Washington, Seattle, WA, USA
3
Cincinanti Children’s Hospital Medical Center, Cincinnati, OH, USA
4
University of Cincinnati, College of Medicine, Cincinnati, OH, USA
5
University of Cincinnati, College of Arts and Sciences, Cincinnati, OH, USA
Introduction communicate information about their status

with respect to various domains of health,
Patient engagement reflects the effective par- including: symptoms, function, quality of
ticipation of patients in their own healthcare. life, health behaviors, and experience with
The Institute of Medicine’s 2001 report on care. PROs translate the patient experience
reforming and improving US healthcare iden- into data. The use of subjective report in
tified patient-centeredness as one of six guid- PROs recognizes that a patient’s health con-
ing aims for healthcare. Patient centeredness cerns may not be perceived by the clinician.
gives rise to the perspective that healthcare is Further, by failing to elicit direct patient
better considered a service rather than a prod- input, clinicians may prioritize topics less
uct. As such, achieving the best patient out- valued by patients. Incorporating PROs into
comes hinges on co-production. With clinical care has the potential to enrich clini-
co-production, the patient (service user) and cian patient communication, as well as pro-
clinician (service professional) are in a collabo- vide data to anchor clinical decisions and
rative relationship, sharing creation of the interventions.
health care service and responsibility for out- Advances in measurement theory and the
comes. This reframing of the patient-clinician use of item response theory (IRT) models in
relationship elevates the importance of meas- measure development have resulted in PROs
uring patient reported outcomes (PROs). that are shorter, more reliable, and more pre-
A PRO is a measure of a patient’s health cise than previous PROs. Nevertheless, the
based entirely on reports that come directly implementation, interpretation, use and eval-
from the patient without interpretation by a uation of PROs pose practical and methodo-
clinician or anyone else. PROs can be self- logical challenges. In this chapter, we explore
reported rating scales, symptoms reports, these challenges in the context of integration
questionnaires, or structured or unstructured of PROs into pharmacoepidemiology research,
interviews that directly record patients’ including clinical trials and observational
responses. PROs enable patients to directly studies.
Patient Reported Outcomes in Routine Car 323
atient Reported Outcomes

P atient Reported Outcomes
P
in Clinical Trials in Routine Care
The patient’s perspective is important during A current challenge to the use of PROs as out-
development of new medication or medical comes in observational studies is that PROs are
products. PROs serve a variety of functions in not routinely collected in clinical care.
new product development. PROs allow one to However, use of PROs in clinical care is starting
capture the patient’s perspective as a clinical to increase. Over the past decade, there has
trial outcome. They can help better understand been a growing body of published evidence on
the impact of a medical condition on the the benefits of incorporating PROs in clinical
patient in the realms of physical, mental, or care including both process and outcomes
social health and uncover (unmet) medical improvement. A number of technical solutions
needs to be addressed by a new medication. for full integration into the electronic health
PROs can also be used to enable estimation of record, used alongside the electronic health
the benefit to the patient of use of a medication record (EHR) or hybrid approaches, have been
during the clinical trial and, by extension, can described. User’s guides have been published
be used to characterize the expected benefit to with step by step instructions for health
the end-user in the clinical setting. Adverse systems considering adoption of PROs to
effects of treatment, or an estimate of risks of facilitate incorporation into clinical care.
use, may also be captured by PROs. The Food Published reports of integration of PROs
and Drug Administration (FDA), as part of the into clinical care have evolved over time
2012 reauthorization of the Prescription Drug from describing acceptability and feasibility,
Use Fee Act (PDUFA V), has been including to demonstration of improved care processes
patient input on the impact of chronic illness and communication, to preliminary evi-
to inform which aspects of illness would be dence of increased clinician satisfaction and
most meaningful to target from the patient even improved outcomes. When PROs are
perspective. collected and reviewed, patients perceive the
After clinical and PROs that are meaningful clinician to have increased awareness of
drug targets from the patient perspective are their symptoms, which otherwise might go
identified, it becomes important to be able to unrecognized and unaddressed. Clinicians
determine whether the drug/treatment/inter- may be better able to identify and diagnose
vention under investigation exerts a clinically patient conditions, including mental health
meaningful effect on the outcomes. Although concerns, if PRO results are routinely col-
a statistically significant change in outcome lected. Using PROs may result in increased
measure scores may be achieved, achieving clinician awareness of the impact of the
statistical significance may not correspond to a health condition on the patients’ health
clinically meaningful benefit. Determination related quality of life and facilitate patient-
that a change is of clinical meaningfulness clinician discussions. Communication,
rests on the idea that the outcome being meas- shared decision making, and collaborative
ured is of relevance and importance to the treatment planning is enhanced when PROs
patient and that the degree of change exerted are incorporated into patient and clinician
creates a clinically appreciable benefit. We conversations. Case Example 19.1 describes
briefly describe methods to address this later in the use of PROs to optimize treatment in
the chapter. clinical care.
324 19 Patient Engagement and Patient Reported Outcomes
Case Example 19.1 Use of Patient Reported Outcomes to Optimize Treatment in Clinical Care
Background and Outcomes Improvement Network)

●● Clinicians have used a treatment strategy registry (https://pr-coin.org).
in several chronic conditions to adjust
Results
medication to reach a treatment target,
Incorporating PROs into routine disease
known as “treat to target.” Examples of
assessment that guides decision making and
possible targets include a lab value
discussing this with the patients can improve
(HgbA1c, lipid), a clinical measure (blood
patient engagement and patient under-
pressure, weight), or a measure of disease
standing of how treatment adjustments are
activity (PROs, or composite measure
made. This could potentially improve adher-
included PRO).
ence to treatment regimens.
●● A “Treat-
to-Target” strategy to gain control
of disease relies on regular, standard Strengths
measurement of disease activity; setting a Use of structured data from clinical practice
target for disease control together with collected on an unbiased sample of patients
the patient; and shared decision making to (every clinic visit, every patient with a given
incorporate patient values and preferences diagnosis) is a resource for understanding
into choice of treatment. comparative effectiveness.
Limitations
Questions
●● There is need for clinician buy-in to collect-
What are some factors that support the use
ing and reviewing standard disease activity
of PROs in patient engagement?
metrics that include the patient perspec-
tives (via PROs), willingness to share treat-
Approach
ment decisions with the patients, and to
Structured data forms, similar to case report
follow standard treatment guidelines rather
forms, embedded in the electronic medical
than their own style of practice.
record enable routine, and regular assess-
●● Additional study is needed into down-
ment of patient disease activity at every clin-
stream effects on outcomes (disease activ-
ical visit. Disease activity scores can be
ity, adherence, etc.).
calculated and scored within the medical
record for review with the patient and com- Key points
pared to treatment target. If the patient is ●● Incorporation of PROs into measures of
not making timely progress toward the tar- disease activity used to inform treatment
get, treatment may be adjusted (this may decisions embeds the patient’s perspective
involve following a clinical decision support, into evaluation of treatment effectiveness
published guideline, or algorithm). Use of and review and use of results fosters
standard data elements for collection at patient engagement.
point of care of disease activity measures ●● Reliable clinical collection of psychometri-
enables data extraction and upload into reg- cally sound PROs; integration of scores as
istries or databases for analysis, coupled discrete, research quality data into the
with extraction of medications. This feeds medical record; and establishing MCIDs in
analysis of comparative effectiveness of PROs will contribute to more efficient, and
treatments. This approach is exemplified by rigorous observational and pharmacoepi-
the PR-COIN (Pediatric Rheumatology Care demiologic research.
Patient Reported Outcomes as Motivation to Develop New Therapeutic Strategie 325
Much of the literature on the effective use of For example, self-management support is a
PROs in clinical medicine has come from the key component of the Chronic Care Model.
fields of oncology and surgery. For example, a Patient engagement with informed, activated
recent compelling report from an oncology patients, is key to productive patient-clinician
randomized trial of PRO use during routine relationships, which affects clinical outcome
cancer treatment showed that integration of improvement. Successful self-management
PROs into the care of patients with metastatic support programs have the following elements:
cancer was associated with increased survival. (i) clinicians communicate the expectation for
This may have been a consequence of expe- the central role of the patient in managing
dited medical care team response to patient their condition, (ii) patients’ self-management
symptoms, particularly adverse symptoms, skills, confidence in management ability, bar-
and, subsequently, the ability to administer riers and supports are assessed regularly, (iii)
chemotherapy for a relatively longer duration. trained staff employ behavior change interven-
Payers recently have begun encouraging tions, (iv) patients co-develop with health care
PRO collection, including the use of financial professionals their own individualized treat-
incentives for tracking PRO data. For example, ment plans, and support is available on an
PROs are used to evaluate the impact and tra- ongoing basis if the patient needs it. Self-
jectory of improvement of surgical interven- management support is a patient-centered,
tions in the case of Medicare reimbursement iterative and ongoing process. Due to the fact
for elective joint replacement. The availability that the majority of chronic illness care occurs
of such data can enhance observational studies outside of the medical office in the interval
by providing prospectively collected informa- (days to months) between office visits, PRO
tion on PROs. assessment has the potential to become an
effective facilitator of in-between visit commu-
nication between patients and clinicians and
may allow patients to better manage their
atient Reported Outcomes
P health conditions.
as Motivation to Develop Further, since medical care at outpatient
New Therapeutic Strategies office visits occurs at relatively infrequent
intervals but health events occur on an ongo-
Therapeutically, incorporating PROs may con- ing basis, the trajectory of illness may not be
tribute to a more targeted conversation on adequately captured at clinic visits. As a result,
issues that concern the patient, which might medical decision making may not incorporate
otherwise go unaddressed. It may help to all relevant health information into treatment
quantify the discomfort a patient feels and sub- decisions. Imagine the situation where a
sequently trigger the use of a therapeutic inter- patient, “Patient A,” begins a new medication
vention. Information on disease control status “Drug A” and experiences rapid improvement
can focus the visit on topics of most relevance in signs and symptoms of disease, which grad-
and be a useful measure of disease burden. ually return to baseline over time. Patient B
Having pre-specified and agreed-upon PRO experiences gradual improvement following
score thresholds that trigger specific evidence- the initiation of “Drug B,” then rapid worsen-
based interventions facilitates action by the cli- ing before their condition returns to baseline
nician. Alternatively, aggregate patient by time of follow up visit. Both patients have
reported data on symptoms, function, quality roughly the same disease activity level at the
of life, or experience related to an intervention follow-up visit. In this scenario, information is
can be shared with an individual patient as lost about the relative lack of effectiveness and
part of a shared decision making discussion. negative impact Patient B experienced being
on Drug B, as more time was spent in a flare data collection. As the primary motivation for
state than with “Patient A” who took “Drug A.” PROs collection may be delivery of optimal
In the latter example, technology can enable patient centered care and better health out-
the capture and transmission of both patients’ comes, with the ability to use the data for
unique experience to the clinical team. Data research as secondary, the clinical team – and
can be collected at home, on computers, patients – must be aligned in perceiving the
mobile devices, or technology enabled patient value of quality PRO data collection. To be most
generated data (e.g. physiologic monitors), and useful clinically, and to allow valid inferences in
then shared with the clinical team. Data collec- research, PROs must be collected routinely (reli-
tion from patient wearable devices requires lit- ably at regular intervals), uniformly (in a consist-
tle or no effort from the patient perspective. ent manner), completely (able to be scored), and
This opens the possibility for patient data from correctly (patients understand the questions and
in-between patient visits to inform action- answers reflect their health status).
oriented interventions. Patients who transmit Once a decision is made to incorporate PROs
data will have the expectation that results are into clinical care, there must be a means to
reviewed, thus workflows will need to be opti- achieve reliable review of the results by the
mized to allow for reliability of review, and clinical team. If clinicians do not understand
prompt response if deterioration in status or the PRO measure, the way it is scored, do not
other change is identified. feel it relevant to their specialty, or feel there is
no intervention or action that can impact the
PRO, they may not be inclined to review it. The
linical Problems to be Addressed
C solution may be cultural and social, and require
by Pharmacoepidemiologic a clinical champion to motivate the colleagues
Research and gain agreement on selection of PROs (more
below). On the other hand, barriers may be
PROs are directly relevant to pharmacoepide- technical, related to location and ease of view-
miology research. They capture information ing of the reports in the medical record. This
central to evaluating the benefits and adverse may require an information technology solu-
effects of medication that may elude physical tion. Another solution may be to create a work-
exam or lab testing. Routine and systematic flow in which discussion of PRO completion is
collection of PROs as part of clinical care will an expected part of the clinic visit and is
facilitate better understanding the impact of included into process of care quality measures.
medications on symptoms and quality of life as There are significant consequences of clini-
part of post market surveillance. While PRO cians not reviewing PROs completed by
collection has benefit to pharmacoepidemiol- patients. When clinicians do not review PROs it
ogy research, the rationale behind incorpora- becomes a threat to co-production of care, risks
tion of the patient voice into clinical care via the patient feeling their time and input was not
the use of PROs is to identify unmet needs of valued and may result in a decreased willing-
patients and to help identify gaps in healthcare ness for PRO completion at subsequent visits. It
and, subsequently, respond to these unmet is also a missed opportunity to check the relia-
needs and/or gaps. bility of the system for PRO administration.
Ensuring PRO Completion PRO Selection, Score Interpretation

and Results Review and Interventions
It is essential that clinical processes are in place Instrument selection is critical for clinical and
so that research quality data results from clinical research applications to ensure the measure is
clinically relevant, addresses an outcome of treatment or lifestyle modification, then track

interest, and is brief and practical to adminis- changes in symptoms to evaluate for possible
ter. Fundamentally, any measure selected for effectiveness. By keeping daily journals or
use should be reliable, valid, and responsive. other means of annotation, patients can see if
These and other considerations such as ease of other triggers or environmental factors may
understanding, ability to foster patient-clinician have led to changes in their health status. This
communication, and value in identifying structured type of intervention allows study of
unmet needs should be reviewed with repre- the impact of medications with relatively rapid
sentative patients and clinicians for input and onset and dissipation of effects (e.g. pain medi-
agreement on importance. In order to be action- cations). In addition, one can study whether
able, scores on measures for a specific health patients who are given access to their outcome
condition should be known, including the nor- data become more activated and engaged in
mal range versus when to intervene for an their health care.
abnormal score. There is need for consensus
and standard setting processes on threshold
Barriers to Measuring PROs
scores to trigger interventions. In order to track
in Clinical Practice and Using PROs
longitudinal change, it is helpful if the minimal
to Guide Interventions
clinically important difference (MCID) for
improvement or worsening has been deter- Barriers to PRO use are varied and start with
mined (discussed below). Ensuring these scores garnering institutional resources to obtain, store
and thresholds for action are known, and iden- and update electronic data collection tools.
tifying evidence based interventions or, where Designing integration of PROs into clinical
such evidence does not exist, gaining consen- workflow is a necessary step, which first
sus agreement among practice clinicians and requires gaining clinician consensus on the pur-
patient stakeholders on recommended inter- pose and value of adding PRO measurement
ventions, will facilitate clinicians and patients and review to the clinical encounter. Clinicians
moving from simple review and discussion of must understand the ways the use of structured
PROs to taking effective action and co- PROs adds value over simply asking the patient
producing treatment plans based in part on about how they feel. Front line staff training on
PRO scores, which are meaningful to them. importance of PRO collection and distribution
of questionnaires or devices of collection is
required for reliability of PRO collection.
Patient Engagement
Clinician training is required on how to use
and Individualized Assessment
PROs to facilitate high quality communication
and Treatment Plans
with patients. This requires training on how to
Mobile health tracking systems offer the pos- interpret PRO results and on how to communi-
sibility of customized measurement (choice of cate about the results with patients. The pro-
PROs, timing of administration), and use of cess could be facilitated by orientating the
individualized measures, which may be of patients themselves to use of PROs and their
great importance in pharmacoepidemiology role in their clinical care. For PROs to serve the
outcome studies. In addition, use of PROs for role in facilitating co-production of care may,
in-between clinic visit care for continuous in some circumstances, take reframing of the
monitoring and data feedback, with custom- patient-clinician relationship, a complex
ized reporting and means to detect signals of endeavor in culture and behavior change.
health status change in a personalized system Buy-in may become easier to obtain as more
lends itself to N-of-1 trials. In this type of study convincing data become available on PRO
an individual can make planned changes to integration into care resulting in improved
outcomes, as barriers to logistics of PRO this challenge is to use composite indices,

collection are lowered, as graphical displays though this is not always a valid approach.
become more intuitive, and as interpretation Composites aggregate multiple scores into a
of data and action steps become more familiar single summary score. A second challenge is
and supported with decision aids. Clinicians determining whether change across time is
may have higher interest in use of PROs if they clinically meaningful. Although there is no
receive feedback on their own patients’ scores, current consensus on the single best approach,
relating to outcomes or experience with care using clinical anchors to determine the small-
(satisfaction) and recognize they can take steps est difference in a score that would prompt a
to improve performance. change in patient management is a common
As use of PROs in clinical trials and pharma- solution. Below, we provide further detail
coepidemiology research becomes more preva- regarding these two challenges and current
lent, PRO endpoints may become more widely approaches to handling them.
accepted goals to measure and monitor treat-
ment efficacy. This may serve as positive rein-
Discordance in Perspectives
forcement to PRO use for monitoring treatment
between Patients, Clinicians
effectiveness in clinical practice settings.
and Researchers
Discrepancies may occur between clinicians’
ethodologic Problems
M and patients’ perceptions of disease activity
to be Solved by level and whether there has been improve-
Pharmacoepidemiologic ment or deterioration in the condition. There
Research may also be lack of concordance between com-
posite measures of disease activity used to
Just as patient engagement and PROs inform assess efficacy in clinical trials and measures
the discussion in the clinical practice setting, that use PROs.
they similarly play an increasing role in the Composite indices are helpful to reduce the
research setting. For example, including presentation of information from multiple
patient relevant outcomes into clinical trials measures into a single summary score. This
has become a priority of the FDA and fostering approach is most valid when the measures
patient engagement in all stages of research included in an index are highly correlated.
has been the genesis of the Patient Centered When a measure does not correlate, or track
Outcomes Research Institute (PCORI) and its well with others, to include it in a composite
significant funding portfolio. However, there index would result in lost or obscured informa-
are methodologic challenges that accompany tion. Studies have found that PROs scores do
PROs in pharmacoepidemiologic research and not always correlate strongly with the compos-
clinical practice. These issues include discord- ite indices. When PROs are relatively inde-
ance between raters and measurement of pendent predictors of treatment response,
within person change over time. their results should be reported separately
rather than included in a composite score. In
one study, composite indices were better at
urrently Available
C detecting flare states though worse than PROs
Solutions at describing low disease activity states.
Composite measures upon which clinicians
Given the subjective nature of PROs, clini- base their assessment of improvement or dete-
cians’, patients’, and caregivers’ ratings may rioration may not necessarily include key
not agree. One currently available solution to aspects of the disease that matter to the patient.
For instance, in the DAS-28 (disease activity prompt a change in patient management. In
score) assessment of rheumatoid arthritis, practice, the terms MID and MCID are some-
28 joints are assessed for tenderness or swell- times used interchangeably, but statistically
ing, but this count does not include the feet or derived and clinically anchored estimates need
ankles. Therefore, if patients have foot involve- not converge. Anchor-based methods use exter-
ment, which can be very painful, it is possible nal indicators considered to be clinically rele-
they may not be satisfied with the degree of vant to the PRO, such as clinical measures (lab
improvement noted on the DAS-28 because it tests, clinician ratings) or patient measures
excludes an important element of their disease (global rating of change) and place subjects on
experience. Using DAS-28 as an outcome in a continuum based on the size of change in the
pharmacoepidemiology studies could there- anchor (large negative change, small negative
fore miss outcomes important to patients and change, no change, small positive change,
result in biased measures of association. large positive change). Ideally, multiple rele-
Further, composite disease activity measures vant anchors should be used, across multiple
may not always translate to decision making samples to confirm responsiveness of the PRO
based on the experience of an individual measure. Another anchor-based technique to
patient. estimate MID employs use of receiver operat-
ing characteristic (ROC) curves to evaluate
group-level criteria for improvement or wors-
Measuring within Person Change
ening of clinical status. Distribution-based
Understanding how to estimate clinically methods are more strictly based on statistics vs.
meaningful change using PROs is important to clinical anchors. The distribution-based
be able to determine the effectiveness of a approach uses scores from a sample to express
treatment or intervention. Unfortunately, the effect in terms of standard-deviation units
determining clinically meaningful change is or standard error of measurement.
complex. There is no consensus on the best Bookmarking and scale judgement of IRT-
approach, and the topic represents an area of based measures: Alternative approaches to
active methodological research. Although measuring meaningful within-patient change
there are statistical approaches to analyzing have been developed based on techniques from
changes in PRO scores over time, the detection the field of educational testing. This approach
of a statistically significant difference may not is applicable to PRO measures developed using
reflect a meaningful clinical difference. There IRT. The general approach is to develop clini-
may be differences in defining a meaningful cal vignettes representing a continuum of IRT-
clinical difference depending on the respond- based scores, present these vignettes to a repre-
ent (e.g. patient, caregiver, or clinician). It may sentative panel of stakeholders (e.g. patients,
depend on the health condition being studied caregivers, clinicians), and have the panel
and whether a patient is experiencing improve- identify thresholds between scores (delineated
ment or deterioration in health. Although by the vignettes), where they would place a
there is not a uniform consensus, we briefly “bookmark” separating different levels of
review some methods for determining mean- severity. Such exercises help to identify clini-
ingful change. cally meaningful cut-points between scores.
Statistical methods: The minimal important Similar qualitative work with panelists can be
difference (MID) represents the smallest used to identify minimal clinically meaningful
change in scores that could be determined differences by presenting PRO items to stake-
important. The MCID is determined based on holders and asking to note how much response
clinical anchors, and is generally regarded as to an item (or items on a scale) would need to
the smallest difference in score that would change for a change in status to be considered
clinically meaningful. Another approach, the training problem-based and experiential with
“scale-judgement” method, entails having video examples and case studies, and includ-
raters compare pre-filled IRT-based PRO ques- ing relevant treatment decision aids and deci-
tionnaires (for example, considering pre- and sion support tools in training.
post- an intervention) and indicating whether Considerations for the display and commu-
or not the person who completed the question- nication of PRO scores have also been pub-
naires had experienced an important differ- lished. Display format preferences tend to vary
ence. These types of approaches require by audience characteristics such as age, educa-
qualitative work with patient and other stake- tion, and role (clinician vs patient). Some stud-
holders across different patient populations ies have identified preferences for line graphs,
(e.g. age, demographics, health conditions) others for bar graphs. Patients tend to prefer
and directions of change in PROs. simpler formats, while clinicians prefer more
Change in perspective. Another factor that data. Directionality of data has been shown to
may complicate the assessment of within- matter, with better health being portrayed as
person change over time is when a person’s higher on the chart, and including lines indi-
perception and valuation of the domain being cating threshold values for normal versus
measured change over time. This phenome- abnormal found to be helpful.
non, termed response shift, means that, across
time, an individual may give the same response
on a PRO (e.g. “Sometimes pain bothers me”) The Future
even though their underlying health status has
changed. Response shift may occur because a Patient engagement in research, advances in
patient’s subjective measurement scale may PRO measurement, and recognition of the
change (e.g. more of the symptom is required importance of garnering direct patient input
before a patient describes it as “sometimes” will result in increased inclusion of the patient
occurring), a patient’s values may change (e.g. voice in the calculus of medication efficacy in
a symptom becomes more important over time clinical studies. There is growing evidence that
leading to different responses to questions electronic PROs may increase the quality of
about the symptom over time), and/or a patient care of processes and clinical outcomes and
may reconceptualize the construct entirely. PROs may become a key part of improving
Statistical and qualitative methods exist to quality of care. PROs may also become increas-
assess response shift. ingly used in comparative effectiveness
research. As more is understood about PRO
development, effective use, interpretation and
Selection of Patient Reported potential applications, the use and new use-
Outcomes and Implementation cases for PROs can be expected to continue to
into Practice grow.
Publications have begun to offer guidance on There is increasing interest in using PRO
training clinicians in use of PROs. data from clinical settings and captured in
Recommendations include: eliciting local bar- electronic health records as structured out-
riers and concerns to address during the train- comes assessment for inclusion in comparative
ing (such as how to deal with patient effectiveness research. This could be a power-
symptoms/concerns outside the specialty, con- ful data source when combined with other
cern about visit time constraints, how to inter- sources of electronic data (see Part II). PROs in
pret results), inclusion of the stakeholders in EHRs may be particular useful when there are
PRO selection and format for presentation, no standardized outcomes assessments pro-
keeping the training relatively brief, making vided by clinicians in the clinical note. The
Further Readin 331
complexities of analysis and interpretation of ●● PROs are measures of health from the per-
longitudinal PRO data require continued study spective of the patient and can come in the
to best leverage such data to make valid infer- form of self-reported symptoms, rating
ences. Cross-cutting PRO measures that could scales, questionnaires or interviews.
be used across conditions and contexts may ●● PROs measure health domains such as
confer advantages such as anticipating in clini- symptoms, function, quality of life and expe-
cal trials the expected outcomes in clinical rience with care.
practice. Such PROs, when collected in a clini- ●● PROs can be used as clinical trial outcomes
cal setting or with technology to support in- to estimate treatment benefits, adverse
between visit data collection, could be used effects, and impacts on health related quality
such that clinical data registries could be com- of life that are not reflected in physical exam
bined with administrative claims data to sup- or lab data.
port comparative effectiveness research. ●● MCID is the change in a PRO score that cor-
Another area of future development is the responds to a substantive impact on the
use of wearable devices and PROs. Coupling patient, whereas a statistically significant
physiologic data with PRO data will aid efforts change in PRO score may not be perceived
to determine clinically relevant change in PRO clinically.
scores. Further, there are open research ques- ●● PROs used in a clinic setting, if reviewed
tions related to the analysis and interpretation with patient and used to inform care deci-
of measures used longitudinally. There are sions, can facilitate patient engagement in
interesting and exciting case examples, and it care.
remains to be seen which model will be scala- ●● Reliable clinical collection of psychometri-
ble and generalizable. Ideally, the culture of cally sound PROs and integration of scores
co-production, patient engagement and self- as discrete, research quality data into the
management will continue to take root and medical record can support EHR based
support the shift toward PRO measurement for observational and pharmacoepidemiologic
meaningful application, evidence-generation research.
and shared decision making. ●● PROs have the potential to support remote
and asynchronous patient monitoring,
Key Points when direct observation of the patient isn’t
possible (e.g. telemedicine, between visit
●● Patient engagement reflects the effective par- measurement).
ticipation of patients in their own
healthcare.
Further Reading
Bantug, E.T., Coles, T., Smith, K.C., et al. (2016). Bantug, E.T., Coles, T., Smith, K.C., et al. (2016).
Graphical displays of patient-reported Symptom monitoring with patient-reported
outcomes (PRO) for use in clinical practice: outcomes during routine cancer treatment: a
what makes a pro picture worth a thousand randomized controlled trial. J. Clin. Oncol. 34
words? Patient Educ. Couns. 99 (4): 483–490. (6): 557–565.
Basch, E. (2017). Patient-reported outcomes - Cella, D., Gershon, R., Lai, J., et al. (2007). The
harnessing Patients’ voices to improve clinical future of outcomes measurement: item
care. N. Engl. J. Med. 376 (2): 105–108. banking, tailored short-forms, and
computerized adaptive assessment. Qual. Life calibration of health-related quality of life

Res. 16 (Suppl 1): 133–141. item banks: plans for the Patient-Reported
Detmar, S.B., Muller, M.J., Schornagel, J.H., et al. Outcomes Measurement Information
(2002). Health-related quality-of-life System (PROMIS). Med. Care 45 (5 Suppl 1):
assessments and patient-physician S22–S31.
communication: a randomized controlled Revicki, D., Hays, R.D., Cella, D., et al. (2008).
trial. JAMA 288 (23): 3027–3034. Recommended methods for determining
Fung, C.H. and Hays, R.D. (2008). Prospects and responsiveness and minimally important
challenges in using patient-reported outcomes differences for patient-reported outcomes. J.
in clinical practice. Qual. Life Res. 17 (10): Clin. Epidemiol. 61 (2): 102–109.
1297–1302. Rotenstein, L.S., Huckman, R.S., and Wagle,
Greenhalgh, J. and Meadows, K. (1999). The N.W. (2017). Making patients and doctors
effectiveness of the use of patient-based happier -the potential of patient-reported
measures of health in routine practice in outcomes. N. Engl. J. Med. 377 (14):
improving the process and outcomes of 1309–1312.
patient care: a literature review. J. Eval. Clin. Santana, M.J., Haverman, L., Absolom, K., et al.
Pract. 5 (4): 401–416. (2015). Training clinicians in how to use
Institute of Medicine (2001). Crossing the Quality patient-reported outcome measures in routine
Chasm: A New Health System for the 21st clinical practice. Qual. Life Res. 24 (7):
Century. Washington, DC: National 1707–1718.
Academies Press. Snyder, C.F., Smith, K.C., Bantug, E.T., et al.
Jayadevappa, R., Cook, R., and Chhatre, S. (2017). What do these scores mean?
(2017). Minimal important differences to infer Presenting patient-reported outcomes
changes in health-related quality of life -a data to patients and clinicians to improve
systematic review. J. Clin. Epidemiol. 89: interpretability. Cancer 123 (10):
188–198. 1848–1859.
Lavallee, D.C., Chenok, K.E., Love, R.M., et al. Thissen, D., Liu, Y., Magnus, B., et al. (2016).
(2016). Incorporating patient-reported Estimating minimally important difference
outcomes into health care to engage patients (MID) in PROMIS pediatric measures using
and enhance care. Health Aff. (Millwood) 35 the scale-judgment method. Qual. Life Res. 25
(4): 575–582. (1): 13–23.
Morgan, E.M., Mara, C.A., Huang, B., et al. User’s Guide to Implementing Patient-Reported
(2017). Establishing clinical meaning and Outcomes Assessment in Clinical Practice.
defining important differences for patient- (2015). International Society for Quality of
reported outcomes measurement information Life Research (prepared by L. Aaronson, T.
system (PROMIS(R)) measures in juvenile Elliott, J. Greenhalgh, et al.). https://www.
idiopathic arthritis using standard setting with isoqol.org/wp-content/uploads/2019/
patients, parents, and providers. Qual. Life 09/2015UsersGuide-Version2.pdf
Res. 26 (3): 565–586. User’s Guide to Integrating Patient Reported
Patient Reported Outcomes (PROs) in Outcomes in Electronic Health Records.
Performance Measurement. (2012) Author: (2017) Prepared for the Patient Centered
National Quality Forum. https://www. Outcomes Research Institute (PCORI) by
qualityforum.org/Publications/2012/12/ Johns Hopkins University. (Eds. C. Snyder
Patient-Reported_Outcomes_in_ and A.W. Wu). http://www.pcori.org/
Performance_Measurement.aspx. document/users-guide-integrating-
Reeve, B.B., Hays, R.D., Bjorner, J.K., et al. patient-reported-outcomes-electronic-
(2007). Psychometric evaluation and health-records.
Further Readin 333
U. S. Department of Health and Human effectiveness research. J. Clin. Epidemiol.

Services, F.D.A., Center for Drug Evaluation 66 (8 Suppl): S12–S20.
and Research, Research Center for Biologics The National Institutes of Health (NIH)
Evaluation and, Center for Devices and supported the development and distribution
Radiological Health. Guidance for Industry. of various IRT-based PROs, including the
(2009). Patient Reported Outcome Measures: Patient Reported Outcomes Measurement
Use in Medical Product Development to Information System (PROMIS). A useful
Support Labeling Claims. reference for PROs developed and evaluated
Wu, A.W., Kharrazi, H., Boulware, L.E., et al. with NIH funding, including PROMIS is the
(2013). Measure once, cut twice--adding Health Measures website. Available at: http://
patient-reported outcome measures to the www.healthmeasures.net/explore-
electronic health record for comparative measurement-systems/promis
334
20
The Use of Meta-analysis in Pharmacoepidemiology

Brenda J. Crowe1, Stephen J.W. Evans2, H. Amy Xia3, and Jesse A. Berlin4
1
Eli Lilly and Company, Indianapolis, IN, USA
2
The London School of Hygiene and Tropical Medicine, London, UK
3
Amgen, Thousand Oaks, CA, USA
4
Johnson & Johnson, Titusville, NJ, USA
Introduction eta-analysis itself may be regarded as an

m
observational study.
Good science seeks to consider all the evidence There are dangers with meta-analysis of
related to a question of interest. The first step is non-randomized studies, because of inherent
assembling all the available evidence and the biases. Nevertheless, they can provide evi-
second is, if appropriate, to provide a numeri- dence beyond that from trials. The Cochrane
cal summary. While both are very important, Collaboration maintains systematic reviews
this chapter concentrates on statistical meth- and meta-analyses of trials publicly available.
ods for providing numerical summaries. Efficacy of treatments is best tested using ran-
We use the definition of meta-analysis from domization but studying harms may require
Working Group X of the Council for non-randomized studies to detect effects not
International Organizations of Medical seen in trials. Given the greater diversity of
Sciences (CIOMS): designs, heterogeneity of patients and wide
possibilities of bias in non-randomized stud-
The statistical combination of quantita- ies, there will be more disagreement than
tive evidence from two or more studies among trials.
to address common research questions, Increased precision is a motivation for meta-
where the analytical methods appropri- analysis of trials having small sample sizes,
ately take into account that the data are unable to provide convincing evidence of the
derived from multiple individual presence or absence of harms. In contrast,
studies. observational studies usually have dramati-
cally larger sample sizes so precision is not
Meta-analysis in medicine has mainly used often the focus, but uncontrolled and unknown
data from randomized trials (we will use “tri- bias and confounding can have a large impact.
als” to refer to randomized clinical trials), but The exploration of reasons for variability of
in pharmacoepidemiology, evidence from results across observational studies can become
observational studies is also relevant. A a main focus of the meta-analysis.
This chapter addresses conceptual and In drug development, data on harms are
methodological issues for meta-analysis, espe- often summarized in an “Integrated Summary
cially for observational studies. of Safety” or similar report just prior to submit-
ting an application for marketing authoriza-
tion. If some type of aggregate safety review,
C which could include formal data integration, is
Addressed by only done when all trials have been completed,
Pharmacoepidemiologic it is a missed opportunity to understand evolv-
Research ing safety. Crude pooling, summing numbers
with and without the harm, by treatment, then
The clinical problems to be addressed by phar- analyzing as if the data came from a single
macoepidemiologic research are usually esti- trial, can be misleading and biased. While the
mating efficacy (or effectiveness) and safety crude approach gives some information, doing
parameters that are unable to be addressed or a proper meta-analysis will always be more
have too much uncertainty in trials. informative.
Requirements around the regulatory assess- For regulatory purposes, periodic aggregate
ment for new therapies or new indications for safety reviews are extremely important. It is
existing therapies, considering overall benefit/ recommended though, that sponsors plan
harm balance, can be especially relevant. repeated, cumulative meta-analyses, with clear
Investigating harms using nonexperimental definitions of adverse events of interest, and to
studies is a major challenge because of con- the extent possible, use standardized data col-
founding (see Chapter 22). Also, rare events lection and study designs (at least from the
will not generally have standardized assess- perspective of safety data collection). During
ment or validation, which makes their evalua- development, ongoing review of blinded trials
tion difficult. Comparing results for a particular is helpful for understanding potential safety
drug on particular harms may then yield con- issues that may require additional data capture
flicting results because of different study or unblinding to determine if additional action
designs or different databases, leaving a confus- is required to protect patients and satisfy regu-
ing picture for clinicians and policy makers. latory requirements. Typically, determination
Meta-analysis using randomized studies may of causal adverse reactions begins after data
be better than relying on potentially biased are unblinded. Better understanding of the
non-randomized studies, but whether rand- safety profile, following a proactive approach
omized or non-randomized studies are used, during development, including periodic updat-
meta-analysis can also help to explain disagree- ing of cumulative meta-analyses, may identify
ments. Disagreements may arise from differ- potential harms earlier.
ences in endpoints, exposure, patient inclusion The exploration of subgroups of patients in
and exclusion criteria, protocols and study whom therapy may be more or less effective is
designs, analysis methods or other reasons a controversial question and, rather than
related to the susceptibility of the constituent focusing on individual trials, meta-analysis
studies to bias. For example, separating studies can be used. A pre-specified protocol is impor-
by whether or not there is possible recall bias in tant in this context.
drug exposure measurement allows assessment Evidence-based medicine (EBM) uses the
of whether recall bias is a problem for a particu- best evidence available in making decisions.
lar therapeutic question. For example, this Meta-analyses have been a key component of
applies with congenital anomalies where ascer- EBM but generally focus on placebo-controlled
tainment of exposure prior to birth and poten- trials, and thus head-to-head comparisons
tial recall bias is important. between therapeutic alternatives are often
336 20 The Use of Meta-analysis in Pharmacoepidemiology
unavailable. However, consumers, health care Susceptibility of the Original

providers and policy makers require compara- Studies to Bias
tive effectiveness evidence to make better
The most important aspect of quality is suscep-
informed decisions – each specific pharmaco-
tibility to bias, and guidelines such as the
logic treatment for a condition/disease should
Preferred Reporting Items for Systematic
be compared with others in terms of safety and
Reviews and Meta-analyses (PRISMA) reflect
efficacy.
this. The Cochrane Collaboration developed a
There are further techniques such as indirect
“Risk of Bias Tool” for assessment of the trials
comparisons and network meta-analyses (also
in a meta-analysis. The “ROBIS” tool assesses
known as mixed treatment comparisons) that
the risk of bias in a meta-analysis itself. A simi-
can combine evidence in a single analysis.
lar approach has been applied to “Risk of Bias
These techniques can allow ranking of treat-
in Non-randomized Studies – of Interventions,”
ments in terms of efficacy and safety and ena-
ROBINS-I.
ble better decisions because they are based on
There are many such tools and while no sin-
more data, but they also require assumptions
gle method of assessing bias is accepted by
that may not be testable.
everyone, those doing and interpreting meta-
analyses in pharmacoepidemiology should be
aware and assess potential bias using one of
M the methods.
to be Solved by Combining poorly done studies can produce
Pharmacoepidemiologic a summary result with misleading precision,
which may have undue credibility and should
Research
not be used as a basis for formulating clinical
or policy strategies. Judgment about suscepti-
The most important problem in studying
bility to bias of a study can be subtly influenced
medicines is obtaining the correct estimates
by its results, so excluding them using subjec-
of their effects on outcomes related to medi-
tive judgment can result in a different, poten-
cal health. Biases can arise in many areas,
tially serious, bias.
especially in non-randomized studies and
there is always statistical uncertainty in the
estimates of effects, especially if they involve
Combinability of Studies
rare outcomes. Meta-analysis should always
address issues of statistical uncertainty and Different studies have different designs, out-
may help with bias. It may be able to address comes, treatments and patients. Combining
the effect of applying different methods of extremely diverse studies can be nonsensical.
studying drug effects. Results from rand- Combining studies of breast cancer with those
omized and non-randomized data may give on coronary heart disease would obviously not
different insights. However meta-analysis be appropriate; the outcomes are not expected
itself has many methodologic issues. The to show the same effect of the same exposure.
problems relate to the process of combining However, it is not always simple. How different
diverse studies and the methods to obtain can studies on different patient populations, or
estimates of effect from all the studies. different regimens for the same treatment, be
Dealing with the original studies in their dif- before combining them is wrong? For example,
fering aspects of design, protocol, analysis, it may be relevant to look at all hormone prep-
reporting in addition to their overall quality, arations as a single group, but for some pur-
is not a trivial task; this section addresses poses it will be important to distinguish types,
some of the main challenges. or even different doses and durations of
t reatment for a single product. In pharmacoep- European Medicines Agency under the aus-
idemiology, risks of adverse effects will often pices of the European Network of Centres
vary notably with duration of treatment and for Pharmacoepidemiology and Pharmaco
length of follow-up. At the very least, it is nec- vigilance (ENCePP®). However, there is no
essary to explore possible variation in the size guarantee that observational studies will
of the effect, especially in terms of how that appear in any registry.
variability in effects might relate to aspects of While not including unpublished studies is
study design, populations included, etc. usual, unpublished data can represent a large
For example, combining all statin trials but proportion of all relevant data. Published
looking just at patients with low LDL (low- results generally are a biased sample.
density lipoprotein) levels at baseline may be Statistically significant, “positive” results (the
reasonable to see if effects are present in that intervention works) are more likely to be pub-
patient group. A more difficult question is lished, published rapidly, published in English,
whether non-randomized studies should be published more than once, published in high
combined with randomized studies. There impact journals, and cited. The contribution
are challenges combining nonrandomized made to the totality of the evidence in system-
studies in a meta-analysis and this is dis- atic reviews by studies with non-significant
cussed later. The treatment of randomized results is as important as those with the statis-
and non-randomized evidence as equivalent tically significant results. If unpublished stud-
in a single analysis is a mistake; some distinc- ies are systematically different from published
tion should be made. Although some have studies in the magnitude and/or direction of
used Bayesian methods to combine evidence the findings, but they are similar in terms of
from both sources in a single analysis, there risk of bias, omitting them yields a biased sum-
is not wide agreement on this approach. mary estimate.
Results should be reported for both sources There is no guarantee that either a published
separately, even if later combined. Within trial or a published meta-analysis will follow
non-randomized studies, studies with differ- the protocol, and bias toward finding “interest-
ent designs, e.g. case-control and cohort, also ing” results is pervasive. Choosing outcomes to
should be evaluated separately (and possibly report after the data are available and analyzed
later combined). is one common way of obtaining biased results
in trials. Investigators do not always keep to
the protocols when publishing. It is one thing
Publication Bias
to have the trial data biased, but then the sys-
Unpublished material cannot be retrieved by tematic reviewers may add to the problem.
literature searches and is likely to be difficult Even in the highly regarded Cochrane Reviews
to find. Publication bias occurs when study there is evidence that reviewers do not keep to
results are not available to be included at the their own protocols and it seems harms are
time when the meta-analysis is conducted, especially likely to be affected by publication
and there are differences between the avail- bias (i.e. studies reporting statistically signifi-
able and unavailable data. Registration of cant increased risk of harms are more likely to
randomized trials, with protocols and other be published).
information, being done prior to results It is clear that the published literature may
being available should reduce the problem. not be as reliable as it should be and it has been
Many journals and regulators require clini- found that review by regulatory authorities
cal trials to be listed in a public registry. may be more reliable, and at the very least
Some registers also contain observational should also be searched for in any meta-
study protocols and results, notably at the analysis based on published data.
Bias in the Abstraction of Data urrently Available

C
Meta-analysis, being retrospective research, is Solutions
subject to the potential biases inherent in
extracting data from reports. It has been shown There are many guidelines on reporting and
that when more than one meta-analysis of a conducting systematic reviews and meta-
treatment for a disease has been done, there analyses and these should be consulted for
can be differences between them, especially in details. Here we give general principles rele-
the inclusion and exclusion of papers to be vant to pharmacoepidemiology.
analyzed. While the results were similar in
some cases, they had occasional extreme disa- Steps Involved in Performing
greement regarding efficacy and some varia- a Meta-analysis
tion in whether results were statistically
significant or not. These disagreements were The main steps in performing a meta-analysis
not easily explainable. In some instances, even are given in this section.
though the same papers are included, the data
extracted can vary, and this can lead to differ- Define the Purpose
ent conclusions about the efficacy of therapy. The primary and secondary objectives of a
Despite efforts to make meta-analysis an objec- meta-analysis should be defined precisely. A
tive, reproducible activity, there is evidently well-formulated question includes a clearly
some judgment involved. defined Patient Population, Intervention,
One other important aspect of meta- Comparator, and Outcome.
analysis is the use of aggregate-level data An example could be “What is the magni-
(usually from published studies) compared tude of the increased risk of gastrointestinal
with using individual patient data (IPD). side effects with NSAIDs (nonsteroidal anti-
Meta-analyses based on IPD have several inflammatory drugs) used for the treatment of
advantages, including the ability to conduct pain, compared with placebo?” For NSAIDs,
proper time-to-event analyses (not always estimating the absolute risk difference (RD)
possible from published data) and the ability (and, thus, the public health implications) as
to perform appropriate subgroup analyses. well as the relative risk (and, thus, the etiologic
Sharing of such data, provided confidentiality implications) might be a secondary objective.
can be preserved, is increasing. The Yale Too broadly defined questions could “mix
University Open Data Access (YODA) Project apples and oranges” and too narrow a focus
is one such example. could lead to finding no, or limited data, or the
Most of the interest in this area is around inability to generalize.
randomized trials, but there are also issues
around the availability of observational data. Perform the Literature Search
Platt and Lieu have noted that there are chal- It is important to have a search strategy sensi-
lenges to overcome despite enthusiasm of the tive enough to identify most of the studies to be
research community for wider availability of included. Just using a computer system may
IPD. They identify three reasons for the chal- not be enough, and it may find too many irrel-
lenges: (i) confidentiality and proprietary evant studies – i.e. the specificity is very low.
concerns, (ii) the cost in terms of time and Searching for the text word “random*” in
effort, required to make raw data usable for MEDLINE will retrieve all randomized con-
analyses, and (iii) the need to create incen- trolled trials (RCTs) but will find many that are
tives for data holders that outweigh the not RCTs and excluding publication types such
disadvantages. as commentaries, editorials, meta-analyses,
reviews, or practice guidelines removes many arguments. However, such exclusions, made
non-relevant citations, without losing any of after having seen the data and the effect of
the relevant trials. The Cochrane Handbook individual studies on the pooled result may
has a chapter devoted to searching for rand- form the basis for legitimate sensitivity analy-
omized trials, including a section on ongoing ses (comparing combined results with and
studies and unpublished data. It also has guid- without that particular study included), but
ance on searching for nonrandomized studies, should not be viewed as primary exclusion
but in practice this is more difficult. Publication criteria.
bias is a major problem for observational stud- Readers often cannot assess whether the
ies since they are not necessarily registered, exclusion criteria were defined after seeing
may not require ethical review, and may be study results, but registration of systematic
untraceable. It seems likely that searching for review protocols helps reduce this problem.
particular adverse event terms is likely to be Registration has also helped with the realiza-
better than using terms related to methods. tion that both reporting of trials and of system-
“Hand” searching, using reference sections of atic reviews is often altered by authors even
relevant retrieved publications or whole after a protocol has been recorded, thus the
searches of relevant journals, may help. potential for bias is considerable. A registry
specifically for systematic reviews is
Establish Inclusion/Exclusion Criteria “PROSPERO” (International Prospective
Rules for including and excluding studies Register Of systematic reviews). Key features
should be defined when the meta-analysis is from the review protocol are recorded and
planned. Limiting to randomized studies with maintained as a permanent record. There also
at least 100 patients has been done where is a registry of studies, maintained by the
many large trials on a topic have been per- European Medicines Agency, under the aus-
formed. With non-randomized studies, one pices of the ENCePP.
might include studies of incident cases only, Studies may generate more than one pub-
when the relationship between exposure and lished paper and choosing which of multiple
outcome differs between incident and preva- papers to include and ensuring that there is no
lent cases. Practical considerations may force double-counting requires care.
changes: e.g. there may be no randomized
studies of a new indication for an existing Collect the Data
therapy. When the relevant studies have been identi-
If broad inclusion criteria are used, then a fied and retrieved, typically, data abstraction
broad hypothesis may be tested. This may performs are developed, pilot tested and revised.
mit examination of the association between A balance is needed between the complete-
design and outcome (e.g. do randomized and ness and time needed to extract that informa-
nonrandomized studies tend to show similar tion. Careful specification in the protocol may
effects?) or the exploration of subgroup effects. help avoid over- or under-collecting informa-
For example, with aspirin given following tion. For randomized trials, it is generally
myocardial infarction, restriction to studies advisable to collect raw data on outcomes by
using more than 75 mg aspirin would not per- group rather than derived measures such as
mit comparison of dose–response effects. odds ratios (ODs). In contrast, for observa-
A priori considerations of original study tional studies, estimates from each of the stud-
design and features should determine inclu- ies adjusted for confounding are best, along
sion and exclusion criteria, not their results. with information about what and how con-
The temptation to justify exclusions post hoc founding factors were included in the
may be strong, making clinically plausible adjustment.
In terms of study quality, it is best to collect studies should contribute more than small
data on the individual aspects of study design studies. Using weights proportional to the
that affect potential bias, such as whether, for inverse of the variance of the within-study OR
example, there was concealment of rand- may introduce bias with rare binary outcomes
omized allocation or “blinded” outcome because the weights depend not only on study
assessment. Such explorations clearly need to size, but on the event rates themselves.
be guided by common sense. For example an The usual basic principle is that within study
outcome of total mortality is less likely to be comparisons between treated and untreated
biased than an outcome such as recurrent are made prior to combination across studies.
chest pain. With randomized trials, this preserves
randomization.
Perform Statistical Analyses Some methods, called “fixed-effect” models,
Odds Ratio, Risk Ratio or Risk Difference, Does It assume there is a single, common effect. Any
Matter? variability among study results is assumed to
There are three summary measures of effect be random and is ignored in producing a sum-
size that can be used in meta-analysis when mary estimate of the treatment effect. Methods
the outcome of interest is binary (e.g. propor- that do not assume this are called “random-
tion of subjects with pain relief): risk ratio effects” models. Between-study variability in
(RR), OR, or RD. Although the summary estimates of treatment effect is taken into
measure used often does not greatly affect the account by random-effects models, incorporat-
statistical significance of the results, the choice ing variability into the weighting scheme for
can affect applicability and interpretability in the summary estimate.
clinical practice. These measures are described Random-effects models produce wider con-
in detail in Chapter 2 on study design. fidence intervals than fixed-effect methods but
RR and RD are easier to interpret than ORs. are not a panacea for unexplained heterogene-
When the baseline (untreated) risk is constant ity. Random-effects models tend to assign rela-
across studies, the RD also allows calculation tively higher weights to small studies than
of relevant public health measures (e.g. a num- fixed-effect models, which may have unwanted
ber of events prevented or caused by a given consequences, particularly when published
treatment). A disadvantage of using RDs in small studies show relatively larger effect sizes
meta-analysis is that it cannot be constant at than unpublished small studies or when there
all levels of baseline risk. If the summary are a large number of small studies that then
measure suggests a decrease of say 10% on an have a combined weight greater than the large
absolute scale, perhaps when the average base- ones. For example, consider an analysis of 10
line risk in the studies is say 30%, in a group trials that all have sample sizes of 500 in both
with a baseline risk of 10% or less, an absolute the treated and control groups. Suppose nine
decrease of 10% will be impossible. studies have event rates of 28% in the treated
ORs have better mathematical properties groups compared with 30% in the control
than RRs; switching the roles of event and groups. In this same analysis, a single study
non-event does not alter ORs; one is the recip- has event rates of 3% in the treated group ver-
rocal of the other (unlike RRs when events are sus 1% in controls. For an inverse-variance
common). weighted analysis of RDs, which are −2% in
the nine studies and +2% in the single study,
Choice of Statistical Method the single study with the low event rates would
In most situations, the statistical method uses get 54% of the weight in the meta-analysis,
some form of weighted average of within- compared with 5.1% of the weight for each of
study results. Intuitively it is clear that large the other nine studies. For an analysis of (log)
relative risks, the single study would get 0.4% relative risks associated with anti-TNF were
of the weight, compared with 11.1% of the 0.99 (95% confidence interval (CI) 0.61–1.68)
weight for each of the other nine studies. for cancers excluding non-melanoma skin
The random-effects model is often imple- cancer (NMSC), and 2.02 (95%CI 1.11–3.95)
mented with DerSimonian and Laird (DL) for NMSC. Relative risks were heterogeneous
methodology but this method of weighting is across the anti-TNF drug groups. The authors
known to be suboptimal in several situations concluded that, despite a reassuring overall
and is not to be recommended in general. short-term risk, they could neither refute nor
Another random-effects method that is worth confirm either an increased or unchanged
considering is the Hartung-Knapp-Sidik- risk. Despite the large numbers of studies
Jonkman (HKSJ) method. While it is straight- and patients included, statistical precision,
forward and can outperform the standard DL differences in baseline cancer risk, and
method, extra caution is needed when there incomplete reporting detail between trials
are 5 studies of very unequal sizes. limited the ability to detect or dismiss
Recently, Stanley and Doucouliagos chal- increases in risk. The authors noted that this
lenged the two core conventional meta- example illustrates the challenges in safety-
analysis methods (fixed- and random-effects) assessments using meta-analyses of RCTs
and proposed a weighted least squares method and suggested that long-term risk assessment
that is neither fixed nor random-effects and requires observational studies.
has some good properties with small numbers
of events. However, unlike the Peto method, Combinability of Results from Diverse Studies:
which is described in more detail below, it is Heterogeneity
unable to deal with zero events in one of the A key question is whether it is clinically and
comparison arms. statistically reasonable to estimate an average
Bayesian statistical methods can take into effect of therapy, either positive or negative.
account the investigator’s prior beliefs about Being too inclusive with studies may mean
the size of an effect or about the factors bias- that the average effect may not apply to any
ing the observed effects. They are particularly particular subgroup of patients. On the other
appealing for meta-analyses that attempt to hand, it may be desirable to allow for some het-
synthesize evidence from multiple sources erogeneity in study design and analysis to
under a unified framework, to make direct increase the generalizability of the results and
probability statements about any hypotheses, to permit the exploration of various factors as
and to handle complex problems. Askling modifiers of the treatment effect. We can dis-
and colleagues used the Bayesian hierarchi- tinguish between aspects affecting variability
cal piecewise exponential survival model to related to modifiable aspects of the conduct
investigate the cancer risk for the anti-tumor and analysis of studies such as choice of sum-
necrosis factor (“TNF”) drug class. All 74 mary measure (e.g. RR vs. RD) or study design
RCTs of TNF inhibitors of at least four weeks features (e.g. blinding in the evaluation of end-
duration were provided to a team of inde- points), and real biological or clinical variation
pendent investigators, for events indicating a in treatment effect. The latter represents the
possible cancer. A Bayesian “piece-wise” potential to target therapy to the appropriate
exponential model was used to analyze the patient populations.
individual patient-level data. One hundred The I2 statistic quantifies among-study vari-
thirty (0.84%) individuals (of 15 418) rand- ability by estimating the proportion of variabil-
omized to anti-TNF therapy were diagnosed ity in point estimates due to heterogeneity
with cancer, compared to 48 (0.64% of 7486) rather than sampling error. I2 is recommended
randomized to comparators. The overall because:
●● it focuses attention on the effect of any het- statistical power, especially with a small num-
erogeneity on the meta-analytic result; ber of studies.
●● its interpretation is intuitive, i.e. the percent- It has been argued that because of the poten-
age of total variation across studies due to tial for bias in observational epidemiologic
heterogeneity; studies, exploring heterogeneity should be the
●● it can be accompanied by an uncertainty main point of meta-analyses of such studies,
interval; rather than producing a single summary
●● it is simple to calculate and can usually be measure.
derived from published meta-analyses; As an example of the type of analysis that
●● it does not inherently depend on the number could be used to investigate study design
of studies in the meta-analysis; and issues, Hennessy and colleagues performed a
●● it may be interpreted similarly irrespective meta-analysis of nonexperimental studies
of the type of outcome data (e.g. time to comparing third generation oral contracep-
event, quantitative, or dichotomous) and tives (those containing gestodene and des-
choice of effect measure (e.g. odds or hazard ogestrel) to second generation pills (those
ratios). containing levonorgestrel) with respect to the
risk of venous thromboembolic events. A
While significant heterogeneity, or a large major issue in these studies has been the pos-
value for I2, means the studies are not all esti- sibility of depletion of susceptibles. Specifically,
mating the same parameter, the magnitude of the concern is that users of the newer drugs
difference may not be great when the compo- might tend to be new users of any oral contra-
nent studies themselves are large (and the ceptives, whereas users of the older, second
within-study variability is small). The generation drugs, would tend to be established
Cochrane Handbook (Section 9.5.2) has quali- users. The risk of venous events tends to be
tative guidelines for interpreting the magni- highest for new users, who have events soon
tude of I2. after beginning pill use. These susceptible indi-
In searching for sources of heterogeneity viduals, the argument goes, would be depleted
one might stratify the studies by patient char- from the ranks of users of second-generation
acteristics or design features and investigate pills, but not from among the third-generation
heterogeneity within and across strata. If strat- pill users, thereby leaving a more susceptible
ification explains the heterogeneity, the com- population of third generation pill users. The
bined results will differ between strata and authors found several studies that had per-
heterogeneity within each stratum would be formed subgroup analyses of new users in
reduced. For example, if the full set of trials their first year of use. When combined, these
includes some studies with severely ill patients subgroups still demonstrated an increased risk
and others with mildly ill patients, one could from third generation pills. The power to look
observe heterogeneity of treatment effects. within subgroups was only available within
Stratification on disease severity could (hypo- the context of the meta-analysis, not within
thetically) show that treatment effects are large any of the individual studies.
in severely ill patients and small in less ill Exploratory analyses of meta-analyzed data
patients. Regression methods, such as weighted may provide insights into biology and/or may
least squares linear regression, can also be generate hypotheses. Figure 20.1 shows a
used to explore sources of heterogeneity. “forest plot.” It is a useful graphical picture of
Graphical methods for meta-analysis can aug- studies, even when no single summary statis-
ment analytical approaches when the focus is tic is derived. This figure shows four
on issues related to heterogeneity. Statistical observational studies on the effect of hydroxy-
tests of heterogeneity often suffer from lack of chloroquine on death or serious outcomes
Observational studies of hydroxychloroquine in Covid-19

Fixed effect, Mantel & Haenszel Method
Study 95% confidence Interval Point estimate %
ID ES (95% CI) Weight
Effect size is hazard ratio
Geleris 1.00 (0.76, 1.32) 29.95
Arshad 0.34 (0.25, 0.46) 26.86
Rosenberg 1.08 (0.63, 1.85) 7.87
Rentsch 1.03 (0.80, 1.33) 35.33
Overall (I-squared = 92.5%, p=0.000) 0.76 (0.65, 0.88) 100.00

Overall summary
diamond shows 95% CI
.25 .5 1 2
Figure 20.1 Forest plot of four recent observational studies on the effect of hydroxychloroquine on death
or serious outcomes from COVID-19. Features of a forest plot are indicated. Source: Data are taken from
Arshad et al. (2020), Geleris et al. (2020), Rentsch et al. (2021), and Rosenberg et al. (2020).
from COVID-19. It is not based on a system- methods and that the inverse-variance-average
atic review. There is variation in the settings, should be avoided.
the detailed outcomes, and some heterogene- With rare events, studies may have no events
ity in the results. Showing the actual values of in one or both of the arms, and relative meas-
the results with their confidence intervals is ures such as relative risk or ORs cannot be cal-
always helpful, even if a summary value is culated, so those studies will be automatically
inappropriate. The sizes and numbers of excluded. RDs can be estimated but in the
events (deaths) in the different studies means presence of rare events produce biased results
that the small studies contribute less (lower and have very limited power. In cases when
“weight”) and larger studies contribute more. there are no events in one arm, relative meas-
The size of the box showing the central esti- ures can be calculated. Some methods, includ-
mate is proportional to the weight. ing the commonly used Mantel-Haenszel
method add “continuity corrections,” typically
Analysis of Rare Events adding 0.5 to all cells in a two-by-two table.
By combining results of trials meta-analysis This leads to bias in the presence of rare events,
can help address challenging problems with and is not necessary, even for the Mantel-
rare events, but many methods for combining Haenszel method. It is better to use the recip-
data are based on large sample approximations rocal of the sample size of the opposite
so may be unsuitable leading to variations in treatment arm.
the overall result depending on the method The Peto method, also known as the “one-
used. Recommendations are based on simula- step” model, is a fixed-effect model that focuses
tion studies in which the “truth” is generated on the observed number of events in the exper-
by the investigators and show that fixed-effect imental intervention and compares it with the
models should be used over random-effects expected number; this has the effect that it can
deal with single arms having zero observed asymmetry, with very few (if any) points
events. It often produces less biased results around the point indicating no effect, for stud-
provided there is no substantial imbalance ies with large variances. This method requires
between treatment and control group sample a sufficient number of studies to permit the
sizes within trials and provided the treatment visualization of a funnel shape to the data. If
effects are not exceptionally large (i.e. the the funnel plot does indicate the existence of
effect size needs to be less than an OR of 5, or publication bias, then one or more of the cor-
greater than an OR of 0.2). Bayesian methods rection methods described below should be
can also be appropriately applied to rare events considered. In the presence of publication
meta-analysis and can deal with zero events to bias, the responsible meta-analyst should also
derive posterior inferences for the treatment evaluate the ethics of presenting a summary
effect estimates. result that is likely to represent an overesti-
Sensitivity analysis is especially important mate of the effect in question.
and recommended to be used and reported Two examples of funnel plots are given in
with rare events, because results may vary with Figures 20.2 and 20.3. These plots represent
choice of statistical methods, scale of measure- studies of psychoeducational programs for sur-
ment, specification of the prior distribution in gical patients. In the first plot, only the pub-
the Bayesian approach and continuity correc- lished studies are represented. The funnel
tion factors. These sensitivity analyses allow appears to have a “bite” taken out of it where
readers to assess the robustness of the results. the small studies showing no effect of these
programs should be. In the second plot, the
Formulate Conclusions and unpublished studies, including doctoral disser-
Recommendations tations, are included, and the former “bite” is
The conclusions of a meta-analysis should be now filled with these unpublished studies.
clearly summarized, with appropriate inter- Sterne and Egger provide guidelines for the
pretation of the strengths and weaknesses and choice of axes in funnel plots of studies with
generalizability. Suggestions for future dichotomous outcomes, recommending that
research and hypotheses generated should be the standard error of the treatment effect
distinguished from conclusions. (e.g. the standard error of the log OR) be used
as the measure of study size and that relative
measures (relative risk, as opposed to RD) be
Publication Bias
used as the treatment effect measures. These
As discussed above under Methodological same authors and a colleague point out that
Problems, when the primary source is pub- publication bias is only one possible explana-
lished data, these may represent a biased sub- tion for funnel plot asymmetry, so that the
set of all the studies that have been done. funnel plot should be seen as estimating
Generally, “significant” studies are more likely “small study effects,” rather than necessarily
to be published than non-significant ones. The publication bias.
“funnel plot,” plotting the effect size (e.g. the Several methods to deal with potential
RD) against a measure of study size, such as unpublished studies have been developed.
the sample size or the inverse of the variance These include formal methods to test for the
of the individual effect sizes, can help. With no presence of publication bias and methods to
publication bias, the points produce a funnel adjust summary estimates to account for
shape, with scattered points centered around unpublished studies, but these methods make
the true value, and with the degree of scatter fairly strong assumptions about the specific
narrowing as the variances decrease. If publi- mechanism producing the publication bias. A
cation bias exists, the funnel would show method called “trim-and-fill” has a fair amount
160
120
Sample size
80
40
2 2
2 2
0
–0.40 0.00 0.40 0.80 1.20 1.60
Effect size
Figure 20.2 Funnel plot for published 34 studies only: analysis of data from Devine and Cook’s review of
psychoeducational programs for surgical patient. Source: Reprinted by permission of the publishers from
Light and Pillemer (1984) by the President and Fellows of Harvard College.
160
120
Sample size
80
40
2 2
2 2
2 2
0
–0.40 0.00 0.40 0.80 1.20 1.60
Effect size
Figure 20.3 Funnel plot for published 34 studies (open boxes) and unpublished (closed triangles):
analysis of data from Devine and Cook’s review of psychoeducational programs for surgical patients. Source:
Reprinted by permission of the publishers from Light and Pillemer (1984) by the President and Fellows of
Harvard College.
of intuitive appeal, although it, too, relies on registration of studies at their inception, prior
assumptions about the missing studies. It is to the availability of results. Others have sug-
based on the funnel plot, focusing on the stud- gested obtaining unpublished data from the
ies that lead to the appearance of funnel plot Food and Drug Administration (FDA), an
asymmetry. Under this approach, a mirror approach used by Turner and colleagues.
image of the studies producing the asymmetry These authors obtained reviews from FDA for
is imputed, using a carefully defined statistical studies of 12 antidepressant agents, conducted
algorithm to determine which studies to mir- a systematic literature search to identify
ror, and the impact of adding those mirror matching publications, and compared the
image studies to the pooled analysis is assessed. results based on published studies with the
As noted earlier, one solution to the problem results based on the FDA data. The analysis
of publication bias is the use of prospective restricted to published literature showed that
94% of the trials were positive. In contrast, the precise estimates. To correctly compare B with
analysis of FDA data showed that only 51% C, we obtain appropriate meta-analytic OR for
were positive. A further review in looking at the comparisons A-B and A-C and obtain OR
other indications for anti-depressants found a (B vs. C) = OR (A vs. B) / OR (A vs. C).
similar bias in the literature. Although the esti- Multiple-treatment comparison techniques
mate of effect size was only increased margin- can easily deal with multiple arms and account
ally, “reporting biases led to significant for correlation between them. They also permit
increases in the number of positive findings in assessment of inconsistency – disagreement
the literature”. The “Open Trials project” and between direct and indirect evidence.
the Yale project “YODA” cited above are
attempts to reduce the bias from using only Assumptions
published literature. Now there is an online Indirect comparison methodologies have simi-
tool (https://fda.opentrials.net/search) that lar assumptions to traditional meta-analysis.
allows FDA documents to be retrieved more The following are the main assumptions.
easily.
i) homogeneity: each of the A-B, B-C and
Going one step further, meta-analyses that
other comparisons should be homogene-
are prospectively planned, with complete pro-
ous enough for those specific comparisons
tocols, including proposed tests of subgroup
to be combined.
effects, prior to having knowledge of the results
ii) similarity: factors affecting response must
of any of the component studies, can be con-
be similarly distributed across the trials
ducted. More on the topic of prospective meta-
(similar patient characteristics, settings,
analysis is presented below.
follow up, and outcomes and methodologi-
cally similar trials).
Indirect Comparison iii) consistency: agreement between direct
and Simultaneous Comparison and indirect evidence needs to be checked.
of Treatments Available for Specific
Conditions
Case Studies of Applications
Decision-makers need to make informed deci-
of Meta-analysis
sions on head-to-head comparisons of treat-
ments, but relevant trials may not exist. When Saving Time and Resources by Using
the treatments have been compared to a com- Meta-analysis
mon comparator, for example placebo, it is Meta-analysis can shorten the time between
possible obtain indirect evidence synthesis – research findings and implementation of
also known as network meta-analysis. change in clinical practice or policy and regu-
Indirect evidence involves using data from lation. A simple but elegant example of the use
trials that have compared medication “A” vs of meta-analysis in the approval context was
medication “B”, and from trials that have com- the use of meta-analysis of ECG data from sev-
pared medication “A” vs medication “C”, to eral clinical pharmacology studies for two drug
draw conclusions about the effect of medica- application submissions. They calculated a
tion “B” relative to medication “C.” It is crucial pooled estimate for the difference between
that when an indirect comparison is estimated, active doses and placebo on QT prolongation,
the analysis respects the randomization. This avoiding the need for a new study to address
means that the analysis must be based on treat- the question.
ment differences within each trial. Simply col- “Cumulative meta-analysis,” i.e. performing a
lapsing results by treatment arm ignores new meta-analysis each time new RCT results
randomization producing biased and overly are available for a given treatment has been
a dvocated. As an example, Antman and col- tests can generate false positive findings (type
leagues analyzed data from 17 trials of 1 error) and this is often ignored. To address
β-blockers for the prevention of death in the that concern, sequential analysis methods
years following a myocardial infarction. In the have been applied, but some suggest that the
left-hand side of Figure 20.4, reproduced from most natural approach is Bayesian, as contin-
their paper, the data are presented as a traditional uous learning and updating our knowledge
meta-analysis, with individual study results pre- over time fits the Bayesian philosophy.
sented along with the summary OR arbitrarily Existing data form the basis for the prior dis-
estimated after 17 trials had been completed. In tribution and new studies update this forming
the right-hand side of Figure 20.4, the same data a posterior distribution, which then becomes
are presented as a cumulative meta-analysis, the new prior distribution. Multiplicity is han-
with an updated summary estimate calculated dled naturally in this framework and conclu-
after the completion of each new trial. The cumu- sions (usually in the form of “credible
lative meta-analysis clearly shows that the intervals”) are expressed as probabilistic state-
updated pooled estimate became statistically sig- ments about findings (which can be flexible in
nificant in 1977 and has remained so ever since. terms of different treatment effect sizes under
Caution is advisable in interpreting cumula- different scales), not as statements about
tive meta-analyses because multiple statistical hypotheses.
Individual RCT and overall meta-analysis results Cumulative Mantel–Haenszei method

Odds ratio (log scale) Odds ratio (log scale)
No.of 0.1 0.2 0.5 1 2 5 10 No.of 0.5 1 2
Year patients patients
1972 77 77
1974 230 307
1974 162 469
1977 3053 3522 Z= –2.29 p<.05
1980 720 4242
1980 111 4353
1981 1884 6237 Z = 3.99 p<.0001
1982 1103 7340
1982 3837 11177 Z = –4.48 p<.00001
1982 1456 12633
1982 560 13193
1983 584 13777
1983 301 14078
1983 529 14607
1984 1741 16348
1987 2395 18743
1988 1395 20138 Z = –4.47 p<.00001
Overal 20138 z=–4.47 p<.00001
Favors treatment Favors control Favors treatment Favors control
Figure 20.4 Results of 17 RCTs of the effect of oral beta-blockers for secondary prevention of mortality in
patients surviving a myocardial infarction presented as two types of meta-analyses. On the left is the
traditional one, revealing any trials with nonsignificant results but a highly significant estimate of the
pooled results on the bottom of the panel. On the right, the same data are presented as cumulative
meta-analyses, illustrating that the updated pooled estimate became statistically significant in 1977 and
has remained so up to the present. Note that the scale is changed on the right graph to improve clarity of
the confidence intervals. Source: Reproduced from Antman et al. (1992) with permission from the American
Medical Association.
There is no consensus on the multiplicity nonexperimental studies, but are not always
problems in cumulative meta-analysis, so confirmed by subsequent, properly designed ran-
some use it mainly as an exploratory tool, pro- domized trials. Consider the case of beta-carotene
viding caveats about the number of analyses in the prevention of cancer. A series of observa-
performed without a formal correction. This tional studies examined the relation between die-
approach is analogous to that for many con- tary intake of foods rich in β-carotene and the risk
ventional safety analyses, for which nominal P of lung cancer. Overall, they showed a relatively
values from hypothesis tests are often provided consistent association between diets rich in
without adjustment, when there are a limited β-carotene and reduced risk of lung cancer.
number of pre-specified outcomes. Subsequent randomized trials of this specific
Another consideration is that estimates of nutrient as a supplement have failed to confirm a
treatment effect may not be stable over time, protective effect against lung cancer.
perhaps due to changing clinical environments.
In the β-blocker example, treatment appears to Cumulative Meta-analysis as a Tool
be less effective in the most recent studies. to Detect Harm Signals earlier
Thus, it may be important to re-evaluate thera- Cumulative meta-analysis also could be used
pies as other treatment strategies evolve. as a tool to detect safety signals earlier. Case
A final caution relates to the continuing need Example 20.1 illustrates some of the potential
for well-designed RCTs. New indications for exist- methodological issues (that are shared by tra-
ing therapies, for example, are often suggested by ditional meta-analysis).
Case Example 20.1 Risk of Cardiovascular Events and Rofecoxib: Cumulative Meta-analysis
(Juni et al. 2004)
Background Strengths
Rofecoxib, a cyclo-oxygenase-2 inhibitor (a Cumulative meta-
analysis potentially can
type of NSAID), was withdrawn from the detect harm earlier than traditional
market because of cardiovascular adverse meta-analysis.
effects.
Question Limitations
Can cumulative meta-analysis of RCTs estab- ●● The validity of pooling of trials that were
lish whether evidence on the adverse effects not clinically homogeneous is questiona-
of rofecoxib was available before its removal? ble. The authors combined the results of
trials with dissimilar control arms (pla-
Approach
cebo, naproxen and non-naproxen
The authors searched bibliographic data-
NSAIDs).
bases and FDA files and included all RCTs in
●● The authors excluded trials that evaluated
patients with chronic musculoskeletal disor-
Alzheimer’s disease. In this case, the inclu-
ders that compared rofecoxib with other
sion of such a trial would have made the
NSAIDs or placebo. Myocardial infarction
early signal disappear.
was the outcome assessed.
Results
Key Points
The adverse cardiovascular effects of
Cumulative meta-analysis is a tool to evalu-
rofecoxib could have been identified several
ate the safety of health interventions.
years earlier.
First, one might question, or at least require Mantel-Haenszel fixed-effect method. They
justification for, the validity of pooling of trials found that 10 of 3179 patients receiving anti–
that are not clinically homogeneous. For exam- IL-12/23 therapies experienced MACEs
ple, the authors combined the results of trials compared with no events in 1474 patients
with dissimilar control arms (placebo, nap- receiving placebo (Mantel-Haenszel risk dif-
roxen and non-naproxen NSAIDs). ference, 0.012 events/person-year; 95% CI,
Second, excluding trials assessing an inter- −0.001 to 0.026; P = 0.12). (NOTE: in the origi-
vention in other indications may be ques- nal paper, the authors use the term “risk differ-
tioned. In Case Example 20.1, trials in chronic ence” but report results in terms of person-time,
musculoskeletal pain were the focus, and trials which would usually require use of rate differ-
in Alzheimer’s disease were excluded. Clearly, ences). They concluded that there was no sig-
one would not combine trials for different indi- nificant difference in the rate of MACEs
cations to assess efficacy, but it is less clear associated with anti–IL-12/IL-23 antibodies,
with harms, although harms could also vary by but that even the meta-analysis may have been
indication. An approach often used in the reg- underpowered to identify a significant differ-
ulatory setting may seem appropriate by ence (because there were only 10 events).
including all indications but stratifying by In a second meta-analysis, Tzellos and col-
indication. leagues also studied anti-IL-12 / 23 biological
Third, one can ask whether efficacy and agents (ustekinumab and briakinumab) with
safety should be evaluated similarly. For effi- respect to risk of MACE, specifically in the set-
cacy, multiple looks at data can lead to false ting of treatment of chronic plaque psoriasis.
positive results and have serious consequences. Studies of psoriatic arthritis were excluded, as
For safety, it could be argued that adjustments in the Ryan meta-analysis. These authors used
should not be as large, if done at all (in the the Peto fixed-effect method to estimate ORs.
interest of remaining sensitive to safety issues They found a “possible higher risk of MACEs”
that might arise). in patients treated with IL-12/23 antibodies
Fourth, it is uncertain whether cumulative compared with placebo-treated patients
meta-analysis can systematically detect harm (OR = 4.23, 95% CI: 1.07–16.75, P = 0.04).
earlier. Rare adverse events, or those occurring
late after exposure may not be seen during Indirect Comparisons: Network
drug development, so cumulative meta- Meta-analysis and Simultaneous
analysis may not always help (or may work Evaluation of Treatment Therapies
best if large post-approval studies are for the Same Indication
conducted). Network meta-analysis has a number of spe-
Ryan and colleagues conducted a meta- cific issues. This is illustrated by an analysis of
analysis of 22 RCTs studying the effects of anti- the efficacy and acceptability of new genera-
IL-12/23 therapies. These are anti-inflammatory tion antidepressants, performed by Cipriani
agents used to treat conditions such as psoria- and colleagues. In one study the authors
sis (the initial indication). The studies included excluded placebo groups where present, after a
10 183 patients. The primary outcome measure careful search for trials that included asking
was major adverse cardiac events (MACE). pharmaceutical companies, regulatory agen-
MACE definitions can vary; in this analysis it cies, and study investigators to supply informa-
was defined as a composite of myocardial tion about details of study design and related
infarction, cerebrovascular accident, or cardio- factors.
vascular death during the placebo-controlled Efficacy was evaluated as the proportion of
portions of the included trials. They chose patients who had a reduction of at least 50%
absolute RDs as their effect measure, using the from the baseline score on standard depression
rating scales or the proportion who scored sponsors to submit all published or unpub-
“much” or “very much” improvement on the lished data and the FDA can use patient-level
clinical global impression. Acceptability of data. The findings from those meta-analyses
therapy was evaluated as the proportion of have been used to support regulatory decisions
patients who terminated the study early for to mandate labeling changes.
any reason during the first eight weeks of treat- For example, the FDA examined antiepilep-
ment. The authors calculated the ORs for each tic drugs and suicidality events. Their review
of the drugs compared to fluoxetine, using a of 199 placebo-controlled trials from 11 anti
random-effects model within a Bayesian epileptic drugs found that there were 1.9 per
framework, resulting in an estimated probabil- 1000 (95% CI: 0.6, 3.9) more antiepileptic drug
ity that each antidepressant was the most effi- patients than placebo patients who experi-
cacious, or the most acceptable, the second, enced suicidal behavior or ideation compared
the third, etc. ranking treatments in terms of to the placebo patients. Based on the findings,
efficacy and acceptability. They also assessed the FDA requested updates to product labels.
the consistency between direct and indirect Not only does meta-analysis sometimes sup-
evidence. port the decision to change or update the cur-
With 117 trials with 25 928 participants and rent labeling of approved drugs, it can also
12 antidepressants included in the analyses, provide evidence as to whether or not to keep a
there was generally consistency between direct drug on the market or withdraw its use for a
and indirect evidence, but not all the antide- particular indication. For example, the FDA’s
pressants were equally efficacious or equally patient-level meta-analysis for rosiglitazone
well tolerated; they were able to report a rank- found about a 40% increase in myocardial
ing for efficacy or acceptability. ischemia among diabetes patients taking insu-
There was both enthusiasm for, and criti- lin or those using nitrates but less evidence
cism of, this study. Exclusion of placebo- when compared with metformin or a
controlled data and publication bias were the sulfonylurea.
main concerns. Another study found that 95%
of published trials were “positive” compared to
only 51% of FDA-registered studies (some of The Future
which were unpublished) so relying primarily
on published data may overestimate benefits. The examples above have raised several impor-
tant issues for the future. When evaluating
Regulators’ Role safety during drug development there is prob-
The U.S. FDA routinely obtains IPD for new lem with multiplicity. Trials have hundreds of
drug applications and may also request such different adverse events and if cumulative
data for other issues. The consequence is that meta-analyses are updated for each trial com-
they are able to do IPD meta-analyses. This has pleted, repeated testing yields further multi-
rarely been done by other regulators, but the plicity problems. Focusing on p-values,
situation is changing and there are signs that ignoring magnitude and clinical importance is
other regulators are interested in following the unwise. More work is needed on sensible ways
FDA example. to deal with the problem.
In recent years, the FDA has used meta- The current situation with respect to regis-
analysis to investigate adverse events associ- tration of clinical trial protocols and results is
ated with the use of certain drugs. While that there is a wide array of registries making it
publication bias is often a major concern in difficult to find all relevant studies. What
conducting a meta-analysis, regulatory author- would be useful is a dedicated search engine,
ities like the FDA have authority to request with low false positive and false negative rates,
Key Point 351
that would be able to find all trials with given FDA to address questions of interest through
characteristics. This would improve meta- specific queries sent to participating data part-
analysis by using all relevant randomized evi- ners. CNODES, in Canada, is a similar kind of
dence and would improve the science. network (see Chapter 9). OHDSI (Observational
When hundreds of categories of events are Health Data Sciences and Informatics) is a
tabulated, specific events will be seen in very global collaboration of investigators perform-
small numbers. Work to date suggests that ing studies in a distributed network but also
more targeted definitions can sometimes lead developing methodology for conducting such
to stronger signals (larger relative risks) and studies. It is unclear whether current meta-
may actually make it more likely that signals analytic methods, e.g. doing separate analyses
will be detected, despite observing fewer events in each database, then using an inverse-
with narrower definitions. variance weighted average, are best suited to
Results of a meta-analysis when there is this distributed network environment. New
substantial heterogeneity are difficult to inter- methods of distributed analysis, which do a
pret. If the heterogeneity is adequately better job of mimicking the results one would
explained in the analysis in terms of subgroup get by having one large dataset with data from
effects, or trial quality, meta-analysis might ALL datasets, need further investigation.
still be an acceptable part of demonstrating In conclusion, while there are no easy
effectiveness or harm. Work is needed to answers to many of the questions presented
establish transparent criteria by which to eval- above, it is clear that meta-analysis will play an
uate situations where inconsistency in results increasingly important role in the formulation
suggests benefit (or harm) in some but not of treatment and policy recommendations.
others. This can be just chance but might have Thus, the design and analytical attributes of
other explanations. the meta-analyses performed, and of the
As the focus of policy and clinical decisions included studies, are of the utmost importance
moves in the direction of comparative effec- and need to be reviewed by the scientific com-
tiveness, which also includes comparative munity in an open, published forum. Meta-
safety, one might wish to make direct compari- analyses, if they are carefully interpreted in
sons across all drugs for a given indication, but view of their strengths and weaknesses, should
this is not easy to implement. Work is needed, prove to be extremely helpful in pharmacoepi-
however, to explore in practice, the conditions demiologic research.
under which indirect comparisons may be
both valid and useful.
The inclusion of observational studies in
meta-analyses, particularly of serious but Key Points
uncommon adverse events, will almost cer-
tainly be a necessity, but reliability of answers ●● Meta-analysis, if carefully done, is a power-
is not guaranteed. Methodological considera- ful method that can be used to identify
tions need to be addressed (e.g. by examining sources of variation among studies and pro-
how methodology relates to study results) to vide an overall measure of effect.
help with interpretation. ●● Combining evidence across diverse study
Distributed data networks have been estab- designs and study populations may lead to
lished within and outside the US and are being generalizable results.
used for large drug safety studies. Using multi- ●● Much care is needed in the interpretation of
ple sources of data for these may mean that meta-analyses, due to issues such as publica-
analysis is best done using meta-analytic tion bias and flaws in the design of compo-
approaches. The Sentinel network is used by nent studies.
●● Meta-analysis plays an increasingly impor- ●● Extended meta-analytic techniques that can

tant role in safety assessment. It might be handle indirect treatment comparisons and
necessary to consider non-randomized stud- more than two interventions in a network
ies, separately from randomized trials, for meta-analysis can strengthen the inference
questions about uncommon adverse events. regarding a treatment.
Further Reading
Antman, E.M., Lau, I., Kupelnick, B. et al. from the International Committee of Medical
(1992). A comparison of results of meta- Journal Editors. N. Engl. J. Med. 351 (12):
analyses of randomized control trials and 1250–1251.
recommendations of clinical experts. DerSimonian, R. and Laird, N. (1986). Meta-
Treatments for myocardial infarction. JAMA analysis in clinical trials. Control. Clin. Trials 7
268 (2): 240–248. (3): 177–188.
Arshad, S., Kilgore, P., Chaudhry, Z.S. et al. Devine, E.C. and Cook, T.D. (1983). Effects of
(2020). Treatment with hydroxychloroquine, psycho – educational interventions on length
azithromycin, and combination in patients of hospital stay: a meta – analytic review of 34
hospitalized with COVID-19. Int. J. Infect. Dis. studies. In: Evaluation Studies Review Annual,
97: 396–403. vol. 8 (ed. R.I. Light), 417–432. Beverly Hills,
Askling, I., Fahrbach, K., Nordstrom, B. et al. CA: Sage.
(2011). Cancer risk with tumor necrosis factor Duan, R., Boland, M.R., Liu, Z. et al. (2020).
alpha (TNF) inhibitors: meta – analysis of Learning from electronic health records across
randomized controlled trials of adalimumab, multiple sites: a communication-efficient and
etanercept, and infliximab using patient level privacy-preserving distributed algorithm. J.
data. Pharmacoepidemiol. Drug Saf. 20 (2): Am. Med. Inform. Assoc. 27 (3): 376–385.
119–130. Geleris, J., Sun, Y., Platt, J. et al. (2020).
Baigent, C., Keech, A., Kearney, P.M. et al. Observational study of hydroxychloroquine in
(2005). Efficacy and safety of cholesterol- hospitalized patients with Covid-19. N. Engl. J.
lowering treatment: prospective meta-analysis Med. 382 (25): 2411–2418.
of data from 90,056 participants in 14 Hartung, J. (1999). An alternative method for
randomised trials of statins. Lancet 366 (9493): meta-analysis. Biom. J. J. Math. Methods Biosci.
1267–1278. 41 (8): 901–916.
Bender, R., Friede, T., Koch, A. et al. (2018). Hartung, J. and Knapp, G. (2001a). A refined
Methods for evidence synthesis in the case of method for the meta-analysis of controlled
very few studies. Res. Synth. Methods 9 (3): clinical trials with binary outcome. Stat. Med.
382–392. https://doi.org/10.1002/jrsm.1297. 20 (24): 3875–3889.
Berlin, I.A. and Colditz, G.A. (1999). The role of Hartung, J. and Knapp, G. (2001b). On tests of
meta-analysis in the regulatory process for foods, the overall treatment effect in meta-analysis
drugs, and devices. JAMA 281 (9): 830–834. with normally distributed responses. Stat.
Cipriani, A., Furukawa, T.A., Salanti, G. et al. Med. 20 (12): 1771–1782.
(2009). Comparative efficacy and acceptability Hartung, J. and Makambi, K.H. (2003). Reducing
of 12 new-generation antidepressants: a the number of unjustified significant results
multiple-treatments meta-analysis. Lancet 373 in meta-analysis. Commun. Stat.Simul.
(9665): 746–758. Comput. 32 (4): 1179–1190.
De Angelis, C., Drazen, I.M., Frizelle, E.A. et al. Hennessy, S., Berlin, I.A., Kinman, I.L. et al.
(2004). Clinical trial registration: a statement (2001). Risk of venous thromboembolism
Further Reading 353
from oral contraceptives containing gestodene hydroxychloroquine on COVID-19 mortality:

and desogestrel versus levonorgestrel: a a population-based cohort study in patients
meta – analysis and formal sensitivity with rheumatoid arthritis or systemic lupus
analysis. Contraception 64 (2): 125–133. erythematosus using the OpenSAFELY
Higgins, I.P. and Green, S. (eds.) (2011). platform. Lancet Rheumatol. 3: e19–e27.
Cochrane Handbook for Systematic Reviews of Rosenberg, E.S., Dufort, E.M., Udo, T. et al.
Interventions Version 5.1.0. Oxford: Cochrane (2020). Association of treatment with
Collaboration. hydroxychloroquine or azithromycin with
IntHout, J., Ioannidis, J.P., and Borm, G.F. in-hospital mortality in patients with
(2014). The Hartung-Knapp-Sidik-Jonkman COVID-19 in New York State. JAMA 323 (24):
method for random-effects meta-analysis is 2493–2502.
straightforward and considerably outperforms Roest, A.M., de Ionge, P., Williams, C.D. et al.
the standard DerSimonian-Laird method. (2015). Reporting bias in clinical trials
BMC Med. Res. Methodol. 14: 25. investigating the efficacy of second –
Juni, P., Nartey, L., Reichenbach, S. et al. (2004). generation antidepressants in the treatment of
Risk of cardiovascular events and rofecoxib: anxiety disorders: a report of 2 meta –
cumulative meta-analysis. Lancet 364: analyses. JAMA Psychiat. 72 (5): 500–510.
2021–2029. Rover, C., Knapp, G., and Friede, T. (2015).
Kearney, P.M., Baigent, C., Godwin, I. et al. Hartung-Knapp-Sidik-Jonkman approach and
(2006). Do selective cyclo‑oxygenase – its modification for random-effects meta-
2 inhibitors and traditional non-steroidal analysis with few studies. BMC Med. Res.
anti-inflammatory drugs increase the risk of Methodol. 15: 99.
atherothrombosis? Meta-analysis of Ryan, C., Leonardi, C.L., Krueger, I.G. et al.
randomised trials. BMJ 332 (7553): 1302–1308. (2011). Association between biologic therapies
Light, R.J. and Pillemer, D.B. (1984). Summing for chronic plaque psoriasis and
Up: The Science of Reviewing Research. cardiovascular events: a meta-analysis of
Cambridge, MA: Harvard University Press. randomized controlled trials. JAMA 306 (8):
Makambi, K.H. (2004). The effect of the 864–871.
heterogeneity variance estimator on some Sidik, K. and Jonkman, J.N. (2002). A simple
tests of treatment efficacy. J. Biopharm. Stat. confidence interval for meta-analysis. Stat.
14 (2): 439–449. Med. 21 (21): 3153–3159.
Metelli, S. and Chaimani, A. (2020). Challenges Sidik, K. and Jonkman, J.N. (2006). Robust
in meta-analyses with observational studies. variance estimation for random-effects
Evid. Based Ment. Health 23 (2): 83–87. meta-analysis. Comput. Stat. Data Anal. 50
Page, M.J., McKenzie, J.E., and Higgins, J.P.T. (12): 3681–3701.
(2018). Tools for assessing risk of reporting Spiegelhalter, D.J., Abrams, K.R., and Myles, J.P.
biases in studies and syntheses of studies: a (2004). Bayesian Approaches to Clinical Trials
systematic review. BMJ Open 8 (3): and Health-Care Evaluation. Wiley.
e019703. Stanley, T.D. and Doucouliagos, H. (2015). Neither
Platt, R. and Lieu, T. (2018). Data enclaves for fixed nor random: weighted least squares
sharing information derived from clinical and meta-analysis. Stat. Med. 34 (13): 2116–2127.
administrative data. JAMA https://doi. Sterne, I.A.C., Egger, M., and Smith, G.D. (2001).
org/10.1001/jama.2018.9342. Investigating and dealing with publication
PROSPERO: https://www.crd.york.ac.uk/ and other biases in meta-analysis. BMJ 323:
prospero 101–105.
Rentsch, C.T., DeVito, N.J., MacKenna, B. et al. Sweeting, M.I., Sutton, A.I., and Lambert, P.C.
(2021). Effect of pre-exposure use of (2004). What to add to nothing? Use and
avoidance of continuity corrections in meta – adverse cardiovascular events in patients

analysis of sparse data. Stat. Med. 23 (9): treated with anti – IL – 12/ 23 biological
1351–1375. agents for chronic plaque psoriasis: a meta-
Sweeting, M.I., Sutton, A.I., and Lambert, P.C. analysis of randomized controlled trials. J.
(2006). Correction. Stat. Med. 25: 2700. Eur. Acad. Dermatol. Venereol. 27 (5):
The Yale University Open Data Access (YODA) 622–627.
Project. http://yoda.yale.edu/ Valsecchi, M.G. and Masera, G. (1996). A new
welcome-yoda-project challenge in clinical research in childhood
Training materials available freely on the ALL: the prospective meta – analysis strategy
internet for using risk of bias tools from the for intergroup collaboration. Ann. Oncol. 7
Cochrane Collaboration (https://training. (10): 1005–1008.
cochrane.org/resource/rob-20-webinar). Verde, P.E. and Ohmann, C. (2015). Combining
Turner, E.H., Matthews, A.M., Linardatos, E. randomized and non-randomized evidence
et al. (2008). Selective publication of in clinical research: a review of methods and
antidepressant trials and its influence on applications. Res. Synth. Methods 6 (1):
apparent efficacy. N. Engl. J. Med. 358 (3): 45–62.
252–260. Ziegler, R.G., Mayne, S.T., and Swanson, C.A.
Tzellos, T., Kyrgidis, A., and Zouboulis, C. (1996). Nutrition and lung cancer. Cancer
(2013). Re-evaluation of the risk for major Causes Control 7 (1): 157–177.
355
21
Studies of Medication Adherence

Julie Lauffenburger1, Trisha Acri2, and Robert Gross3
1
Brigham and Women’s Hospital and Harvard Medical School, Boston, MA, USA
2
Courage Medicine, Philadelphia, PA, USA
3
University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA
Introduction relationship where the patient implements the

provider’s recommendations.
The underuse of essential medications imposes Another reason that adherence has been dif-
significant clinical and financial burdens on ficult to define is that it is not a single behavior
healthcare systems. As many as half of patients but instead encompasses a set of behaviors
do not take their medications as prescribed, over time. One common taxonomy developed
resulting in an estimated $100 billion in excess by a scientific consensus group classifies
annual spending in the US alone. Without adherence along three phases: (i) initiation, (ii)
accurate measurements of adherence for implementation, and (iii) persistence (see
research and practice, the problem will remain Figure 21.1).
underappreciated and poorly addressed. In Initiation describes initial engagement with
this chapter, we describe the importance the prescribed medication. As many as 30% of
of adherence in pharmacoepidemiologic newly-prescribed therapies are never actually
research, methods for measuring adherence, filled by patients, which is often referred to as
methodologic issues that arise once adherence primary non-adherence. Implementation rep-
has been measured, and future directions. We resents how well patients follow the prescribed
focus here on examples from HIV and cardio- regimen after they have begun treatment.
metabolic diseases because these areas have While varying greatly across diseases, approxi-
been major focuses of adherence research. mately 50% of patients are thought to not cor-
Despite its importance, medication adher- rectly follow prescribed regimens. Persistence
ence is difficult to define. Earlier research has refers to how long the patient continues to fol-
used the term Compliance, or “the extent to low the regimen. Poor adherence can occur
which the patient’s dosing history conforms to along any of these phases.
the prescribed regimen.” However, this term This taxonomy helps distinguish between
implies that patients passively “conform” to patients who never initially fill a prescription
the prescriber’s directions, so the term adher- from patients who occasionally forget to take
ence is now strongly preferred. Adherence bet- doses and those who take a medication regu-
ter conveys the idea of a patient–provider larly at first but then later discontinue.
356 21 Studies of Medication Adherence
Clinician Patient Patient

Patient fills
writes adheres to continues
prescription
prescription therapy therapy
Phase: (1) Initiation Phase: (2) Phase: (3) Persistence

Implementation
Synonyms: Acceptance, Synonym: Non-
Primary adherence Synonyms: Execution discontinuation
Figure 21.1 Phases and taxonomy of adherence.
Classifying all these patients simply as “non- can also be the specific focus of pharmacoepi-
adherent” ignores that they may differ with demiologic research.
respect to treatment outcomes and likely have Non-adherence can be intentional or unin-
different adherence barriers requiring differ- tentional. Studies have identified many poten-
ent interventions. tial barriers to adherence, broadly categorized
as patient, system, and medication-specific fac-
tors. Common patient barriers consist of forget-
ting to take the medication, lack of knowledge
C
or health literacy, and psychosocial factors such
Addressed by as depression and lack of social support. System
Pharmacoepidemiologic barriers include costs of the medication as well
Research as logistical difficulty in obtaining medication,
including drug shortages in some settings. Key
Adherence research confronts the truism medication-specific factors include regimen
attributed to former US Surgeon General C. complexity and adverse effects. “Pill fatigue”
Everett Coop, MD that “drugs don’t work in can also occur, in that adherence can decrease
patients who don’t take them.” Measuring over time. This may be due to, for example,
adherence is essential to address several issues being emotionally overwhelmed by taking
in the interpretation of studies of beneficial medication or no longer having a sense of
and adverse effects of medications. In rand- urgency about the medical issue, particularly
omized trials, poor adherence to the drug being when patients are observed for periods longer
tested can lead to underestimates of drug effi- than in typical trials and they have not experi-
cacy. Information about adherence also allows enced disease complications prevented by the
for a more accurate assessment of drug safety medications themselves. It is also common that
because those who do not take the drug cannot the optimal adherence seen early in therapy
experience its toxicity. Poor adherence may may decrease over time.
itself also be a marker of toxicity or adverse While missed doses are a more common
events. adherence problem, taking extra doses can also
Once a medication is marketed, information occur. Extra doses of drugs with a narrow ther-
from clinical trials gives only a limited view of apeutic window, such as warfarin for antico-
how drugs are used by patients. Patients who agulation, may result in toxicity. Patients may
volunteer for clinical trials are often more also take extra doses of opioids prescribed for
motivated than those in usual care, so assess- the treatment of pain because of inadequate
ing adherence in observational studies pro- relief or for recreational purposes.
vides a more “real world” estimate of Measuring adherence can also be useful for
adherence. Finally, because adherence itself is determining thresholds of how much medica-
a major determinant of treatment outcomes, it tion must be taken to obtain desired clinical
outcomes, which likely differ by drug and dis- shortly after initiation. Many studies focus on
ease. In hypertension, taking at least 80% of incident users, but there are situations in
prescribed medication has been an acceptable which studying prevalent users is more impor-
standard for blood pressure control. However, tant because new initiators are only a small
in HIV, higher thresholds are often necessary. proportion of all patients on a therapy at a
Even though 80% of doses taken may not be given time. Regardless of measurement
the optimal universal cut-point for acceptable approach, as discussed later, the discovery of
adherence, this threshold persists across non-adherence in clinical settings can be
research and quality measures. Therefore, the embarrassing for patients. Thus, knowledge
default adherence goal should be to encourage that one’s adherence is being monitored risks
the patient to take as many prescribed doses as influencing the behavior it is measuring (i.e. a
possible, and future research should focus on Hawthorne effect). When selecting a method
identifying more empiric and robust dose– and describing results, it is highly recom-
response thresholds for various drug-disease mended to use standardized reporting
settings. approaches based on the adherence taxonomy
presented in Figure 21.1.
M
to be Addressed by urrently Available
C
Pharmacoepidemiologic Solutions
Research
Specific Techniques for Measuring
Adherence
While the gold standard for measuring adher-
ence to pharmacotherapy is directly observed Self-Reports
therapy, this approach is only practical in lim- The most common adherence measurement
ited settings, such as the administration of a method has been patient self-report or asking
novel agent in a controlled environment. There respondents about their adherence behaviors.
are many different methods for measuring Self-reported metrics are simple, relatively
medication adherence, and each method has inexpensive, quick and feasible and can be
strengths and weaknesses. The most appropri- obtained over the telephone, in person, or with
ate method depends upon the situation and paper or electronic surveys. Several validated
precision needed for the measurement. Some methods for assessing self-reported adherence
methods require more intensive patient-level are described below.
contact while others provide more granular Self-reported adherence measures range
data about timing of dose-taking. For example, from one-item questions inquiring about the
prospective clinical trials can use many differ- frequency of missed doses to longer multi-item
ent methods. However, options are more lim- assessments evaluating beliefs associated with
ited in retrospective studies using databases. adherence and identifying barriers to non-
Thus, the use of multiple measures or sources adherence. Most self-reported measures
of data may be useful to confirm findings. involve count or estimation-based recall
For all approaches, the interpretation of focused on the implementation phase, in
adherence findings may also change depend- which respondents report the number of doses
ing on whether incident users or prevalent missed or taken within an interval or to esti-
users of medication are examined, as adher- mate their overall execution of adherence.
ence tends to be higher among prevalent users, Numerous validated self-reported adherence
in part because discontinuation is highest scales exist in the English language. Perhaps
the most common self-reported adherence tool a dherence in pharmacoepidemiology. Data

historically used is the eight-item Morisky can be obtained from health insurers (i.e.
Medication Adherence Questionnaire (MMAS- administrative claims data) or from pharma-
8), but it requires licensing fees. The adult cies or patients directly. Pharmacy dispensing
AIDS Clinical Trials Group (ACTG) adherence data are generally considered to be accurate
questionnaire, three-item tool by Wilson et al., measures of drug dispensing because the dis-
and the Brief Medication Questionnaire are penser (e.g. a pharmacy) would not receive
other examples of common, publicly-available insurance reimbursement if the medication
tools that explore both behaviors and barriers fills are not recorded and guilty of fraud if bill-
to adherence. Overall, the choice of measure ing for medications not actually dispensed.
may depend on its purpose (e.g. clinical use or There are several different methods for
research), burden to patients, and disease measuring adherence using pharmacy dis-
states in which it has been validated. Self- pensing data. In all approaches, adherence is
report may also be the easiest method for clini- measured indirectly based on patterns of medi-
cians to administer and best way to identify cation dispensings (using medication name,
reasons for poor adherence for targeting dispensing date, and days supplied) by gener-
interventions. ating a “drug supply diary” that strings
Self-reported adherence measures are mod- together consecutive medication dispensings
erately correlated with methods using elec- based on the dates on which medication are
tronic drug monitors (EDM) or pharmacy dispensed to the patient and the duration of
dispensing data (described later), though con- the supply dispensed. This supply diary can
cordance can vary depending on the patient’s adjust for overlapping fills (e.g. truncating the
level of adherence or the measurement win- days supplied for medications which are
dow. However, because of potential over- refilled before the medication supply from the
reporting, self-reported measures generally prior dispensing would have been exhausted)
have high specificity and low sensitivity; that and any known interruptions that may have
is, self-reported non-adherence is generally occurred (e.g. by hospitalization). When gener-
accurate, while high self-reported adherence ating the supply diary, researchers generally
may not be accurate. consider medications that are chemically
Self-reported adherence measurements have related and not intended for use in combina-
other limitations, such as being limited by the tion to be interchangeable (e.g. two beta-
ability to recall missed doses and being subject blockers). For example, patients may initiate
to social desirability bias (i.e. over-reporting one beta-blocker and later switch to a different
adherence to please providers or researchers), beta blocker. In this case, beta-blocker adher-
which can be potentially reduced by adminis- ence is often measured continuously, rather
tering questions at a kiosk or tablet/computer. than separately measuring adherence to each
Reading instructions and questions aloud and medication, to generate one continuous expo-
including photographs of medicines can sure episode. Sometimes, medications within
reduce literacy barriers. Self-report is also the same disease state but chemically different
unlikely to provide a precise measurement of (e.g. beta-blockers and calcium channel block-
timing or patterns of dose-taking. ers) could be considered interchangeable.
Several types of adherence metrics can be
Pharmacy Dispensing Data calculated using these data, including imple-
Pharmacy dispensing measurement was pio- mentation and persistence measures (see
neered in the late 1980s with wide use in Figure 21.2). In the most common approach,
chronic diseases and is one of the most the proportion of days that patients had an
commonly used modalities for measuring available supply of medication, or the
Start of End of
follow-up follow-up
Patient 1 Fill Fill Fill Fill Fill Fill
Patient 2 Fill Fill Fill Fill Fill Fill
Assume here each fill is 30 days, follow-up is 365 days, and supply diary is adjusted for overlap
Measuring implementation (e.g., Proportion of Days Covered [PDC])

- Patient 1: PDC = (30 + 30 + 30 + 30 + 30 + 30)/365=0.49 or 49.3%
- Patient 2: PDC = (30 + 30 + 30 + 30 + 30 + 30)/365=0.49 or 49.3%
Measuring persistence
Refill gap method (90-day)
- Patient 1: “Persistent”
- Patient 2: “Non-persistent”
Treatment anniversary method
Figure 21.2 Example of implementation and persistence measures in pharmacy dispensing data.
roportion of days covered (PDC), is calculated.

p whether a dispensation overlaps with the end
The PDC is calculated by dividing the number of a follow-up period (i.e. the “treatment anni-
of days with an available supply of medication versary method”) or a “refill gap method” that
by the number of days in the interval. Other measures the availability of drug supply at a
approaches include calculating the medication fixed time after the last dispensing (e.g. whether
possession ratio (MPR). MPR is calculated as patients have a gap of at least 30 or 60 days with
the quotient of the total number of days sup- no medication after the supply is exhausted).
plied of all dispensings in each interval for the Whichever method is chosen, investigators
medication under investigation and the total should conduct sensitivity analyses of the “gap
number of days in the interval. The primary dif- rule.” However, it is difficult to disentangle
ference between PDC and MPR adherence met- clinically-directed medication discontinuation
rics is how overlapping days supplied of the from discontinuation against provider
same medication are handled; the MPR focuses recommendation.
on days supplied while PDC focuses on days Compared with self-report, pharmacy dis-
covered. The specific approach is typically pensing data are not biased by poor recall and
determined by researcher preference; however, can be obtained from computerized records.
PDC is becoming increasingly recommended Another advantage is that the data are availa-
by healthcare quality organizations and may be ble on large numbers of patients (often mil-
more accurate for measuring adherence to lions in a single database). However, the
multiple medications. Regardless, the different quality of data may be less accurate when
approaches yield similar results and are simi- tracking is less crucial for reimbursement or if
larly associated with clinical outcomes. prescriptions are obtained outside of insurance
Approaches for persistence include evaluat- plans. To overcome this, some approaches
ing whether clinically-meaningful treatment compile data from pharmacies directly to cap-
gaps or discontinuations are observed in the ture all medications dispensed to patients and
dispensing data. Approaches include evaluating not just those paid by insurers. However, if
patients use different pharmacies and the data can be done impulsively before a visit.
are not compiled, then the records may be Unannounced pill counts, in person or by tel-
incomplete and logistically more challenging ephone, are alternatives to mitigate this type
to acquire. of misclassification. During visits, subjects
Conversely, for questions related to one-time review the contents of each of their pill bot-
prescriptions (e.g. short courses of antibiotics), tles. Of course, this approach is also suscepti-
while viable to study initiation, these data may ble to intentional deception; however, the
not be useful for implementation since repeat estimated adherence from pill bottle review
dispensings are required to calculate the has been shown to be associated with treat-
amount of medication consumed. Also, in the ment response. The time for both staff and
US, pharmacy dispensing data are generally participants is a potential disadvantage of pill
only accurate for outpatient medications, counting and additional source of error. In
because medications are not paid for sepa- addition, missing data can result when
rately during hospitalizations. Furthermore, patients do not bring in their pill bottles or
pharmacy dispensing data also do not measure have them available during telephone calls.
actual pill consumption, so they cannot be Reinforcing the importance of accuracy with
used when the timing of missed doses is piv- staff is vital to ensure validity.
otal. However, the estimation of adherence
with pharmacy dispensing data has been Medication Diaries
shown to be valid for chronic medications Although the measures described above sum-
where measuring overall exposure between marize how much medication was taken over a
refills is clinically relevant. Moreover, phar- specified time period, they provide no detail on
macy dispensing measures of adherence have the timing of missed doses. Missing doses may
also been shown to be strongly associated with have different consequences depending on
clinical outcomes. whether they were missed consecutively or if
they were missed at separate times (see
Pill Counts Figure 21.2). One solution is medication dia-
While less commonly used, adherence can also ries where participants keep a record of the
be measured indirectly by pill counts, which date and time of each dose of medication and
involves counting the number of pills available its timing with food. These data can be col-
for a patient. Pill counts are like pharmacy dis- lected handwritten or electronically. Diaries
pensing data in that percent adherence is cal- are regularly used in pediatric patients and are
culated by dividing the days’ supply consumed particularly useful for medications like insulin
by the number of days observed. Data collected or inhalers that are otherwise difficult to meas-
include the dispensing date, quantity dis- ure. However, diaries are susceptible to both
pensed, number of pills per dose, and number overreporting and underreporting of adher-
of pills left in the bottle, adjusted for doses ence. Social desirability results in patients list-
taken that day and any additional pills left over ing doses even though they were not taken, but
from the last count. the potential is lessened somewhat by the bur-
Like adherence measures estimated using den of creating a detailed falsified record. In
the medication dispensing date and days sup- fact, the risk of underreporting may be greater
plied (e.g. PDC), pill count data cannot deter- because of the burden of tracking each dose. It
mine if the medication was consumed. is also not easy to employ this method at scale
However, they do provide direct evidence that for larger studies. Newer approaches are
the medication was not taken when pills are exploring the use of apps on enabled smart
left over. Pill counts are susceptible to phones to track more nuanced medication-
deception since “dumping” pills is simple and taking patterns.
Electronic Drug Monitoring Technology ingestion. However, the use of drug concentra-
Electronic drug monitors (EDMs) have similar tions to measure adherence is limited by vari-
advantages as medication diaries but are less ability across patients (i.e. absorption,
susceptible to deception or forgetting to docu- distribution, metabolism, and clearance) and
ment doses taken, because they provide time- the need to potentially collect concentrations
stamped data about doses taken. While there frequently to gain a fuller picture of adherence
are several different hardware options, EDMs behaviors. Measurement of drug concentra-
employ electronic date/time stamp technology tions in hair using liquid chromatography and
triggered by opening a container (i.e. pill bot- confirmed by mass spectrometry can be a use-
tle), puncturing a blister pack, or ingesting a ful indicator of long-term medication expo-
dose, and data are uploaded to a computer or sure. Unfortunately, many assays are
smartphone via hardwire or wireless linkage. unavailable commercially. Furthermore,
While EDMs are less susceptible to decep- serum drug levels may not be the relevant
tion than self-report, they could theoretically measure for many drugs when the site of
be more susceptible than pharmacy dispensing action is elsewhere (e.g. intracellularly rather
data. However, it is highly unlikely that sub- than in serum or hair). Cost and patient incon-
jects will open and close the monitor to record venience may also be limitations.
doses over long periods of time without actu-
ally taking the medication, though this does Measuring Primary Adherence
occasionally happen accidentally. Though Each of the approaches described above have
EDM technology is advancing rapidly, the focused on later adherence phases (e.g. imple-
packaging and cost of EDMs can still be bur- mentation and persistence). Without linkage
densome and expensive. For example, they to other types of data (e.g. electronic health
often preclude the use of pillboxes by generally records that include provider medication
requiring that the medication remain in the orders), it can be difficult to evaluate medica-
package until taken. However, EDMs could be tion initiation (primary adherence) in dispen-
used even when the medication is not kept in sation data without knowledge of what was
the container, where participants are asked to prescribed. Newer approaches are beginning to
open the empty bottle whenever they take link these data sources to allow better assess-
medications from a pillbox. Newer approaches ment of the full cascade. On their own, elec-
integrate EDMs with text messaging technol- tronic health record data limited to physician
ogy to remind patients when they miss doses. orders are not useful at evaluating adherence
In 2015, the US Food and Drug because they do not provide information about
Administration approved the first ingestible actual patient behaviors. By contrast, self-
sensor technology that measures actual intake report may allow for easier study of primary
time through ingestion of a medication that adherence.
communicates with an adhesive patch and
sends a signal to the team monitoring adher- Measuring Adherence to Non-pill
ence. Other research is exploring the utility Formulations
and accuracy of adherence measures in which Measuring adherence to non-pill formulations
patients take date and time stamped photo- can be difficult, largely because they are gener-
graphs of themselves or their pills each time ally administered with a variable dosing sched-
they take a dose. ule. Injectable medications like insulin may be
administered on a sliding scale, with doses
Drug Concentrations adjusted as needed, so measuring adherence
Identifying the presence of a drug in plasma or using indirect dispensing data may be impre-
other tissues provides direct evidence of drug cise. For these, persistence-based measures
may presenting be the most accurate. For in intention-to-treat analyses. To compensate,

inhaled medications, dispensing data could be Phase III trials may inflate sample sizes to
used for ones with specific schedules (e.g. tio- account for variability in drug exposure or
tropium). Medication diaries and self-report incorporate run-in periods to minimize poor
could theoretically be used but have the same adherence. Of course, medication adherence
issues as pill formulations. EDMs have been itself can also be a primary or secondary out-
used for metered dose inhalers and ophthalmic come in randomized trials, particularly for
solutions; while the monitors increase the size studies of interventions. The inclusion of
of packaging, the devices cannot be taken out adherence data in analyses of trials is particu-
of the package unlike pill formulations. For larly important when a treatment fails.
transdermal formulations in patches (e.g. nico- Reasons for failure might be attributable to
tine, testosterone), adherence based on dispen- either lack of biological effect or non-
sation data is a viable option because the adherence, so unless adherence is measured,
supply is typically fixed. However, for creams the results of the trial will be only partly use-
and ointments, because the amount used at ful. (See Case Example 21.1)
each application varies by the size of the lesion Similarly, observational studies of compara-
being treated or the size of the individual, self- tive effectiveness and safety of medications
reports and medication diaries may be the only often benefit from measuring the relationship
currently viable options. Measuring adherence between adherence and treatment response.
will continue to be a challenge for newer non- “As-treated” analyses of safety evaluations
pill formulations, including biologics, and in often censor follow-up in patients who discon-
disease states in which both oral and injectable tinue therapies for reasons other than toxicity
formulations are used interchangeably (e.g. to decrease bias toward the null. Secondly,
osteoporosis). marginal structural modeling approaches
often include adherence as a time-varying
exposure. Exploring the relationship between
nalysis Issues
A adherence observed between comparators may
enrich the conclusions from these studies.
in Adherence
Use of Adherence Data in Clinical Selecting Adherence Intervals

Trials and Comparative
For all adherence measures, one must choose a
Effectiveness Studies
pre-specified window for assessing and evalu-
In analyzing trials, the standard approach ating adherence. While the choice of interval
remains intention-to-treat. This approach lim- length depends on the research goals, in gen-
its the introduction of bias and makes the eral, monitoring adherence over shorter inter-
results more generalizable. In clinical trials, vals would be desirable, because interventions
adjusting results for adherence is complicated can be more rapidly implemented. Selecting
by the fact that the behavior of being adherent the interval depends on two important factors:
itself is associated with better outcomes (i.e. a pharmacokinetics/pharmacodynamics and
healthy user effect). While clinical trial par- granularity of the adherence measurement.
ticipants may be more motivated to adhere to For drugs with short half-lives and onset of
treatments than those in clinical practice, action, short intervals are likely to be more
non-adherence occurs for all types of self- clinically relevant than when drugs have
administered therapies. Missed doses will typ- longer half-lives and onsets of action. Shorter
ically make the active drug less effective and intervals can be calculated for techniques that
diminish observed differences versus placebo can accurately assess adherence over short
Analysis Issues in Adherenc 363
Case Example 21.1 Non-adherence as a Key Factor for the Success of HIV Pre-Exposure
Prophylaxis Trials
Background ence was 84% or greater and was meas-

●● Pre-exposure prophylaxis (PrEP) may pre- ured by pill counts and self-report.
vent HIV transmission, especially in sero- ●● Trials with lower success had lower adher-
discordant couples (e.g. HIV- positive ence rates or did not report adherence.
patients with HIV-negative partners). ●● Variations in adherence patterns (such as
●● Several antiretroviral-based PrEP drug in Figure 21.2) or administration prefer-
therapies, including oral and vaginal gel ences, such as for oral or gel therapies, was
formulations, have been studied since the also thought to influence adherence.
late 2000s in clinical trials in sub-Saharan
Strength
Africa and other regions of clinical need.
This was a comprehensive summary of fac-
Question tors for observed results across trials.
Four early trials of antiretroviral-based PrEP
Limitations
showed success in reducing HIV transmis-
Different techniques of measuring adher-
sion and were also well-tolerated by patients.
ence (e.g. pill counts and self-report) limited
However, three subsequent trials found sur-
the ability to make clear comparisons across
prisingly negative results of no effect of PrEP,
trials.
including some that were stopped for
futility. Key Points
●● Non- adherence resulted in surprisingly
Approach
negative trials of a therapy that was ulti-
Researchers in 2012 reviewed these seven
mately found to be efficacious, which
trials’ data on adherence to PrEP therapies
almost halted efforts to continue to study
and determined the degree of variation in
PrEP.
adherence and potential explanations for
●● Efforts to ensure optimal adherence
the mixed findings.
appear critical to the success of PrEP.
Results ●● Additional efforts to expand PrEP formula-
●●Adherence to PrEP was found to vary sub- tions, such as injectables, are underway;
stantially depending on the trial. In trials these options may help reduce non-
that demonstrated effectiveness, adher- adherence once trials are completed.
periods of time, such as electronic data moni- to observe variation since, by definition,
tors. By contrast, for measures derived from patients are classified as fully adherent (100%)
dispensing data, adherence analysis intervals for the first 30 or 90 days. In general, choices
must be longer because adherence is based on for an adherence interval should be made
evaluating patterns between medication dis- based on pharmacokinetic/pharmacodynamic
pensing dates in conjunction with the days data.
supplied per dispensing (e.g. 30 days, like in
Figure 21.2). The vast majority of dispensings
Statistical Analysis
for chronic medications in the United States
are for 30-day supplies, and increasingly The simplest approach to summarizing adher-
90 days. Accordingly, measuring adherence in ence is the percent of doses taken (or missed).
intervals shorter than 180 days make it difficult For electronic monitors, because the timing of
each dose is available, percent of doses taken components of the regimen. Another common
“on time,” standard deviation of time between approach is to measure adherence at the thera-
doses, or maximum time gap between doses peutic class levels individually (e.g. measuring
can be calculated. For metrics in pharmacy dis- adherence to beta-blockers and calcium chan-
pensing data, analyses focus on percentage of nel blockers separately) and then “average”
available medication or gaps between dispen- adherence across the entire chronic condition
sations. Self-report typically focuses on the for patients exposed to any medication for that
proportion of doses the patients have taken. disease state (e.g. hypertension).
Whichever metric is used, one must choose
whether to include adherence as a continuous
Time-Varying Nature of Adherence
or dichotomous variable. As previously
and Trajectory Modeling
described, dichotomous thresholds must con-
sider both the likelihood of failure and clinical Adherence is a non-static behavior, and meth-
consequences of treatment failure. Few thresh- ods are needed to capture changes in adher-
olds have been established based on evidence, ence over time. This phenomenon has
yet in research and quality improvement historically been ignored in studies that meas-
efforts, to dichotomize these adherence varia- ure adherence only once or over short inter-
bles, patients are often defined as fully adher- vals. Even when measured longitudinally,
ent if they take 80% of prescribed doses. So, adherence data are often averaged. However,
for example, both patients in Figure 21.2 would patients may experience substantial increases
be “non-adherent” because they had a and decreases in adherence not fully captured
PDC < 80%. Often, evaluating differences in by composite measures. Consider, for example,
adherence on a continuous scale would be the two patients presented in Figure 21.2,
clinically more meaningful than binary meas- where Patient One takes the medication regu-
ures, although adherence measures are typi- larly but with intermittent gaps while Patient
cally not normally distributed. In addition, the Two takes the medication perfectly for the first
presence of non-adherence values of zero will six months but then discontinues. Both
make it impossible to log transform continu- patients had the same calculated adherence
ous adherence data. Alternatively, when nei- (~50%) but very different medication use pat-
ther dichotomous nor continuous measures terns. Composite, cross-sectional measures
capture the clinically-relevant dose–response obfuscate the potential for each patient to
relationship, assigning ordinal adherence cat- experience different health outcomes and
egories (e.g. <60%, 60–<80%, 80–<100% etc.) require different adherence interventions.
may be sometimes preferable. Other research- Advanced statistical methods are beginning
ers have sometimes used quantile regression to to take advantage of repeated measurements in
overcome assumptions of linearity. adherence data, particularly in dispensing
In addition, evaluating regimens with multi- data, to enhance analysis beyond composite
ple medications poses analysis challenges. A measures. One such method applied is group-
final consideration is how to accurately char- based trajectory modeling which estimates
acterize adherence to multiple medications for changes in an outcome that is measured
the same condition (e.g. anti-hypertensives). repeatedly over time and identifies individuals
Some classify adherence based on optimal with similar longitudinal patterns. This
adherence to at least one medication for that approach fits a semiparametric (discrete) mix-
disease state (e.g. hypertension) to be fully ture model and assigns groupings in longitudi-
adherent, although this misclassifies individu- nal data (e.g. monthly PDC) based on
als who are non-adherent to some but not all probability distributions for a pre-specified
Key Point 365
number of groups. The probability of belong- The Future

ing to each potential group is modeled as a
multinomial logistic regression, and within Adherence studies are likely to advance in sev-
each group, adherence is modeled as a func- eral ways. Because optimal adherence thresh-
tion of time. The statistical output includes olds may differ across individuals and diseases,
each individual’s estimated probabilities of researchers are beginning to explore personal-
group membership and estimated trajectory ized adherence targets. Novel approaches
curve of adherence over time for each group. using other types of data are likely to emerge as
However, trajectories provide general patterns well, including the use of more advanced
for adherence behaviors; that is, no one indi- microelectronic technology, often linked with
vidual follows the exact pattern described by communication systems that both identify and
the trajectory of the group to which they are report non-adherence, or the enhancement of
assigned. mobile and smartphone technology for track-
ing and intervening on adherence. Refinements
Prediction of Adherence to currently available electronic monitors will
for Interventions also likely include more convenient packaging
that can both help with adherence (e.g. a
Unfortunately, low rates of adherence have per- reminder or organizer system) and provide
sisted despite extensive efforts to identify and two-way personalized communication with
predict patients at risk of poor adherence with patients.
the goal of developing interventions to improve Hopefully, with greater recognition of the
adherence. Despite the expansion of databases importance of non-adherence, more research
with rich patient data, prediction of future will be conducted over the next several decades
adherence remains poor. Traditional to solve some of these problems as well as
approaches have focused on clinical and demo- develop better approaches to improving adher-
graphic factors at the time of medication initia- ence so that evidence-based medications can
tion, with discriminative ability that is modest be optimally used.
at best even with dozens of predictor variables
(e.g. c-statistics generally ranging between
0.60–0.75). Even machine learning, with the
capability of measuring complex interactions Key Points
among predictor variables, has not led to drasti-
cally improved prediction, likely because the ●● Nonadherence occurs in as many as half of
true factors associated with poor adherence are patients, resulting in avoidable costs and
usually not observable in databases. One of the adverse clinical outcomes. Accurate adher-
more successful approaches has been evaluat- ence measurement must be incorporated
ing patterns of medication filling shortly after into research and clinical practice or the
initiation. For example, in pharmacy dispens- problem will remain underappreciated and
ing data, researchers have found that failing to poorly addressed.
refill in the second and third months after ini- ●● While missing doses is the more common
tiation is highly predictive of poor adherence adherence problem, taking extra doses can
over the following year (i.e. past adherence pre- also be an issue in some settings, such as
dicts future adherence). Predictions of adher- anticoagulation and pain medication.
ence by providers has also been shown to be no ●● Potential barriers to adherence include
better than chance, so they should not be used patient-level factors, system-level factors,
routinely in adherence studies or in practice. and medication-specific factors.
●● Currently available techniques for measur- of the appropriate duration, time period, and
ing adherence include self-report, pharmacy metrics for adherence.
refill measures, pill counts, medication dia- ●● Adherence is a non-static behavior, and
ries, electronic drug monitors, and drug con- methods are needed to capture changes in
centrations. Measuring adherence to adherence over time.
multiple medications and non-pill formula- ●● The inclusion of adherence data in analyses
tions has special challenges. of clinical trials and studies of comparative
●● Measurement of adherence can be used to effectiveness is particularly important when
determine the threshold of how much medi- a treatment fails.
cation must be taken to obtain the desired ●● The effectiveness of existing adherence
clinical outcome. interventions remains modest. Better meth-
●● Because adherence behavior varies over ods for detecting and addressing nonadher-
time, analysis of data requires consideration ence will be welcome developments.
Further Reading
Acri, T., Grossberg, R., and Gross, R. (2010). approach to classifying and predicting
How long is the right interval for assessing long-term medication adherence. Med. Care
antiretroviral pharmacy refill adherence? 51: 789–796.
JAIDS 54 (5): e16–e18. Gross, R., Bilker, W.B., Friedman, H.M., and
Alfin, S.D., Pradipta, I.S., Hak, E., and Denig, P. Strom, B.L. (2001). Effect of adherence to
(2019). A systematic review finds newly initiated antiretroviral therapy on
inconsistency in the measures used to plasma viral load. AIDS 15: 2109–2117.
estimate adherence and persistence to Gross, R., Bellamy, S.L., Chapman, J. et al. (2013).
multiple cardiometabolic medications. J. Clin. Managed problem solving for antiretroviral
Epidemiol. 108: 44–53. therapy adherence: a randomized trial. JAMA
Conn, V.S. and Ruppar, T.M. (2017). Medication Intern. Med. 173: 300–306.
adherence outcomes of 771 intervention trials: Grossberg, R., Zhang, Y., and Gross, R. (2004). A
systematic review and meta-analysis. Prev. time-to-prescription-refill measure of
Med. 99: 269–276. antiretroviral adherence predicted changes in
Choudhry, N.K., Shrank, W.H., Levin, R.L. et al. viral load in HIV. J. Clin. Epidemiol. 57:
(2009). Measuring concurrent adherence to 1107–1110.
multiple related medications. Am. J. Manag. Lauffenburger, J.C., Balasubramanian, A.,
Care 15: 457–464. Farley, J.F. et al. (2013). Completeness of
De Geest, S., Zullig, L.L., Dunbar-Jacob, J. et al. prescription information in US commercial
(2018). ESPACOMP medication adherence claims databases. Pharmacoepidemiol. Drug
reporting guideline (EMERGE). Ann. Intern. Saf. 22: 899–906.
Med. 169: 30–35. Lauffenburger, J.C., Isaac, T., Bhattacharya, R.
Franklin, J.M., Shrank, W.H., Lii, J. et al. (2016). et al. (2020). Prevalence and impact of having
Observing versus predicting: initial patterns of multiple barriers to medication adherence in
filling predict long-term adherence more nonadherent patients with poorly controlled
accurately than high-dimensional modeling cardiometabolic disease. Am. J. Cardiol. 125:
techniques. Health Serv. Res. 51: 220–239. 376–382.
Franklin, J.M., Shrank, W.H., Pakes, J. et al. Lo-Ciganic, W.H., Donohue, J.M., Thorpe, J.M.
(2013). Group-based trajectory models: a new et al. (2015). Using machine learning to
Further Readin 367
examine medication adherence thresholds assessment using centralized pharmacy

and risk of hospitalization. Med. Care 53: records: description and validation. Med. Care
720–728. 26: 814–823.
Mehta, S.J., Asch, D.A., Troxel, A.B. et al. (2019). Van der Straten, A., Van Damme, L., Haberer,
Comparison of pharmacy claims and J.E., and Bangsberg, D.R. (2012). Unraveling
electronic bottles for measurement of the divergent results of pre-exposure
medication adherence among myocardial prophylaxis trials for HIV prevention. AIDS
infarction patients. Med. Care 57: e9–e14. 26: F13–F19.
Nguyen, T.M., La, C., and Cottrell, N. (2014). Vrijens, B., De Geest, S., Hughes, D.A. et al.
What are validated self-report adherence (2012). A new taxonomy for describing and
scales really measuring?: a systematic review. defining adherence to medications. Br. J. Clin.
Br. J. Clin. Pharmacol. 44: 427–445. Pharmacol. 73: 691–705.
Osterberg, L. and Blaschke, T. (2005). Adherence Wilson, I.B., Lee, Y., Michaud, J. et al. (2016).
to medication. N. Engl. J. Med. 353: 487–497. Validation of a three-item self-report measure
Steiner, J.F., Koepsell, T.D., Fihn, S.D., and Inui, for medication adherence. AIDS Behav. 20:
T.S. (1988). A general method of compliance 2700–2708.
368
22
Advanced Approaches to Controlling Confounding

in Pharmacoepidemiologic Studies
Sebastian Schneeweiss1 and Samy Suissa2
1
Harvard Medical School and Brigham & Women’s Hospital, Boston, MA, USA
2
McGill University and Jewish General Hospital, Montreal, Quebec, Canada
Introduction design and analytic strategies. This chapter

provides an overview of selected options that
The past two decades have witnessed advances fit typical pharmacoepidemiologic data sources
in the design and analysis of epidemiologic and study questions.
studies. In this chapter, we introduce some of
these approaches with a focus on confounding
control, one of the major methodological chal- Methodological Problems
lenges in pharmacoepidemiology. to be Addressed by
Pharmacoepidemiologic
Research
C
Addressed by The availability of large longitudinal patient-
Pharmacoepidemiologic level healthcare databases make the new-user
Research cohort design a natural design choice as a start-
ing point that mimics the classical parallel
Pharmacoepidemiologic analyses are in princi- group controlled trial, except of course for the
ple not different from analyses in other subject randomized treatment assignment (Figure 22.1).
areas within epidemiology. They are typically Efficient sampling within such cohorts, includ-
concerned with valid estimation of the causal ing case–control, case-cohort, and 2-stage sam-
effects. Some issues specific to pharmacoepi- pling designs are important extensions.
demiology stem from the constraints of sec- Bias can be reduced by appropriate design
ondary data sources, in particular large choices. Considerations about the sources for
electronic longitudinal healthcare databases exposure variation will lead to decisions on the
(see Chapter 10). Another difference is the appropriate study design. In a hypothetical
close interdependency of treatment choice causal experiment, one would expose a patient
with health status, severity of disease, and to an agent and observe the outcome, then
prognosis. Pharmacoepidemiologists try to rewind time, leave the patient unexposed, and
reduce bias by appropriate choices of study keep all other factors constant to establish a
Figure 22.1 Principle of the new First-line use drug A

No drug use
user design and its variations when
studying second line therapies. S
Source: Reproduced from Schneeweiss
(2010) with permission from John First-line use drug B
Wiley & Sons Ltd.
Switch to 2nd line drug C

Common 1st line use drug A
S
Switch to 2nd line drug D
Add drug C to A
Common 1st line use drug A
S
Add drug D to A
Figure 22.2 Study design choice Exposure variation

by source of exposure variation. within patients
Source: Reproduced from Schneeweiss yes no
(2010) with permission from John
Wiley & Sons Ltd.
Case-crossover Exposure variation
study between patients
yes no
Crossover trial
Cohort study Exposure variation
between providers
yes
Randomized
controlled trial Instrumental
variable analysis
Cluster
randomized trial
counterfactual experience. Since this experi- allocation of treatments between different

ment is impossible, the next logical expansion patients. For most pharmacoepidemiologic
is to randomly introduce or observe exposure studies, we utilize variation in exposure among
variation within the same patient but over time individual patients, and we will therefore use a
(Figure 22.2). If we observe sporadic drug use cohort study design. Exposure variation among
resulting in fluctuations of exposure status higher-level entities (provider, region, etc.) can
within a patient over time, if that drug has a be exploited using instrumental variable (IV)
short washout period, and if the adverse event analyses (described below).
of interest has a rapid onset, then we may In a cohort design, there are several advan-
consider a case-crossover design or related tages to identifying patients who start a new
approach (see below). Another option is random drug. As patients in both the study group and
370 22 Advanced Approaches to Controlling Confounding in Pharmacoepidemiologic Studies
the comparison group have been newly started potential confounders and apply efficient sam-
on medications, they have been evaluated by pling designs to reduce the time and resources
physicians who concluded that these patients it takes to complete the study, and (ii) analytic
might benefit from the newly prescribed drug. approaches that try to make better use of the
This makes treatment groups more similar in existing data with the goal of improved control
characteristics. The clear temporal sequence of of confounding.
confounder ascertainment before treatment
initiation in an incident user design also avoids Efficient Sampling Designs within
mistakenly adjusting for consequences of treat- a Cohort Study
ment. Studying new users is also useful when
comparing newly marketed to existing medica- In any cohort study, the resources needed to col-
tions, which is prone to bias because patients lect data on all cohort members can be prohibi-
who stay on treatment for a longer time may be tive. Even with cohorts formed from
less susceptible to the study outcome. computerized databases, there may be a need to
A common criticism of the incident user supplement and validate data with information
design is that excluding prevalent users reduces from hospital records and other sources. When
study size. While true, researchers should be the cohort size is large, such additional data
aware that by including prevalent users they gathering can become a formidable task.
might gain precision at the cost of validity. Moreover, even if no additional data are needed,
Identifying incident users in secondary data- the data analysis of a cohort with multiple and
bases is not costly. In some situations, particu- time-dependent drug exposures can be techni-
larly studies of second-line treatments in chronic cally infeasible, particularly if the cohort size
conditions, we can only study patients who and number of outcome events are large.
switch from one drug to another, as very few To counter these constraints, designs based
patients will be treatment naive. Such switching on sampling subjects within a cohort exist.
is often not random. A fairer treatment compari- These designs are based on the selection of all
son may be achieved by comparing new switch- cases with the outcome event from the cohort,
ers to the study drug with new switchers to a but differ in the selection of a small subset of
comparison drug (Figure 22.1). Nevertheless, “non-cases.” Generally, they permit the pre-
prevalent new-user cohort designs are being cise estimation of relative risk measures with
developed to minimize bias in situations where negligible losses in precision. Below, we dis-
precision is needed and one needs to include as cuss structural aspects of cohorts and present
many new users of the study drug as possible or three sampling designs within a cohort: nested
when a comparison drug is not identifiable. case– control, multi-time case–control, and
Even with appropriate designs, however, all case–cohort.
observational pharmacoepidemiologic studies
still must consider carefully how to approach Structures of Cohorts
potential confounding. Figure 22.3 illustrates a hypothetical cohort of
21 newly diagnosed diabetics over the period
1995–2010. This cohort is plotted in terms of
Currently Available calendar time, with subjects ranked according
Solutions to their date of entry, which can correspond to
the date of diagnosis or treatment initiation.
The solutions available to minimize confound- An alternative depiction of this same cohort
ing in pharmacoepidemiologic database stud- could be based on disease onset. In this
ies can be broadly categorized into: (i) instance, the illustration given in Figure 22.4
approaches that collect more information on for the same cohort, using follow-up time as
Figure 22.3 Illustration of a

calendar-time cohort of 21 subjects
followed from 1995 to 2010 with 4
cases (•) occurring and related
risk-sets (–).
1995 2010
Calendar time (years)
Figure 22.4 Illustration of follow-

up-time cohort representation after
rearranging the cohort in Figure 22.1,
with the new risk-sets (–) for the 4
cases.
0 15
Follow-up time (years)
the new time axis, is significantly different that Figures 22.3 and 22.4 produce distinct
from the previous one. In these follow-up-time risk-sets for the same cases in the same cohort,
cohorts, the same subjects are ranked accord- as illustrated by the different sets of subjects
ing to the length of follow-up time in the study crossed by the vertical broken line for the same
with zero-time being the time of diagnosis or case under the two forms of the cohort. In
treatment start. Figure 22.3, for example, the first chronological
The question of which of the two forms one case to occur has in its risk-set only the first six
should use rests on one’s judgment of the more subjects to enter the cohort, while in
relevant of the two time axes. This decision is Figure 22.4, all 21 cohort members belong to
important, since it affects the demarcation of its risk-set at the time that the first case arises.
“risk-sets,” which are fundamental to the anal- While the second form based on disease dura-
ysis of data from cohorts and consequently the tion is often used, because drug exposure can
sampling designs within cohorts. A risk-set is vary substantially over calendar time, the first
formed by the members of the cohort who are form may be as relevant for the formation of
at-risk of the outcome event at a given point in risk-sets and data analysis as the second form.
time, namely they are free of the outcome and Regardless, an advantage of having data on the
members of the cohort at that point in time entire cohort is that the primary time axis can
called the index date. Drug exposure measures be changed according to the study question,
are then anchored at this index date. It is clear using calendar time for one analysis, duration
of disease or drug exposure for another, with While the nested case–control design leads
respective adjustment in the analysis for the to the computation of the odds-ratio as an esti-
effect of the other time axis. mate of the rate or hazard ratio, its extension
called the “quasi-cohort approach,” provides a
The Nested Case–Control Design construct to estimate crude and adjusted rate
The modern nested case–control design differences. Moreover, the nested case–control
involves four steps: design can be used to contrast exposure to a
drug versus no exposure by comparing the rate
1) Defining the cohort time axis, as above,
of outcome to that of an external population.
2) Selecting all cases in the cohort, i.e. all sub-
The resulting standardized incidence ratio
jects with an outcome event of interest,
(SIR) calculation must take into account that
3) Forming a risk-set for each case, and
cohort members with the longest follow-up
4) Randomly selecting one or more controls
have a greater chance of being selected in the
from each risk-set.
nested case–control sample, since they belong
Figure 22.5 illustrates the selection of a to all the risk-sets (Figure 22.5). The appropri-
nested case–control sample from a cohort, with ate method to perform external comparisons
one control per case (1 : 1 matching). It is clear using data from a nested case–control design
from the definition of risk-sets that a future has been developed and uses knowledge about
case is eligible to be a control for a prior case, as the sampling structure to yield an unbiased
illustrated in the figure for the fourth case (the estimate of the standardized rate.
circle occurring last in time), and that a subject
may be selected as a control more than once. The Multi-time Case–Control Design
Bias is introduced if controls are forced to be The multi-time case–control design was intro-
selected only from the non-cases and subjects duced as an alternative strategy to improve the
are not permitted to be used more than once. precision of the odds ratio in a case–control
The property leading to subjects possibly study with transient time-varying exposures,
being selected more than once in the sample in a setting where increasing the number of
may be challenging when the exposure and control subjects is too costly. This approach is
covariate factors are time-dependent, particu- based on increasing the number of observa-
larly when the data are obtained by question- tions per control subject, by measuring drug
naire where the respondent would have to exposure at many different points in time.
answer questions regarding multiple time Indeed, several case–control studies will col-
points in their history (see also Chapter 13). lect extensive data on time-dependent expo-
Figure 22.5 Nested case–control

sample of one control (■) per case
(•) from cohort in Figure 22.4.
0 15
sures but use only a portion of these data in 10 controls per case (rate ratio 2.13; 95% CI:
estimating the rate ratio. 1.48–3.05). Alternatively, keeping only one
For example, the International control patient per case, but increasing the
Agranulocytosis and Aplastic Anemia Study number of control time windows per subject
(IAAAS) assessed the risk of agranulocytosis from 1 to 10 months (prior to the index date)
associated with the use of analgesics using a also improved the precision, with a rate ratio of
case–control study of 221 cases of agranulocy- 1.99 (95% CI: 1.36–2.90).
tosis and 1425 controls. While the study col-
lected data on exposure for four weeks prior to The Case–Cohort Design
the index date, only one week’s worth of data The case–cohort design involves two steps:
was used in the analysis. The multi-time case–
1) selecting all cases in the cohort, i.e. all sub-
control approach allows the use of all available
jects with an adverse event; and
exposure data during the four weeks (i.e. four
2) randomly selecting a sample of predeter-
control person-moments) rather than only one
mined size of subjects from the cohort, irre-
week (i.e. one control person-moment) to
spective of case or control status.
improve the precision of the odds ratio esti-
mate, which must however be corrected for Figure 22.6 depicts the selection of a case–
within-subject correlation. cohort sample of six subjects from the illustra-
This design increases the number of control tive cohort. Note it is possible that some cases
observations per case, thus potentially also selected in step 1 are also selected in the step 2
increasing the power of the study without add- sample, as illustrated in the figure for the third
ing additional subjects. For example, in a case.
nested case–control study within a cohort of The case–cohort design resembles a reduced
12 090 patients with chronic obstructive pul- version of the cohort, with all cases from the
monary disease (COPD), there were 245 inci- full cohort included. The method of analysis
dent cases of acute myocardial infarction for a case-cohort sample takes into account the
(AMI) that occurred during follow-up, for overlap of cohort members between successive
whom 1 and 10 controls per case were identi- risk-sets induced by this sampling strategy.
fied. The rate ratio of AMI associated with use The first advantage of the case–cohort design
of antibiotics in the month prior to the index is its capacity to use the same sample to study
date was 2.00 (95% CI: 1.16–3.44) with one multiple types of events. In contrast, the nested
control per case, with improved precision with case– control design requires different control
0 15
Figure 22.6 Case-cohort sample with six controls (■) from cohort in Figure 22.4.
groups for each type. For example, a nested drug and of the drug under study, which inher-
case–control study of the risks of β-agonists ently includes the subjects who switched from
had two distinct control groups, one of size 233 the comparator to the study drug as well as
for the 44 asthma deaths, the other of size 422 those who initiated the study drug de novo.
for the 85 asthma near-deaths. Another useful These latter subjects can be directly matched to
advantage is that the case–cohort design per- contemporaneous initiators on covariates or
mits one to change the primary time axis of propensity scores (PSs). For the subjects who
analysis from calendar to disease time and vice switched from the comparator to the study
versa, depending on either the assumed model drug, comparators can be selected from the
or the targeted outcome. This is not possible base cohort by matching conditional on expo-
with the nested case–control study, where the sure sets. Time-based exposure sets are defined
primary time axis must be set a priori to permit by the time from the first prescription of the
the risk-set construction. Yet another advan- comparator drug up to the point of switching
tage is its simplicity in sampling, which has (with a time interval such as ±1 month), while
benefits in both comprehensibility and com- prescription-based exposure sets are defined
puter programming. by the number of prescriptions of the compar-
The nested case–control design does have ator drug received up to the point of switching.
some advantages over the case–cohort design. Thus, each switcher to the study drug will
The first is the simplicity of statistical power belong to an exposure set that includes sub-
calculation. The nested case–control design is jects of similar duration or prescription history
independent of the size of the cohort, while for with a dispensing of the comparator drug. The
the case–cohort design knowledge about over- importance of the exposure sets is that a visit
lap in risk-sets is essential, thus greatly compli- occurred where the physician decided to either
cating these calculations. Second, data on continue the comparator treatment or switch
time-dependent exposure and covariates need to the new study drug, providing equivalent
only be collected up to the time of the risk-set time points in the disease course at which con-
for the nested case–control study, while the col- founding patient characteristics can be meas-
lection must be exhaustive for the case–cohort. ured and controlled for.
Time-conditional propensity scores (TCPS)
Prevalent New-User Cohort Designs can be used to identify and match, within the
A common situation in pharmacoepidemiology exposure sets, the comparator drug users most
involves the study of the effect of a new drug similar to the patients who switched to the
entering the market, with the best comparator study drug. The time-dependent Cox propor-
being an older drug. Most often, patients pre- tional hazards model or, alternatively, condi-
scribed the new drug will have been switched tional logistic regression if the exposure sets
from the older comparator drug. An incident are too large, can be used to compute the “pro-
new-user cohort design based on incident new pensity” of switching to the study drug, versus
users of the study and comparator drugs, includ- continuing on the comparator drug, as a func-
ing only patients who were treatment-naïve to tion of the time-varying patient characteristics
both drugs, would be optimal. However, it measured at the point of the exposure set, thus
would exclude the possibly large number of conserving the matching induced by the expo-
subjects who switched from the older to the new sure set and avoiding adjusting for causal
drug, a clinically relevant subset. The prevalent intermediates. For the purposes of the positiv-
new-user cohort design provides an approach to ity assumption, the TCPS of the switcher
include these switchers. should lie within the range of the TCPS of the
A prevalent new-user cohort is formed from members of the corresponding exposure set.
the base cohort of all users of the comparator To emulate the randomized trial, the selection
process can be initiated with the first chrono- either an exposed or unexposed status, which
logical index study drug subject and repeated represents for data analysis the first column of
sequentially. Thus, each subject who initiated a 2 × 2 table, one for each case. Since each case
the study drug will have a comparator user, will be matched to itself for comparison, the
matched on the TCPS. Cohort entry is taken as analysis is matched and thus we must create
the date of the first prescription of the study separate 2 × 2 tables for each case (see
drug and the corresponding date for the Maclure 1991 for further details).
matched comparator. With respect to control information, the data
This approach is also useful for studies hav- on the average drug use pattern are necessary
ing a “non-use” comparator, by using a physi- to determine the typical probability of expo-
cian visit or prescription for any drug other sure during the time window of effect. This is
than the study drug as the comparator. Several done by obtaining data for a sufficiently stable
questions remain regarding this design, in par- period of time. For example, we may wish to
ticular the potential bias from using the preva- study the risk of ventricular tachycardia in
lent users as comparators, which should be association with the use of inhaled beta-
investigated by stratification on the incident/ agonists in asthma, where prolonged Q-T
prevalent new-user status. intervals were observed in patients in the four-
hour period following drug absorption.
Within-Subject Designs Table 22.1 displays data for 10 cases of ventric-
When dealing with the study of transient ular tachycardia, including the average num-
effects on the risk of acute adverse events, ber of times a day each case has been using
Maclure asserts that the best representatives of β-agonists in the past year. Note that there are
the source population that produced the cases six four-hour periods (the duration of the effect
would be the case subjects themselves: this is period) per day. Such data determine the pro-
the premise of the case-crossover design. This portion of time that each asthmatic is usually
is a design where comparisons between expo- spending in the effect period and thus poten-
sures are made within subjects, thus removing tially “at risk” of ventricular tachycardia. This
confounding by factors that remain constant proportion is then used to obtain the expected
within subject. An extension to the case- exposure on the basis of time spent in these “at
crossover design, the case–time–control design risk” periods, for comparison with the actual
is also presented here. exposure in the cases observed during the last
four-hour period. This is done by forming a
The Case-Crossover Design 2 × 2 table for each case, with the correspond-
The case-crossover study is simply an observa- ing control data as defined above, and combin-
tional crossover study in the cases only. The ing the tables using the Mantel–Haenszel
subjects alternate at varying frequencies technique.
between exposure and non-exposure to the To carry out a case-crossover study, three
drug of interest, until the adverse event occurs, critical points must be considered. First, the
which happens for all subjects in the study study outcome must be an acute event that is
sample, since all are cases by definition. With hypothesized to be the result of a transient
respect to the timing of the adverse event, each drug effect. Thus, drugs with chronic or regu-
case is investigated to determine whether lar patterns of use which vary only minimally
exposure was present within the presumed between and within individuals are not ame-
effect period. This occurrence is then classified nable to this design. Nor are latent adverse
as having arisen either under drug exposure or events. Second, since a transient effect is under
non-exposure on the basis of the presumed study, the presumed effect period must be pre-
effect period. Thus, for each case, we have cisely stated. For example, in a study of the
Table 22.1 Hypothetical data for 10 subjects with ventricular tachycardia included in a case-crossover
study of the risk of ventricular tachycardia in asthma associated with the four-hour period after β-agonist
exposure.
β-agonist use in last Usual β-agonist use Periods of exposure Periods of no exposure
Case # 4 hoursa (Ei) in last year (N1i) (N0i)
1 0 1/day 365 1825

2 1 6/year 6 2184
3 0 2/day 730 1460
4 1 1/month 12 2178
5 0 4/week 208 1982
6 0 1/week 52 2138
7 0 1/month 12 2178
8 1 2/month 24 2166
9 0 2/day 730 1460
10 0 2/week 104 2086
Note: Rate ratio estimator is ( Ei N0i)/( (1 − Ei)N1i)

a
Inhalations of 200 mcg: 1 = yes, 0 = no.
possible acute cardiotoxicity of inhaled β- associated with the use of inhaled β-agonists.
agonists in asthmatics, this effect period can be Using a cohort of 12 301 asthmatics followed
hypothesized to be four hours after having during 1980–1987, 129 cases of fatal or near-fatal
taken the usual dose. An incorrect specifica- asthma and 655 controls were identified. The
tion of this time window can have important amount of β-agonist used in the year prior to
repercussions on the relative risk estimate. the index date was used for exposure. Table 22.2
Third, one must be able to obtain reliable data displays the data comparing low (<12 canisters
on the usual pattern of drug exposure for each per year) with high (>12) use of β-agonists.
case, over a sufficiently long period of time. The crude odds ratio for high β-agonist use was
4.4 (95% CI: 2.9–6.7). Adjustment for all avail-
The Case–Time–Control Design able markers of severity lowered the odds ratio
One of the limitations of the case-crossover to 3.1 (95% CI: 1.8–5.4).
design is the assumption of the absence of a To apply the case–time–control design, expo-
time trend in the exposure prevalence. An sure to β-agonists was obtained for the one-year
approach that adjusts for such time trends is current period and the one-year reference
the case–time–control design, an extension of period prior to the current period. First, a case-
the case-crossover analysis that uses, in addi- crossover analysis was performed using the dis-
tion to the case series, a series of control sub- cordant subjects among the 129 cases, namely
jects to adjust for exposure time trends. By the 29 who were current high users of β-
using cases and controls of a conventional agonists and low users in the reference period,
case–control study as their own referents, the and the 9 cases who were current low users of
case–time-control design addresses the time β-agonist and high users previously. This analy-
trend assumption. sis is repeated for the 655 controls, of whom
The approach is illustrated with data from there were 90 discordant in exposure; that is,
the Saskatchewan Asthma Epidemiologic 65 were current high users of β-agonists and
Project, conducted to investigate the risks low users in the reference period, and 25 were
Table 22.2 Illustration of a case-time-control analysis of data from a case–control study of 129 cases
of fatal or near-fatal asthma and 655 matched controls, and current beta-agonist use.
Cases Controls OR 95% CI
High Low High Low
Current beta-agonist use (case–control) 93 36 241 414 3.1a 1.8–5.4

b
Discordant use (case-crossover) 29 9 3.2 1.5–6.8
Discordantb use (control-crossover) 65 25 2.6 1.6–4.1
Case-time-control 29 9 65 25 1.2 0.5–3.0
a
Adjusted estimate from case–control analysis.
b
Discordant from exposure level during reference time period.
current low users of β-agonists and high users confounding are available. Several approaches
previously. The case–time–control odds ratio, fit key characteristics of longitudinal health-
using these discordant pairs frequencies for a care databases well and address important
paired-matched analysis, is given by (29/9)/ concerns in pharmacoepidemiologic analyses.
(65/25) = 1.2 (95% CI: 0.5–3.0). This estimate,
which does not account for potential confound- Propensity Score Analyses
ing by asthma severity that varies over time, Propensity score analysis has emerged as a con-
indicates a minimal risk for these drugs. venient and effective tool for adjusting large
The case–time–control approach can pro- numbers of confounders. In an incident user
vide an unbiased estimate of the odds ratio in cohort design, a PS is the estimated probability
the presence of confounding by time-invariant of starting medication A versus starting medica-
factors, including indication, despite the fact tion B, conditional on all observed pretreatment
that the indication for drug use (in our exam- patient characteristics. Propensity scores are a
ple, intrinsic disease severity) is not measured, multivariate balancing tool that balance large
because of the within-subject analysis. It also numbers of covariates in an efficient way even if
controls for time trends in drug use. the study outcome is rare, which is frequent in
Nevertheless, its validity is subject to several pharmacoepidemiology. Estimating the PS
assumptions, including the absence of time- using logistic regression is uncomplicated, and
varying confounders, such as increasing strategies for variable selection are well
asthma severity over time (an important prob- described. Variables that are only predictors of
lem, since new drugs may be more likely to be treatment choice but are not independent pre-
implemented when disease is most severe), so dictors of outcome will lead to less precise esti-
that caution is recommended in its use. mates and in some extreme situations to bias.
Selecting variables based on p-values is not
helpful as this strategy depends on study size.
Analytic Approaches for Improved
Once a PS is estimated based on observed covar-
Confounding Control
iates there are several options to utilize it in a
Balancing Patient Characteristics second step to reduce confounding. Typical
Confounding caused by imbalance of patient strategies include adjustment from quantiles of
risk factors between treatment groups is a the score with or without trimming, regression
known threat to validity in nonrandomized modeling of the PS, or matching on PSs.
studies. Many analytic options for reducing Matching illustrates the working of PSs well.
Fixed ratio matching on PSs like 1 : 1 match- ally reduce residual confounding. Another
ing has several advantages that may outweigh advantage is that the multivariate balance of
its drawback of not utilizing the full dataset in potential confounders can be demonstrated by
situations where not all eligible patients match. cross-tabulating observed patient characteris-
Such matching will exclude patients in the tics by actual exposure status after fixed ratio
extreme PS ranges where there is little clinical matching. Matching with a fixed ratio in cohort
ambivalence in treatment choice (Figure 22.7). studies does not require matched analyses, but
These tails of the PS distribution often harbor variable ratio matching does. Analytic tech-
extreme patient scenarios that are not repre- niques that condition on the matching sets and
sentative for the majority in clinical practice may be used in this setting include conditional
and may be due to residual confounding. logistic regression or stratified Cox regression,
Trimming these extreme PS values will gener- depending on the data model.
(a) Figure 22.7 Two hypothetical

Patients Patients propensity score distributions before
never always and after matching. (a) Before matching:
treated treated two propensity score distributions
% of with study with study
subjects partially overlap indicating some
drug drug similarities between the comparison
groups in a multivariate parameter
space. (b) After 1 : 1 matching on
propensity score: Not all patients found
matches that were similar enough in
their multivariable characteristics. Areas
of non-overlap between PS
distributions drop out entirely. Source:
Reproduced from Schneeweiss (2010)
0 with permission from John Wiley &
0 0.5 1 Sons Ltd.
Exposure propensity score
= treated with study drug
= treated with comparison drug
(b)
Patients Patients
never always
treated treated
% of with study with study
subjects drug drug
0
0 0.5 1
Exposure propensity score
= treated with study drug
= treated with comparison drug
In summary, PS analyses are convenient f undamental choice in a pharmacoepidemio-

tools to adjust for many covariates when study logic study design and may influence results
outcomes are rare. Extensive confounding dramatically. Ideally, we want to restrict the
adjustment is central in most pharmacoepide- comparison population to patients who have
miologic applications, and in secondary the identical indication as the users of the
healthcare databases we can often define many study agent. Rosiglitazone and pioglitazone
covariates. In contrast to traditional outcome are such a medication pair. They were mar-
models, PS matching allows the investigator to keted around the same time, were both indi-
achieve covariate balance achieved in the final cated for second line treatment of diabetes,
study sample. PS estimation is well developed come from the same class of compound, and in
for comparing two agents using logistic regres- the early marketing phase were thought to
sion to predict treatment choice. When more have a similar effectiveness and safety profiles.
than two agents or several dose categories are This should make treatment choice largely
compared, multinomial regression models are random with regard to patient characteristics
used to estimate the PS and either pragmatic and treatment groups comparable by design,
pairwise matching to a common reference resulting in almost overlapping PS distribu-
group or multidimensional matching is tions and little confounding.
applied. Of importance, PS analyses still can Limiting to incident users: By restricting
only adjust for measured variables, although the study population to new users of the study
they can be used to adjust for many at the same agent or a comparator agent we implicitly
time. Further, one loses the ability to see the require that both groups have been recently
effects of adjusting for one variable at a time. evaluated by a physician. Based on this evalua-
In situations where exposure is rare, disease tion, the physician has decided that the indi-
risk scores, an alternative to PS analysis, might cating condition has reached a state where a
be more suitable. They estimate the association pharmacologic treatment should be initiated.
between patient factors and the study outcome Therefore, such patients are likely to be more
in an unexposed population using multivariate similar in observable and unobservable char-
regression and summarize the relationship in acteristics than comparing incident users ver-
each patient’s estimated probability of the out- sus non-users or versus ongoing users of
come independent of exposure. another drug.
Matching on patient characteristics:
Focusing on the Analysis Multivariate PSs demonstrate areas of non-
of Comparable Patients overlap where no referent patients with com-
Restriction is a common and effective analytic parable baseline characteristics can be
tool to make drug user groups more compara- identified. It is recommended to remove those
ble by making populations more homogene- patients from the analysis as they do not con-
ous, which leads to less residual confounding. tribute to the estimation and may introduce
Some restrictions are quite obvious since they bias. Such a restriction can be achieved by
are made by explicit criteria, for example, lim- trimming these patients from the study popu-
iting the study population to elderly patients lation (see Figure 22.7b) or by matching
with dementia to study the safety of antipsy- patients on the PS or on specific key patient
chotic medications used to control behavioral characteristics of importance.
disturbances in this population. Other restric- While restriction is an important tool to
tions are more implicit and blur the line improve internal validity it will reduce gener-
between design and analytic strategies. alizability of findings. However, in pharma-
Choice of comparator group: picking a coepidemiology we usually place higher value
comparator group is arguably the most on internal validity even if that comes at the
price of reduced external validity. Investigators thereof. The hundreds of proxies that will be
need to be aware of this tradeoff and make identified can then be adjusted for in a large PS
choices accordingly. model. Collinearity may likely occur but is
irrelevant, as the individual parameters esti-
Unobserved Patient Characteristics mated in the large PS regression will not be
and Residual Confounding interpreted but only used for predicting treat-
Once a study is implemented, strategies to ment. Such a high-dimensional PS approach
reduce confounding further are limited to has been shown empirically to improve con-
observable disease risk factors. Secondary data, founding adjustment in many settings,
like electronic healthcare databases often lack although it is not yet fully evaluated. Although
critical details on health state and risk factors, adjusting for variables that are only related to
which can lead to residual confounding. the exposure and not to the outcome (an IV)
could theoretically introduce bias, in practical
Proxy Adjustment scenarios the advantage of adjusting for poten-
Longitudinal electronic health care databases tial confounders outweighs the risk of adjust-
are as much a description of medical sociology ing for the rare instrument.
under financial constraints as they are records
of delivered health care and can be analyzed as Exploiting Random Aspects
a set of proxies that indirectly describe the in Treatment Choice Via Instrumental
health status of patients. This status is pre- Variable Analysis
sented through the lenses of health care pro- As explained above, we are interested in iden-
viders recording their findings and tifying residual random exposure variation
interventions via coders and operating under after adjusting for observable confounders to
the constraints of a specific health care system. more completely account for confounding.
For example, old age serves as a proxy for many However, in secondary data, not all clinically
factors including comorbidity, frailty, and cog- relevant risk factors of the outcome may be
nitive decline; use of an oxygen canister is a recorded. To attempt to address this limitation,
sign of frail health; having regular annual we can try to identify naturally occurring
check-ups is indicative of a health-seeking life- quasi-random treatment choices in routine
style and increased adherence. Adjusting for a care. Factors that determine such quasi-
perfect surrogate of an unmeasured factor is random treatment choices are called IVs, and
equivalent to adjusting for the factor itself. IV analyses can result in unbiased effect esti-
Frequently used proxies in pharmacoepidemi- mates even without observing all confounders
ologic analyses are the number of prescription if several assumptions are fulfilled.
drugs dispensed, the number of physician vis- An instructive example of an instrument is a
its, and hospitalizations before the index drug hospital drug formulary. Some hospitals list
exposure. Such measures of healthcare inten- only drug A for a given indication and other
sity, while not perfect surrogates, are useful hospitals list only drug B. It is a reasonable
proxies for general health and access to care assumption that patients do not choose their
and have been shown to meaningfully help preferred hospital based on its formulary but
adjust for confounding. rather based on location and recommendation.
Proxy adjustment can be exploited by algo- Therefore, the choice of drug A versus drug B
rithms that systematically search through should be independent of patient characteris-
recorded codes for diagnoses, procedures, tics in the hospitals with these restricted for-
equipment purchases, and prescription drug mularies. Thus, comparing patient outcomes
dispensings before the initiation of study drug from drug A hospitals with patient outcomes
use to identify potential confounders or proxies from drug B hospitals should result in unbiased
effects of drug A versus drug B, using the after a sudden change in treatment patterns
appropriate analytic tools. An example of such may then be a reasonable instrument.
a study is a study on the risk of death from
aprotinin, an antifibrinolytic agent given to Sensitivity Analyses
reduce bleeding during cardiac surgery. The A series of sensitivity analyses can help inves-
study identified surgeons who always used tigators to better understand how robust a
aprotinin and compared their outcomes to sur- study’s findings are to a set of structural
geons who always used aminocaproic acid, an assumptions. Some of the sensitivity analyses
alternative drug. If physician skill level and suggested below are generic and others are
performance are on average equal between specific to database analyses.
institutions, independent of drug use, this will An important but underutilized diagnostic
result in valid findings. On the other hand, of tool for the impact of unobserved confounders
course, such an assumption may not be valid, on the validity of findings in nonrandomized
e. g., if academic hospitals allow less restrictive studies is quantitative sensitivity analyses.
formularies, are more likely to see sicker Basic sensitivity analyses of residual confound-
patients, and have skilled physicians, all of ing try to determine how strong and how
which may be true. imbalanced a confounder would have to be
Instrumental variable analyses rely on the among drug categories to explain the observed
identification of a valid instrument, a factor effect. Such an “externally” adjusted relative
that is assumed to be related to treatment, but risk (RRadj) can be expressed as a function of
neither directly nor indirectly related to the the unadjusted relative risk (RRunadj), the
study outcome. As such, an IV is an observed independent RR of the unmeasured con-
variable that causes (or is a marker of) varia- founder on the disease outcome (RRCD), and
tion in the exposure similar to random treat- the prevalence of the confounder in both drug
ment choice. Typically, the following three exposure categories (PC|E):
assumptions need to be fulfilled for valid IV
RRunadj
estimation: (i) an IV should affect treatment or RRadj.
be associated with treatment choice by sharing PC E 1 RRCD 1 1
a common cause. The strength of this associa- PC E 0 RRCD 1 1
tion is also referred to as the instrument
strength, (ii) an IV should be a factor that is as As an example, a recent cohort study could
good as randomly assigned, so that it is unre- not find the expected association between use
lated to patient characteristics, and (iii) an IV of TNF (tumor necrosis factor) alpha inhibi-
should not be related to the outcome other tors, an immunomodulating agent, in treating
than through its association with treatment. rheumatoid arthritis, and the incidence of seri-
As such, an IV analysis sounds very much like ous bacterial infections. There was a concern
a randomized trial with noncompliance. The that physicians may have prescribed the agent
flip of a coin determines the instrument status selectively in patients with more progressive
(treat with A vs. treat with B) and the amount disease. A sensitivity analysis demonstrated
of random noncompliance determines the the direction and strength of any such bias and
strength of the instrument. In nonrandomized concluded that it would be unlikely to change
research, however, identifying valid instru- the clinical implications of the study. This type
ments is difficult and successful IV analyses of sensitivity analysis is particularly helpful in
are infrequent. In principle, treatment prefer- database studies, but is underutilized.
ence can be influenced by time if treatment Spreadsheet software is available for easy
guidelines change rapidly and substantially. A implementation of such sensitivity analyses
comparison of patient outcome before versus (drugepi.org). Lash and Fink proposed an
approach that considers several systematic to be improved to reliably achieve unbiased

errors simultaneously, allowing sensitivity estimates of causal treatment effects that will
analyses for confounding, misclassification, carry the weight of medical decision making.
and selection bias in one process. Several developments are promising. One is
As another example: When using retrospec- the use of IV analyses utilizing the multilevel
tive databases, it is usually cumbersome or structure of healthcare systems. Another is the
impossible to contact patients and ask when expanded use of PS methods including its com-
they began using a drug for the first time in bination with machine learning for high-
order to implement an incident user cohort dimensional proxy adjustment. A development
design. Therefore, incident users are identified that is gaining importance is the enrichment of
empirically by a drug dispensing that was not existing data environments with supplemental
preceded by a dispensing of the same drug for clinical data linked from electronic medical
a defined time period. This washout period is records, from disease registries, from patient
identical for all patients. A typical length is surveys, and/or from laboratory test result
six months. In sensitivity analyses, this interval repositories. While this information will pro-
could be extended to 9 and 12 months. In a vide an opportunity for improved confounding
study on the comparative safety of antidepres- adjustment, it comes with equally large meth-
sant agents in children in British Columbia, odological challenges, as information is col-
this interval was extended from one year to lected in routine care and may have been
three years to ensure that the children in the requested/recorded selectively in patients who
study were treatment-naïve before their first were thought to benefit most. Clearly there is
use, which helped balance comparison groups still plenty of work to be done to find satisfac-
and reduce confounding. Although increasing tory and scalable solutions for the control of
the length of the washout increases the likeli- confounding.
hood that patients are truly incident users, it
also reduces the number of patients eligible for
the study. This tradeoff is particularly worth Key Points
noting in health plans with high enrollee
turnover. ●● Pharmacoepidemiologic studies must be
There is often uncertainty about the correct efficient by offering rapid information of the
definition of the exposure risk window based utmost validity.
on the clinical pharmacology of the study ●● Computerized databases provide valuable
agent. This is further complicated in health- data sources for pharmacoepidemiologic
care databases, since the discontinuation date studies with unique methodological
is imputed through the days’ supply of the last challenges.
dispensing/prescription. Varying the exposure ●● Target trial thinking and study design
risk window is therefore insightful and easy to choices, like the new-user active-
accomplish in cohort studies. comparator cohort design, are paramount
in limiting the effect of confounding and
other biases.
Conclusion ●● Epidemiologic designs such as nested case–
control and case-crossover are efficient
Minimizing confounding in pharmacoepide- approaches to assess the risks and benefits of
miologic research is an ongoing development drugs if additional data are collected.
that is context and data-source specific. While ●● Confounding bias can be assessed efficiently
great progress has been made in analyzing lon- using subsets of the data with enriched
gitudinal healthcare databases, much remains information.
Key Point 383
Case Example 22.1 (Confavreux, C. et al.)
The occurrence of relapses in multiple scle- ●● Large clinical population with extensive
rosis is highly variable and unpredictable. computerized information.

Vaccines, particularly for hepatitis B, have ●● Efficient study design using only cases for
been suspected to induce relapses in multi- this acute event and transient drug
ple sclerosis. exposure.
●● Confounding factors that do not change
Question
over time are inherently controlled for by
Does vaccination increase in the rate of
the within-subject matched analysis.
relapse in multiple sclerosis?
Limitations
Approach
●● Low vaccination prevalence in this clinical
A case-crossover study within the European
population does not permit assessment of
Database for Multiple Sclerosis network.
the risk for shorter effect periods, such as a
Cases with a relapse after a 12-month
week.
relapse-free period were questioned on vac-
●● Confounding by factors that change over
cinations. Exposure to vaccination in the
time, such as infections, could not be con-
two-month risk-period immediately preced-
trolled for.
ing the relapse was compared with that of
the four previous two-month control periods Key Points
to estimate the relative risk. ●● Multiple sclerosis is highly variable over
time and thus not easily amenable to
Results
cohort or case–control study designs.
●● The prevalence of exposure during the
●● The case-crossover design is an efficient
two-month risk period was similar to that
approach to study vaccine safety.
of the four control periods.
●● The relative risk of relapse associated with
exposure to any vaccination was thus

unity.
Case Example 22.2 (Schneeweiss S., Patrick AR., et al.)
Background Approach
A meta-analysis of randomized placebo- A PS-matched new user cohort study of chil-
controlled trials has shown an increased risk dren and young adults (10–18 years of age)
of suicides among children initiating antide- using health care utilization databases from
pressants (AD), however, the study could not the province of British Columbia linked with
elucidate the comparative safety, i.e. whether vital statistics information. Suicidal actions
this risk varies meaningfully between anti- were defined as emergency hospitalizations
depressant agents. for intentional self harm and completed sui-
cides. First exposure to a tricyclic AD with no
Question
AD exposure in previous three years was
Do tricyclic AD have similar risk of suicidal
compared to new use of fluoxetine. Follow-up
actions including completed suicide com-
started the day after the first AD dispensing
pared with selective serotonin reuptake
inhibitors (SSRIs).
(Continued)
and ended 14 days after their last exposure to ●● Confounding factors may bias results
the cohort-qualifying drug exposure. due to strong channeling of sicker
patients to the safer SSRIs but nonparsi-
Results
monious adjustment with high-
●● Unadjusted and insufficiently adjusted
dimensional PSs can remedy the issue in
analyses showed a spurious up to 50% rela-
this example.
tive risk decrease for tricyclic ADs com-
pared with SSRIs, suggesting that tricyclic Limitations
agents are avoided in patients with suicidal ●● Despite the large population size, the num-
thoughts as they are known to be poison- ber of outcomes remains limited, leading
ous in high doses. to less precise estimates and making
●● There was no difference in the risk for sui- adjustment for many potential confound-
cidal actions between tricyclic ADs and ers in outcome regression models
fluoxetine after nonparsimonious high- perilous.
dimensional PSs matching. ●● Confounding by outcome risk factors that
channel patients into the treatment groups

Strengths
can be strong and hard to control.
●● Large stable population of children and
young adults with information on all health Key Points

service encounters and vital statistics. ●● Two major antidepressant classes have
●● The new users design ensures that all similar risks of suicidal actions in newly
patient characteristics are assessed before treated children and young adults.
exposure starts and that suicidal actions ●● The new user cohort design is a flexible
shortly after initiation will be accounted for and robust approach for studies that rely
and the duration of use-dependent risk entirely on secondary healthcare databases
function can be plotted and illustrate the when combined with nonparsimonious PS
depletion of the most susceptible patients. adjustment.
Further Reading
Blais, L., Ernst, P., and Suissa, S. (1996). Confavreux, C., Suissa, S., Saddier, P. et al. (2001).
Confounding by indication and channeling Vaccinations and the risk of relapse in multiple
over time: the risks of beta-agonists. Am. J. sclerosis. New Engl. J. Med. 344 (5): 319–326.
Epidemiol. 144: 1161–1169. Lash, T.L., Fox, M.P., MacLehose, R.F. et al.
Brookhart, M.A., Rassen, J.A., and Schneeweiss, (2014). Good practices for quantitative bias
S. (2010). Instrumental variable methods in analysis. Int. J. Epidemiol. 43: 1969–1985.
comparative safety and effectiveness research. Maclure, M. (1991). The case-crossover design: a
Pharmacoepidemiol. Drug Saf. 19: 537–554. method for studying transient effects on the
Brookhart, M.A., Schneeweiss, S., Rothman, K.J. risk of acute events. Am. J. Epidemiol. 133:
et al. (2006). Variable selection for propensity 144–153.
score models. Am. J. Epidemiol. 163: Ray, W.A. (2003). Evaluating medication effects
1149–1156. outside of clinical trials: new-user designs.
Collet, J.P., Schaubel, D., Hanley, J. et al. (1998). Am. J. Epidemiol. 158: 915–920.
Controlling confounding when studying large Rothman, K.J., Greenland, S., and Lash, T.L.
pharmacoepidemiologic databases: a case (eds.) (2008). Modern Epidemiology, 3rde.
study of the two-stage sampling design. Philadelphia, PA: Lippincott Williams &
Epidemiology 9: 309–315. Wilkins.
Further Reading 385
Schneeweiss, S., Patrick, A.R., Solomon, D.H., Sturmer, T., Rothman, K.J., Avorn, J., and Glynn,
Dormuth, C.R., Miller, M., Mehta, J., Lee, J.C., R.J. (2010). Treatment effects in the presence
and Wang, P.S. (2010). Comparative safety of of unmeasured confounding: dealing with
antidepressant agents for children and observations in the tails of the propensity
adolescents regarding suicidal acts. Pediatrics. score distribution–a simulation study. Am. J.
125 (5): 876–88. Epidemiol. 172: 843–854.
Schneeweiss, S. (2018). Automated data-adaptive Suissa, S. (1995). The case–time–control design.
analytics for electronic healthcare data to Epidemiology 6: 248–253.
study causal treatment effects. Clin. Suissa, S. (2015). The quasi-cohort approach in
Epidemiol. 10: 771–788. pharmacoepidemiology: upgrading the
Schneeweiss, S. and Avorn, J. (2005). A review of nested case–control. Epidemiology 26 (2):
uses of health care utilization databases for 242–246.
epidemiologic research on therapeutics. J. Suissa, S., Dell’aniello, S., and Martinez, C.
Clin. Epidemiol. 58: 323–337. (2010). The multitime case–control design for
Schneeweiss, S. (2010). A basic study design for time-varying exposures. Epidemiology 21 (6):
expedited safety signal evaluation based on 876–883.
electronic healthcare data. Suissa, S., Edwardes, M.D., and Boivin, J.F.
Pharmacoepidemiol. Drug Saf. 19: 858–868. (1998). External comparisons from nested
Schneeweiss, S., Patrick, A.R., Sturmer, T. et al. case–control designs. Epidemiology 9:
(2007). Increasing levels of restriction in 72–78.
pharmacoepidemiologic database studies of Suissa, S., Moodie, E.E., and Dell’Aniello, S.
elderly and comparison with randomized trial (2017). Prevalent new-user cohort designs for
results. Med. Care 45 (10 Suppl 2): S131–S142. comparative drug effect studies by time-
Schneeweiss, S., Seeger, J.D., Maclure, M. et al. conditional propensity scores.
(2001). Performance of comorbidity scores to Pharmacoepidemiol. Drug Saf. 26 (4):
control for confounding in epidemiologic 459–468.
studies using claims data. Am. J. Epidemiol. Wacholder, S. (1991). Practical considerations in
154: 854–864. choosing between the case–cohort and nested
Spitzer, W.O., Suissa, S., Ernst, P. et al. (1992). case–control designs. Epidemiology 2:
The use of beta-agonists and the risk of death 155–158.
and near death from asthma. N. Engl. J. Med. Walker, A.M. (1996). Confounding by indication.
326: 501–506. Epidemiology 7: 335–336.
387
Part IV
Special Applications and the Future of Pharmacoepidemiology

389
23
Special Applications of Pharmacoepidemiology
David Lee1, Björn Wettermark2, Christine Y. Lu3, Stephen B. Soumerai3, Robert T.

Chen4, Sharon-Lise T. Normand5, Art Sedrakyan6, Danica Marinac-Dabic7, Daniel B.
Horton8, Sonia Hernandez-Diaz9, Tamar Lasky7, Krista F. Huybrechts10, Claudia
Manzo7, Emil Cochino11, Hanna M. Seidling12, David W. Bates10, Bennett Levitan13,
Rachael L. DiSantostefano13, and Scott Evans14
1
(Retired) Center for Pharmaceutical Management, Management Sciences for Health, Arlington, VA, USA
2
Disciplinary Domain of Medicine and Pharmacy, Uppsala University, Uppsala, Sweden
3
Harvard Medical School and Harvard Pilgrim Health Care Institute, Boston, MA, USA
4
Brighton Collaboration, Task Force for Global Health, Decatur, GA, USA
5
Harvard Medical School and Harvard School of Public Health, Boston, MA, USA
6
New York Presbyterian Hospital and Weill Cornell Medical College, New York, NY, USA
7
US Food and Drug Administration, Silver Spring, MD, USA
8
Rutgers Robert Wood Johnson Medical School, Rutgers Center for Pharmacoepidemiology and Treatment Science, Rutgers School of
Public Health, New Brunswick, NJ, USA
9
Harvard T.H. Chan School of Public Health, Boston, MA, USA
10
Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA
11
European Medicines Agency, Amsterdam, The Netherlands
12
Cooperation Unit Clinical Pharmacy, University of Heidelberg, Heidelberg, Germany
13
Department of Epidemiology, Janssen Research & Development, Titusville, NJ, USA
14
Biostatistics Center, The George Washington University, Rockville, MD, USA
Note: The views expressed in this section are those of the authors, and not necessarily those of the US Food and Drug Administration
or the European Medicines Agency.
In this chapter, we will present selected special problems, examples of solutions, and perspec-
applications of pharamcoepidemiology, which tives on the future. Each application section
include studies of drug utilization; evaluating then ends with a case example and key points.
and improving prescribing; special methodo-
logical issues in pharmacoepidemiologic stud-
ies of vaccine safety; epidemiologic studies of
Studies of Drug Utilization
implantable medical devices; research on the
effects of medications in pregnancy and in
Introduction
children; risk management; the pharmacoepi-
demiology of medication errors; and benefit– Drug utilization was defined by the World
risk assessments of medical treatments. We Health Organization (WHO) as the “market-
present this information using a standard for- ing, distribution, prescription and use of drugs
mat, focusing on clinical and methodological in a society, with special emphasis on the
390 23 Special Applications of Pharmacoepidemiology
resulting medical, social, and economic conse- either be descriptive or analytical assessing
quences.” Studies of drug utilization address factors influencing drug utilization. Some
not only the medical and nonmedical aspects drug utilization studies may also be qualita-
influencing prescribing, dispensing, adminis- tive to gain a deeper understanding of drug
tration, and taking of medication, but also the prescribing, dispensing, or consumption of
effects of drug utilization at all levels of the medicines.
health care system. Drug utilization review studies assess the
Drug utilization research (DUR), in the appropriateness of drug utilization, usually
more recent European literature, refers to an by linking prescription data to the reasons
“eclectic collection of descriptive and analyti- for drug prescribing. Explicit predetermined
cal methods for quantifying, understanding, criteria are created, which aspects of the
and evaluating the processes of prescribing, quality, medical necessity, and appropriate-
dispensing, and consumption of medicines ness of drug prescribing may be compared.
and for testing interventions to enhance the Drug use criteria may be based upon such
quality of these processes.” This new defini- parameters as indications for use, daily
tion was introduced to illustrate that DUR dose, and length of therapy, or others such
may include both qualitative and quantitative as: failure to select a more effective or less
studies and emphasize the importance of hazardous drug if available, use of a fixed
intervention studies to promote quality use of combination drug when only one of its com-
medicines. In North America, DUR usually ponents is justified, or use of a costly drug
refers to drug utilization review (discussed in when a less costly equivalent drug is availa-
a separate section). In drug utilization studies, ble. In North America, these studies are also
the objective is to quantify the present state, known as drug use evaluation (DUE) studies.
developmental trends, and time course of For example, a large number of studies have
drug usage at various levels of the health care documented the extent of inappropriate pre-
system, whether national, regional, local, or scribing of drugs, particularly antibiotics,
institutional. Routinely compiled drug statis- and the associated adverse clinical, ecologi-
tics or drug utilization data that result from cal, and economic consequences. Other
such studies can be used to estimate drug uti- studies have been conducted in the elderly
lization in populations by age, sex, social class, (e.g. using Beers criteria for Potentially
morbidity, and other characteristics, and to Inappropriate Medication Use in Older
identify areas of possible over- or underutili- Adults).
zation. They also can be used as denominator Quantitative DUR and DUE studies, aimed at
data for calculating rates of reported adverse detecting and quantifying problems, combined
drug reactions (see Chapter 7); to monitor the with methods to understand their underlying
utilization of specific therapeutic categories factors, are usually one-time projects (rather
where particular problems can be anticipated than routinely conducted), provide for only
(e.g. narcotic analgesics, hypnotics and seda- minimal, if any, feedback to the involved pre-
tives, and other psychotropic drugs); to moni- scribers and, most importantly, do not include
tor the effects of informational and regulatory any follow-up measures to ascertain whether
activities (e.g. adverse events alerts, delisting any changes in drug therapy have occurred
of drugs from therapeutic formularies); as (see Table 23.1). A DUR or DUE program, on
markers for very crude estimates of disease the other hand, is an intervention in the form
prevalence (e.g. anti-Parkinsonian drugs for of an authorized, structured, and ongoing sys-
Parkinson’s disease); to plan for drug importa- tem for improving the quality of drug use
tion, production, and distribution; and to esti- within a given health care institution (see
mate drug expenditures. Such studies can Table 23.1). DUR studies may then be used
Studies of Drug Utilizatio 391
Table 23.1 Drug utilization studies in perspective: operational concepts.
Drug utilization review

Drug statistics Drug utilization study program
Synonyms (therapeutic) Drug utilization data Drug utilization review or Drug audit
drug utilization review study
Descriptive quantitative Yes Yes Usually
methods
Analytic quantitative No Yes Maybe
methods
Qualitative methods No Maybe No
Continuous (ongoing) Usually No Yes
after the implementation of the program to adult and pediatric hospital admissions. The
assess its effectiveness. situations that may lead to preventable
Qualitative drug utilization studies apply adverse drug reactions and drug-induced ill-
qualitative methods. Such studies originate ness include the use of a drug for the wrong
from social science and gather information indication; the use of a potentially toxic drug
that may not be compiled in numerical form. when one with less toxicity risk would be as
These studies include focus-group discussions, effective; the concurrent administration of
open-ended questionnaires, in-depth inter- an excessive number of drugs, thereby
views, and observations. They are valuable to increasing the possibility of adverse drug
explore the perceptions of prescribers, phar- interactions; the use of excessive doses, espe-
macists, and patients dealing with medicines cially for pediatric or geriatric patients; and
in gaining a deeper understanding of social continued use of a drug after evidence
phenomena in drug utilization. becomes available concerning important
toxic effects. Many contributory causes have
been proposed: excessive prescribing, failure
Clinical Problems to be Addressed
to define therapeutic endpoints, the increased
availability of potent prescription and non-
Research
prescription drugs, increased public exposure
For a drug to be approved for market, it must to information and marketing about drugs,
be shown that it can effectively modify the the availability of illicit preparations, and
natural course of disease or alleviate symptoms prescribers’ lack of knowledge of the phar-
when used appropriately– that is, for the right macology of the prescribed drugs. Medication
patient, with the right disease, in the proper error (discussed in the “Medication Errors”
dosage and intervals, and for the appropriate section of this chapter), poor patient adher-
length of time. Used inappropriately, drugs ence (see Chapter 21), discontinuation of
often fail to live up to their potential, with the therapy, and problems in communication
potential for consequent morbidity and resulting from modern day fragmentation of
mortality. Even when used appropriately, all patient care also may contribute to increased
drugs have the potential to cause harm. morbidity and mortality.
However, many of their adverse effects are Therapeutic practice, as recommended by
predictable and preventable. relevant professional bodies, academic
Adverse drug reactions and nonadherence to researchers, and opinion leaders is initially
therapy are important causes of preventable based largely on data from premarketing
c linical trials. Complementary data from clini- number of patients in a defined population
cal experience and studies in the post- who use a drug inappropriately during a
marketing period may result in changes in particular time frame among all those who
therapeutic indication (e.g. an antibiotic received the drug in that population during
becoming no longer a choice because of anti- that time frame. There has been a large growth
microbial resistance), treatment duration (e.g. in databases in all countries, but available drug
short-course antibiotic treatment of exposure and diagnosis data are still suboptimal
community-acquired pneumonia in children in most settings. Also, the criteria to be used to
under five years old), regimen (e.g. changes define “appropriate” are often arbitrary.
due to tolerance to oral hypoglycemic agents), Since most drug consumption statistics were
and precautions and contraindications (e.g. compiled for administrative or commercial
gastrointestinal bleeding with nonsteroidal reasons, the data are usually expressed in
anti-inflammatory agents) among others. As terms of cost or volume. Data on drug
therapy recommendations are updated utilization can be available in several different
through guidelines and other approaches, drug quantities: One is total costs or unit cost, such
utilization studies must address the relation- as cost per package, tablet, dose, or treatment
ship between therapeutic practice as recom- course. Although such data may be useful for
mended and actual clinical practice. measuring and comparing the economic
impact of drug use, these units do not provide
information on the amount of drug exposure
Methodological Problems to be
in the population. Moreover, cost data are
Addressed by
influenced by price fluctuations over time,
Pharmacoepidemiologic Research
distribution channels, inflation, exchange rate
DUR may be quantitative or qualitative. fluctuations, price control measures, etc.
Quantitative studies aim to establish quantities, Official data on drug expenditures in hospitals
presented in numeric figures in categories or may also be false if there are separate rebates
rank order and measured in various units. and risk-sharing arrangements between payers
These studies collect data to determine and pharmaceutical companies.
associations between variables and differences Another quantity, volume data, are available
between categories, using different statistical from manufacturers, importers, or distributors,
methods. On the other hand, qualitative as the overall weight of the drug sold or the
studies attempt to examine, analyze, and unit volume sold (e.g. the number of tablets,
interpret observations for the purpose of capsules, or doses sold). This is closer to the
discovering underlying meanings and patterns number of patients exposed. However, tablet
of relationships. Drug use data may be obtained sizes vary, making it difficult to translate
from several sources, and their usefulness weight into the number of tablets. Prescription
depends on the purpose of the study at hand. sizes also vary, so it is difficult to translate
All have certain limitations in their direct number of tablets into the number of exposed
clinical relevance. For quantitative studies, the patients.
ideal is a count of the number of patients in a The number of prescriptions is the measure
defined population who ingest a drug of most frequently used in drug utilization
interest during a particular time frame. The studies. However, different patients receive a
drug utilization data available are only different number of prescriptions in any given
approximations of this, and raise many time interval and there may be large differences
questions about their presentation and between countries in prescription regulations
interpretation. For studying the quality of and reimbursement. To translate the number
medicine use, one ideal is a count of the of prescriptions into the number of patients,
one must divide by the average number of pre- between countries, but there has been a large
scriptions per patient, or else distinctions must growth in access to them over time everywhere.
be made between first prescriptions and refill Aggregated sales data have been used in
prescriptions. The latter is, of course, better for drug utilization research (DUR) for decades.
studies of new drug therapy, but will omit indi- Today, most countries in Europe keep some
viduals who are receiving chronic drug ther- records of drug sales, at a regional or national
apy. Additional problems may be created by level. These data can be obtained from health
differences in the number of drugs in each pre- authorities as well as from private companies
scription. Finally, it should be noted that all such as IQVIA, a well-known commercial
these units represent approximate estimates of source of drug utilization data. The
true consumption. The latter is ultimately Pharmacoepidemiological Research on
modified by the patients’ actual drug intake Outcomes of Therapeutics by a European
(see degree of adherence, Chapter 21). ConsorTium (PROTECT) indicates that
From a quality of care perspective, to inter- aggregate sales data are widely collected across
pret drug utilization data appropriately, it is Europe. In the US, the IQVIA National Sales
necessary to relate the data to the reasons for Perspective database documents sales data for
the drug usage. Data on morbidity and mortal- prescription drugs, over-the-counter (OTC)
ity may be obtained from national registries products, and some self-administered
(general or specialized); national samples diagnostic products. Data collected include
where medical service reimbursement schemes volume of dollars and quantities moving from
operate; ad hoc surveys and special studies; manufacturers in various outlets within all
hospital records, physician records; and patient states. In Canada, the IQVIA CompuScript
or household surveys. Appropriateness of use database contains data on prescriptions sold
must be assessed relative to indication for treat- from about two-thirds of Canadian retail
ment, patient characteristics (age-related phys- pharmacies. Sales data collected by authorities
iological status, sex, habits), drug dosage or by companies such as IQVIA may be the
(over- or under-dosage), concomitant diseases only secondary source available in studies
(that might contraindicate or interfere with conducted in regions where other databases
chosen therapy), and the use of other drugs are not yet established or more accessible. This
(interactions). However, no single source is is the case in most low- and middle-income
generally available for obtaining all this infor- countries, but it is important to acknowledge
mation. Moreover, because of incompleteness, that large scale databases are created also in
the medical record may not be a very useful many of these countries.
source of drug use data. Individual-level dispensing databases are
increasingly available. These administrative
claims databases may be kept by pharmacy
Examples of Currently Available
chains, health authorities or reimbursement
Solutions
agencies. An example of such a dispensing
Current Data Sources database is the Swedish Prescribed Drug
Drug utilization studies have been conducted Register, which contains data with unique
using a large variety of data sources, including patient identifiers for the entire population of
sales registries, procurement records, 10 million inhabitants for all dispensed
reimbursement/claims databases, medical prescriptions in ambulatory care. This registry
records, pharmacy dispensing records, includes data on the patient (age, sex, personal
pharmacy stock records, disease-based identification number, place of residence), dis-
registries, and population health surveys. The pensed drug (Anatomic Therapeutic Chemical
availability of such data varies substantially [ATC] classification code, defined daily dose
[DDD] number, prescribed dose, package, estimates of the number of persons exposed to
reimbursement, date of prescribing and dis- a particular medicine or class of medicines.
pensing), prescriber (profession, specialty, The DDD is the assumed average daily
workplace), and pharmacy (identifier, loca- maintenance dose for a drug for its main
tion). It can often be linked to many other reg- indication in adults. Expressed as DDDs per
isters, including clinical quality registers with 1000 inhabitants per day (DDD/TID), for
diagnoses and outcome data as well as socio- chronically used drugs, it can be interpreted as
economic data on migration, family situation, the proportion of the population that may
income, education, and country of birth. receive treatment with a particular medicine
Large databases are also derived from on any given day. For use in hospital settings,
electronic health records. For example, the the unit is expressed as DDDs per 100 bed-days
General Practice Research Database (GPRD) (adjusted for occupancy rate); it suggests the
in the United Kingdom is based on medical proportion of inpatients that may receive a
records from general practitioners (GPs). DDD. For medicines that are used for short-
Hundreds of GPs contribute anonymized term periods, such as antimicrobials, the unit
patient information to a central database that is expressed as DDDs per inhabitant per year;
now contains millions of patients. Included this provides an estimate of the number of
are prescriptions issued by the GP but with no days for which each person is treated with a
information from the pharmacy to permit particular medication in a year.
assessment of adherence. Although primarily The DDD methodology is useful for working
used for studying drug safety, such databases with readily available gross drug statistics and
have also been used to study drug utilization. is relatively easy and inexpensive to use.
Although the use of health insurance However, the DDD methodology should be
databases has also been reported in countries used and interpreted with caution. The DDD is
outside North America, Europe and East Asia, not a recommended or a prescribed dose, but a
medical and pharmaceutical databases are technical unit of comparison; it is usually the
generally not available in most low- and result of literature review and available
middle-income countries. The International information on use in various countries. Thus,
Network for Rational Use of Drugs (INRUD) DDDs may be high or low relative to actual
and WHO have developed approaches that use prescribed doses. For example, children’s doses
standardized criteria/indicators to measure are substantially lower than the established
changes in medicines prescribing, dispensing, DDDs. If unadjusted, this situation will lead to
and patient care, which has facilitated the an underestimation of population exposures,
study of drug utilization in low- and middle which may be significant in countries with a
-income countries, including adherence to large pediatric population. Pediatric DDDs
chronic therapy. The approaches include have also been proposed, but the concept and
recommendations on core and complementary its application have not yet been incorporated
indicators, minimum sample sizes, sampling into the WHO methodology. Finally, DDDs do
methods, and data collection techniques, not take into account variations in adherence.
depending on study objectives. The prescribed daily dose (PDD) is another
unit, developed as a means to validate the
Units of Measurement DDDs. The PDD is the average daily dose
The DDD methodology was developed in prescribed, as obtained from a representative
response to the need to convert and standardize sample of prescriptions. Problems may arise in
readily available volume data from sales calculating the PDD due to a lack of clear and
statistics or pharmacy inventory data to exact dosage indication in the prescription and
medically meaningful units and to make crude dosage alteration via verbal instructions
between prescribing events. For certain groups eventually lead to preparation of drug
of drugs, such as the oral anti-diabetes medi- utilization statistics.
cines, the mean PDD may be lower than the The US uses the Iowa Drug Information
corresponding DDDs. Up to twofold variations System (IDIS), which is a hierarchical drug
in the mean PDD have been documented in coding system based on the three therapeutic
international comparisons. Although the DDD categories of the American Hospital Formulary
and the PDD may be used to estimate popula- Society (AHFS), to which a fourth level was
tion drug exposure “therapeutic intensity,” the added to code individual drug ingredients.
methodology is not useful to estimate inci- Other coding systems, such as the National
dence and prevalence of drug use or to quan- Drug Code and the Veterans’ Administration
tify or identify patients who receive doses Classification, do not provide unique codes for
lower or higher than those considered effective drug ingredients.
and safe. In the United Kingdom, British National
The Infectious Diseases Society of America Formulary (BNF) codes are widely used for
and the Society for Healthcare Epidemiology drug utilization studies. The BNF provides
of America (IDSA/SHEA) have recommended monographs for drugs available in the UK. The
days of therapy (DOT) for expressing numbering system is produced by NHS
antimicrobial drug use. DOT is the number of Prescription Services, part of the NHS Business
days when at least one dose of a medication Services Authority in England.
was administered irrespective of dose or route The International Classification of Diseases
of administration. They are not affected by is a system of diagnostic codes for classifying
dose adjustments and can be used in both adult diseases and other health problems. The ICD is
and pediatric populations. Similar to PDDs, published by the WHO and used worldwide in
expressing drug use in the number of DOTs morbidity and mortality statistics, drug
requires patient level use data, which may not reimbursement systems, and automated
be feasible at every facility. decision support in healthcare. The system
includes categories relating to medicinal
Classification Systems substances, but in the context of adverse
Classification systems are used to categorize outcomes, and often in quite broad terms. It
drugs into standardized groups. For example, does not include codes suitable for recording
the ATC classification system is generally used and classifying drug utilization.
in conjunction with the DDD methodology. The Systematized Nomenclature of Medicine
The ATC system consists of five hierarchical (SNOMED) provides a core general
levels: a main anatomical group, two terminology for use in various medical fields.
therapeutic subgroups, a chemical–therapeutic SNOMED clinical terms (CT) contain more
subgroup, and a chemical substance subgroup. than 311 000 active concepts in clinical settings,
Medicinal products are classified according to organized in different hierarchies. An
the main therapeutic indication for the individual number represents each concept,
principal active ingredient. Use of the ATC and several concepts can be used in
classification system is recommended for combination to describe a complex situation.
reporting drug consumption statistics and Clinical finding/disorder and procedure/
conducting comparative DUR. The WHO intervention are examples of the main levels in
International Drug Monitoring Program uses SNOMED CT. Substance and pharmaceutical/
the system for drug coding in adverse drug biologic product hierarchy was also introduced
reaction monitoring, and some developing as a top-level hierarchy in order to distinguish
countries have begun to use the ATC system to drug products from their chemical constitu-
classify their essential drugs, which may ents (substances). It contains multiple levels of
granularity, used to support a variety of pur- for optimizing drug utilization. Questions yet
poses, including electronic prescribing and for- to be addressed through proper methods deal
mulary management (See www.ihtsdo.org). with the role of printed drug information such
as drug bulletins; the duration of effect of
educational interventions such as group
The Future
discussions, lectures, and seminars, each in
DUR has expanded in all countries across the the outpatient and inpatient settings; and the
globe, ranging from early descriptive studies generalizability of face-to-face methods. We
to advanced studies combining different data also need more research on whether the
sources to further understand medicine use benefits and savings achieved with intervention
in the population. Addressing the medical, strategies outweighed the costs of performing
social, and economic aspects of drug utiliza- the intervention.
tion remains relevant for future research. More clinically applicable approaches to
However, the types of drugs in focus will dif- drug utilization review programs, such as the
fer, with 42% of the substances in drug devel- computerized screening of patient-specific
opment being biologics, compared to 8% on drug histories in outpatient care to prevent
the market currently. The growing pressures drug-induced hospitalizations, still require
on all healthcare systems with aging popula- further development and assessment. Patient
tions, rising patient expectations, stricter outcome measures and process measures of
clinical targets, and expensive new medicines quality of drug utilization have to be included
will further increase the need for drug utili- in such studies.
zation studies to monitor that resources are The availability of large computerized
used wisely and that new medicines are pre- databases that allow the linkage of drug
scribed to those who may benefit most from utilization data to diagnoses is contributing to
them. expand this field of study. The WHO/INRUD
From a public health perspective, the indicator-based approach to drug utilization
observed differences in national and interna- studies facilitates the conduct of DUR in low-
tional patterns of drug utilization require much and middle-income countries. Drug utilization
further study. The medical consequences and review programs, particularly approaches that
the explanations for such differences are still take into primary consideration patient
not well documented. Analysis of medicine use outcome measures, merit further rigorous
by gender and age group may suggest impor- study and improvement. Opportunities for the
tant associations. The increasing availability of study of drug utilization are still under-
population-based data resources will facilitate explored, but the political issue regarding the
studies of incidence and prevalence of medi- confidentiality of medical records and
cine use by age and gender. limitations in funds and manpower will
Numerous studies have addressed factors determine the pace of growth of DUR.
influencing drug prescribing. However, the
relative importance of the many determinants
Key Points for Studies of Drug
of appropriate prescribing still remains to be
Utilization
adequately elucidated. Many strategies aimed
at modifying prescribing behavior have been ●● Drug utilization studies can be performed to
proposed and adopted. Current evidence quantify, identify problems in drug
indicates that mailed educational materials utilization, monitor changes in utilization
alone are not sufficient to modify prescribing patterns, understand their determinants, or
behavior. For interventions shown to be evaluate the impact of interventions.
effective in improving drug prescribing, there ●● Drug utilization studies may be conducted
is a need to further define their relative efficacy on an on-going basis in programs for
and proper role in a comprehensive strategy improving the quality use of medicines.
Case Example 23.1 Prevalence of Potentially Suboptimal Medication Use in Older Men
and Association with Adverse Outcomes
Background ●● Reported use of one or more potentially
Medication-related symptoms often underlie inappropriate medicines was associated with
presentations to primary care services and greater hazard of admission to hospital.
emergency departments and are a common ●● Hospital admissions for falls were not
cause of hospital admission, morbidity and associated with any of the markers of sub-
mortality. Among frail older people, falls and optimal prescribing.
confusion (geriatric syndromes) and hip frac-
Strengths
tures may be medicine-use related. Adverse
●● Community-based sampling of study
drug reactions and polypharmacy are also
population.
common in the elderly.
●● Large sample size.
Question ●● Reliable data on selected morbidity and
How prevalent is suboptimal use of medi- mortality endpoints captured through
cines in men over 65 years old and is this comprehensive data linkage system.
associated with adverse outcomes?
Limitations
Approach
●● Study was based on volunteers, recruited
●● Prospective cohort study of community-
from randomly selected subjects of a pre-
dwelling older men.
vious population-based study.
●● Use of a comprehensive population- based
●● Potential underestimation of adverse
data linkage system, combined with self-
events; e.g. falls that do not result in hos-
reported retrospective data from a health-
pitalization were not captured in the com-
in-men survey that included biochemical
prehensive database, and may not be
and hormone analysis of blood samples.
recalled in self-reports.
●● Adverse outcomes included self- reported
●● Accuracy of self- reported medication his-
or documented history of falls and
tories was not validated.
database-recorded hospital admissions
●● Unavailability of medication dosage data
due to geriatric syndromes, cardiovascular
limits determining inappropriateness due
events, and death.
to excessive dosing.
●● Markers of suboptimal medication use
●● Potential underutilization focused only on
were defined for potential over-utilization,
certain cardiovascular conditions and
potential under-utilization, and potentially
treatments.
inappropriate medicines use.
●● Medication use variables did not account
Results for all valid medication indications/
●● Use of potentially inappropriate medicines contraindications.
(48.7%), polypharmacy (≥5 medications,
35.8%), and potential under-utilization Key points
(56.7%) were highly prevalent, and overall ●● Additional data collection approaches may
8.3% of participants reported some form be combined with use of comprehensive
of potentially suboptimal medication use. data linkage systems.
●● Study results suggest that both medication
●● Polypharmacy was associated with all
cause admission to hospital, cardiovascu- overuse and underuse occur frequently in

lar events, and all cause mortality over older men and may be associated with sig-
4.5 years of follow-up. nificant adverse clinical outcomes.
●● Reducing under-utilization is just as
●● Potential medication under-utilization was
associated with subsequent cardiovascular important as reducing over-utilization and
events. inappropriate medication use.
●● Assessing appropriateness of drug utiliza- Some of the factors responsible for suboptimal
tion requires data on indication for treat- prescribing include the failure of clinicians to
ment, patient characteristics, drug regimen, keep abreast of important new findings on the
concomitant diseases, and concurrent use of risks and benefits of medications; excessive
other medications. Even then, the criteria to promotion of some drugs through
be used to define “appropriate” are arbitrary. pharmaceutical company advertising, sales
●● When assessing quality of care, drug representatives, or other marketing strategies;
utilization studies must often rely on simple errors of omission; patient and family
multiple sources of data. demand for a particular agent, even when it is
not scientifically substantiated; and clinical
inertia. These diverse influences suggest the
valuating and Improving
E need for tailoring intervention strategies to the
Prescribing key factors influencing a given behavior based
on models of behavioral change and knowledge
One of the most important, but perhaps under- translation. Poor adherence to medications is
appreciated, goals of pharmacoepidemiology is another important factor contributing to the
to foster ways to improve evidence-based pre- care-gap (discussed in see Chapter 21).
scribing. It is clear, however, that there is a sig-
nificant disconnect between the available
Methodological Problems to be Addressed
evidence for treatment (“what we know”) and
by Pharmacoepidemiologic Research
everyday clinical practice (“what we do”). This
so-called care-gap in prescribing needs to be Internal Validity
urgently addressed. If physicians and other Poorly controlled studies (e.g. one-group post-
health practitioners fail to update their knowl- only or pre– post designs without a control
edge and practice in response to new and clini- group) produce misleading (and usually exag-
cally important evidence on the outcomes of gerated) estimates of the effects of a variety of
specific prescribing patterns, then the “fruits” medication prescribing interventions. Many
of pharmacoepidemiologic research may have nonintervention factors can affect medication
little impact on clinical practice. Thus, the sci- use over time. Indeed, the “success” of many
ence of assessing and improving clinical prac- uncontrolled studies is often due to the attribu-
tices has grown rapidly in importance. The tion of preexisting trends in practice patterns
rapid growth of this new field (sometimes rather than to the studied intervention.
referred to as translation research) is based on Because randomized controlled trials (RCTs)
the recognition that passive knowledge dissem- are sometimes not feasible (e.g. because of
ination (e.g. publishing articles, distributing contamination of controls within a single insti-
practice guidelines) is generally insufficient to tution) or ethical (e.g. unacceptability of with-
improve clinical practices without supplemen- holding quality assurance programs from
tal behavioral change interventions based on controls), other quasi-experimental designs
relevant theories of diffusion of innovations, (e.g. interrupted time-series with or without
persuasive communications, adult learning comparison series, pre–post with concurrent
theory or social cognitive theory. comparison group studies) should be used
instead of weak one-group post-only or pre–
post designs that do not generally permit valid
Clinical Problems to be Addressed by
causal inferences.
Issues related to underuse, overuse, and mis- Regression Toward the Mean
use of medications all contribute to the subop- The tendency for observations on popula-
timal utilization of pharmaceutical therapies. tions selected on the basis of exceeding a
Evaluating and Improving Prescribin 399
redetermined threshold level to approach

p Detecting Effects on Patient Outcomes
the mean on subsequent observations is a Few large well-controlled studies have linked
common and insidious problem. This argues changes in prescribing to improved patient
once again for the need to conduct RCTs and outcomes (e.g. a link between improvements
well-controlled quasi-experimental studies to in processes of care and patient outcomes
establish the effectiveness of interventions such as adverse clinical events). Sample sizes
before they become a routine part of quality may need to be enormous to detect even
improvement programs. modest changes in patient outcomes.
However, process outcomes (e.g. use of rec-
ommended medications for acute myocardial
Unit of Analysis
infarction from evidence-based practice
A common methodological problem in studies
guidelines) are often sensitive, clinically rea-
of physician behavior is the incorrect use of
sonable, and appropriate measures of the
the patient as the unit of analysis. This violates
quality of care, and improvements in pro-
basic statistical assumptions of independence
cesses of care may be important in and of
because prescribing behaviors and outcomes
themselves, and some are associated with
for individual patients are likely to be correlated
better clinical outcomes.
within each physician’s practice. Such
hierarchical “nesting” or statistical “clustering”
often leads to accurate point estimates of effect Examples of Currently Available
but inappropriately low p-values and narrow Solutions
confidence intervals compared with the correct
Conceptual Framework
unit of analysis, such as the physician or
A useful starting point for designing an inter-
practice or facility. Consequently, interventions
vention to improve prescribing is to develop a
may appear to lead to “statistically significant”
framework for organizing the clinical and non-
improvements in prescribing practices because
clinical factors that could help or impede
of mistakenly inflated sample sizes.
desired changes in clinical behaviors. The
Fortunately, methods for analyzing clustered
“Theory of Planned Behavior” or “PRECEDE”
data are available that can simultaneously
(Predisposing, Reinforcing, and Enabling
control for clustering of observations at the
Constructs in Educational Diagnosis and
patient, physician, and facility levels.
Evaluation) is an example of such a frame-
work. This model – PRECEDE – was devel-
Ethical and Legal Problems Hindering oped for adult health education programs, and
the Implementation of Randomized proposes factors influencing three sequential
Clinical Trials stages of behavior change: predisposing, ena-
It has been argued that there are ethical and bling, and reinforcing factors. Predisposing
legal problems related to “withholding” inter- variables include such factors as awareness of
ventions designed to improve prescribing. This a guideline, knowledge of the underlying evi-
explicitly assumes that the proposed interven- dence, or beliefs in the efficacy of treatment.
tions will be beneficial. In fact, the effectiveness However, while a mailed drug bulletin or
of many interventions is the very question that e-mail alert may predispose some physicians to
should be under investigation. Others have new information (if they read it), behavior
argued that mandating interventions without change may be impossible without new ena-
adequate proof of benefit and lack of unin- bling skills (e.g. skills in administering a new
tended consequences is unethical. What is therapy, or overcoming patient or family
important is to demonstrate that such interven- demand for unsubstantiated treatments). Once
tions are safe, efficacious, and cost-effective a new pattern of behavior is tried, multiple and
before widespread adoption. positive reinforcements (e.g. through peers,
reminders, and feedback) may be necessary to leader involvement, this approach includes
fully establish the new behavior. Such a brief orientation to research findings, printed
framework explains a common observation: educational materials, and encouragement to
namely, that multifaceted interventions that implement guidelines during informal “teach-
encompass all stages of behavior change are able moments.” However, opinion leader stud-
more likely to improve physician prescribing ies have shown mixed results and their
than are uni-dimensional interventions that cost-effectiveness remains to be determined.
only predispose, enable, or reinforce. Another popular approach to improving
physician performance is providing physicians
Empirical Evidence on the with clinical feedback regarding their prescrib-
Effectiveness of Interventions ing practices, either in the form of comparative
to Improve Prescribing practice patterns with peers or predetermined
There are numerous research syntheses that standards such as practice guidelines. Most
have evaluated the effectiveness of the most types of feedback have a clinically minimal
commonly studied interventions, including: effect on prescribing and are unlikely to offset
educational interventions; monitoring and the costs of the programs themselves.
feedback; multimedia campaigns; and Drug utilization review programs attempt to
formulary interventions and financial review the appropriateness of medication pre-
incentives; and so on. scribing for individual patients (e.g. drug inter-
Distributing printed educational materials actions and dosage). Results from controlled
aimed at improving prescribing practice trials do not support the effectiveness of this
remains the most ubiquitous form of approach.
prescribing education. Unfortunately, use of Computerized reminders can enable physi-
disseminated educational materials alone may cians to reduce errors of omission by issuing
affect some predisposing variables (e.g. alerts to perform specific actions in response to
knowledge or attitudes), but have minimal patient-level information such as laboratory
effect on actual prescribing practice. findings or diagnoses. However, excessive
Academic detailing involving face-to-face reminders may create “reminder-fatigue.”
visits by pharmacists, physician counselors or Computerized decision support systems (CDSS)
peer‑leaders have been shown to be effective in integrated with electronic health records, a
promoting evidence-based practice and major component of health information tech-
improving patient outcomes. Reinforcement is nology, succeed beyond a “secretarial reminder”
needed to sustain improvement. What sets function. They can support physicians’ prescrib-
academic detailing apart from industry ing decisions in more complex areas such as
detailing is that the messengers and the dosage, schedule, suboptimal choices, and pre-
messages of the former are presumably inde- vention of adverse drug events. CDSS might
pendent, objective, and evidence-based. moderately improve process of care such as
Group education methods such as small rates of lab monitoring and prescribing deci-
group discussions conducted by clinical leaders sions, and they may help reduce the length of
have been shown to improve prescribing. hospital stay compared with routine care while
Traditional large-group, didactic continuing comparable or better cost-effectiveness is
medical education seminars have not been as achieved, but there is no evidence that CDSS
successful, by themselves, in improving have fulfilled their promised effect on health-
physician practice. care costs, mortality or other clinical adverse
Identifying local opinion leaders is another events. (See Case Example 23.2.)
approach to help in the adoption of new Studies have shown that a broader warning
pharmacological agents. In addition to opinion‑ campaign involving the medical and popular
Evaluating and Improving Prescribin 401
Case Example 23.2 For Evaluating and Improving Physician Prescribing
Because many experts have assumed that ●● Cluster randomized trial design eliminated
computerized decision support will be the selection and volunteer bias, while con-
magic bullet that improves physician pre- trolling for any secular improvements in
scribing, vast resources are being committed quality of care.
to widespread implementation before rigor- ●● Overcomes almost all of the flaws in
ous evaluation. design and analysis of previous studies of

computerized decision support.
Question
●● Conducted in a “real world” setting rather
Can computerized decision support improve
than with house-staff or within academic
quality of ambulatory care for chronic condi-
practices or in the hospital setting.
tions such as ischemic heart disease or asthma?
Approach Limitations
●● Cluster randomized controlled trial of 60 ●● Newer technologies or better delivery sys-
primary care practices in the United tems may be more acceptable to primary
Kingdom; 30 practices randomized to care physicians.
heart disease and 30 to asthma, with each ●● Inadequate attention may have been paid
practice acting as a control for the non- to the nontechnological factors that influ-
assigned condition. Practices already had ence the acceptability and use of an inter-
computerized health records and many vention (e.g. lack of up-front buy in from
had electronic prescribing. end-users, lack of incentives for using the
●● Compared about 40 measures of guideline system, lack of participation in the actual
adherence (prescribing, testing, patient guideline development process).
reported outcomes) between intervention
and control practices one year after imple- Key points
mentation of decision support. ●● Just as with any drug or device, interven-
tions to improve prescribing should be
Results tested in controlled studies before wide-
●● No effect of computerized decision sup- spread adoption.
port on any measure of guideline adher- ●● Results of interventions (including deci-
ence for either heart disease or asthma. sion support) conducted at specialized
●● Hypothesized reason for lack of effect was academic centers or in the hospital set-
extremely low levels of use of, and dissat- ting may not necessarily be applicable to
isfaction with, the decision support tools the “real world” of busy primary care
by physicians. practice.
press, internet, newspapers, television, and when the adverse effects of marketed drugs are
radio may be effective in changing prescribing severe and preventable, alternative agents exist,
patterns in large populations. However, because and the messages are simple enough to convey
information may be oversimplified and dis- in mass communications.
torted when communicated in the media, this Finally, in response to escalating health care
type of intervention should probably be used spending, formulary interventions (such as
referred drug lists) and financial incentives

p prescribing problems associated with the over-
(e.g. pay for performance) have been used to use, misuse, and underuse of medications (as
change the way that physicians practice medi- discussed in the previous section on “Drug utili-
cine. Numerous quasi-experimental studies zation” in this chapter). Finally, studies examin-
suggest such interventions are associated with ing the economic outcomes of interventions as
little cost-savings or improvements in quality of well as studies that include patient reported out-
care but may have unintended consequences comes would advance the field.
(e.g. treatment discontinuity). There is also a tremendous need for carefully
controlled research of interventions to improve
prescribing, and how best to combine various
The Future
strategies to allow for rapid and effective
In general, the achievement of long-term implementation of prescribing guidelines.
changes in practice depends on inclusion of New models are needed to predict the most
multiple strategies that predispose, enable, and effective types of intervention for specific prob-
reinforce desired behaviors. The following char- lem types and various broader questions still
acteristics recur in successful interventions: need to be answered, including issues related
to opportunity costs and cost-effectiveness.
●● Use of theoretical or conceptual frameworks
Important effects of medications on many
to help identify key factors that influence
health outcomes have been demonstrated in
prescribing decisions, informed by surveys,
clinical trials; therefore, it is reasonable to
focus groups, or interviews.
hypothesize that more appropriate use of some
●● Targeting physicians in need of education
medications could reduce morbidity and mor-
(e.g. review of prescribing data).
tality, increase patient functioning, and improve
●● Recruitment and participation of local opin-
quality-of-life. Whether improved prescribing is
ion leaders.
a surrogate measure, or an outcome that directly
●● Use of credible and objective messengers
leads to improved health outcomes, it remains a
and materials.
critically important area for future study.
●● Face-to-face interaction, especially in pri-
mary care settings.
●● Repetition and reinforcement of a limited Key Points for Evaluating
number of messages at one time. and Improving Prescribing
●● Provision of acceptable alternatives to the
●● Quality problems in prescribing exist at the level
practices that are to be extinguished.
of medication overuse (e.g. antibiotics for viral
●● Use of multiple evidence-based strategies to
respiratory tract infections), misuse (e.g. non-
address multiple barriers to best practice.
steroidal anti-inflammatory drugs without gas-
●● Emphasis on the goal of improvement in the
tric protection for patients at very high risk of
quality of prescribing and patient safety, not
upper gastrointestinal bleeding), and underuse
just cost minimization in the guise of quality
(e.g. bisphosphonates in the secondary preven-
improvement.
tion of osteoporosis-related fractures or inhaled
Although we know that prescribing problems corticosteroids for reactive airways disease).
exist, we still know surprisingly little about their ●● Passive interventions, such as dissemination
prevalence or determinants. This paucity of of printed or emailed guidelines, drug utili-
data is all the more remarkable considering that zation reviews, or traditional continuing
three-quarters of all physician visits end in the medical education lectures, are unlikely to
prescription of a drug. Future research efforts improve practice.
need to describe in greater detail the nature, ●● More active interventions (e.g. one-to-one edu-
prevalence, rate of prescribing, and severity of cation, point-of-care reminders, achievable
Special Methodological Issues in Pharmacoepidemiologic Studies of Vaccine Safet 403
benchmarks with audit and feedback), espe- to persons due to ill health. A higher standard of
cially when combined together to overcome safety is required for vaccines because of the
barriers at the level of the system, the physi- large number of persons who are exposed, some
cian, and the patient, are able to modestly of whom are compelled to do so by law or regu-
improve the quality of prescribing. lation for public health reasons. These issues are
●● Interventions (e.g. financial incentives) may the basis for strict regulatory control and other
have unintended consequences that need to oversight of vaccines by the Food and Drug
be investigated. Administration (FDA) and the WHO. These
●● For most interventions that aim to change concerns also often lead to investigation of
prescribing, there is little evidence on their much rarer adverse events following vaccina-
economic outcomes or patient-reported tions than for pharmaceuticals. However, the
outcomes. cost and difficulty of studying AEFI’s increase
●● Just as with the adoption of drugs and devices, with their rarity, and it is difficult to provide
interventions to improve physician prescribing definitive conclusions from epidemiologic stud-
need to be tested in rigorous controlled trials ies of such rare events. Furthermore, high
before widespread and expensive implementa- standards of accuracy and timeliness are needed
tion. In particular, investigators should con- because vaccine safety studies may have impor-
sider “mixed-method” evaluations (i.e. tant impacts on vaccination policies.
quantitative data, prescriber surveys, qualita- Unlike many classes of drugs for which
tive inquiry about barriers and facilitators to other effective therapy may be substituted, vac-
adoption) of their interventions to better under- cines generally have few alternative choices,
stand what works or does not work -and why. and the decision to withdraw a vaccine from
the market may have wide ramifications.
Establishing whether an AEFI is causally due
pecial Methodological Issues
S to a vaccine or not, and if so, a prompt defini-
in Pharmacoepidemiologic Studies tion of the attributable risk are critical in plac-
of Vaccine Safety ing AEFI’s in the proper risk/benefit
perspective. Vaccines are relatively universal
Vaccines are among the most cost-effective exposures. Therefore, despite the relative rarity
public health interventions available. However, of serious true vaccine reactions, ~50 000
no vaccine is perfectly safe or effective. reports of AEFI’s are received annually in the
Concerns over adverse events following immu- US (∼8% are serious), few reports can be caus-
nizations (AEFIs) can result in vaccine hesi- ally linked to vaccination. Recommendations
tancy and destabilize immunization programs. for immunization requires a dynamic balanc-
As immunization programs “mature” with ing of risks and benefits. As vaccine preventa-
high vaccine coverage and near elimination of ble diseases approach eradication, information
target vaccine-preventable diseases, there may on complications due to vaccination relative to
be increased prominence of vaccine-induced that of the wild disease (that the vaccine pre-
and vaccine-coincidental AEFI’s, particularly vents) may lead to a perception that the vac-
in the modern media. cine complications outweigh the benefits and
therefore lead to a discontinuation or decreased
use of the vaccine.
Research in vaccine safety can help to distin-
guish true vaccine reactions from coincidental
The tolerance for AEFI’s to vaccines given to events, estimate an attributable risk, facilitate
healthy persons – especially healthy babies – is vaccine injury compensation, identify risk fac-
lower than for medical products administered tors that can inform the development of valid
precautions and contraindications, and if the insufficient to rule out other potential causes
pathophysiologic mechanism becomes known, of the adverse event. To overcome these limita-
develop safer vaccines. tions, epidemiology and creative methodology
All of these issues have been highlighted have been vital in providing a rigorous scien-
during the COVID19 pandemic. An unprece- tific approach for assessing the safety of
dented number of new vaccines have been vaccines.
developed (many using technologies with lim-
ited or no prior use in humans) and approved Signal Detection
for emergency use within ~300 days (vs. typi- High profile vaccine adverse events, such as
cal decade) of identification of the target SARS intussusception following rotavirus vaccina-
Coronavirus 2 pathogen. New adverse events tion (see Case Example 23.3), demonstrate the
of special interest [AESI; e.g., thrombosis with need for surveillance systems able to detect
thrombycytopenia syndrome] have been iden- potential aberrations in a timely manner.
tified after certain vaccines and are under However, some factors make identification of
study. In the meantime, challenging benefit- true signals difficult; for example, many vac-
risk policy decisions for various subpopula- cines are administered early in life, at a time
tions have had to be made as the entire global when the baseline risk for many adverse health
population likely needs to be immunized, pos- outcomes is constantly evolving. Until the
sibly periodically. The heterogeneity in quality recent advent of systematic analyses of auto-
of pre-and post-introduction safety data avail- mated data including data mining of spontane-
able for each COVID19 vacine presents oppor- ous reports, identification of a vaccine safety
tunities and challenges for pharmacovigilance, signal occurred as much due to a persistent
especially in low- and middle-income patient as from data analysis.
countries.
Assessment of Causality
Assessing whether any adverse event was
actually caused by vaccine is generally not
Addressed by
possible unless a vaccine-specific clinical syn-
drome (e.g. myopericarditis in healthy young
An Institute of Medicine (IOM) review of vac- adult recipients of smallpox vaccine), recur-
cine safety in 1991 found that the US knowl- rence upon rechallenge (e.g. alopecia and
edge and research capacity has been limited hepatitis B vaccination), or a vaccine-specific
by: (i) inadequate understanding of biologic laboratory finding (e.g. Urabe mumps vaccine
mechanisms underlying adverse events, (ii) virus isolation) can be identified. When the
insufficient or inconsistent information from adverse event also occurs in the absence of
case reports and case series, (iii) inadequate vaccination (e.g. seizure), epidemiologic stud-
size or length of follow-up of many population- ies are necessary to assess whether vaccinated
based epidemiologic studies, and (iv) limita- persons are at higher risk than unvaccinated
tions of existing surveillance systems to persons. The latter is complicated by limited
provide persuasive evidence of causation. unvaccinated populations (particularly
These limitations were cited again in a more among children), potential confounding due
recent IOM review (2011), Adverse Events of to differences between those vaccinated and
Vaccines: Evidence and Causality. This report not vaccinated, and determining whether
noted that even very large epidemiologic stud- events are attributable to particular vaccines
ies may not detect or rule out rare events, and since frequently combination vaccines or
that using case reports is complicated by the more than one vaccine are administered
wide variation of available information often together.
Case Example 23.3 For Special Methodological Issues in Pharmacoepidemiologic Studies

of Vaccine Safety
Background 1999) with 95 confirmed intussusception
●● Five cases of intussusception occurred cases following RRV-TV (confirmed by
among 10 054 recipients from three phase medical record review).
three trials [each with its own data safety ●● Case–control study found infants receiv-
monitoring board] of a rhesus-human ing RRV-TV were 37 times more likely to
rotavirus reassortant-tetravalent vaccine have intussusception three to seven days
(RRV-TV), vs one case in 4633 controls, a after the first dose than infants who did
difference that was not statistically signifi- not receive RRV-TV (95% confidence inter-
cant. However, this analysis failed to take val [CI] = 12.6–110.1).
into account that three of the five vacci- ●● Retrospective cohort study using large
nated intussusception cases had their linked databases (LLDB) among 463 277
onset within one week of vaccination. children in managed care organizations
●● RRV-TV was licensed on 31 August from the Vaccine Safety Datalink (VSD)
1998 with intussusception listed as a project demonstrated that those receiving
potential adverse reaction needing post- the vaccine, 56 253 infants, were 30 times
licensure surveillance. more likely to have intussusception three
to seven days after the first dose than
Question infants who did not receive the vaccine
Does a vaccine containing four live viruses, a (95% CI = 8.8–104.9).
rhesus rotavirus (serotype 3) and three ●● Causal link between RRV-TV receipt and
rhesus-human reassortant viruses (serotypes intussusception suggested in the postmar-
1, 2, and 4), increase the risk of intussuscep- keting period at a frequency detectable by
tion in RRV-TV vaccine recipients? current surveillance tools (approximately
1/5000–1/10000 vaccinees).
Approach
●● On 16 July 1999, CDC recommended tem-
●● Conduct post-licensure vaccine safety sur-
porarily suspending use of RRV-TV follow-
veillance with both passive and active
ing initial 15 VAERS reports. When the
systems.
VAERS findings were substantiated by pre-
●● Analyze Vaccine Adverse Event Reporting
liminary findings of the more definitive
System (VAERS) spontaneous reporting
studies, the manufacturer voluntarily
surveillance for reports of intussusception
recalled the vaccine. The US Advisory
(new code need to be added).
Committee on Immunization Practices rec-
●● Conduct active surveillance on large popu-
ommendations for RRV-TV vaccination
lations with two studies (a case–control
were withdrawn in October 1999.
and a cohort study) on RRV-TV recipients
and controls. Key Points
●● Quantify risk associated with RRV-TV ●● Safety assessment during pre-licensure
receipt and intussusception. trials can benefit from oversight by a sin-
gle data safety monitoring board with at
Results least one member with rare disease epide-
●●15 initial VAERS reports (total of 112 miology skills.
VAERS intussusception reports from licen- ●● Passive surveillance systems such as
sure on 31 August 1998 to 31 December VAERS are subject to multiple limitations,
(Continued)
including underreporting and biased was available more quickly and efficiently
reporting, reporting of temporal associa- than the traditional case–control study
tions or unconfirmed diagnoses, lack of (teams sent to 19 states to manually col-
denominator data, and unbiased compari- lect data). Pre-organized LLDB’s are now
son groups; VAERS alone cannot usually the preferred mode for routine post-
assess causality. licensure pharmacovigilance.
●● VAERS can successfully provide a timely ●● A “rapid cycle” initiative for more timely
vaccine adverse event alert (hypothesis detection of vaccine safety signals has
generation) and overcome its limitations been formed by the CDC VSD project; this
when paired with rigorous hypothesis project successfully simulated and retro-
testing active surveillance systems that spectively “detected” the RRV-TV intussus-
can assess causality. Both contribute to a ception signal within the VSD by mid-May
functional vaccine safety surveillance 1999.
system. ●● Subsequent clinical trials and post-
●● Similar results for both active surveillance marketing studies of next generation rota-
studies help confirm the causal relation- virus vaccine used standardized Brighton
ship between RRV-TV and intussusception. Collaboration case definition for intussus-
●● The hypothesis testing results from the ception to show these new vaccines are
VSD project using pre-organized LLDBs safer than RotaShield.
Measurement Accuracy quently exposed (e.g. subpopulations with spe-

Misclassification of exposure status (vaccina- cial indications). For studies of rare outcomes,
tion) may occur if there is poor documentation case–control and self-control designs are the
of vaccinations. Documentation of exposure most efficient. Case–control designs typically
status has been fairly good through school age, sample the source population of cases, identify
but difficulty has been encountered in ascer- an appropriate control group, and assess the
taining vaccination status in older persons. exposure status of both groups to estimate the
Misclassification of outcomes can also occur in risk associated with exposure (see Chapter 2).
observational studies that rely on ICD-9 codes Self-controlled designs usually find vaccinated
from computerized databases. Such diagnosis cases in the source population and compare
codes are often validated with a manual medi- incidence rates of post-vaccination adverse
cal records review. Vaccines in the US (and events during defined periods following vacci-
some other countries) are now routinely nation (also known as the risk interval) with
shipped with 2D barcodes to help solve this the rates of the adverse event in the same per-
problem. son during varying control periods (either pre-
or post-vaccination), which do not occur in
Sample Size close relationship to the time of vaccination.
Because adverse health events of concern (e.g.
encephalopathy) are often extremely rare, it Confounding and Bias
may be a challenge to identify enough cases for Because childhood vaccines are generally
a meaningful study. The difficulty with ade- administered on schedule and children may
quate study power is further compounded in have developmental dispositions to particular
assessing rare events in populations less fre- events, age may confound exposure-outcome
relations, e.g. MMR vaccine with febrile sei- ease of implementation and interpretation,
zures or pneumococcal vaccine with sudden this proportional reporting rate ratio (PRR) has
infant death syndrome (SIDS). Consequently, been widely used for vaccine safety signal
such factors must be controlled in the study detection in spontaneous reporting systems
design and analysis, which is often done by such as the US VAERS, (see Chapter 7). Of
matching. course, any signals need to be confirmed in for-
More difficult to control are factors leading mal epidemiologic studies.
to delayed vaccination or non-vaccination.
Such factors (e.g. low socioeconomic status) Epidemiologic Studies
may confound studies of AEFIs and lead to Historically, ad hoc epidemiologic studies have
underestimates of the true relative risks. Those been employed to assess potential AEFIs.
who have not been vaccinated may differ sub- However, automated, large-linked databases
stantially from the vaccinated population in provide a more flexible framework for hypoth-
risks of AEs and thus be unsuitable as a refer- esis testing than ad hoc epidemiologic studies.
ence group in epidemiologic studies. The The Centers for Disease Control and Prevention
unvaccinated may be persons for whom vacci- (CDC) initiated the VSD project in 1990 and
nation is medically contraindicated, or they has since become a standard for vaccine safety
may have other risks for the outcome being surveillance and research. The VSD project
studied (e.g. they may be members of low soci- prospectively collects vaccination, medical
oeconomic groups). Similarly, vaccinated per- outcome (e.g. hospital discharge, outpatient
sons may be preferentially targeted for visits, emergency room visits, and deaths), and
vaccination because of their underlying medi- covariate data (e.g. birth certificates, census)
cal condition, potentially over-estimating the under joint protocol at multiple managed care
true relative risk. In addition, some children organizations (MCOs). The VSD project also
may be unvaccinated due to parental choice. provides safety signals near-real time using
These children may have different health care “rapid cycle” data sets that are updated weekly
utilization patterns than fully vaccinated chil- or monthly. This surveillance is usually
dren, which in turn could bias study results. employed to monitor newly licensed vaccines
and existing vaccines that may have new rec-
ommendations. Pre-specified events of interest
(usually chosen based on pre-licensure clinical
Solutions
trials, signals from other sources, or historical
Signal Detection concerns, such as Guillain-Barre Syndrome)
Identifying a potential new vaccine safety are tested at regular intervals (weekly or
problem (“signal”) pre- or post-licensure monthly) for increased risk of an event follow-
requires a mix of clinical and epidemiologic ing the vaccine under surveillance using an
expertise. Hypothesis clarification via stand- appropriate comparison group, either histori-
ardized case definitions or clinical case series cal data or a concurrent cohort. New “Tree
may be needed. Data mining is often used to Scan” methods also allow real time surveil-
assess disproportional reporting in spontane- lance for unspecified outcomes. The VSD can
ous reporting systems. One disproportionality also validate surveillance findings and test new
assessment tool for comparing safety profiles ad hoc vaccine safety hypotheses using tradi-
of vaccines involves comparing the proportional epidemiologic methods and recently
tions of particular symptoms out of the total more frequent use of self-control designs that
number of symptoms reported for a given vac- can avoid potential confounding from person-
cine to that observed among reports for another level factors and comorbidities (assuming they
vaccine or group of vaccines. Because of the do not change over time). Due to the high
c overage attained in the MCOs for most vac- Furthermore, proving that an AEFI may be
cines, few unvaccinated controls are available, coincidental can be difficult, particularly when
and thus the VSD is limited in its capacity to the event is rare. On the other hand, combin-
assess associations between vaccination and ing adversomics, systems biology, and Big Data
adverse events with delayed or insidious onset may allow eventual shift from “one size fits all”
(e.g. neurodevelopmental or behavioral to personalized vaccinations. Vaccine safety
outcomes). surveillance and research requires persistence
The VSD provides an essential, powerful, in providing rigorous science, educating the
and relatively cost-effective complement to public, and providing reassurance that robust
ongoing evaluations of vaccine safety in the vaccine safety systems are in place.
US. Similar systems have since been developed
in other high income countries such as
Key Points for Special
Denmark, the UK, and Taiwan. Similar capac-
Methodological Issues
ity is needed in low and middle-income coun-
in Pharmacoepidemiologic Studies
tries, where first introduction of new vaccines
of Vaccine Safety
targeting locally prevalent pathogens are
increasingly occurring. ●● There are still substantial gaps and limita-
tions to our knowledge of many vaccine
safety issues.
The Future
●● A high standard of safety is required for vac-
Although considerable progress has been cines due to the large number of persons
achieved in vaccine safety monitoring and who are exposed, some of whom are com-
research, there are still several challenges, both pelled to do so by law or public health
scientific and non-scientific. One analytic regulations.
challenge is identifying optimal risk windows ●● New research capacity, such as the Vaccine
following vaccination, which requires under- Safety Datalink, provides powerful tools to
standing the biologic mechanisms of the AEFI address many safety concerns. Similar
but can be somewhat arbitrary; data-driven capacity needs to be expanded globally, espe-
approaches to defining risk intervals are cially for new vaccines (e.g. COVID-19).
underway. Detecting lifetime dose responses to
multiple exposures of the same vaccine or vac-
cine components, determining the feasibility pidemiologic Studies
E
of studying vaccine safety of combined and of Implantable Medical
simultaneous vaccinations, and data-mining Devices
for unknown AEFI using electronic medical
records are all challenging areas in vaccine Recent decades have seen an explosion in
safety research. Vaccine safety science also medical device technologies worldwide. The
faces the challenge of credibility; as vaccine global medical devices market reached a
preventable diseases continue to decline, many value of nearly $423.8 billion in 2018 and is
are skeptical about the need for vaccination expected to grow to nearly $521.64 billion by
and are suspicious of the motives of both gov- 2022 and $612.7 billion by 2025. Groundbreak
ernments (since many vaccines are mandated) ing innovations in the areas of transcatheter
and manufacturers. The VSD is studying interventions, nanotechnology, telemedicine,
health outcomes after “alternative” (vs rou- robotic procedures, sophisticated health
tine) vaccination schedules. Faulty research information technology software, and smart
may arise from those seeking to prove that the applications continue to offer new diagnos-
motives for vaccination are questionable. tic and therapeutic options to patients and
Epidemiologic Studies of Implantable Medical Device 409
clinicians. Recent approvals of the medical events, and uncertainties surrounding the
devices that use artificial intelligence to aid in effects of long-term exposure. Implantable
diagnosis are good examples of the dynamic medical devices comprise an important device
landscape where software as a medical device category in the very heterogeneous world of
will play a more prominent role in the deliv- medical devices. As true of other devices,
ery of health care. implantables share characteristics that distin-
guish them not only from other devices, but
also from regulated drugs. Implantable devices,
What Is a Medical Device and how
most of which are in US class III (highest risk),
Is it Different from a Drug?
can be further differentiated.
The US government defines a medical device as Implantable devices often have a long (years
an instrument, apparatus, implement, machine, or even decades) product life cycle (from
contrivance, implant, in vitro reagent or other design to device removal) although incremen-
similar or related article, including any compo- tal changes occur over time. Implantable
nent part or accessory which is: (i) recognized in devices may consist of multiple components
the official National Formulary or United States (such as a total hip implant) or a single compo-
Pharmacopeia or any supplement of them, (ii) nent (such as a pacemaker lead). Exposure to
intended for use in the diagnosis of disease or such devices is typically chronic, with the
other conditions or in the cure, mitigation, onset of exposure clearly defined at time of
treatment or prevention of disease in men or implantation. Exposure typically ends at the
other animals, or (iii) intended to affect the time of device removal, but in practice may
structure or any function of the body of man or continue if part of the device remains in the
other animals, and which does not achieve its human body (e.g. silicone leakage from rup-
primary intended purposes through chemical tured breast implants).
action within or on the body of man or other Outcomes associated with implantable
animals and which is not dependent upon being devices are affected not only by underlying
metabolized for the achievement of any of its patient factors and device factors (such as bio-
principal intended purposes. materials), but also by user interface (e.g. oper-
There are many similarities regarding mediator technique, operator experience). Adverse
cal device definitions and classifications, but effects of implantable devices are typically
differences exist in requirements for approval. localized but may sometimes be systemic (e.g.
International Medical Device Regulators secondary to toxic, allergic, auto-immune
Forum (IMDRF) emerged in 2011 as a driver of effects). Additional hazards may be related to
harmonization and convergence efforts among human factors (e.g. improper programming of
regulatory authorities on medical devices. This pacemakers) and interference (e.g. magnetic
voluntary organization brings together regula- resonance imaging interaction with deep brain
tors from Australia, the European Union, stimulator leads). Lastly, malfunctions may
Canada, the United States, Singapore, China, derive from several sources, including manu-
Japan and Brazil. IMDRF has made major steps facturing problems, design-induced errors,
toward the convergence of activities in the area and anatomic or engineering effects.
of adverse event reporting, patient registries,
software as a medical device, and implementa-
tion of Unique Device Identification (UDI),
Implantable Medical Device Epidemiologic
among other efforts.
Research
In this section, we concentrate on implanta-
ble devices because of their significant clinical The following key issues should be considered
and public health impact, high risk for adverse when planning and conducting medical device
epidemiologic research: (i) assessment of ben- performance in the sub-populations of interest

efits and harms in real world settings, and (ii) in routine care. The utility of observational
assessment of long-term outcomes. studies has been increasing with advances in
medical device data capture in medical records,
Benefits and Harms Profile in a electronic databases, and prospective regis-
Real-Word Setting tries, and with the development of and dissem-
Real-world implantable device evaluation ination of analytical tools.
relies predominantly on observational research
methods. One of the main reasons for such Long-Term Safety and Effectiveness
focus is related to limitations of RCTs. In the Premarket device clinical trials are typically of
RCTs of surgical devices, the participating short duration (e.g. one to two years), and gen-
operators are typically early adopters, highly erate limited information on long-term safety
skilled, and quick learners, which affects the and effectiveness. Due to the inherent com-
learning curve. Traditionally, the learning plexity of implantable devices, it is often diffi-
curve is studied using observational studies of cult to predict fully their long-term safety and
the volume-outcome relationship. In the past, effectiveness based solely on the preclinical
some volume-outcome studies have testing and premarket clinical trials.
demonstrated that increased surgical volume Postmarket attention is therefore increasingly
has an inverse linear relationship with the directed toward ensuring that studies/surveil-
incidence of adverse outcomes; while others lance of enough size and length of follow-up
have identified a volume threshold for are conducted in the postmarket setting to bet-
procedures above which increasing volume is ter illustrate and assess problems occurring
no longer associated with improved outcomes. long-term. Many countries have established
There are three distinct components of the national registries of procedures involving
volume-outcome relationship that can be implanted medical devices that collect long-
studied: (i) lifetime experience (operator’s term patient outcomes and device perfor-
volume), (ii) operator’s volume per unit of mance (e.g. orthopedic registries in Sweden,
time (e.g. year), and (iii) hospital volume the United Kingdom, Australia, Canada and
where operators practice. Other factors related other countries). During the past decades
to learning curve might include type of proce- the US had seen a growth in the number
dure and practice setting (e.g. academic hospi- of national registries (e.g. National
tal). Adequate study of learning curves can Cardiovascular Data Registry [NCDR], Kaiser
establish thresholds for proficiency based on Permanente National Joint Replacement
background expertise related to physicians’ Registry [NJRR], American Joint Replacement
specialties. For example, thresholds for stent- Registry [AJRR], National Breast Implant
ing of carotid artery vary by operator across Registry [NBIR], and Vascular Quality
specialties (e.g. radiologists, cardiologists, and Initiative [VQI].
neurosurgeons).
In addition to highly selected clinicians,
RCTs often involve homogeneous, nonrepre-
Solved by Implantable Medical
sentative patient populations. Pre-marketing
Device Epidemiologic Research
device trials often lack sufficient representa-
tion of important patient populations (women, Evidence generation for implantable medical
children, elderly, racial and ethnic minorities, devices requires accounting for unique issues
etc.), which diminishes the generalizability of that typically do not arise when evaluating
results. Well-designed observational studies benefits and risks of drugs. Key issues arise
can provide more information on device from the interaction of device, operator, patient
and the setting (e.g. hospital, outpatient clinic) robust, widely incorporated medical device
in which the device is being used. Furthermore, nomenclature will significantly further safety
device design, complexity, and its specific bio- surveillance and epidemiologic studies of
material and mechanical characteristics, can medical devices.
be as important to outcomes as the device’s
clinical applications such as type of the lesion Challenges in National Population
being treated, severity of the disease, and con- Exposure Assessment
comitant therapy. In commonly used device The challenges include incorporation of
epidemiologic research databases, these details UDIs into data systems, including electronic
are often only partially available, and some- health records, and routine documentation
times are missing. of device use and patient problems associ-
ated with that use. In the US, population
Challenges in Individual Patient medical device exposure data must be
Exposure Assessment derived from a variety of sources including
The UDI captures critical device informa- electronic health records, administrative
tion, such as the name of the manufacturer, claims data, registries and coordinated regis-
brand, version or model, and device group try networks, national surveys, nationally
terms. The UDI system has only recently representative samples of health providers,
been established, and the UDI is now enter- and marketing data. These data sources dif-
ing data systems within the US, unlike phar- fer in their level of device specific granular-
maceuticals where the National Drug Code ity, design (retrospective versus prospective),
(NDC) Directory has a long history and is and data collection (patient reports, sales,
broadly used. Consequently, many databases etc.). While these sources differ in the level
still lack specific device identifiers, making of completeness and reliability, they may
exposure assessments challenging. For exam- complement each other.
ple, procedure codes may capture device
groups (such as hip implants or types of hip Challenges in Comparative Studies
articulation systems), but lack specificity to Epidemiologic research relies on non-
the manufacturer-level. Characterization of experimental data to develop evidence about
the sensitivity and specificity of device iden- the safety and effectiveness of medical prod-
tifiers found in medical records, clinical reg- ucts. While limitations of non-randomized
istries, or insurance claims databases will be studies are well-known, two facts must be rec-
important to understand errors in device ognized. First, of the methodological
exposure. Promoting routine documentation approaches that are recognized as key compo-
of UDIs in medical records and in other nents of high internal validity in pharmaceuti-
health databases will contribute to a better cal studies (e.g. randomization, allocation
understanding of device-specific perfor- concealment, masking/blinding, withdrawal/
mance. Another challenge associated with follow-up, and intention-to-treat analyses),
medical device epidemiology relates to device not all can be applied to evidence develop-
complexity – devices are frequently approved ment for medical devices. Aside from recog-
or used as systems involving several compo- nized limitations of clinical trials (e.g. select
nents. Device components are often used in study subjects, small sample sizes, short dura-
combination with components of the same or tion), the need for data observed in routine
different brands. Thus, experience with cap- care arises because of learning curve issues,
turing complete device exposure information product modifications, and risks of unex-
is far more complex for devices than it is for pected adverse events related to mechanical
drugs. Once completely adopted, such a failure of the products.
Addressing Sample Size and Real We have good tools to address unequal dis-
World Performance Issues tribution in observed baseline patient charac-
Premark clinical trials are designed to have teristics. Several analytical methods are
sufficient statistical power for effectiveness available to handle selection factors and con-
outcomes. Powering the RCTs for less common founding. These methods involve stratifica-
or rare but serious side effects is not feasible in tion, regression models, or a combination of
most instances. RCTs of devices, because of the two using propensity scores. Machine
small sample size and participant selection, learning approaches are promising as they can
often lack generalizability (or external validity), accommodate many more confounders than
which is defined as the extension of research traditional models and are less reliant on para-
findings and conclusions from a study con- metric assumptions. Each approach relies on a
ducted on a sample population to the popula- set of statistical assumptions that may or may
tion at large. not be appropriate in the setting. When it is felt
Systematic reviews with meta-analyses that there is unmeasured confounding present
attempt to capitalize on the detailed data collec- beyond that accounted for in the collected
tion within each study. Systematic reviews are information, another potential approach is
one mechanism to address the small study that of instrumental variable-based methods
problems of the RCTs. Systematic reviews with which have assumptions and limitations of
meta-analysis assume that most of the individ- their own.
ual RCTs of devices and surgery carefully
record relevant clinical outcomes and offer a
good opportunity to conduct evidence appraisal
Solutions
and synthesis when a reasonable number of
studies are available. Adverse Event Reporting Systems
Well-designed observational studies are Once approved, manufacturers must monitor
often large and involve consecutive patient the safety of their products, including forward-
enrollment and comprehensive data collec- ing reports of adverse events to regulatory
tion. They are the best suited tools to evaluate authorities. In the US, manufacturers are
the safety and effectiveness of devices in rou- required to submit reports of device-related
tine care. With large observational studies, deaths, serious injuries, and malfunctions to
one can evaluate relevant subgroup effects as the FDA. Healthcare providers and consumers
well as rare safety and effectiveness end- submit reports voluntarily. These reports,
points that cannot usually be captured by obtained through passive surveillance are
RCTs. housed in the Manufacturer and User Facility
Device Experience (MAUDE) database, estab-
Ensuring Comparability of Study lished in 1996. Most reports in MAUDE are
Groups from manufacturers, with a small percentage
Cohort designs offer the opportunity to create from user facilities, voluntary sources, and
groups of patients exposed to different devices importers.
of interest. Optimally, these designs are based Passive reporting systems have notable
on prospective and consecutive patient enroll- weaknesses including: (i) data may be incom-
ment, prospective data collection, and a study plete or inaccurate and are typically not inde-
that is hypothesis driven. Such observational pendently verified, (ii) data may reflect
studies should use statistical approaches to reporting biases driven by event severity or
adjust for measured confounders and methods uniqueness or by publicity and litigation, (iii)
that help characterize the impact of unmeas- causality generally cannot be reliably inferred
ured confounding on results. from any individual report, and (iv) events are
generally underreported and this, in combina- been used to monitor device outcomes. The
tion with lack of denominator (exposure) data, DELTA was validated and applied in the sur-
precludes determination of event incidence or veillance of a spectrum of medical devices
prevalence. The latter point is particularly including medical devices, coronary stents,
important for implantable devices, because peripheral vascular stents, and other implant-
reports may capture device-associated events able devices. The system is compatible with a
(such as thrombosis, infection, stroke, revision broad array of potential data sources and sup-
or replacement) for which estimation of inci- ports a variety of statistical methods, allowing
dence is of paramount importance. for both unadjusted and risk-adjusted safety
To enhance the usefulness of reported data, monitoring for prospective and retrospective
statistical tools are used to assist in detecting analyses.
safety signals. Bayesian and other data mining
methods are employed to estimate the relative Mandated Post Marketing Studies
frequency of specific adverse event-device com- The FDA has a unique statutory authority to
binations as compared to the frequency of the mandate postmarketing studies either as a
event with all other devices in the same group. condition of approval or “for cause” later in the
In addition to MAUDE, the Medical Product postmarketing period. A major regulatory/
Safety Network (MedSun) was established to public health challenge the FDA is facing is to
provide additional reporting system based on find appropriate balance for obtaining clinical
the subset of 350 user facilities in the US. The data premarketing to prevent delays in device
MedSun network helps promote dialogue approval and ensure that only safe and effec-
between the FDA and clinical community, tive devices enter the marketplace. The appro-
refine potential safety signals in real time priate postmarketing questions answerable in
through targeted surveys, problem solving, a mandated post-approval study include long-
and posting of reports. term safety and effectiveness, a real-world
experience of the device as it enters broader
Signal Detection/Outlier Identification user populations (clinicians and patients),
Using Variety of Methodologies effectiveness of training programs and learn-
In many countries, national registries of proce- ing curve effects, and the device performance
dures involving implanted medical devices have in certain subgroups of patients not well stud-
significantly augmented their national surveil- ied in the premarketing clinical trials.
lance efforts using a variety of methodologies. Designing scientifically sound but practical
For example, a likelihood-based scoring method studies, and then achieving adequate patient
of calculation of CUSUM is used by the Scottish\ and physician recruitment rates can be chal-
Orthopedic registry described as part of lenging for implantable device studies.
International Consortium of Orthopedics
(ICOR) series. In another example, the Registries
Australian orthopedic registry process identified Recognition that RCTs cannot fill all the gaps in
the ASR artificial hip as outlier device using this clinical evidence for implantable devices is not
method followed by proportional-hazards mod- new, but has garnered renewed interest as reg-
eling to calculate the hazard ratios and adjust for istries have emerged as powerful resources to
age and sex in order to conduct a comparative harness a full potential of observational stud-
analysis of revision rate between groups. ies. The International Medical Device
Regulator’s Forum (IMDRF) definition of med-
Automated Surveillance ical device registry system is as follows:
Automated surveillance using Data Extraction “Registries are organized systems with a pri-
and Longitudinal Time Analysis (DELTA) has mary aim to increase the knowledge of medical
devices contributing to improve the quality of of vulnerable populations, leading to increased

patient care that continuously collect relevant external validity (generalizability). The large
data, evaluate meaningful outcomes, and com- number of diverse patients present the opportu-
prehensively cover the population defined by nities to study device effect heterogeneity and to
exposure to particular device(s) at a reasonably advance methods such as high-dimensionality
generalizable scale (e.g. international, national, propensity scores and instrumental variables.
regional, and health system).” Registries pro- Current limitations of administrative databases
vide important infrastructure for conducting include the lack of UDIs, potential inaccuracy
large-scale medical device studies. In the of coding of diagnosis, difficulty separating
absence of a unique device identification code, comorbidity from complications, and type of
the added value of registries for medical devices revision procedure performed. The lack of clini-
surveillance include capturing brand/model cal information in the administrative billing
specific information crucial for signal identifi- data can be supplemented by linking the billing
cation and comparative effectiveness/safety data to data from registries or other clinically
studies. The complexity and scientific rigor of a rich data from other data sources including
registry can vary from those designed to evalu- electronic health records.
ate quality of health care delivered, those spe-
cifically established to study sustained Coordinated Real-World and Registry
effectiveness and safety of a specific procedure, Networks (CRNs)
and those designed to systematically collect In 2015 the US National Medical Device Registry
long-term data on many different types of treat- Task Force (NMDRTF) recommended the stra-
ment including risk factors, clinical events, and tegically Coordinated Registry Networks
outcomes in a defined population. Once the (CRNs) as an approach to build the national sys-
framework of a registry is in place, studies with tem for medical devices. To implement that rec-
various designs can be performed using the reg- ommendation, Medical Device Epidemiology
istry data (e.g. cohort, case–control, cross sec- Network (MDEpiNet), through cooperative
tional, quasi-experimental). agreement with the FDA, has been advancing
Major limitations of the registries include various sources of real-world data and address-
their often voluntary nature and the short ing the needs of device research and surveil-
duration of follow-up of patients. For implant- lance for multiple stakeholders. The CRN
able medical devices in particular, the modes approach circumvents the limitations of tradi-
of follow-up are critical. These registries can be tional registries, claims, EHRs, and other rele-
linked to other databases including adminis- vant data by promoting interoperability and
trative billing data. The linkage of clinically harmonization and by building linked data sys-
rich procedural and intra-hospital data cap- tems from these multiple sources. MDEpiNet
tured by the registry to the follow-up data from has been developing the CRN-based learning
administrative databases (such as Medicare healthcare communities to speed the develop-
and Medicaid databases) can substantially ment and maturity of the networks. CRNs are
augment the value of registries. facilitated and supported by the MDEpiNet
Coordinating Center at Weill Cornell Medicine
Administrative Claims Data which is establishing working groups charged
The use of administrative databases for epide- with advancing multiple maturity domains
miologic research has the strengths of studying through developmental work and implementa-
large numbers of patients with diverse charac- tion. Each CRN focuses on broad and balanced
teristics and wide varieties of clinical settings, as stakeholder participation which leads to strong
well as inclusion of longitudinal data from the stakeholder engagement and sustainability.
real-world clinical care, and good representation Within the community, the CRN leaders can
leverage medical device ecosystem stakeholders Several measurable characteristics of devices

currently engaged with the MDEpiNet. have been shown to be predictive of device use
and outcome (device variation). For instance,
Methodological Framework the types of weight bearing surface in hip
for Implantable Medical Device replacement systems are related to revision
Outcome Evaluation rates. In particular, hard on hard bearing sur-
A methodological framework for implantable faces, such as metal-on-metal or ceramic-on-
device epidemiology and surveillance involves ceramic, result in higher revision rates.
understanding factors impacting the decision Additionally, large diameter femoral head size
to implant the device, identifying the may result in lower dislocation rates. The pro-
comparison group(s), and estimating the safety cess of implantation fixation to the bone also
and effectiveness of the device compared to the results in variations in clinical outcomes. Hip
alternative strategies. In the context of multi- systems can be implanted with bone cement
ple clinical issues and methodological chal- that helps position the implant within the bone
lenges noted previously, a critical issue in or the systems may have a porous surface that
addressing these goals relates to the multiple permits bone to grow into its surface.
sources of variability that exist with implanta-
ble devices: systematic and random variation
The Future
due to the patient, to the surgeon or operator,
to the healthcare center, and to the device Epidemiology, Digital Health
itself. and Patient-Generated Health Data
Measurable patient characteristics may pre- The broad scope of digital health includes cat-
dict what type of device is implanted as well as egories such as mobile health (mHealth),
clinical and device outcomes (patient varia- health information technology (IT), wearable
tion). For instance, in the case of total hip devices, telehealth and telemedicine, and per-
replacements, advanced age, comorbidities sonalized medicine. These technologies open
such as heart failure and diabetes, and non- new opportunities for patients and consumers
elective admissions are associated with infe- to better manage and track their health and
rior patient outcomes. However, advanced age wellness through greater access to informa-
is also associated with increased use of metal- tion. The interface of these devices and health
on-polyethylene hip systems compared to hip care will continue to open new opportunities
systems constructed from other bearing for medical device epidemiology to lead
surfaces. patient-centered evidence generation, synthe-
Surgeon and surgical center skills (provider sis and appraisal.
variation) may have a large impact on the type
of implantable device used, such as the type of Epidemiology and Evidence-Informed
hip replacement surgery and clinical out- Practice and Policy
comes. Several features of the surgical proce- Epidemiology is becoming the essential link
dure in which the device is implanted can also between an exploding demand for the knowl-
vary. For example, some orthopedic surgeons edge derived from diverse evidence and the
may use less invasive approaches/access when decisions made in health care policy and prac-
implanting a total hip replacement system tice settings. In the larger public health con-
than other surgeons. Complications and device text, the imminent future of device
failures can increase if the surgeon is early in epidemiology will be to integrate and infer
his/her learning curve and annual surgical vol- from massive amounts of heterogeneous and
umes of the surgeon and of the center may be multidimensional data available from dispa-
associated with procedural success. rate data sources. In doing so, medical device
epidemiology will continue to draw from which provide new and promising opportuni-
advances in electronic health records, elec- ties for the epidemiologic study of medical
tronic data capture, standard taxonomy, devices. The intent is to have a comprehensive,
unique device identifiers, global patient iden- up-to-date risk–benefit profile of specific med-
tifiers, integrated security, and privacy ser- ical devices at any point in its life cycle so that
vices. Contemporary device epidemiology will optimally informed decisions can be made and
be able to mobilize the advances of transla- provide more useful information to practition-
tional health research sciences through new ers, patients, and industry. Evolution of pub-
methods that combine basic science and clini- lic–private partnerships will drive the
cal data, leading to the choice of best available collaborative knowledge sharing between
treatment targeted for specific groups of members of the ecosystem.
patients.
Epidemiology and International
Infrastructure
Translational Epidemiology
The accelerating pace of emerging medical
Epidemiologic data have an enormous
technologies worldwide will continue, and
potential to help guide basic science investi-
the information science applications are
gations (e.g. guiding the development of bio-
expected to further shape IT-based health
markers for detection of patient risks for
care dealing with new demands for storage,
development of adverse responses to
transmission, management, and analysis of
implantable devices). In addition, when
patient data. The future global impact of epi-
combined with preclinical and other data
demiology on our understanding of implant-
sources (e.g. genetic, histology, explant
able devices will depend on technological and
retrieval), epidemiologic findings could sig-
policy solutions for international collabora-
nificantly broaden evidence synthesis. In
tion to achieve consistency between global
addition, epidemiology could leverage pre-
data sources, regulations, and methodologic
existing implant-related data from observa-
approaches for various medical device
tional data sources, in individuals with and
implant applications.
without implant-related adverse outcomes to
Collaborative research efforts can particu-
improve our understanding of implant safety
larly help to fill a major gap via international
and effectiveness. These types of interdisci-
consortia. One such example of a collabora-
plinary application of epidemiology could
tive effort is ICOR, International Consortium
lead to more effective identification of candi-
of Vascular Registries (ICVR) and
date biomarkers predictive of certain
International Coalition of Breast Registries
implant-related responses (both local and
Associations (I-COBRA). Development of
systemic) in different patient subgroups. For
international infrastructure creates opportu-
instance, in silico approaches could combine
nities for novel methods developments for epi-
epidemiologic and other data sources to help
demiologic studies. The methods for
guide development of biomarkers.
harmonization, sharing, and combining data
are not well developed and require innovative
Epidemiology and Public–Private approaches. Such international collaborations
Partnerships coupled with increasing regulatory conver-
Public–private partnerships will continue to gence driven by IMDRF present unprece-
bring together expertise and diverse data dented opportunity for influencing the clinical
sources such as registries, electronic health and policy decision making with enormous
records, and patient generated health data, public health implications.
Case Example 23.4 For epidemiologic Studies of Implantable Medical Devices

(See Chughtai et al. BMJ)
Background ●● New York State longitudinal records of

●● Surgical mesh is approved for treatment of hospitalization are available. Limiting data
pelvic organ prolapse (POP) since 1996 to New York State residents enables good
and its use has been increasing over time. follow-up particularly for short-term
●● In 2008 and 2011 FDA issued safety com- outcomes.
munications and highlighting fivefold ●● Repeat operation is likely reliably coded
growth of adverse event reports over time. for patients receiving outpatient or in-
●● The evidence mostly came from analyses patient procedures.
of Manufacturer and User Facility Device ●● Propensity score matching enabled
Experience (MAUDE) database and exten- adjustment of differences in baseline
sive literature review. characteristics between mesh and no
●● The safety of surgical mesh has been scru- mesh groups.
tinized by the media and led to an unprec-
edented number of lawsuits worldwide. Results
●● There were 27 991 women undergoing
Issues POP repairs during the three-year study
●● No major studies of mesh outcomes were period (2008–2011). More than 24% of
conducted, and due to the relatively low patients received mesh and mesh use did
chance of complications large scale and not decrease over time.
real-world study is needed. ●● 90-day complications related to urinary
●● Determine the risk complications after retentions occurred more after mesh-
mesh use when compared to control group based repairs (risk ratio 1.33; 95% CI,
at 90 days. 1.18–1.51).
●● Determine the risk of repeat operations ●● The risk of repeat operation was also sig-
(re-operations) after mesh use when com- nificantly higher after mesh-based repairs
pared to control group at one year after at one year (risk ratio, 1.47; 95% CI,
surgery. 1.21–1.79).
●● Hospital claims data analyses can provide ●● There was interaction of surgery with age
good evidence but the code for erosion (cutoff at 65 years). Mesh use was not asso-
was introduced in 2008. ciated with statistically significant risk uri-
nary retention among the younger women
Approach but led to much higher risk of repeat oper-
●●Mesh use is frequent but there are many ation in this age group
surgeons who are able to avoid the “short-
cut” and use native tissues for POP repairs. Strengths
Most of these operations are conducted in ●● Large and diverse real-world patient
the hospitals and the exposure informa- population.
tion appears in US hospital claims data ●● Ability to specify the exposure groups and
since 2008. Repeat surgery also requires outcomes.

hospitalization and is expensive: hence ●● Billing codes facilitated reasonable risk
this outcome information appears in US classification of patients despite some

longitudinal hospital claims data. limitations.
(Continued)
Limitations longitudinal records are released by data

Reliability and completeness of outcome owners.
codes unknown but bias should be equally ●● It is important to focus exposures and out-
distributed. comes of interest that are codes using bill-
POP disease severity and vaginal and ing claims. Collaboration with surgeons
abdominal mesh is hard to determine using trained in research, robust study design
codes but use of mesh is known to be related and analytic approaches can reduce the
surgeon training/preference and the impact impact of measured and unmeasured
of unmeasured confounding is limited confounders.
●● Claims data can be used for medical device
Key points research when codes are available and
●● Claims data can be used for medical device
longitudinal records are released by data
research when codes are available and owners.
Key Points for epidemiologic Studies The historical exclusion of pregnant women
of Implantable Medical Devices and children from clinical trials leave these
populations more susceptible to the potential
●● Medical devices are of great public health
risks of drugs used off-label and without high-
importance.
quality evidence of efficacy, effectiveness, or
●● Medical devices and their users have diverse
safety. Pharmacoepidemiology has an impor-
and unique characteristics, creating differ-
tant role in supplying critical missing evidence
ent challenges compared with studying
about the beneficial and harmful effects of
pharmaceuticals.
drugs in pregnant women and children.
●● Existing data sources have limited utility for
However, pharmacoepidemiologic research
medical device epidemiology because com-
in these populations presents numerous meth-
plete documentation of device use is not
odologic and practical challenges related to
routine.
changing disease epidemiology and treatment
●● The lack of a detailed identification system
patterns in different stages of pregnancy and
(analogous to the National Drug Code) for
childhood; infrequent uses of medications,
medical devices presents a barrier to under-
rare outcomes, and subgroup considerations
standing device performance.
requiring large sample sizes; and bias from
●● Due to increasing availability of electronic
confounding and other sources.
data sources, methodology for extracting
and assessing the information from diverse
data sources will be required. Clinical Problems to be Addressed by
Unique Biology and Epidemiology
esearch on the Effects
R PREGNANT WOMEN: Biologic processes in
of Medications in Pregnancy pregnancy lead to rapid changes in baseline
and in Children risks of maternal conditions, including indica-
tions for treatment (e.g. nausea). Moreover,
Note: This section reflects the views of the some outcomes are unique to pregnant women
authors and should not be construed to repre- (e.g. pre-eclampsia) or their offspring (e.g.
sent FDA’s views or policies. birth defects), and their etiologically sensitive
Research on the Effects of Medications in Pregnancy and in Childre 419
period changes throughout gestation (e.g. criti- attention deficit/hyperactivity disorder) has
cal portions of organogenesis occur in the first risen over time. Drug formulation and adher-
trimester). Therefore, investigators must con- ence may also affect drug effectiveness and
sider the fluctuating timelines of both the need safety in children.
for and the potential effects of treatment.
CHILDREN: The timing of birth and biologic Evidence to Inform Clinical Practice,
changes due to the growth and development of Role of Pharmacoepidemiology
children from infancy through adolescence Because of the typical exclusions of pregnant
influence their risks of disease (e.g. necrotizing women and children from pre-approval clini-
enterocolitis, chlamydial infection) and pat- cal trials, there is usually very little premarket
terns of treatment. Some medical conditions information on drug safety and effectiveness in
exclusively affect children or specific subgroups pregnant or pediatric populations. Given that a
(e.g. retinopathy of prematurity), while others drug’s safety can rarely be predicted based on
manifest and are treated differently from adults its structure and function alone, animal stud-
(e.g. depression). The states and rates of growth ies are often used to identify pregnancy-related
and development provide critical windows into or pediatric toxicity. However, animal studies
pediatric health and can represent or influence can be poor predictors of teratogenic and other
key variables (e.g. exposures, outcomes). toxic effects in humans. Therefore, most infor-
mation regarding the benefit/risk profile of
drugs in these populations is collected after a
Treatment Responses and Patterns drug’s initial approval; post-approval con-
PREGNANT WOMEN: The physiological trolled observational studies provide the pri-
changes associated with pregnancy can alter mary approach for identifying potential
pharmacokinetics. Extrapolating conclusions teratogenic and toxic effects in pregnant and
regarding dosing, efficacy and safety from non- pediatric populations.
pregnant populations will therefore often be
incorrect. Drug utilization patterns and medi-
cation adherence also vary more markedly
Solved by Pharmacoepidemiologic
around pregnancy. Some medications may be
Research
discontinued due to perceived or real risks
associated with use during pregnancy (e.g. Defining the Population
lithium, valproate), while others are indicated PREGNANT WOMEN: The peculiarities of
specifically during pregnancy (e.g. treatment pregnancy research start when defining the
for hyperemesis gravidarum). Nonadherence target population, since the unit of observa-
could lead to under-treatment of diseases (e.g. tion may be the mother, the pregnancy (sib-
asthma) that could adversely affect outcomes ling clusters within mother), or the fetus
of pregnancy; and to misclassification of expo- (multifetal clusters within pregnancy).
sure (e.g. if drug use is assessed solely based on Sometimes parity or twinning are outcomes of
prescriptions) that could bias studies. interest themselves; when they are not the
CHILDREN: Pharmacokinetics change rap- outcome, the analysis needs to account for the
idly in childhood as various organs needed to correlation within mother and within preg-
absorb, distribute, metabolize, and excrete nancy. Moreover, restriction of the study pop-
drugs mature. Children differ from adults in ulation to livebirths may have consequences
body composition, skin surface area, and other for risk and relative risk estimation and
factors that affect drug absorption and metabo- interpretation.
lism. While children are mostly healthy and CHILDREN: The age definition for pediatric
use fewer medications than adults, chronic populations varies across countries, organiza-
medication use in children (e.g. for asthma, tions, and agencies, with cutoffs mostly around
ages 17–22. Additionally, age-related develop- about recall bias or differential misclassifica-
mental changes and heterogeneity of pediatric tion of exposure. In theory, the birth of a mal-
populations require consideration of study- formed child or a severe pediatric condition
specific subgroups in pediatric research, such may affect recall of prior, remote exposures
as infants (premature and term), children (pre- (e.g. during pregnancy or infancy). More com-
school and school age), and adolescents. plete exposure recall among mothers of cases
Varying definitions of pediatrics and pediatric would create a false positive association
sub-groups make comparisons across studies between the drug and the birth anomaly or
and regulatory bodies more challenging. While pediatric condition, or overestimate an associ-
birthdate is sometimes withheld from large ation if it exists. One approach to reducing this
databases as an identifier, precise age measure- bias is improving accuracy by using well-
ments are vital for studying children in the ear- designed interviews with highly structured
liest months and years of life. Researchers questions to maximize recall and minimize
studying infants and young children may need errors in exposure assessment.
special waivers or permissions to access birth- Pregnancy and childhood are highly
date data. dynamic times in medication management,
including use of OTC medications. It is there-
Sample Size Requirements fore important for investigators to define expo-
and Challenges sure during the etiologically relevant window
An important practical hurdle when studying (e.g. the first few weeks after conception for
the effects of medications in pregnancy and in neural tube defects) with high specificity since
children is the ability to attain adequate sam- misclassifying subjects who are unexposed
ple sizes because, often, the use of specific during the transient sensitive periods as
medications is uncommon, and diseases and exposed would dilute the association, if one
outcomes of interest are correspondingly truly existed.
uncommon. Therefore, in the absence of large PREGNANCY: Given the variability of expo-
effects, the study size needed rapidly becomes sures and risks across stages of pregnancy,
prohibitive. identification and consideration of gestational
age is important for many research questions.
Exposure Ascertainment, Timing, Moreover, not all pregnancies are 40 weeks in
and Misclassification duration, and outcomes of interest may be asso-
To ascertain exposures, we can either rely on ciated with shorter gestational length (e.g.
secondary data (such as records of medication spontaneous abortion). In those instances, one
dispensings) or primary data collection (such must avoid defining the exposure window in a
as interviews). Users of secondary data sources way that creates differential opportunity for
should be aware of the potential disconnect exposure in affected and unaffected pregnan-
between prescribed, dispensed, and consumed cies (e.g. exposure during the first trimester
medications. Studies with primary data collec- would be less likely for miscarriages), as this
tion may rely on interviews for drug exposure will bias the association measure. A few differ-
information, which is often the only feasible ent strategies are available to avoid this bias,
way to obtain information on OTC drug use including defining exposure at start of follow-
and verify the intake of medications. This up (e.g. use at conception) and using a time-
approach raises concerns about the overall varying exposure definition (e.g. use in last
accuracy of recall. Moreover, researchers often 30 days).
conduct such interviews retrospectively after CHILDREN: Because pediatric dosing is fre-
the outcome of interest has occurred (e.g. birth quently weight-based, studies of dose effects in
anomalies, pediatric cancer), raising concern younger children may be more challenging in
secondary data sources without weight data. development and other long-term outcomes
One may categorize dose based on dose distri- also face multiple potential sources of con-
butions within suitable age groups or from founding (see Chapter 22).
estimated dose per weight imputed from PREGNANCY: While teratogenesis has
appropriate pediatric growth charts (e.g. received special attention as a rare but dra-
national, WHO). However, imputed weight- matic outcome, pregnancy researchers and
based doses may be less valid in children with regulators more recently have expanded their
conditions (e.g. malnutrition) whose weight attention to include other obstetric and neona-
distributions substantially deviate from the tal outcomes such as fetal losses and long-term
source population or when granular dosage is consequences in the child. Major birth defects
desired. are typically defined as those that are life
Researchers ascertaining exposures through threatening, require major surgery, or present
interviews or surveys should understand that a significant disability. Depending on inclu-
parents do not always know what medications sion criteria and ascertainment windows, the
adolescent children are or are not taking. risk can range from 1% to greater than 10%.
Therefore, investigators should strongly con-
Outcome Definition and Ascertainment sider including an internal reference group
Validation of outcomes in pregnant and pedi- with consistent outcome definition and data
atric populations may involve consultation collection for major malformations. Moreover,
with expert clinicians, patients, and families. given the etiologic heterogeneity of malforma-
As with other populations, studies using tions, combining multiple malformations into
administrative claims or EHR data may use a single outcome may lack a sound embryo-
restrictive algorithms or validate outcomes logic basis. A more appropriate approach may
using medical records. be to create categories that reflect the embryo-
Differential misclassification of outcomes logic tissue of origin or teratogenic mecha-
with respect to exposure leads to additional ana- nism, when known. However, researchers
lytic challenges. Diagnostic bias may occur if sometimes lump together various fetal malfor-
exposed pregnant women or children receive mations, partially for conservation of statisti-
more testing because of suspicion of drug- cal power. Of note, birth certificates can be
induced effects, resulting in more complete inaccurate records of birth anomalies and are
diagnosis or over-diagnosis of subclinical condi- not recommended as gold standards.
tions (e.g. minor anomalies or viral illnesses). CHILDREN: Conceptualization and opera-
Evaluation of long-term outcomes following tionalization of study outcomes may differ in
prenatal or early life exposures (e.g. neuropsy- pediatric populations because of differences in
chiatric disorders) is particularly challenging disease manifestation, symptoms, and diag-
for outcomes that are rare (e.g. pediatric malig- nostic findings. The sensitivity, specificity, and
nancies), have less readily available measures predictive values of outcome definitions will
(e.g. cognitive impairment), do not consist- also be different in children than in adults
ently come to medical attention (e.g. autism because of differences in disease prevalence.
spectrum disorders), or in settings that limit Growth (e.g. height velocity, weight z-score)
long-term follow-up (e.g. certain automated and development (e.g. motor, behavioral, or
databases, transitions to adult care). Loss to pubertal milestones) can serve as important
follow-up could compromise statistical power outcomes in pediatric studies.
and lead to bias if dropout is not random
and informative censoring is not appropri- Confounding
ately handled. Evaluations of the effects of Confounding can occur if children or pregnant
prenatal or early life drug exposure on child women who receive a drug are more likely to
have risk factors for the outcome. Available important to understanding the indications
approaches to minimize confounding include and effects of treatment in early life, especially
(i) restricting to individuals with the in premature infants. States of growth (e.g.
indication(s) and using an alternative thera- underweight or obesity) and development (e.g.
peutic strategy as the reference, (ii) adjusting skeletal turnover, pubertal stage) could also be
for potentially unbalanced risk factors for the sources of confounding or effect modification.
outcome(s) of interest (e.g. severity of the
underlying condition), (iii) comparing contin- Selection Bias
uers with discontinuers of the medication of When selection into or retention in the study is
interest, and (iv) sibling discordance study to directly or indirectly affected by the exposure
control for stable family factors (see also and the outcome, selection bias may distort
Chapter 22). estimates of risk.
PREGNANCY: Although many treatment PREGNANCY: The ideal pregnancy cohort
indications are not traditional risk factors for begins at or before conception, if not at the
adverse outcomes of pregnancy, certain indica- time of first exposure. Most studies enroll
tions may be confounders due to strong asso- women after pregnancy is confirmed, which
ciations with other conditions or behaviors may underestimate risks of early pregnancy
that are risk factors for the outcome. For exam- events (e.g. miscarriages). Follow-up of
ple, women treated with antidepressants for exposed and unexposed pregnancies should
depression or anxiety may also be more likely start at comparable gestational ages to avoid
to have comorbidities or habits predisposing to bias. Similarly, in studies of birth defects,
adverse pregnancy outcomes. In addition, women should be enrolled before prenatal
women with anxiety utilize more health care screening for major malformations is com-
resources, including fetal or early-life testing pleted to avoid biased selection.
(e.g. echocardiography), than unaffected coun- Unique to the study of birth anomalies is the
terparts. Hence, anxious women are more possibility of pregnancy losses, whether spon-
likely to have infants diagnosed with mild car- taneous or induced. Studies of liveborn infants
diac malformations that might have gone clini- underestimate the risk of lethal and prenatally
cally undetected in children of other women, detectable anomalies resulting in termination.
such as small muscular ventricular septal Bias may occur in instances where exposed
defects. Failure to account for such sources of and reference groups have different propor-
confounding and surveillance bias when stud- tions of terminations of affected fetuses.
ying the safety of psychotropic medications Sensitivity analyses can be used to assess the
might bias results. uncertainty around the relative risk estimates
CHILDREN: Potentially important sources due to this potential selection bias.
of confounding in pediatric studies may be dif- Case–control studies are susceptible to selec-
ficult to ascertain in certain settings, such as tion bias resulting from inappropriate control
administrative claims. Such confounders selection. In some studies of birth anomalies,
include second-hand smoke exposure, paren- controls comprise infants with malformations
tal income and occupation (e.g. measures of besides those affecting cases, in order to reduce
socioeconomic status or environmental expo- the opportunity for differential recall of expo-
sure), parenting behaviors (e.g. seeking testing sure. This approach is valid as long as the expo-
or antibiotics), early childhood feeding (e.g. sure being evaluated does not increase the risk
breastmilk or formula), familial medical con- of the control malformations. Whether mal-
ditions (genetics), and vaccinations. Birth formed or not, controls should be sampled
weight and gestational age at birth, also miss- from the same population that gave rise to the
ing in some databases, can be particularly cases, e.g. same hospital catchment area.
Lastly, one can introduce selection bias dur- remature birth inception cohorts) or children
p
ing the analysis by adjusting for variables that with new-onset disease, particularly rare pedi-
share common causes with the outcome or are atric diseases. Unfortunately, even large incep-
affected by it. For example, adjustment for low tion cohorts are often too small to examine the
birth weight is unwarranted when the analytic risks of less common outcomes (e.g. specific
goal is to estimate the total effect of prenatal birth anomalies) related to specific exposures.
variables, such as maternal drug use, on infant
mortality, or when the goal is to estimate the Registries
direct effect but there is an unmeasured com- PREGNANT WOMEN: For new or infrequently
mon cause of low birth weight and mortality. used drugs, it is more efficient to assemble
CHILDREN: Among pediatric populations, cohorts of exposed women and follow them to
research in very premature infants, who are at determine outcomes including maternal,
high risk of poor outcomes, may be particularly obstetric, fetal, and infant outcomes (known as
subject to selection bias. For instance, in studies exposure pregnancy registries). While critically
of infants within highly specialized neonatal important for the detection of major adverse
intensive care units, selection bias may result effects (e.g. isotretinoin teratogenicity), the
when referral relates to both the risk of expo- small size of most registries prohibits identifi-
sure (e.g. indomethacin) and outcomes (e.g. cation or disproof of small to moderate effects
intraventricular hemorrhage). Furthermore, in involving rare outcomes. Selective inclusion or
the most medically fragile neonatal popula- retention may affect the generalizability of
tions, failure to account for mortality risk in the absolute risk estimates and may bias the rela-
first weeks of life (e.g. through time-varying or tive risk if related to both exposures and out-
competing risk models) may lead to survivor comes. To avoid selection bias, women should
treatment selection bias (immortal time bias). be enrolled into a pregnancy registry before
pregnancy outcomes are known. In any phar-
macoepidemiologic study, reference groups
should be comparable. Pregnancy registries
Solutions
should therefore compare women exposed to a
Below we review the main pharmacoepidemi- drug of interest with other women with similar
ologic designs used for pregnant and pediatric indications, whether untreated or treated with
populations to quantify the risk/benefit profile alternative drugs. When feasible, multi-drug
of medication exposure during pregnancy, pregnancy registries allow comparisons among
infancy, or childhood, and the solutions they drugs from the same class or indication.
offer to common forms of bias. CHILDREN: Like other prospective disease
cohorts, pediatric registries of rare pediatric
Prospective Cohorts diseases can serve as settings for drug safety
Prospective inception (or follow-up) cohorts and effectiveness research, providing rich
offer the advantage of identifying drug expo- information about study variables. Rare pedi-
sure before adverse outcomes are recognized. atric disease registries may also collect biologic
For pregnancy cohorts, women are identified specimens that facilitate molecular pharma-
at the time of pregnancy planning or shortly coepidemiology. Population-based pediatric
after conception; the periodic collection of registries that comprehensively ascertain
information on demographics, exposures, and affected children minimize selection bias and
potential confounders; as well as formal evalu- yield generalizable knowledge. In contrast,
ation of offspring at birth (or fetal death) and registries relying on voluntary participation
ideally throughout childhood. A similar may be subject to selection bias and reduced
approach can be applied to newborns (e.g. external validity.
Retrospective Cohorts and Nested inappropriate control selection, limited focus

Case–Control Studies within on one outcome, and inability to estimate
Automated Health Care Databases absolute risks (unless studies are nested within
Population-based automated healthcare data- defined cohorts). Case–control studies on birth
bases (e.g. national registries, administrative anomalies are often based on interviews,
claims and EHR databases) offer detailed, allowing collection of information often miss-
longitudinal real-world data on health care ing from other data sources (e.g. non-
utilization, diagnoses, procedures, and treat- prescription drugs, lifestyle variables).
ments across various healthcare settings. Examples include the Birth Defects Study to
Clinical care reflects real-world practices, and Evaluate Pregnancy exposureS (BD-STEPS)
study populations may include minorities and EUROmediCAT.
and other marginalized populations often
excluded from volunteer-based studies. While Newer Designs
their size makes them excellent settings to Epidemiologists continue to explore more
study rare exposures or outcomes, some auto- valid and efficient approaches to study preg-
mated databases have substantial limitations nant women and children. In specific circum-
due to lack of child–mother linkages or rou- stances, when carefully conducted with clearly
tinely collected data on OTC drugs and other stated assumptions and interpretation of esti-
potentially important variables (see mates, novel designs may bring advantages to
Chapter 22). When only a small fraction of the field.
pregnant women or children are exposed to To avoid between-person confounding, one
treatments of interest, even large cohorts may might study the risk of birth anomalies using a
be insufficient. Multi-site collaborations may self-controlled design or a sibling discordance
allow identification of exposed pregnancy or study. For example, to study whether flu vac-
pediatric cohorts nested in multiple large cines triggers miscarriage, one could conduct a
databases with the goal to conduct surveil- self-controlled study comparing the frequency
lance and/or etiologic research on medication of vaccinations during the month before mis-
safety. Examples include the Medication carriage and a one-month control window
Exposure in Pregnancy Risk Evaluation three months before miscarriage. In sibling
Program (MEPREP), the International discordance studies, one compares outcomes
Pregnancy Safety Study (InPreSS) consor- in siblings born to the same parents but who
tium, the Comparative Effectiveness Research differ with respect to exposure status during
through Collaborative Electronic Reporting pregnancy or early childhood, therefore
(CER2) Consortium, and PEDSnet. accounting for unmeasured genetic and envi-
ronmental factors.
Case–Control Studies
Case–control studies identify individuals with
The Future
the outcome of interest (e.g. specific birth
anomaly) and compare their frequency of Professional organizations and governments
exposures to that in a control group without have increasingly supported policies and regu-
this outcome. This design can facilitate evalua- lations that prioritize research on medicines
tion of associations between prenatal or early and devices for pregnant women and children.
life exposure to relatively common medication Large-scale, longitudinal, collaborative
and risks for rare events. However, important research using multiple data sources across
challenges include retrospective ascertain- multiple countries will become increasingly
ment of exposure, inability to evaluate infre- common and important for generating gener-
quently used medications, potential for alizable, actionable evidence for these
nderstudied populations. Collaborative net-

u data collection (including biospecimens for
works and pooled resources, potentially pharmaco-omic studies) from patients and
through distributed data models and sharing families. Future efforts must help build capac-
of protocols, data, and analytic code and tools, ity, expertise, and infrastructure to conduct
will enable the conduct of robust research on pharmacoepidemiologic studies in under-
rare exposures and rare outcomes in pregnant served settings and low- and middle-income
women and children. Linkages between com- countries. Such approaches will be necessary
plementary data types – e.g. automated data- to address the changing epidemiology of dis-
bases, registries, patient-generated data, eases affecting pregnant women and children,
genomic and other -omic data – will enable including the rise in obesity and associated
discovery and validation of personalized treat- ailments, emerging infectious diseases (e.g.
ment regimens, as well as outcome validation Zika virus, SARS-CoV-2), improving survival
and better confounding control (e.g. from life-threatening diseases (e.g. cystic
propensity-score calibration). Given the limits fibrosis, inborn errors of metabolism), and
of large health care databases to address all many others. More research and better meth-
research questions about pregnant and pediat- ods are also needed for pharmacoepidemio-
ric populations, we should continue to use, logic research in lactating mothers and
improve, and teach field methods for primary breastfed infants.
Case Example 23.5 For Studies of Drug-Induced Birth Defects (see Huybrechts et al, JAMA
2018)
Background Approach
●● Prior experiences with thalidomide ●● Researchers conducted a cohort study
(delayed regulatory action in response to nested in the nationwide Medicaid Analytic

safety signal) and other antiemetic drugs eXtract to evaluate the association
(including potential over-reactions) raised between ondansetron exposure during the
important questions about the potential first trimester of pregnancy and the risk of
teratogenicity of treatments for nausea cardiac malformations and oral clefts;
and vomiting in pregnancy. using propensity score (PS) stratification
●● By 2009, ondansetron was the most fre- to control for treatment indications and
quently prescribed drug for nausea and other potential confounders.
vomiting in pregnancy in the US.
Results
●● Among 1 816 414 pregnancies,
Issue
88 467 women filled prescriptions for
●● Available evidence on the risks of cleft
ondansetron during the first trimester.
palate and heart defects following ondan-
●● Absolute risks of cardiac malformations
setron exposure in early pregnancy was
were 94.4 and 84.4 per 10 000 exposed and
limited and conflicting, with concerns
unexposed liveborn infants, respectively.
about systematic errors in prior research.
●● Absolute risks of oral clefts were 14.0 and
●● Ondansetron can cross the placenta and
11.1 per 10 000 exposed and unexposed
disrupt serotonin pathways by blocking
liveborn infants respectively.
5-HT3 receptors, making an association
●● PS-adjusted relative risks were 0.99 (95%
biologically plausible.
CI, 0.93–1.06) for cardiac malformations
(Continued)
and 1.24 (95% CI, 1.03–1.48) for oral clefts ●● Concerns about residual confounding, mit-
(corresponding to three additional cases of igated through use of high-dimensional
oral cleft per 10 000 women treated). propensity scores, alternate reference
●● Findings persisted across multiple sensi- groups, and a negative control analysis.
tivity analyses. ●● Severe congenital malformations resulting
in pregnancy losses or terminations are
Strengths missed in cohorts restricted to livebirths;
●● Large cohort size allowing associations to quantitative bias analysis was conducted.
be estimated with great precision.
●● Prospective exposure ascertainment based Key points
on filled prescriptions, free of recall bias. ●● Healthcare utilization databases are
●● Estimation of absolute risks and risk increasingly being used to complement
differences. exposure pregnancy registries and case–
●● Rich patient-level information allowing for control studies.
extensive confounding control. ●● Cohorts nested in healthcare utilization
●● Additional comparisons of ondansetron- databases facilitate the study of rare expo-

exposed women with women exposed to sure and outcomes.
other antiemetics, yielding consistent ●● Large population- based cohorts ensure
results. representation of populations that are fre-
●● Use of validated outcome definitions with quently underrepresented in clinical trials
high positive predictive value. and volunteer cohort studies but are most
●● Robustness of findings tested across mul- at risk of adverse pregnancy outcomes.
tiple sensitivity analyses. ●● Rich information on potential confounders
(e.g. underlying indication and severity,

Limitations maternal comorbidities, and concomitant
●●Differences between filled prescriptions medication use) allow for extensive con-
and medication consumption; sensitivity founding control.
analyses limited to women with ≥2 filled ●● Sensitivity analyses can address potential
prescriptions or intravenous administra- limitations and test robustness of

tion led to similar results. findings.
Key Points for Research on the

coepidemiologic studies in pregnant and
Effects of Medications
pediatric populations include changing pat-
in Pregnancy and in Children
terns of drug utilization and adherence dur-
●● Because pregnant women and children are ing pregnancy and childhood, rapid changes
often excluded from clinical trials, pharma- in risks of outcomes and etiologically sensi-
coepidemiologic research is an important tive periods, timing of enrollment relative to
source of real-world evidence on the use, critical events (e.g. conception, prenatal
effectiveness, and safety of drugs used dur- screening, birth), selection of survivors
ing pregnancy and childhood. when only livebirths are included, missing
●● In addition to random error and confound- information (e.g. structural malformations
ing, methodological challenges for pharma- in fetal losses, referral patterns of neonates),
Risk Managemen 427
and differential opportunities for drug In the context of human medicines in the
exposure due to variable durations of United States, the FDA has defined risk man-
pregnancies. agement as:
●● Exposure pregnancy registries (small ad hoc
cohorts that oversample exposed) are useful an iterative process of (1) assessing a
to evaluate the safety to new drugs, identify product’s benefit-risk balance, (2) devel-
high-risk teratogens, and can collect real oping and implementing tools to mini-
use, sociodemographic information and bio- mize its risks while preserving its
logic samples, but are underpowered to benefits, (3) evaluating tool effective-
identify lesser risks. ness and reassessing the benefit-risk
●● Pediatric disease registries can facilitate balance, and (4) making adjustments, as
research on drug safety and effectiveness in appropriate, to the risk minimization
rare diseases but may be subject to selection tools to further improve the benefit-risk
bias and reduced external validity if enroll- balance. This four-part process should
ment is voluntary. be continuous throughout a product’s
●● Large cohorts nested within large healthcare lifecycle, with the results of risk assess-
databases may provide adequate power for ment informing decisions regarding
less frequent exposures or outcomes and con- risk minimization.
tain detailed clinical information on
population-based samples but frequently lack In the European Union (EU), the concept of
information on OTC drugs, adherence, out- risk management is established in legislation.
comes that do not trigger claims, and impor- Article 1 (28b) of Directive 2001/83 EC as
tant covariates such as BMI, smoking, early amended, defines a risk management system
childhood feeding, and pediatric growth as: “a set of pharmacovigilance activities and
parameters. interventions designed to identify, characterize,
●● Case–control studies can efficiently estimate prevent or minimize risks relating to a medici-
associations with the outcome collected, nal product including the assessment of the
provided drugs are relatively commonly effectiveness of those interventions.” Thus, in
used, selection of controls is valid, and retro- the EU, risk management incorporates (i) the
spective collection of information is not identification or characterization of the safety
biased by differential recall. profile of the medicinal product, with emphasis
●● Under certain assumptions, self-controlled on important identified potential risks and
designs and sibling discordance studies may missing information, and also on which safety
be valuable alternative study designs. concerns need to be managed proactively or
further studied (the “safety specification”), (ii)
the planning of pharmacovigilance activities
Risk Management aimed at characterizing and quantifying clini-
cally relevant risks, and identifying new adverse
For medicines, risk management is used to reactions (the “pharmacovigilance plan”), and
ensure that the potential benefits of a medicine (iii) the planning and implementation of risk
exceed its potential risks, and to minimize minimization measures, including the evalua-
those risks throughout the lifecycle of the prod- tion of the effectiveness of these activities (the
uct. Current understanding of the risks of med- “risk minimization plan”). As a result, both in
icines is based on the premise that the risk of a the US and the EU, risk management measures
medicine derives not only from the inherent are iterative processes frequently leading to the
properties of the medicine, but also from how generation of similar data needs and conceptu-
the medicine is used in actual clinical practice. ally similar risk management tools.
Clinical Problems to be Addressed imparts a risk that would not be present under
by Pharmacoepidemiologic a different condition of use. For example, if
Research drug A, when used in combination with drug
B, results in an unacceptable risk that is not
All medicines have risks. The traditional tools
present when either drug is used alone, this
used to manage the risks of prescription medi-
unacceptable risk is preventable by ensuring
cines have been the prescription status itself
that drug A and drug B are never coadminis-
(i.e. whether the drug was approved for pre-
tered. Other sources of preventable adverse
scription only use or whether it could be
events are medication errors and, occasionally,
obtained without a prescription), professional
injury from product quality defects. Because
labeling, and the requirement that manufac-
they are preventable, medication errors are
turers monitor and report to regulatory author-
well suited to risk management efforts.
ities adverse events that occur with use of the
Because medication errors can occur any-
medicine once it is marketed. In the past few
where in the medication use system, efforts to
decades, additional minimization strategies
minimize the risk of medication error must
have been undertaken to manage more proac-
involve multiple stakeholders.
tively the risks of certain medicinal products.
Unavoidable risks are those that might occur
These measures have included increased com-
when all the known necessary conditions for
munication to patients as well as to healthcare
safe use of a product are followed. In these cir-
professionals, and measures to restrict, in vari-
cumstances, risk minimization activities might
ous ways, the usage of certain medicines.
be directed toward identifying the adverse con-
The Complexities of the Medication sequences as early as possible with the aim of
Use System preventing more serious harm. For example, a
The medication use system is a complex net- drug may be known to cause liver damage but
work of stakeholders, including patients, their its occurrence in a specific patient may not be
families, physicians, nurses, pharmacists, predictable or preventable. In this case, risk
other health professionals, health care organi- minimization activities might be directed
zations and healthcare facilities, payors, man- toward regular monitoring of liver enzyme lev-
ufacturers, and regulatory agencies. Not only els to identify any hepatic damage as early as
does each stakeholder have a role in ensuring possible and thus to stop or modify the treat-
the safe use of a medicine, the interactions ment to prevent serious hepatitis or hepatic
among them do as well. failure.
In this context, the approach to risk manage- Removing all risks from the use of all medi-
ment must consider all environments of where cines is not the overall goal of managing the
the medicine will be used (e.g. hospitals, long- risks of medicines. Rather, careful considera-
term healthcare facilities, physicians’ offices, tion of benefit–risk balance both for the indi-
outpatient home care). vidual patient and for the target population is an
important consideration of risk management.
The Sources of Risk from Medical
Products Risk Management Strives to be
There are several sources of risks from medical Scientifically Driven
products. The known risks of a product are The scientific approach to risk management
based on prior experience or, in some cases, on requires integrating data from various studies
the pharmacologic or other properties of the and disciplines that, when taken together, can
medicine. promote the safe and effective use of a medi-
Preventable risks can occur when a product cine. The scientific approach also compels
is administered under a condition of use that manufacturers and regulators to examine the
Risk Managemen 429
critical gaps in knowledge that exist. Such gaps risks is clearly much less intense. In the middle
may concern the pharmacologic properties of of this spectrum are the vast majority of medi-
the medicine, clinical outcomes related to its cines, mainly prescription medicines, for
use, including that in higher risk populations, which a measured approach to risk manage-
or the way the medicine is used in actual prac- ment must be taken.
tice. Any of these areas could lead to further For most prescription medicines, the most
post-approval studies, the results of which common side effects are generally not life-
would lead to changes in labeling or other threatening. Rather, many are mild and self-
changes that could enhance the safe and effec- limited. Others are bothersome, and some are
tive use of the medicine. so clinically significant that they require the
medicine to be discontinued. OTC medicines
Risk Management Proceeds have been found to be safe and appropriate for
throughout a product’s Lifecycle use without the supervision of a health care
Knowledge about a product’s safety profile is provider, and they can be purchased by con-
limited to some extent at the time of product sumers without a prescription. When OTC
approval, because of recognized practical limi- medicines are taken properly, most of their
tations in the drug development process. Even side effects are generally mild. However, there
after a product has been marketed for a decade can be serious, even life-threatening or fatal,
or more, uncertainties will remain. For exam- side effects of OTC medicines when they are
ple, a study of new molecular entities approved not taken properly. For example, acetami-
for use by the US FDA between 2002 and nophen (paracetamol), one of the most widely
2014 indicated that safety-related labeling used OTC analgesics, is generally very safe
changes were being made as long as 13 years when taken as recommended on the product’s
after the products were approved. Because of label. Overdose, however, can result in acute
this lifecycle approach, all stakeholders – severe liver injury, which can lead to acute
patients, practitioners, manufacturers, and liver failure, and sometimes the need for liver
regulators – must remain vigilant about the transplantation or even death.
benefit–risk profile of a medicine. Such vigi-
lance is critical for informed decision making, Risk Management Is a Proactive Process
which is an important component of the safe Risk management must be proactive to be opti-
and effective use of medicinal products. mally effective. The ability to identify risks in
the pre-approval period allows manufacturers
Risk Management Applies to all to work with regulators on risk management
Medicines planning during the drug development phase.
All medicines have risks. The magnitude, fre- A proactive approach in the post-approval
quency, and severity of risks vary from medi- phase demands that manufacturers, regula-
cine to medicine. For example, at one end of tors, and practitioners agree on a system to
the spectrum, neutropenia is commonly asso- identify new risks, manage known risks, assess
ciated with chemotherapy and other immuno- the effectiveness of the risk management
suppressive agents and is a major risk factor efforts, and modify them as needed. A care-
for development of infections. Strategies to fully designed risk management plan can iden-
monitor, prevent, and manage these infections tify or further characterize risks, communicate
can lead to improved outcomes and will ensure and manage risks using evidence-based tools
the benefits of these drugs continue to out- when possible, and assess the effectiveness of
weigh the risk. At the other end of the spec- these efforts in a proactive way. The proactive
trum, many topical OTC medicines have very nature of risk management planning demands
few side effects. The management of these the constant vigilance of all stakeholders.
Risk Management Activities experimental data or on clinical trial data.

Managing the risks of medicines is not a single Non-experimental data include individual case
activity or the province of a single profession reports of suspected adverse drug reactions
or stakeholder group. Rather, it is an iterative (spontaneous reports), case series of such
process that involves a set of inter-related reports, databases of spontaneous reports,
activities, including risk assessment, risk mini- disease-based registries, drug-based registries,
mization, and evaluation of risk minimization electronic medical records systems, adminis-
strategies with adjustments, as appropriate, to trative claims databases, drug utilization data-
the risk minimization strategies to optimize bases, poison control center databases, and
the benefit–risk balance of the medicinal other public health databases that track usage
product. of the medicine.
It is also important for risk assessments to
Risk Assessment identify medication errors. Proactive risk
Risk assessment occurs throughout a product’s assessments that reflect human and environ-
lifecycle and consists of identifying, character- mental factors in medicine’s use should be
izing, and quantifying the risks associated with employed from the earliest stages of product
the use of a medicine, and evaluating their design to help anticipate potential medication
importance in relation to the benefit–risk bal- errors. After approval, the identification of
ance. Pre-approval risk assessment is an exten- medication errors must focus on identifying
sive process that involves preclinical safety the specific reasons for, or causes of, the event.
assessments, clinical pharmacology assess- New risks of a medicine will continue to be
ments, and clinical trials. Animal toxicology recognized after the drug is on the market.
studies are performed prior to the first human Some of these risks will be sufficiently serious
exposure to a new medicine to establish the to alter the benefit–risk balance of the medi-
general toxicity profile of the drug and to guide cine, such that post-approval regulatory action
initial human dosing. Clinical pharmacologic will be needed.
studies establish the pharmacokinetic profile
of the medicine, exposure-response relation- Risk Minimization
ships, and can be used to assess drug–drug Risk minimization or mitigation refers to a set
interactions. Pre-approval clinical trials proof interventions intended to prevent or reduce
vide the efficacy and safety information that the occurrence, or the severity of adverse
form the basis for an approval decision. The events associated with exposure to a medicine.
pre-approval safety assessment generally The range of risk mitigation activities varies
quantifies and characterizes the common from one country or region to the next, but cer-
adverse events associated with a medicinal tain common themes emerge.
product. Depending on the number of subjects The very fact that a medicine must be
exposed prior to approval, less common reviewed before approval is, in many ways, the
adverse events might also be detected. most fundamental risk mitigation activity, in
Because even large clinical development that it prohibits the marketing of medicines
programs cannot identify all risks associated that have not been judged to be safe and effec-
with a product, it is imperative that risk assess- tive, thus virtually eliminating the risks of
ment continue in the post-approval period, medicines being legally marketed for which
when large numbers of persons will be exposed there is no demonstrated benefit. The require-
to the medicine, including many with comor- ment that certain medicines be available only
bid conditions or on concomitant medications by prescription is another form of risk mitiga-
not present in clinical trials. Post-approval risk tion. The premise underlying the prescription-
assessment can be based on either non- only status of a medicine is that some
Risk Managemen 431
medicines are potentially harmful, or the medicine. Specific risk information may also
method of their use is not safe without the be targeted to healthcare professional societies
involvement of a health care provider, whose to share with their members. The types of com-
judgment can be used to ensure that, for a par- munication tools can include letters, prescriber
ticular patient, the potential benefits outweigh checklists, or educational brochures.
the potential risks. Communication tools for patients can include
a dosing card for medicines with complicated
Risk Communication dosing instructions or a patient alert card. In
Communicating information about the bene- certain cases, these additional communication
fits and risk of medicines is central to minimiz- measures may be required as part of a formal
ing the risks of medicines. The principal form Risk Minimization Plan in the EU or risk eval-
of communication to healthcare professionals uation and mitigation strategy (REMS) in the
in the US is the product’s approved profes- US.
sional labeling. In the EU, this professional Regulatory agencies have also been engaging
information is known as the Summary of in increasing efforts to communicate the risks
Product Characteristics (SmPC). of medicines. FDA’s primary tool for communi-
Additional communications to healthcare cating important new and emerging safety
professionals come in the form of so-called information about a medicine is through a Drug
“Dear Health Care Provider Letters” or “Direct Safety Communication or DSC. In the EU, the
Healthcare Profession Communication.” competent regulatory authorities communicate
These letters, typically issued by a medicine’s drug safety information using different meth-
manufacturer, are usually one to a few pages in ods, which depend upon what has been estab-
length, and generally focus on specific, newly- lished for the individual country. The European
identified safety information. Medicines Agency (EMA) plays a central role in
Labeling directed to patients and consumers coordinating the communications.
is also a risk communication tool in that it
highlights basic information necessary for the Additional Risk Minimization Strategies
safe use of the product, and often provides A variety of other minimization strategies can
instructions for actions that patients should be employed when product labeling and other
take when certain symptoms are present. In the forms of risk communication are not suffi-
US, FDA-approved patient labeling includes a cient. In the US, FDA reviews proposed propri-
Medication Guide, a Patient Package Insert, or etary names of medicinal products to ensure
instructions-for-use. In the EU, all medicines that these names are not similar in spelling or
are required to have a patient information leaf- pronunciation to the proprietary or established
let, which must be provided to the patient as names of other medicines. In addition, FDA
part of the product packaging, or in exceptional reviews the proposed container labels, carton
circumstances as a separate leaflet or even as labeling, packaging, and product design to
online information. This leaflet is based upon ensure that these do not have features that
the information provided in the SmPC but writ- could cause or contribute to medication errors.
ten in patient-friendly language. Similarly, in the EU, medicines authorized
If additional communication measures are through the centralized procedure have their
utilized, they are generally designed to address invented (or brand) name approved by the
one, or at most a few, specific important risks (invented) name review group who checks that
associated with a medicine and may include there are no products licensed with similar
focused risk information targeted to practition- names in the EU, which could lead to confu-
ers that are likely to prescribe the medicine or sion. The layout, format, and wording on the
to care for patients that are treated with the immediate and outer packaging of the product
are also reviewed as part of the evaluation pro- tions of the initial measures, if warranted, to
cedure of the medicine and these form part of improve the risk minimization strategy in the
the authorization. context of an iterative process of evaluation,
There are a variety of other strategies that correction, and re-evaluation throughout the
can be used to mitigate risks associated with a lifecycle of a medicinal product.
medicine. For health care providers, this might In the EU, the pharmacovigilance legislation
include specialized training or materials that explicitly requires the active monitoring of the
facilitate discussions between health care pro- outcome of risk minimization activities. In the
viders and patients about the risks and safe use US, the metrics of a REMS assessment plan is
of medicine such as prescriber-patient agree- approved in advance of REMS implementa-
ments. For teratogenic medications, these may tion, and assessment reports are submitted by
include strategies to prevent fetal exposure to a the manufacturers at pre-defined intervals and
medicine, including required pregnancy test- additionally, if needed.
ing prior to prescribing or dispensing the medi- The following principles apply to measuring
cine to individuals that could become pregnant, the effectiveness of risk minimization:
as well as contraceptive counseling. Some
●● Robust risk minimization evaluation is lon-
medicines may require patient monitoring or
gitudinal in nature.
periods of observation by a healthcare profes-
●● A multi-faceted assessment is needed for a
sional after administration or it may require
comprehensive risk minimization evaluation.
that a medicine is administered in a certain
There are some key elements aimed to evalu-
type of healthcare setting that is equipped to
ate the implementation of the risk minimi-
manage the serious adverse event.
zation, such as:
In the EU and the US, a controlled distribu-
1) enablers and barriers for optimal pro-
tion system may be used to minimize an impor-
gram delivery and success.
tant risk. The patients’ access to the medicine
2) stakeholders’ knowledge, attitudes and
in such a system is contingent on fulfilling
perception of risk.
strict requirements before the medicinal prod-
3) intended and observed clinical behavior.
uct is used. Since a controlled access program
●● Safety outcome data define the ultimate suc-
has large implications for all stakeholders, the
cess of a risk minimization program.
use of these types of programs are generally
●● The unintended consequences of a risk min-
limited and are guided by a clear therapeutic
imization measures should be taken into
need for the product based on its demonstrated
account.
benefit, the nature of the associated risk, and
the likelihood that this risk can be managed by First, evaluation of risk mitigation activities
such program. can assess if risk mitigation recommendations
are being followed (e.g. the proportion of
Evaluation of Risk Minimization patients who receive the required information,
and Mitigation Measures are aware of it, or using the tools provided).
Evaluation of risk minimization and mitiga- This can be measured using target audience
tion activities is a critical component of a risk surveys or via proxy indicators such as health-
management system; this aims to ensure that care professionals’ requests for refills of con-
the objectives of the risk mitigation measures sumable tools (e.g. checklists and forms).
are fulfilled and that the activities in place are Second, the evaluation can focus on under-
proportionate. Such evaluation is closely standing of the purpose of the risk minimiza-
related to the risk assessment activities, but it tion tools and their key messages (e.g. the
also differs in the way that enables modifica- proportion of correct responses in a test on
Risk Managemen 433
such risk minimization tools). Scientifically Methodological Problems to be

rigorous survey methods should be applied; Addressed by
comprehensive guidelines for research are Pharmacoepidemiologic Research
available in published literature.
The Roles of Pharmacoepidemiology
Next, the evaluation can determine if rec-
in Risk Management
ommended behaviors are being followed (e.g.
Pharmacoepidemiology can play several roles
if patients that have read the information they
in risk management. The most fundamental
are given, do the specific actions the informa-
role is to identify and quantify the risks of a
tion recommends). Questionnaire-based sur-
medicine using a variety of pharmacoepide-
veys are not well suited to assess behavioral
miologic techniques, including clinical trials,
modification because they rely on the
spontaneous reports, case series, and observa-
respondent’s self-reporting, which have a
tional pharmacoepidemiologic studies. Use of
large impact on the validity of the study find-
these techniques for risk assessment is
ings. Therefore, this evaluation could rely on
described in Parts I and II of this book.
time-trends analyses of data from electronic
An important use of pharmacoepidemiology
medical records.
is to measure how medications are used in
The ultimate measures of success of a risk
practice, especially if they are used under con-
minimization program are the safety outcomes.
ditions that can lead to adverse outcomes.
This evaluation can demonstrate whether the
Examples of pharmacoepidemiologic findings
introduction of a risk mitigation strategy leads
that could signal that a product is not being
to a decrease in frequency or the severity of the
used appropriately include a finding that a
adverse events. Incidence rates or cumulative
medication is being prescribed concomitantly
incidence are most appropriate to be used,
with a contraindicated medication, a finding
while reporting rates should only be used with
that a drug is being used in a population of
caution (e.g. for rare events), due to the well-
patients for whom the potential benefits do not
known underreporting. The incidence of the
outweigh the potential risks, and a finding that
adverse events can be evaluated in cohort stud-
a medication is frequently prescribed for a
ies using information from healthcare data-
duration of treatment that is associated with
bases or registries. Disease registries may be
an increased risk of serious adverse events. For
more suitable to evaluate the risk minimization
these analyses, drug utilization databases,
measures as they may contain comparator
electronic medical record systems, and other
groups, which could provide background rates
administrative healthcare data, especially
for the events.
those with longitudinal patient-level data, are
It is also important, while challenging, to
often useful.
identify potential barriers to patient access to
A third application of pharmacoepidemiol-
the medicine related to the risk minimization
ogy is to provide population-based assess-
strategies (e.g. providers in the patient’s area
ments of the causes and contexts in which
may choose not to prescribe the drug because
known harm from medications can occur. For
they may be unwilling to follow the risk mini-
these analyses, one or more public health
mization recommendations).
databases may be helpful to estimate the bur-
Pharmaceutical manufacturers usually fund
den of a given drug-related toxicity in the
and/or conduct the evaluations of risk mitiga-
population. Because they are designed for
tion strategies, and regulators review the
the public health purposes of quantifying
results of those evaluations. In some instances,
health and harm in society, projected national
regulators may conduct independent assess-
level estimates are often available. They are
ments of drug safety.
especially useful for characterizing and quan- Examples of Currently Available

tifying known drug risk, rather than identify- Solutions
ing new risks.
In both the US and the EU, specific legislation
A fourth emerging role of pharmacoepide-
has been enacted to formalize managing the
miology in the field of risk management is the
risks of medicines. The legal and regulatory
assessment of risk mitigation efforts. For an
frameworks in each of these jurisdictions are
effective evaluation, the risk mitigation activ-
beyond the scope of this chapter. In the US, the
ity must have a clearly defined goal that is rel-
legislation specifies that FDA can require a
evant and measurable. Goals that are based on
REMS when certain criteria are met. In the
vague or imprecise metrics generally cannot
EU, a Risk Management Plan (RMP) is required
be measured, and even if they are measurable,
for all medicinal products. Despite their differ-
interpretations of the findings would be diffi-
ences, REMS and RMPs share the common
cult. As noted above, assessing the effective-
features of being able to use, as appropriate,
ness of a risk mitigation strategy can be
and allowed by law or regulation, communica-
conducted at several levels, including pro-
tion to patients, communication to healthcare
cesses, behaviors, and health outcomes. While
professionals, and certain restrictions to man-
the traditional methods of pharmacoepidemi-
age the risks of medicines. In the US, it is
ology may be used to assess the observed
required that REMS be assessed to ensure that
behavior and health outcomes, it is quite likely
they are meeting their goals. In the EU, meas-
that additional methods, such as those used in
uring the effectiveness of risk minimization
social sciences and health policy and manage-
activities and interventions is included in the
ment fields, may be needed to assess process
definition of a risk management system.
and behavior. It is important to understand
the relationship between each component of
the risk mitigation strategy and the desired
The Future
health outcomes. It is possible that practition-
ers and patients adhere to the processes and Managing the risk of medicinal products is an
exhibit the behaviors desired by the risk miti- evolving area involving multiple stakeholders
gation strategy, but that the health outcome of in the complex medication use system.
interest is not improved or is difficult to meas- One critical area for future development is to
ure. Alternatively, it is possible that practition- continue to improve the way risk mitigation
ers and patients do not adhere to the processes activities are being implemented. Many of the
or exhibit the desired behaviors, but the risk mitigation tools have relied in whole or in
desired health outcome (e.g. a reduction in the part on communicating a risk associated with
specific risk) is achieved, perhaps because of a medicinal product to increase the stakehold-
other interventions or factors that were not ers’ awareness and knowledge. The goal being
part of the risk mitigation strategy. In either that awareness of a particular risk will impart
case, a critical examination of the risk mitiga- knowledge and influence prescribing practices
tion strategy would be necessary. of the medicinal product or how the practition-
Another role for pharmacoepidemiology is ers will monitor patients once treatment with
in the area of assessing risk communication. the product has begun.
Approaches may include a survey of patients’ Risk management plans are designed to
and healthcare providers’ understanding of work within a complex medication use sys-
the risks and safe use of a medicine. Other tem. A current challenge for risk manage-
approaches may include focus groups, ques- ment systems is that they be developed in
tionnaires, interviews, and other methods used ways that can integrate, with minimal diffi-
to assess readability and/or understanding. culty, into the current medication use systems
Risk Managemen 435
and in a manner that is least burdensome for mitigation strategies. Measurement of this
healthcare providers. Risk mitigation strate- impact is important, because it allows policy
gies that require documentation of safe use makers and other stakeholders to determine if
conditions need refinement. A future chal- the goals of the strategy are being met. The
lenge is to develop a quality systems approach challenges for this field include developing
to implementing this type of risk mitigation models to relate risk mitigation strategies to
strategy in a manner that is seamless for prac- health outcomes, as well as ways to identify the
titioners and within the scope of their usual contribution of individual components of the
clinical practice. strategy to the overall outcome. A further chal-
Another area for future refinement is to con- lenge is to assess if there are negative conse-
tinue to gather evidence of the impact of risk quences of risk mitigation strategies.
Case Example 23.6 For Pharmacoepidemiology and Risk Management

Risk Management Example in the EU – SmPC, a Physician Prescribing checklist was
Bupropion/Naltrexone (Mysimba) implemented to help:
use Mysimba only as approved, consider-
Background
●●
ing any risk factors

Bupropion/naltrexone (Mysimba) was
consider concomitant conditions when
authorized in the EU in March 2015. Mysimba
●●
evaluating individual risk–benefit balance,

is a medicine used together with diet and
before the treatment decision.
exercise to treat obesity; it can also be given
to very overweight patients who have Results
weight-related complications. Post-marketing, a review of all data for hepa-
totoxicity led to regulatory actions: product
Issue information recommendations updates. The
The main safety and tolerability concerns additional risk minimization activities con-
were related to central nervous system, gas- tinued, with re-evaluation planned integrat-
trointestinal adverse events, and longer-term ing results from the drug utilization study
cardiovascular outcomes. Both additional and physician survey.
pharmacovigilance and risk minimization
activities were required. Strengths
The comprehensive set of actions addressed
Approach limitations in safety data at approval and
Additional pharmacovigilance activities: The managed identified risks. Additional data
manufacturer agreed to conduct two phase- collected, including spontaneous reporting,
4 randomized clinical trials, evaluating major informed further regulatory decisions in an
adverse cardiovascular events, a drug utiliza- iterative process.
tion study, and a physician survey, looking at Limitations
the real-world use and the potential off- Additional activities focus on most impor-
label use. tant risks of the product considering the lim-
Risk minimization activities: The main risk ited time available to prescribers in the
minimization strategy was to prevent treatment act, as well as striking a balance
Mysimba use by patients with an increased between the need to know more and what is
risk of seizures and suicide. In addition to the reasonable to request from a manufacturer.
Key Points for Risk Management adverse drug event (ADE), and these errors
may be intercepted before they reach the
●● The risks of a medicine derive both from its
patient, or may reach the patient without con-
inherent pharmacological properties as well
sequence. However, a small fraction of medi-
as from the way the medicine is used.
cation errors indeed reaches the patient and
●● Because the risks of medicines can occur at
result in patient harm, typically described as
any point in the complex medication use sys-
an ADE. An ADE would be considered pre-
tem, managing the risks of medicines
ventable if a medication error is associated
requires that the entire medication use sys-
with the ADE.
tem be involved.
Given the prevalence of prescription medi-
●● Risk management is an iterative process
cation use, it is not surprising that preventable
involving multiple, related activities that
ADE are one of the most frequent types of pre-
proceed throughout the product’s lifecycle.
ventable iatrogenic injuries. The IOM report,
●● Risk mitigation refers to a set of activities
To Err Is Human suggested at least 44 000–
designed to minimize the risks of a medicine
98 000 deaths in the US are from iatrogenic
while preserving its benefits.
injury. If accurate, this would mean that there
●● Risk communication is an important com-
are about 8000 deaths yearly from ADE and
ponent of risk management.
1 million injuries from drug use.
●● It is critical that risk management activities
be evaluated.
Safety Theory
One prominent theory for human errors that
he Pharmacoepidemiology
T also applies for medicine was promoted by
of Medication Errors James Reason. He differentiated the person
approach from the system approach for which
Medications are the most commonly used he made use of the image of the Swiss cheese
form of medical therapy today. For adults, (“Swiss cheese model”). In the Swiss cheese
about 75% of office visits to GP and internists model, a system has set in place several barri-
are associated with the continuation or initia- ers, defenses and safeguards – pictured as
tion of a drug, while in the hospital multiple cheese slices – that should prevent errors.
medication orders tend to be written for each However, every defense has its shortcomings
patient daily. Medication errors have been (“holes in the cheese”), which is why there are
defined as “any error in the process of order- usually several combined, and under unfavora-
ing, dispensing, or administering a drug” ble circumstances, the barriers might fail alto-
regardless of whether an injury occurred or the gether, allowing an error to arise and slip
potential for injury was present. through the defenses. Hence, Reason empha-
Mechanistically, medication errors are trig- sizes, that an error likely always results from
gered from errors in planning actions (i.e. “active failures and latent conditions.” Over
knowledge-based mistakes or rule-based mis- the years, this model has been adopted and
takes) or errors in executing correctly planned refined for specific situations in medicine, e.g.
actions (i.e. action-based slips or memory- for administration errors. In summary, it high-
based lapses). In clinical practice, a medication lights that although “human errors” occur
error may occur at any stage of drug therapy, commonly, the true cause of accidents is often
including drug prescribing, transcribing, man- the underlying systems that allow a person
ufacturing, dispensing, administering, and error to result in an accident. Root cause analy-
monitoring. Medication errors with potential sis can be used to define the cause of the defect.
for harm are called near-misses or potential Only relatively infrequently are individuals
The Pharmacoepidemiology of Medication Error 437
responsible for clusters of errors, and, most and benefits of drugs, they can also be used to
often, errors resulting in harm are made by study medication errors and preventable ADEs
workers whose overall work is good. To make (i.e. those due to errors). Approaches to detect-
the hospital a safer place, a key initial step is to ing medication errors include manual or auto-
eliminate the culture of blame, and instead matic screening of claims data, administrative
build a culture of safety. Errors and adverse databases, medical records, electronic health
outcomes should be treated as opportunities records, incident reports mostly by providers
for improving the process of care through sys- in hospitals, patient monitoring, direct obser-
tem changes, rather than a signal to begin dis- vation often by pharmacists, and spontaneous
ciplinary proceedings. (self-reporting) approaches. All of these
Indeed, systems changes for reducing errors approaches have inherent advantages and pit-
can greatly reduce the likelihood of error, and falls and there is no single approach that is
probably in turn, of adverse outcomes. Within considered the gold standard for detecting
medicine, much of the research has come medication errors or ADEs. Factors which
from anesthesia, which has made major might influence the identification of medica-
improvements in safety. Examples of success- tion errors and ADEs include the setting
ful systems changes in medication delivery (ambulatory vs inpatients; routine care vs
include implementation of safer working con- research studies), the expected types of medi-
ditions that allow for concentrate and efficient cation errors (prescribing vs administration
work (e.g. staffing, facilities) but also imple- errors), and the projected costs of detection. In
mentation of system changes by use of infor- addition, the type of detection method influ-
mation technology such as computerized ences which types of medication errors are
physician order entry (CPOE), clinical deci- found (e.g. only those resulting in patient
sion support systems (CDSS), unit dosing, bar- harm) and with which frequency.
coding of medications, and implementation of Screening of claims data, administrative
“smart pumps” that can recognize what medi- databases, medical records, and electronic
cation is being delivered. These technologies health records is used to evaluate large data
can track medication use, and more impor- sets, but is generally done retrospectively. The
tantly, the frequencies and types of warnings quality of the available information, however,
as they alarm. varies between different data sources (see
Overall, the area of safety has a different phi- Chapters 8, 9, and 10) which restricts opportu-
losophy and several different tools than classic nities to comprehensively detect medication
epidemiology. For improving safety, culture is errors, more of some types than others.
extremely important, and tools such as root Especially in the outpatient setting, claims
cause analysis and failure mode and effects data can be obtained for very large numbers of
analysis – which can be used to project what individuals. Weaknesses include that it cannot
the problems with a process may be before be determined with certainty whether or not
they occur – are highly valuable. When com- the patient actually consumed the medication,
bined with epidemiologic data, such tools may and, if not linked to other information sources,
be extremely powerful for improving the safety clinical detail is often limited (e.g. information
of care. on weight or renal function but also informa-
tion on dosage is typically missing), making it
hard to answer questions that relate to a
Patient Safety Concepts as Applied
patient’s clinical condition.
to Pharmacoepidemiology
In the inpatient setting, manual chart review
While pharmacoepidemiology techniques is a well-established method to detect ADEs
have most often been used to study the risks and medication errors. With most relevant
patient information at hand, the appropriate- observation has been successfully and reliably
ness of drug prescribing and administration used to classify complex medication errors,
can be assessed, although documentation may and it is particularly useful at stages that are
still be incomplete, especially for assessing not sensitive to other detection methods (e.g.
issues such as appropriateness. The main prob- drug preparation and drug administration).
lems with chart review are that it is time- Many of the early medication error and ADE
consuming and expensive. When electronic studies were performed in the hospital setting.
health records are in place, the manual screen- In the inpatient adult setting, patients are vul-
ing of paper-based information can be replaced nerable to medication errors because of the
by semi-automated approaches. The level of severity of their illness, the complexity of their
standardization and the extent to which clini- disease process and medication regimens, and
cal information is stored by using controlled at times because of their age (e.g. the elderly
vocabulary determines the feasibility and are particularly susceptible). In pediatric drug
effectiveness of automated, algorithm-based use, the system-based factors that may contrib-
data analyses of ADEs and medication errors. ute to a higher rate of near misses include the
When electronic health records include elec- need for weight-based dosing and dilution of
tronic prescribing applications with clinical stock medicines, as well as decreased commu-
decision support, data from these applications nication abilities of young children.
can readily be used to detect many types of Knowledge about errors in the ambulatory
medication errors at the stage of prescribing. setting is increasing (see Case Example 23.7),
However, the specificity of these systems will although research lags behind the inpatient
also depend on the availability of information setting due to the difficulties of accessing
accessible via the electronic health records. patients once they leave a doctor’s office.
Spontaneous reporting (self-reported) of Increasingly, pharmacies are used for medica-
medication errors is comparatively easy to be tion safety studies, and a lot of research has
set in place and to maintain, both in inpatient been done about errors at the point of transi-
and outpatient settings. However, both ADEs, tion from hospitals to ambulatory settings and
medication errors and also critical incidents vice versa. Overall, comparisons among stud-
(i.e. near misses) are substantially underre- ies are challenging because of variations in
ported (see also Chapter 7). This may be due to data quality and methodology.
the fact that a certain situation must first be Of note, it is important to acknowledge that
recognized and evaluated as an incident, that while every setting has its particular risk for
reporting takes time, and that despite a nonpu- medication errors, new risks might also arise
nitive policy of the hospital, reporting is feared when new systems or processes are imple-
to be associated with disciplinary actions. mented in a specific environment. Hence,
Thus, spontaneous reporting is useful for get- while it is the obvious motivation of imple-
ting samples of errors and learning about error mentation of medication safety strategies to
etiology, but cannot be used to assess the mitigate the risk of medication errors, one
underlying rate of medication errors in a sam- must bear in mind that also new errors might
ple. However, patient monitoring using their become evident due to new risks.
self-reporting for ADEs has been successful,
and can identify more ADEs than chart review.
Direct observation is primarily done during
research studies at inpatient sites and offers a
comprehensive assessment of medication dis- Medication errors can occur at any stage of the
pensing and administration errors. While medication use process, including prescribing,
being both cost and personnel intensive, direct transcribing, dispensing, administering, and
Case Example 23.7 For the Use of Pharmacoepidemiology to Study Medication Errors
Most of the data on the frequency of adverse The key strength of this approach was that,
drug events have come from the inpatient by calling patients, it was possible to identify
setting, and many outpatient studies have many adverse drug events that were not
relied on chart review or claims data to noted in the chart.
detect ADEs. Gandhi’s 2003 study on the fre-
quency of adverse drug events in a Limitations
community-living population used a differ- The key weakness of this approach is that
ent approach. many of the effects patients attributed to
their medications may not have been due to
Issue the medications at all, but due to other
The goal of the study was to assess the fre- things such as their underlying conditions.
quency of adverse drug events in an ambula- The authors attempted to address this by
tory primary care population. asking the patient’s physician in each
Approach instance whether they believed the symp-
The frequency of adverse drug events was toms related to the medication.
assessed by calling patients after a visit at
Key points
which medications were prescribed, to deter-
●● Calling patients – though expensive and
mine whether or not an adverse drug event
time-consuming – identifies many adverse
had occurred, and in addition to review the
drug events that are not identified through
chart at three months.
chart review.
Results ●● Almost none of the visits was associated
●● Adverse drug events occurred at a rate of with an ICD-9 code suggesting the pres-
20.9 per 100 patients. ence of an ADE, suggesting that claims
●● About eight times as many adverse drug data should not be used to estimate the
events were identified by calling patients frequency of ADEs of all types in the out-
as by reviewing charts. patient setting.
●● While the severity of the ADEs overall was ●● More work is needed to facilitate assess-
fairly low, about a third were preventable, ment of whether a specific patient com-
and 6% were both serious and preventable. plaint is related to a medication.
monitoring. Of these stages, prescribing errors Important types of errors include dosing,
in the hospital have been documented to cause route, frequency, drug-allergy, drug–drug
the most harm, although errors at any stage interaction, drug–laboratory (including renal
can do so, and monitoring errors (i.e. errors dosing), drug–patient characteristic, and drug
caused by lack of proper monitoring) are quite administration during pregnancy. Although
prominent outside the hospital. The greater these errors occur most frequently at the drug
proportion of harmful errors at the drug pre- ordering stage, they can occur at any stage in
scribing stage may be a consequence of the the medication use process.
data collection methods employed in these In several studies, dosing errors have repre-
studies, which were multipronged but sented the most frequent category. To deter-
excluded direct observation, the most sensitive mine whether or not a dosing error is present,
technique for administration error detection. most often some clinical context is needed, for
example the patient’s age, gender, weight, level typically cannot be detected with claims data,
of renal function, prior response to the medica- since allergy information on patients is not
tion (if it has been used previously), response available. Thus, these errors have to be detected
to other similar medications, clinical condi- either through chart review, which is labori-
tion, and often the indication for the therapy. ous, or more often through electronic medical
While many of these data elements can be record data.
obtained from review of the medical chart, Drug–drug interaction exposures represent
many are not typically available from claims an interesting and difficult area, both for
data alone. research and interventions, to decrease errors.
Administration errors also represent a com- While many interactions have been reported,
mon type of error. Many drugs can be given by the severity varies substantially from minor to
one or a few routes and not by many others. life-threatening. If a conscious decision is
Some such errors – such as giving benzathine made to give patients two medications despite
penicillin that contains suspended solids intra- the knowledge that they may interact, this can-
venously instead of intramuscularly – would not be considered an error except in very lim-
often be fatal, and though they have caused ited circumstances, for example with
fatalities, are fortunately very rare. Other route meperidine and monoamine oxidase inhibi-
errors – such as grinding up a sustained tors. Also, it is legitimate to give many medica-
released preparation (which can negate the tions together despite clear interactions with
slow release properties of the drug) to give it important consequences if there are no good
via a tube – are much more frequent, and can alternatives, or if dose alterations are made, or
have serious consequences. Route errors are if additional monitoring is carried out (for
especially problematic at the administration example, with warfarin and many antibiotics).
stage of the medication use process and often However, the necessary alterations in dosing
happen in the context of patient-administered or additional monitoring are often omitted,
drugs. Unfortunately, administration errors which can have severe consequences. It is pos-
are both difficult to detect and much less often sible in large claims data sets to detect situa-
intercepted than prescribing errors. tions in which simultaneous exposures appear
Frequency errors can occur either at the pre- to have occurred, but not possible to determine
scribing, dispensing, or administration stage. if this actually occurred, as a physician may
While these errors probably cause less harm give patients instructions to cease the use of
cumulatively than dose or route errors, they one of the drugs.
can be problematic. Some frequency errors at Drug–laboratory errors (e.g. monitoring of
the prescribing or dispensing stage can be potassium) represent an important category of
detected even with claims or prescription data. errors, but can be difficult to detect electroni-
Such errors have greater potential for harm cally because of poor interfaces between labo-
when drugs are given with a greater frequency ratory and pharmacy information. Such errors
than intended. However, the therapeutic ben- are relatively straight-forward to identify when
efit may not be realized when given with too large pharmacy and laboratory databases can
low frequency, and extremely negative effects be linked, although again assessment of clini-
can occur for some drugs, for example with cal outcomes is difficult unless these data are
antiretrovirals, to which resistance develops if also available.
they are given at a low frequency. Renal dosing errors represent a specific sub-
Allergy errors represent a particularly serious type of drug–laboratory errors and are espe-
type of error, even though most of the time cially important; these errors can also be and
when a drug is given to a patient with a known often are considered dosing errors. In one large
allergy, the patient does well. Allergy errors inpatient study (Chertow et al. 2001), nearly
40% of inpatients had at least mild renal insuf- Many medications are contraindicated in
ficiency, and there are many medications that pregnancy. Here, the greatest difficulty for
require dosing adjustment in the presence of the investigator is assessing whether or not the
decreased glomerular filtration. In that study, patient is actually pregnant at the time of the
without clinical decision support, patients exposure, although this can be assessed retro-
received the appropriate dose and frequency of spectively by identifying the date of birth,
medication only 30% of the time. assuming a term pregnancy, and then working
Many studies of drug–patient characteristic backward. The outcomes of interest are often
checking have focused on the use of medications not represented in ways that make it easy to
in the presence of specific diseases. However, in perform analyses, although data on medica-
the future, genomic testing will undoubtedly tion exposures and on births are readily avail-
dominate, as many genes have profound effects able and can often be linked.
on drug metabolism (see Chapter 14). Currently, Another important piece of clinical informa-
few large data sets can be linked with genotype tion for pediatrics is a child’s weight. Most
information, but this is becoming increasingly pediatric medications are dosed on the basis of
frequent in clinical trials and a number of weight. Standardized documentation of this
cohorts are being established as well. information is unavailable, hindering not only
analyses of pediatric dosing but also actual
dosing by pediatricians.
A final issue is the coding of allergies. It is
Addressed by
important for both clinical care and research
that allergies are differentiated from sensitivi-
Information Bias ties or intolerances through codes rather than
In performing drug analyses, the present con- free text. Continued drug use in the presence
ventions preclude the determination of total of drug sensitivity may be perfectly appropri-
daily dose in several ways. Physicians may pre- ate, whereas the same treatment in the pres-
scribe a greater amount of medicine than is ence of an allergy is likely an error. It is
required for the time period prescribed. For particularly important that severe reactions,
example, if a patient requires 50 mg of atenolol such as anaphylaxis, are clearly coded and
per day, the doctor may actually write a pre- identifiable in the medical records. New aller-
scription for 100 mg of atenolol per day and gies need to be captured in better ways. The
verbally convey instructions to the patient to eventual aim is to have one universal allergy
divide the pills. This is particularly problem- list in an electronic format for each patient,
atic with drugs that must be titrated to the rather than multiple disparate lists.
appropriate therapeutic dose (e.g. warfarin). If
either physicians or pharmacists were required Sample Size Issues
to document an accurate total daily dose, this Sample sizes are often small in medication
would improve the ability to perform research. error and ADE studies if direct observation is
Another important methodological issue is used as detection method (high costs of data
measurement of patient adherence to medica- collection). Electronic databases will be an
tions (see also Chapter 20). Since prescribing important tool to improve sample sizes in a
and dispensing data are seldom jointly availa- cost effective manner. Computerized physician
ble, determining patient adherence is order entry systems, electronic health records,
extremely difficult. Improving clinician access test result viewing systems, computerized
to data from pharmacy benefit managers might pharmacy systems, bar-coding systems, phar-
be very useful, as might availability of elec- macy benefit managers, and claims systems
tronic prescription data to pharmacies. will all be important sources of such data.
There will be important regulatory issues that ●● In most studies, about one to two third of
will need to be addressed before actual con- adverse drug events are preventable.
struction and use of these systems. ●● The epidemiology of medication errors and
adverse drug events has been fairly well
Generalizability described for hospitalized adults, but less
Many existing medication error studies have information is available for specific popula-
limited generalizability due to setting or meth- tions, and for the ambulatory setting.
ods. For example, many studies have been per- ●● It is possible now to detect many medication
formed in tertiary care academic hospital errors using large claims databases, and as it
settings. It is unclear how findings from this becomes possible to link these data with
setting translate to other settings. Also, meth- more types of clinical data including espe-
odologies vary widely from study to study, hin- cially laboratory and diagnosis data, it will
dering comparisons. be feasible to more accurately assess the fre-
quency of medication errors across
populations.
The Future
●● The increasing use of electronic health
The future of pharmacoepidemiologic research records potentially linked to data collected
will include large databases that allow linking from wearables should have a dramatic effect
of prescription information with clinical and on our ability to do research in this area using
claims data. These types of databases will facil- pharmacoepidemiologic techniques.
itate the studies of medication errors and
ADEs. They will also be critical for detecting
rare ADEs. Sources of data for these databases enefit–Risk Assessments
B
will include systems of computerized physi-
of Medical Treatments
cian order entry, computerized pharmacy, bar-
coding, pharmacy benefit managers, and
Introduction
electronic health records. Standardized coding
of data, that is, the uniform coding of drug Assessing the benefit–risk (B-R) balance of
names, as well as doses and concentrations, medical treatments has always been an inte-
will be an important advancement to allow gral part of drug development, regulatory, and
easy analysis. public health decisions. However, methodol-
Other important issues that must be ogy and regulatory policies for B-R have
addressed are representing prescriptions in advanced considerably in the last decade.
ways that allow determination of total daily Advancements include the application of
dose, joint documentation of prescriptions and structured B-R frameworks and consideration
dispensing data to allow determination of of the patient perspective in regulatory sub-
patient adherence, clear documentation of missions and review. In parallel, numerous
conditions like pregnancy or weights of pediat- B-R initiatives have been led by pharmaceuti-
ric patients, and improved coding of allergies. cal and device trade organizations, patient
advocacy groups, public–private partnerships,
and academic groups.
Key Points for the
While definitions vary, benefit–risk is gener-
ally defined as weighing the benefits of a treat-
of Medication Errors
ment against its harms for its expected use.
●● Medication errors are very common, com- The term “risk” is ambiguous in the field of
pared to adverse drug events, and relatively benefit–risk and may refer to the general
few result in injury. nature of the harmful effect, its frequency, its
Benefit–Risk Assessments of Medical Treatment 443
severity or a combination of these concepts. To assessing the patient perspective, ranging from
lessen this ambiguity, the terms “harm” or qualitative focus groups and structured inter-
“unfavorable effect” and the analogs of “bene- views through quantitative preference studies.
fit” or “favorable effect” are used, however, the We address patient preference studies and
phrase “benefit-risk” is still used due to its their application in B-R decision making.
ubiquity. The goals for benefit–risk assess-
ments vary. Examples include go/no-go deci-
sions by a medical product development
Addressed by
company, reviews by Institutional Review
Boards and data monitoring committees, regu-
latory reviews by a health authority, reconsid- Methods for gathering, synthesizing, and com-
eration of a treatment after gaining new municating a B-R assessment based on availa-
information post-approval, and therapeutic ble data are the key tools for decision-making.
decisions by clinicians and patients. B-R
assessments are applied to pharmaceuticals, Identifying Appropriate Data
vaccines and medical devices, and consumer for Benefit–Risk Assessment
health products, and they are a critical compo- The considerations for data source selection and
nent of point-of-care decision making by a good pharmacoepidemiology practices apply to
patient and physician. B-R assessment. Data for B-R assessment come
from a variety of sources including randomized
controlled trials (see Chapter 17), spontaneous
reports and observational data sources (see Part
II Sources of Pharmacoepidemiologic Data,
Systematic Approach to B-R Chapters 7–11). When there are multiple data
Assessment sources, this can result in discordant findings,
Varying approaches to B-R assessment have differences in outcome definitions and ascer-
been used, resulting in considerable differ- tainment, important differences between popu-
ences in the depth, transparency, and clarity of lations, and different treatment comparators.
the assessment. What has emerged in recent The validity in identifying and characterizing
years is the use of structured, framework the important benefits and harms must be care-
approaches to rationally and defensibly frame, fully evaluated through assessment of confound-
conduct and communicate a B-R assessment. ing, bias, study design, and generalizability. The
We describe these frameworks and their principles for choosing among the available data
application. sources in pharmacoepidemiology are discussed
in Chapter 12. The validity of exposure and diag-
Incorporating the Patient Perspective nostic data is further discussed in Chapter 13.
Traditionally, the judgments required in the Design-based and analytic approaches to adjust
design and conduct of clinical research have for confounding and bias in observational stud-
been the province of physicians, scientists, and ies are discussed in Chapter 22.
regulators. Increasingly, regulatory agencies,
patient advocacy groups and industry have Integrating Benefits and Risks
advocated for patient engagement. While the Synthesizing the evidence in a structured man-
idea of patient-focused benefit–risk is well- ner is challenging. Typical research studies can
accepted, the challenge is to determine what have multiple efficacy and safety endpoints of
information should be obtained from patients, interest. Some outcomes may favor one
how it should be obtained, and how it can be treatment, while others may favor the compar-
used. There are numerous techniques for ator. Approaches that integrate benefits and
harms should be sufficiently inclusive to frameworks are a set of principles, processes

account for the frequency, clinical impact and and tools to guide decision-makers in select-
uncertainty of potentially many benefits and ing, organizing, analyzing and communicating
harms considered under multiple conditions evidence for B-R decisions. Frameworks lead
including important patient subgroups. to transparency, consistency and discipline in
B-R decisions.
Communicating Benefit–Risk There are several well-known B-R frame-
Assessment works: the FDA B-R framework, the
Communicating information about the bene- Pharmaceutical Research and Manufacturers of
fits and risk of medicines is central to both B-R America Benefit–Risk Action Team (BRAT)
assessment and risk management. This can be framework, the Multi-criteria Decision Analysis
challenging due to its complexity and the (MCDA) framework suggested by EMA for com-
diversity of stakeholders who will receive this plex B-R decisions, and the framework embed-
information. Structured B-R frameworks ded in the International Council for
improve the ability to communicate in a clear Harmonisation’s (ICH) PSUR/PBRER and
and consistent manner. Clinical Overview templates. While the frame-
works differ in focus and methodology, they
share many steps similar to those in the BRAT
Currently Available Solutions
framework presented in Table 23.2.
Structured Approaches to B-R : B-R Given the ubiquity of B-R frameworks and
Frameworks their incorporation into the regulatory guid-
The most significant recent advance in B-R has ance, we structure this currently available solu-
been the introduction of B-R frameworks. B-R tions section based on these framework steps.
Table 23.2 Steps in the BRAT benefit–risk framework.
Step Definition
1) Define decision Summarize the nature of disease, medical need for treatment, disease and
context treatment epidemiology, study treatment, dose/formulation, indication(s),
patient population, critical subgroups, comparator(s), time horizon for
outcomes, relevant decision-making bodies.
2) Identify and define Identify and define all important outcomes, define a preliminary set of measures
outcomes for each outcome, document rationale for outcomes to be included and excluded.
3) Identify and Determine and document all data sources, extract raw data, summarize over data
summarize source sources, assemble effects table.
data
4) Customize the Modify the outcome list and their definitions based on review of the data and
framework clinical expertise. May include tuning of outcomes not considered relevant to a
particular B-R assessment or stakeholder group.
5) Assess importance If applicable, assess outcome clinical impact, weight or preferences from the
of outcomes perspective of patients, decision makers or other stakeholders.
6) Integrated B-R Summarize data into tabular and graphical displays (e.g. effects table) to aid
assessment: analysis interpretation, identify and fill any information gaps, interpret summary
and visualization information, potentially conduct sensitivity analyses to assess the impact of
uncertainty on clinical or preference data.
7) Expert judgment and Render and communicate a decision.
communication
Define Decision Context fibrillation, the primary efficacy outcome was

The main role of the decision context is to stroke (ischemic or hemorrhagic) or systemic
ensure decision-makers are aligned on key ele- embolism, and the primary safety outcome
ments of the assessment. Characterizing the was major bleeding. Major bleeding here was
severity of the disease and the medical need defined as the composite of fatal bleeding,
are particularly important, as the greater the symptomatic critical organ bleeding (e.g. hem-
severity or medical need, the more allowances orrhagic stroke, intracerebral hemorrhage),
typically may be made for treatment-related transfusions 2 units packed red blood cells or
harms. These background elements are also an whole blood, or hemoglobin drops 2 g/dl.
avenue for incorporating the patient perspec- Fatal and non-fatal hemorrhagic stroke events
tive into B-R. The context forces up-front are included in both the primary efficacy and
agreement on an appropriate comparator, dose safety outcomes. This double counting can
and population, the choices of which are often lead to confusion, as some events will count as
not straightforward in post-approval contexts. a both a benefit and a harm. Double counting
For example, in post-approval settings, the also occurs between all-cause death and other
most appropriate comparator may be the one efficacy outcomes, as strokes, myocardial
used in recent clinical trials (e.g. placebo) or infarctions and systemic emboli may be fatal.
the standard of care or comparators available Counting these deaths twice can potentially
in claims or electronic health records. distort the findings. Finally, major bleeding
Examples of decision contexts can be found includes a mix of events that are fatal, cause
in the FDA’s “Voice of the Patient” reports and irreversible harm or are transient without
in recent drug approvals, searchable at https:// sequelae, yet each event is weighted equally in
www.accessdata.fda.gov/scripts/cder/daf. the composite outcome. This large range of
clinical impacts under one outcome compli-
Identify and Define Outcomes cates the comparison between benefits and
The second step of the BRAT B-R Framework harm.
is to identify and define the important out- The value trees in Figure 23.1 show one way
comes; i.e. benefits, harms and other treatment to resolve these issues. Figure 23.1a shows the
properties. While studies measure many out- key outcomes of a typical atrial fibrillation
comes, some are highly correlated with each trial. Figure 23.1b includes the same events,
other or are causally dependent, some double- but with ischemic events classified as benefits
count events, and some have a wide range of and hemorrhagic events classified as risks.
clinical impacts. Generally, B-R requires iden- This separation avoids double-counting occur-
tifying a smaller set of key outcomes that drive ring between benefits and harms. Additionally,
the assessment. A tool often used to select and efficacy outcomes are defined to avoid double-
depict the outcomes used in B-R is a value tree, counting. Finally, benefits and harms are clas-
a hierarchic graph in which outcomes are sified by whether they are fatal or generally
grouped by anatomic, functional or clinical result in irreversible harm. By separating the
impact (see Figure 23.1). fatal and irreversible events from less impact-
The value tree can be utilized to understand ful ones, a first pass at the B-R assessment can
problems caused by the interrelationships be made with clinically comparable benefits
between outcomes. Outcomes that are easily and harms, then less impactful events can be
interpreted individually can be problematic included (Figure 23.1b). In some cases, having
when considered collectively for B-R. For multiple value trees can be helpful, particu-
example, in the Randomized Evaluation of larly when different decision-makers have dif-
Long-Term Anticoagulation Therapy (RE-LY) ferent views on which events are most
study, the pivotal trial of dabigatran for atrial important.
(a)
Death All-cause death
Ischemic stroke
Stroke
Benefits Hemorrhagic stroke
Embolism Systemic embolism
Benefit-
Risk MI Myocardial infarction
Balance
Major bleeding
Risks Bleeding
Non-major clinically-relevant bleeding
(b)
Ischemic death
Non-fatal ischemic stroke

Fatal / irreversible
Benefits
ischemic events Non-fatal myocardial infarction
Non-fatal systemic embolism

Benefit-
Risk
Balance Fatal bleeding
Fatal / irreversible
hemorrhagic
events Non-fatal critical organ bleeding
Risks
Reversible Transfusions and hemoglobin drops

hemorrhagic
events Non-major clinically-relevant bleeding
Figure 23.1 Example value trees for treatment of atrial fibrillation. (a) Outcomes from a typical atrial
fibrillation statistical analysis plan. (b) Modified value tree with one approach for benefit–risk assessment.
Identify and Summarize Source Data i mportant outcomes from the value tree have
Approaches for handling the choice of the been captured and where confounding and
database, study design, and analysis that apply selection bias have been rigorously adjusted
to pharmacoepidemiology in general also (see Chapter 22).
apply to B-R assessment. The most robust data In medical product development, the main
for B-R are from randomized controlled trials data sources are the pivotal clinical trials
(see Chapter 17) and observational data where the protocol and statistical analysis plan
sources (See Chapters 7–12), where the pre-specify the efficacy outcomes, populations
and means for pooling data from multiple disabling vs transient major bleeds. If the avail-
studies (see Chapter 17). Specific safety out- able observational data sources cannot ade-
comes may be collected based on anticipated quately measure these outcomes, the tree will
safety effects due to the mechanism of action need to be modified. In some cases, an algo-
(for example, bleeding events in an anticoagu- rithm can approximate a desired outcome. For
lant trial). Otherwise, there is passive collec- example, principal hospital discharge diagno-
tion and reporting of safety data (adverse sis codes can be used to identify bleeding-
events) by investigators. related hospitalization and to differentiate
Following a medical treatment’s approval, upper and lower gastrointestinal bleeding.
data informing B-R accumulate in a larger Another approach is to simplify the value tree
sample of subjects in clinical practice and post- to two composite outcomes, potential throm-
approval studies mainly through multiple boembolic events and intracranial hemor-
observational sources. Post-approval studies rhage, both defined by ICD-10-CM or other
are more likely to compare two or more active medical codes. An advantage of this approach
treatments and observational studies generally is a more intuitive interpretation of the study
do not contain a placebo arm, unlike many outcomes compared to a large value tree.
randomized controlled trials for registration. Disadvantages are that the composite out-
Definitions of exposure and efficacy and safety comes are mixtures of events with varying
outcomes in the observational settings are severity (ranging from fatal to transient with
dependent upon the means of collection, and no sequelae) and there is the potential to miss
ascertainment may vary in quality and validity important underlying differences between
(See Chapter 13). treatments.
For the structured B-R assessment, the deci-
sions on the inclusion or exclusion of studies Assess Importance of Outcomes – Value
should be based on a pre-specified analysis Judgments and Patient Preferences
plan and documented for transparency in Step 5 of the BRAT framework (Table 23.2) is
decision-making. If comparisons are made to assess the relative importance of the out-
between effect estimates across studies or comes. Benefit–risk assessment is a combina-
meta-analysis is considered (see Chapter 20), tion of data-based probability assessment and
the extent to which study populations, defini- value judgments. Statistical analyses provide
tions, study designs, and analysis approaches the probability of events prevented and caused,
are comparable should be described including but they do not indicate how important those
the potential for bias and confounding. This is events are to decision-makers, including
particularly important in B-R given that multi- patients. In B-R, these value judgments are
ple outcomes are assessed. typically called “weights” or “preferences.”
Often, these value judgments are based on
Customize the Framework clinical judgment. FDA and EMA B-R assess-
After assessing source data, there may be a ments reflect this approach. For example, sev-
need to revisit the outcomes for B-R (Table 23.2, eral FDA anticoagulant B-R assessments are
step 4). For example, new AEs may emerge or similar to that described above, in which clini-
there may be differences between the out- cal judgment partitions events into two catego-
comes of interest vs those available in the ries: events that are fatal or cause irreversible
source data. In these cases, B-R assessment harm, and events that cause reversible harm.
may require a revised set of outcomes. This partitioning gives two effective weights
Consider the value tree in Figure 23.1b. The (roughly “high” and “low”). Similarly, the
outcomes in this tree require distinctions such EMA’s guidance on review of drug applications
as fatal vs non-fatal myocardial infarctions and stresses the need to describe the value
judgments used and notes that “a ‘descriptive’ many unanswered questions on their use for
approach with explicit considerations about regulatory applications. Ongoing initiatives to
the importance of the different effects and how address them are described below.
trade-offs are weighed will generally be
appropriate.” Putting it all Together – Integrated B-R
While most B-R problems can be assessed Assessment
with clinical judgment, there are many that We discuss integrated B-R assessment and
benefit from formal, rigorous studies to assess communicating the assessment, Steps 6 and 7
how patients weigh benefits and harms. These of the BRAT B-R framework (Figure 23.1),
“patient preference studies” measure what together.
attributes of a treatment are important to There are numerous approaches to combin-
patients, the relative importance of these attrib- ing data on medical need, outcome data,
utes, what tradeoffs between them patients uncertainty, clinical judgment, preferences,
would accept, and the heterogeneity of these heterogeneity, etc. into a B-R assessment.
results among patients. Preference studies are Qualitative approaches use textual descrip-
taking on increasing roles in pharmaceutical tive summaries to make a B-R argument.
and medical device companies, patient groups They are most applicable when the assess-
and regulatory agencies. An increasing number ment is self-evident from the data, e.g. statis-
of regulatory submissions include preference tically significant benefit and no appreciable
information, the FDA has issued guidance on adverse events. Semi-quantitative approaches
roles for preference information in medical use a combination of tabular and graphical
device reviews, and both the FDA and EMA are displays, potentially coupled with preference
conducting preference studies. There are many or weight information. Most B-R in regula-
methods to elicit patient preferences, ranging tory applications is descriptive or semi-
from qualitative interviews through quantita- quantitative. Quantitative approaches
tive surveys. Overviews of these methods are compute summary metrics by combining the
listed under Further Reading. data and potentially preferences from multi-
There have been numerous preference stud- ple outcomes. These tend to be applied in the
ies for anticoagulants, in both patients and academic literature.
other stakeholders. For example, in a prefer- A common means to display B-R data is an
ence survey for acute coronary syndrome, US effects table, which summarizes information
patients regarded non-fatal disabling stroke as on all key benefits and harms (Table 23.3). The
equal importance to death, or greater in impor- columns vary, but may include the outcome
tance to death in another survey, and they were name, a brief outcome definition and units,
willing to accept considerable probability of the outcome value for each treatment, esti-
bleeding in exchange for reductions in the mates of between treatment differences (e.g.
chance of death or disabling stroke. In addi- risk difference) with associated uncertainty
tion to population averages, preference studies (e.g. 95% confidence intervals), brief notes on
can measure heterogeneity in preferences. strength of evidence and a link to data sources.
Some patients will be very tolerant of risks, While relative measures such as relative risk,
while others will be very risk averse. Preference odds ratio or hazards ratio can also be included,
studies can measure how widely patients’ pref- these measures are less useful for B-R. Because
erences vary and whether there are distinct the baseline rates for outcomes may be dispa-
subgroups of patients with important differ- rate, a large relative risk may correspond to a
ence in preferences. very small absolute difference in the number
While patient preference studies are increas- of events when the baseline rate is low, while
ingly popular for B-R applications, there are small relative risks may correspond to large
Table 23.3 Effects table for atrial fibrillation (simulated data).
Event rate (/10 000

person-years)
Study Rate difference / 10 000 NNT or

Outcome drug Comparator person-years (95% CI) NNH
Efficacy
Ischemic death, ischemic stroke, MI 430 512 −87 (−143, −30) −115
or systemic embolism
Ischemic death 189 221 −34 (−72, 3) −294
Non-fatal ischemic stroke 154 172 −19 (−51, 13) −526
Non-fatal myocardial infarction 83 100 −17 (−43, 8) −588
Non-fatal systemic embolism 4 20 −16 (−26, −7) −625
Safety
Major bleeding 398 345 49 (0, 99) 204
Fatal and critical organ bleeding 119 125 −6 (−33, 20) −1667
Fatal bleeding 37 38 −1 (−15, 13) −10000
Non-fatal critical site bleeding 82 87 −5 (−31, 21) −2000
Transfusions > = 2 units or Hbg drop 308 245 60 (17, 103) 167
>−2 gm/dl
Clinically relevant non-major bleeding 1192 1137 44 (−50, 138) 227
All outcomes are measured per 10 000 person-years. CI = confidence interval, MI = myocardial infarction,
NNT = Number needed to treat, NNH = Number needed to harm. Positive NNTs or NNHs indicate more events on
comparator. Negative NNTs or NNHs indicate more events on the study drug.
absolute differences when the baseline rate is a measure of uncertainty but not for statistical
high. hypothesis testing. Number needed to treat
Table 23.3 shows an effects table for the (NNT) and number needed to harm (NNH) are
atrial fibrillation value tree in Figure 23.1b. also included in the table. NNT and NNH are
The data are simulated but realistic. All out- calculated as the reciprocal of the correspond-
comes in the example presented are in person- ing rate differences. NNT (NNH) is interpreted
year rates, though effects tables in general can as the number of person-years of exposure
include any type of risk or rate calculation and with a treatment vs the comparator needed to
can show categorical or continuous outcomes. prevent (or cause) one additional harmful
Because the incidence rate of most outcomes is event. When comparing one benefit to one
low, the data are scaled to a hypothetical popu- harm, NNT and NNH are often useful.
lation (10 000 person-years) to simplify repre- However, when considering many outcomes at
sentation and comprehension of these data. once, NNT and NNH are generally less helpful
The rate differences represent the additional and can require complex mathematics and
number of events caused or prevented by using weights. Additionally, confidence intervals for
one treatment compared to another treatment NNT and NNH can be difficult to interpret
with 10 000 person-years of exposure. With when associations are not statistically
many outcomes considered in B-R, we significant. For these reasons, NNT and NNH
generally show the 95% confidence interval as have limited application in B-R assessment.
When considering the full set of outcomes, a Figure 23.2 can sorted with the non-composite
forest plot of the rate differences in Table 23.3 outcomes placed in order of decreasing weight.
is very helpful, particularly when communi- If the outcomes with largest weight all favor
cating a B-R assessment (Figure 23.2). The one treatment, benefits outweigh harms for
interpretations of this effects table and forest that treatment. Another class of approaches
plot are addressed in the case example. are net clinical benefit (NCB) measures. While
the terminology is not standard, we define
Weighting and the Patient Perspective NCB approaches as those that use a weighted
The integrated B-R approaches reviewed above sum of risk (or rate) differences between study
used clinical judgment to make a defensible drug and comparator to summarize the differ-
B-R decision. The example in Table 23.3 and ence between treatments. Typically, beneficial
Figure 23.2 lends itself to such approaches. events have positive weights and harmful
When the B-R tradeoff is more complex; i.e. events have negative weights. The larger the
when some outcomes favor the study drug and absolute value of the weight or the larger the
other outcomes favor the comparator, weight- risk difference for an outcome, the more that
ing and preference assessments can be criti- outcome contributes to NCB. Sensitivity analy-
cal – both to incorporate the patient perspective ses can be conducted on NCB measures, where
and to make a B-R assessment. distributions for the weights are propagated
There are numerous approaches by which into uncertainty in the NCB result, allowing
the weighting or patient preferences can be assessing more complex metrics such as the
incorporated into B-R. Graphic techniques probability that benefit exceeds risk.
show the clinical and preference data in uni- More general, and complex, approaches to
son. For example, the risk differences in B-R assessment include multi-criteria deci-
–87
Death, stroke, MI & non-CNS systemic embolism
–34
Ischemic death
–19
Non-fatal ischemic stroke
–17
Non-fatal myocardial infarction
–16
Non-fatal non-CNS systemic embolism
49
Major bleeding
–6
Fatal & critical organ bleeding
–1
Fatal bleeding
–5
Non-fatal critical organ bleeding
60
Transfusions & Hgb decreases > 2 gm / dl
44
Non-major clinically-relevant bleeding
–150 –100 –50 0 50 100 150

Rate difference (/10,000 person-years)
Favors study drug Favors comparator
Figure 23.2 Forest plot showing rate differences per 10 000 person-years for key benefits and harms in
atrial fibrillation treatment (simulated data). Diamonds are points estimates. Bars show 95% confidence
intervals. Efficacy outcomes are in orange. Safety outcomes are in green.
sion analysis and stochastic multi-criteria preference studies. The Quantitative Sciences
acceptability analysis, for which references in the Pharmaceutical Industry Benefit–Risk
are included under Further Reading. Finally, Working Group has developed novel statistical
preference studies can be used to address the approaches. The International Society for
questions raised at the outset of this chapter– Pharmacoeconomics and Outcomes Research,
what probability of harm is acceptable in the Society for Medical Decision Making and
exchange for a given degree of benefit? A pref- the International Academy of Health
erence study can assess benefit–risk tradeoffs, Preference Research develop standards for B-R
such as the maximum acceptable risk, from decision-making and the conduct of patient
the perspective of patients; i.e. the maximum preference studies.
probability or level of severity of a harm that a
patient will accept in exchange for a given Advanced Methods on the Horizon
benefit. For example, in atrial fibrillation, in Benefit–risk assessment is an evolving field in
exchange for reducing the chance of non-fatal which assessments will become more struc-
disabling stroke by one percentage point, US tured and quantitative. Advances will include:
patients on average would accept over a 6.0% (i) refinement of B-R and patient preference
chance of non-major clinically relevant bleed- assessment tools, (ii) development of method-
ing and up to about a 2.0% chance of extracra- ologies with greater clinical applicability, and
nial major bleeding. (iii) transitioning B-R assessment from a post-
hoc exercise to a design and conduct issue
having pre-specified plans. Communication
The Future
of B-R will evolve with emphasis on risk (or
Initiatives, Guidances and Partnerships rate) difference summaries and graphical
There are many recent and on-going initiatives communications.
on B-R methods and policy. The FDA and A key goal of B-R is to contrast the benefits
EMA implemented structured approaches to and harms of therapeutic alternatives as expe-
communicating B-R assessment. New FDA rienced by patients. Typical approaches ana-
guidance for drugs and biologics is forthcom- lyze benefits and harms independently. B-R
ing. Additional guidance will focus on system- assessment based on combining the separate
atic approaches to utilizing patient and marginal effects of each outcome has limita-
caregiver input. The Council for International tions, failing to (i) account for outcome
Organizations in the Medical Sciences has dependencies, (ii) systematically incorporate
forthcoming updates to its publication on B-R the relative importance of combinations of
evaluation. Public–private partnerships align outcomes, (iii) summarize the cumulative
stakeholders on methodology and policy. For nature of various outcomes, and (iv) effectively
example, the Innovative Medicines Initiative deal with competing risks.
(IMI) developed recommendations for B-R For example, suppose 100 patients are
methodology and communication (IMI treated with a new treatment vs placebo, and
PROTECT), and the Medical Device Innovation one efficacy outcome and one safety outcome
Consortium developed a framework on the use are measured and considered equally impor-
of patient preference studies in medical device tance. Further suppose the efficacy and adverse
development. The IMI PREFER project will event rates for the new treatment are both 50%,
generate recommendations on the use of while the rates for both are zero for placebo. If
patient preference studies. Other professional the 50 patients who experience a treatment
societies have developed innovative methods benefit are the same patients who experienced
and standards for B-R decision making and harm, then the net clinical benefit is zero for
all patients (Figure 23.3a). However, if the 50 “ partial credit” to DOOR outcomes, ranging
patients with benefit are different than the from the most desirable (100) to least desirable
patients that experienced the harm, then the net (0). The amount of partial credit may be
clinical benefit is positive for half the patients obtained through patient preference studies or
and negative for half the patients (Figure 23.3b). a survey of expert clinicians. Sensitivity analy-
Separate marginal summaries cannot distin- ses to varying partial credit scoring permits
guish these two scenarios. personalized analyses.
One method that accounts for these depend- In the future, the bond between B-R assess-
encies is desirability of outcome ranking ment and precision medicine will grow. The
(DOOR). Conceptually, DOOR uses outcomes goal in practice is not to identify who will ben-
to analyze the patients rather than the patients efit or who will be harmed, but who has a posi-
to analyze the outcomes. A key to utilizing tive B-R profile. B-R assessment will transition
DOOR is to determine how to analyze one from a post-hoc activity to one that begins in
patient before analyzing many. study design and proceeds through trial con-
DOOR begins with construction of a global duct, is pre-specified, systematic, and a pri-
ordinal outcome, where a DOOR outcome is a mary focus. For example, structured B-R
composition of benefits and harms that occur assessment may become a standard assess-
in a single patient. In one approach to DOOR, ment of data monitoring committees. When
one can estimate the probability that a ran- formal B-R hypotheses cannot be prespecified,
domly selected patient randomized to one the planned methodology for assessing B-R
treatment has a better DOOR than a randomly can be prespecified. Patient preference studies
selected patient on the other treatment. This and the collection of quality-of-life information
has an intuitive appeal from a clinical perspec- will become more prevalent and will be
tive given the connection to the clinical included in clinical trials more often, rather
decision-making question of what is the prob- than being evaluations in a separate popula-
ability that a patient will have a better overall tion. Sensitivity analyses to patient preferences
response on one treatment relative to another. will become common, assessing the robustness
An alternate approach to DOOR assigns of the results.
(a) (b)
Benefit No benefit Benefit No benefit
Harm 50% … Harm … 50%
No harm … 50% No harm 50% …
Figure 23.3 Demonstration of the impact of dependencies between outcomes. Both tables show risk
differences for an equally weighted benefit and risk. In Figure 23.3a the treatment benefits no one (NCB = 0
for all patients); In Figure 23.3b, the treatment benefits half of patients (NCB > 0 for 50 patients and NCB < 0
for 50 patients). However, the marginal distributions of benefit and harm of the two tables are identical.
Case Example 23.8 For Benefit–Risk Assessment

Background Results
●● A novel favor Xa inhibitor anticoagulant ●● The decision context is to assess B- R of
(study drug) was studied in a pivotal phase the study drug vs the comparator in sub-
3 randomized clinical trial, comparing the jects with non-valvular atrial fibrillation,
study drug to an active comparator in over a period of up to four years from ini-
subjects with non-valvular atrial fibrilla- tial dose, for regulatory decision-makers.
tion (ECG-confirmed). 10 000 subjects ●● An effects table and a forest plots using
age ≥ 18 years. Study duration was up to a exposure-time rates per 10 000 person-
maximum of four years. years were generated (Table 23.3,
●● The primary efficacy outcome was the Figure 23.2). Focusing first on events that
composite of stroke and non-central nerv- are fatal or cause irreversible harm, per
ous system embolism (NCSE). The primary 10 000 person-years, there are 87 (95% CI
safety outcome was major bleeding, the 30, 143) fewer events per 10 000 person-
composite of fatal bleeding, critical organ years in the efficacy composite of ischemic
bleeding (e.g. hemorrhagic stroke, intracer- death, ischemic stroke, MI or systemic embo-
ebral hemorrhage), transfusions ≥2 units lism; and there are 6 (95% CI -20, 33) fewer
packed red blood cells or whole blood, or fatal and critical organ bleeding events on
hemoglobin drops ≥2 g/dl. the study drug. The composite efficacy out-
come is nominally statistically significant
Question favoring the study drug, while the fatal and
For the health authority submission, do ben- critical organ bleeding shows little differ-
efits outweigh harms for the study drug vs ence; suggesting benefits outweigh harms.
active comparator? ●● The forest plot makes clear visually that
death, ischemic stroke, myocardial infarc-

Approach
tion and non-CNS systemic embolism each
●● Utilize the BRAT framework to develop the
contribute meaningfully to efficacy. Major
decision context, value tree, consider the
bleeding favors the comparator, but this
importance of outcomes and develop an
difference is driven primarily by less
integrated B-R assessment.
impactful bleeds of transfusions and
●● The value tree based on clinical trial out-
hemoglobin reductions. The B-R tradeoff is
comes (Figure 23.1a) was modified for B-R
preventing 87 (95% CI 30, 143) fatal/irre-
assessment to remove double-counting
versible harm ischemic events vs causing
and avoid composite outcomes that
60 (95% CI 17, 108) reversible bleeding
include events with a wide range of clini-
events and 44 (−40, 13) clinically relevant
cal impact (Figure 23.1b). The efficacy
non-major bleeds. Even without taking
composite outcome is ischemic events
patient preferences into account, the ben-
that are generally fatal or cause irreversi-
efits appear to outweigh the harms.
ble harm. The safety composite outcome is
●● This B- R analysis can also be performed
hemorrhagic events that are generally
with NNT and NNH. For example, one fatal/
fatal or cause irreversible harm. Additional
irreversible harm ischemic event is
hemorrhagic events are included under a
prevented for every 115 person-years of
composite outcome “reversible hemor-
exposure to the study drug vs comparator
rhagic events.”
(Continued)
(NNT), while 204 person-years exposure is Limitations

needed to see an excess major bleeding ●● The outcomes are considered independ-
event (NNH) on study drug, most of which ent. A more through B-R assessment
are far less clinically impactful than the would assess whether there is a depend-
ischemic events, also strongly suggesting ency between key outcomes, especially
that benefit exceeds harm. between ischemic events and hemor-
rhagic events.
Strengths
●● While the transfusions, hemoglobin reduc-
●● The use of outcomes that avoid double-
tions and clinically relevant non-major
counting, clearly distinguish between
bleeds are far less impactful than the ben-
harmful events caused vs harmful events
efits and harms in the primary B-R assess-
prevented, and the use of a pair of compos-
ment, these less impactful bleeds may
ite outcomes with similar clinical impact
cause patients to discontinue treatment,
to simplify assessment of whether benefits
indirectly favoring the comparator.
outweigh harms.
●● The use of risk differences to clearly indi-
cate the number of events caused and Key points

prevented. Use of a structured B-R framework enables a
●● Concise tabular and graphic depictions flexible, transparent and defensible approach
that clearly depict the B-R assessment. to B-R assessment
Key Points for Benefit-Risk have developed much more rigorous B-R
Assessments of Medical regulatory policies relating to regulatory
Treatments approval and post-approval decisions.
●● These advances have led to transparent
●● Assessing the B-R balance of medical treat-
and rigorous systematic approaches to
ments has always been an integral part of
B-R which incorporate the patient per-
drug development, regulatory, and public
spective and can be communicated clearly
health decisions.
and succinctly to clinical and patient
●● Methodology for B-R has advanced consid-
audiences.
erably, and health authorities worldwide
Further Reading
Studies of Drug Utilization MD: American Society of Health System

Pharmacists.
Beer, C., Hyde, Z., Almeida, O.P. et al. (2011). Dartnell, J.G.A. (2001). Understanding,
Quality use of medicines and health outcomes Influencing and Evaluating Drug Use.
among a cohort of community dwelling older Australia: Therapeutic Guidelines Limited.
men: an observational study. Br. J. Clin. Elseviers, M., Wettermark, B., Almarsdottir, A.B.
Pharmacol. 71: 592–599. et al. (eds.) (2016). Drug Utilization Research:
Brown, T.R. (ed.) (2006). Handbook of Methods and Applications. Chichester, West
Institutional Pharmacy Practice, 4e. Bethesda, Sussex: Hoboken NJ: Wiley.
Further Reading 455
Hennessy, S.M., Bilker, W.B., Zhou, L. et al. Cabana, M.D., Rand, C.S., Power, N.R. et al.
(2003). Retrospective drug utilization review, (1999). Why don’t physicians follow clinical
prescribing errors, and clinical outcomes. practice guidelines? JAMA 282: 1458–1465.
JAMA 290: 1494–1499. Christensen, M. and Lundh, A. (2016).
Kauffman, D.W., Kelly, J.P., Rosenberg, L. et al. Medication review in hospitalised patients to
(2002). Recent patterns of medication use in reduce morbidity and mortality. Cochrane
the ambulatory adult population of the United Database Syst. Rev. 2: CD008986.
States: the Slone survey. JAMA 287: 337–344. Cochrane Effective Practice and Organization of
Kidder, D. and Bae, J. (1999). Evaluation results Care Group. What study designs should be
from prospective drug utilization review: included in an EPOC review and what should
medicaid demonstrations. Health Care Financ. they be called? 2017; http://epoc.cochrane.org/
Rev. 20: 107–118. sites/epoc.cochrane.org/files/public/uploads/
World Health Organization (1993) How to Resources-for-authors2017/what_study_
Investigate Drug Use in Health Facilities: designs_should_be_included_in_an_epoc_
Selected Drug Use Indicators. WHO/ review.pdf. Accessed November 29, 2017.
DAP/93.1. Geneva: World Health Davey, P., Marwick, C.A., Scott, C.L. et al. (2017).
Organization. Interventions to improve antibiotic
World Health Organization (2009) Medicines prescribing practices for hospital inpatients.
Use in Primary Care in Developing and Cochrane Database Syst. Rev. 2: CD003543.
Transitional Countries: Fact Book Donner, A., Birkett, N., and Buck, C. (1981).
Summarizing Results from Studies Reported Randomization by cluster – samples size
between 1990 and 2006. WHO/EMP/ requirements and analysis. Am. J. Epidemiol.
MAR/2009.3. Geneva, Switzerland: World 114: 906–914.
Health Organization. Flodgren, G., Parmelli, E., Doumit, G. et al. (2011).
World Health Organization International Local opinion leaders: effects on professional
Working Group for Drug Statistics practice and health care outcomes. Cochrane
Methodology, WHO Collaborating Centre for Database Syst. Rev. 8: CD000125.
Drug Statistics Methodology, WHO Giguere, A., Legare, F., Grimshaw, J. et al. (2012).
Collaborating Centre for Drug Utilization Printed educational materials: effects on
Research and Clinical Pharmacology (2003). professional practice and healthcare outcomes.
Introduction to Drug Utilization Research. Cochrane Database Syst. Rev. 10: CD004398.
Geneva: World Health Organization. Gillaizeau, F., Chan, E., Trinquart, L. et al.
(2013). Computerized advice on drug dosage
to improve prescribing practice. Cochrane
Database Syst. Rev. 11: CD002894.
Evaluating and Improving Godin, G., Belanger-Gravel, A., Eccles, M., and
Prescribing Grimshaw, J. (2008). Healthcare professionals’
intentions and behaviours: a systematic
Arditi, C., Rege-Walther, M., Durieux, P., and review of studies based on social cognitive
Burnand, B. (2017). Computer-generated theories. Implement. Sci. 3: 36–46.
reminders delivered on paper to healthcare Gonzales, R., Steiner, J.F., Lum, A., and Barrett,
professionals: effects on professional practice P.H. (1999). Decreasing antibiotic use in
and healthcare outcomes. Cochrane Database ambulatory practice: impact of a
Syst. Rev. 7: CD001175. multidimensional intervention on the
Auerbach, A.D., Landefield, C.S., and Shojania, treatment of uncomplicated acute bronchitis
K.G. (2007). The tension between needing to in adults. JAMA 281: 1512–1519.
improve care and knowing how to do it. N. Greer, A.L. (1988). The state of the art versus the
Engl. J. Med. 357: 608–613. state of the science: the diffusion of new
medical technologies into practice. Int. J. winding road. J. Am. Coll. Cardiol. 43:
Technol. Assess. Health Care 4: 5–26. 1738–1742.
Grimshaw, J.M., Shirran, L., Thomas, R. et al. Majumdar, S.R., Tsuyuki, R.T., and McAlister,
(2001). Changing provider behavior: an F.A. (2007). Impact of opinion leader-
overview of systematic reviews of endorsed evidence summaries on the quality
interventions. Med. Care 39 (suppl 2): 2–45. of prescribing for patients with cardiovascular
Grudniewicz, A., Kealy, R., Rodseth, R.N. et al. disease: a randomized controlled trial. Am.
(2015). What is the effectiveness of printed Heart J. 153 (22): e21–e28.
educational materials on primary care Mason, J., Freemantle, N., Nazareth, I. et al.
physician knowledge, behaviour, and patient (2001). When is it cost-effective to change the
outcomes: a systematic review and meta- behavior of health professionals? JAMA 286:
analyses. Implement. Sci. 10: 164. 2988–2992.
Han, Y.Y., Carcillo, J.A., Venkataraman, S.T. Mostofian, F., Ruban, C., Simunovic, N., and
et al. (2005). Unexpected increased mortality Bhandari, M. (2015). Changing physician
after implementation of a commercially sold behavior: what works? Am. J. Manag. Care 21:
computerized physician order entry system. 75–84.
Pediatrics 116: 1506–1512. Nanji, K.C., Slight, S.P., Seger, D.L. et al. (2014).
Isaac, T., Weissman, J.S., Davis, R.B. et al. Overrides of medication-related clinical
(2009). Overrides of medication alerts in decision support alerts in outpatients. J. Am.
ambulatory care. Arch. Intern. Med. 169: Med. Inform. Assoc. 21: 487–491.
305–311. O’Brien, M.A., Rogers, S., Jamtvedt, G. et al.
Ivers, N., Jamtvedt, G., Flottorp, S. et al. (2012). (2007). Educational outreach visits: effects on
Audit and feedback: effects on professional professional practice and health care
practice and healthcare outcomes. Cochrane outcomes. Cochrane Database Syst. Rev. 4:
Database Syst. Rev. 6: CD000259. CD000409.
Johnson, M.J. and May, C.R. (2015). Promoting Rousseau, N., McColl, E., Newton, J. et al.
professional behaviour change in healthcare: (2003). Practice based, longitudinal,
what interventions work, and why? A qualitative interview study of computerized
theory-led overview of systematic reviews. evidence based guidelines in primary care.
BMJ Open 5: e008592. BMJ 326: 314–322.
Lu, C.Y., Ross-Degnan, D., Soumerai, S.B., and Serumaga, B., Ross-Degnan, R., Avery, A.J.
Pearson, S.A. (2008). Interventions designed et al. (2011). Has pay-for-performance
to improve the quality and efficiency of improved the management and outcomes of
medication use in managed care: a critical hypertension in the United Kingdom? BMJ
review of the literature – 2001-2007. BMC 342: 322–329.
Health Serv. Res. 8: 75. Serumaga, B., Ross-Degnan, D., Avery, A.J. et al.
Luiza, V.L., Chaves, L.A., Silva, R.M. et al. (2011). Effect of pay for performance on the
(2015). Pharmaceutical policies: effects of cop management and outcomes of hypertension
and co-payment on rational use of medicines. in the United Kingdom: interrupted time
Cochrane Database Syst. Rev. 2015: CD007017. series study. BMJ 342: d108.
Majumdar, S.R., Guirguis, L.M., Toth, E.L. et al. Simon, S.R., Majumdar, S.R., Prosser, L.A. et al.
(2003). Controlled trial of a multifaceted (2005). Group versus individual academic
intervention for improving quality of care for detailing to improve the use of
rural patients with type 2 diabetes. Diabetes antihypertensive medications in primary care:
Care 26: 3061–3066. a cluster-randomized controlled trial. Am. J.
Majumdar, S.R., McAlister, F.A., and Furberg, Med. 118: 521–528.
C.D. (2004). From knowledge to practice in Soumerai, S.B., McLaughlin, T.J., and Avorn, J.
chronic cardiovascular disease–a long and (1989). Improving drug prescribing in primary
Special Methodological Issues in Pharmacoepidemiologic Studies of Vaccine Safety 457
care: a critical analysis of the experimental Howson, C.P., Howe, C.J., and Fineberg, H.V.
literature. Milbank Q. 62 (2): 268–317. (eds.) (1991). Adverse Effects of Pertussis and
Soumerai, S.B., McLaughlin, T.J., Gurwitz, J.H. Rubella Vaccines: A Report of the Committee to
et al. (1998). Effect of local medical opinion Review the Adverse Consequences of Pertussis
leaders on quality of care for acute myocardial and Rubella Vaccines. IOM (Institute of
infarction: a randomized controlled trial. Medicine). 1991 Washington, DC: National
JAMA 279: 1358–1363. Academy Press.
Huang, W.T., Chen, R.T., Hsu, Y.C. et al. (2017).
Vaccination and unexplained sudden death
Special Methodological Issues risk in Taiwanese infants. Pharmacoepidemiol.
in Pharmacoepidemiologic Studies Drug Saf. 26 (1): 17–25.
of Vaccine Safety Izurieta, H.S., Zuber, P., Bonhoeffer, J. et al.
(2013). Roadmap for the international
Bate, A. and Evans, S.J.W. (2009). Quantitative collaborative epidemiologic monitoring of
signal detection using spontaneous ADR safety and effectiveness of new high priority
reporting. Pharmacoepidemiol. Drug Saf. 18: vaccines. Vaccine 31 (35): 3623–3627.
427–436. Kramarz, P., France, E.K., DeStefano, F. et al.
Cashman, P., Macartney, K., Khandaker, G. et al. (2001). Population-based study of rotavirus
(2017). Participant-centred active surveillance vaccination and intussusception. Pediatr.
of adverse events following immunisation: a Infect. Dis. J. 20: 410–416.
narrative review. Int. Health 9 (3): 164–176. Larson, H.J., Cooper, L.Z., Eskola, J. et al. (2011).
Centers for Disease Control and Prevention Addressing the vaccine confidence gap. Lancet
(1999). Intussusception among recipients of 378 (9790): 526–535.
rotavirus vaccine– United States, 1998–1999. McNeil, M.M., Gee, J., Weintraub, E.S. et al.
MMWR Morb. Mortal. Wkly Rep. 48: (2014). The vaccine safety datalink: successes
577–581. and challenges monitoring vaccine safety.
Fine, P.E. and Chen, R.T. (2001). Confounding in Vaccine 32 (42): 5390–5398.
studies of adverse reactions to vaccines. Am. J. Murch, S.H., Anthony, A., Casson, D.H. et al.
Epidemiol. 136: 121–135. (2004). Retraction of an interpretation. Lancet
Glanz, J.M., McClure, D.L., Xu, S. et al. (2005). 363: 750.
Four different study designs to evaluate Murphy, T.V., Gargiullo, P.M., Massoudi, M.S.
vaccine safety were equally validated with et al. (2001). Intussusception among infants
contrasting limitations. J. Clin. Epidemiol. 59: given an oral rotavirus vaccine. N. Engl. J.
808–818. Med. 344: 564–572.
Glanz, J.M., McClure, D.L., Magid, D.J. et al. Poland, G.A., Ovsyannikova, I.G., and Kennedy,
(2009). Parental refusal of pertussis R.B. (2018 Aug 28). Personalized vaccinology:
vaccination is associated with an increased a review. Vaccine 36 (36): 5350–5357.
risk of pertussis infection in children. Schonberger, L.B., Bregman, D.J., Sullivan-
Pediatrics 123: 1446–1451. Bolyai, J.Z. et al. (1979). Guillain–Barre
Gruber, M.F. and Marshall, V.B. (2018). syndrome following vaccination in the
Regulation and testing of vaccines. In: National Influenza Immunization Program,
Vaccines, 7e (eds. S. Plotkin, W.A. Orenstein, United States, 1976–1977. Am. J. Epidemiol.
P. Offit and K.M. Edward), 1547–1565. 110: 105–123.
Philadelphia, PA: W.B. Saunders. Stratton, K., Ford, A., Rusch, E., and Clayton,
Halsell, J.S., Riddle, J.R., Atwood, J.E. et al. E.W. (eds.) (2011). Adverse Effects of Vaccines:
(2003). Myopericarditis following smallpox Evidence and Causality. IOM (Institute of
vaccination among vaccinia-naive US military Medicine). Washington, DC: The National
personnel. JAMA 289: 3283–3289. Academies Press.
Wang, S.V., Stefanini, K., Lewis, E. et al. (2020). https://www.globenewswire.com/news-

Determining which of several simultaneously release/2019/09/19/1918062/0/en/Global-
administered vaccines increase risk of an Medical-Devices-Market-Report-2019-2022-A-
adverse event. Drug Saf. (in press). 521-Billion-Opportunity-Analysis.html.
Xu, S., Zhang, L., Nelson, J. et al. (2010). Published 2019. Accessed August 1, 2020.
Identifying optimal risk windows for self- International Medical Device Regulators Forum
controlled case series studies of vaccine safety. (IMDRF) Patient Registries Working Group.
Stat. Med. 30: 142–152. Principles of International System of
Yih, W.K., Kulldorff, M., Dashevsky, I., and Registries Linked to Other Data Sources and
Maro, J.C. (2019 Jul 1). Using the self- Tools. http://www.imdrf.org/docs/imdrf/
controlled tree-temporal scan statistic to final/technical/imdrf-tech-160930-principles-
assess the safety of live attenuated herpes system-registries.pdf. Published 2016.
zoster vaccine. Am. J. Epidemiol. 188 (7): Accessed May 1, 2018.
1383–1388. International Medical Device Regulators Forum
(IMDRF) Patient Registries Working Group.
Tools for Assessing the Usability of Registries
Epidemiologic Studies in Support of Regulatory Decision-Making.
of Implantable Medical Devices http://www.imdrf.org/docs/imdrf/final/
consultations/imdrf-cons-registries-n46-pd1-
Adhikari, S., Rose, S., and Normand, S.-L. 170817.pdf. Published 2017. Accessed May 1,
(November 2019). Nonparametric Bayesian 2018.
instrumental variable analysis: evaluating International Medical Device Regulators Forum
heterogeneous effects of coronary arterial (IMDRF) Patient Registries Working Group.
access site strategies. J. Am. Stat. Assoc.: 1–14. Methodological Principles in the Use of
https://doi.org/10.1080/01621459.2019.1688663. International Medical Device Registry Data.
Chughtai, B., Mao, J., Buck, J. et al. (2015). Use http://www.imdrf.org/docs/imdrf/final/
and risks of surgical mesh for pelvic organ technical/imdrf-tech-170316-methodological-
prolapse surgery in women in New York principles.pdf. Published 2016. Accessed May
state: population based cohort study. BMJ 1, 2018.
350: h2685. https://doi.org/10.1136/bmj. Rose, S. and Normand, S.-L. (2019). Double
h2685. robust estimation for multiple unordered
Fortune Business Insights. Medical Devices treatments and clustered observations:
Market Size, Share and Industry Analysis By evaluating drug-eluting coronary artery stents.
Type (Orthopedic Devices, Cardiovascular Biometrics 75 (1): 289–296. https://doi.
Devices, Diagnostic Imaging, IVD, MIS, org/10.1111/biom.12927.
Wound Management, Diabetes Care, Shuren, J. and Califf, R.M. (2016). Need for a
Ophthalmic Devices, Dental & Nephrology), National Evaluation System for Health
End User (Hospitals & Ambulatory Surgica. Technology. JAMA 316 (23): 2429–2437.
https://www.fortunebusinessinsights.com/ Tsougranis, G., Eldrup-Jorgensen, J., Bertges, D.
industry-reports/medical-devices- et al. (2020). The Vascular Implant
market-100085. Published 2019. Accessed Surveillance and Interventional Outcomes
August 1, 2020. (VISION) Coordinated Registry Network: An
Global News Wire. Global Medical Devices effort to advance evidence evaluation for
Market Report 2019–2022 -A $521+ Billion vascular devices. J Vasc Surg. 72 (6):
Opportunity Analysis. Research and Markets. 2153–2160.
Research on the Effects of Medications in Pregnancy and in Children 459
Research on the Effects Hernandez-Diaz, S., Werler, M.M., Walker, A.M.,

of Medications in Pregnancy and in and Mitchell, A.A. (2000). Folic acid antagonists
during pregnancy and the risk of birth defects.
Children
N. Engl. J. Med. 343 (22): 1608–1614.
Hernández-Díaz, S., Bateman, B.T., Palmsten, K.
de Bie, S., Coloma, P.M., Ferrajolo, C. et al.
et al. (2019 Dec 20). Using nationally
(2015). The role of electronic healthcare
representative survey data for external
record databases in paediatric drug safety
adjustment of unmeasured confounders: an
surveillance: a retrospective cohort study. Br. J.
example using the NHANES data.
Clin. Pharmacol. 80 (2): 304–314.
Pharmacoepidemiol. Drug Saf. https://doi.
Contopoulos-Ioannidis, D.G., Seto, I., Hamm,
org/10.1002/pds.4946. [Epub ahead of print].
M.P. et al. (2012). Empirical evaluation of age
Horton, D.B., Haynes, K., Denburg, M.R. et al.
groups and age-subgroup analyses in pediatric
(2017). Oral glucocorticoid use and
randomized trials and pediatric meta-
osteonecrosis in children and adults with
analyses. Pediatrics 129 (Suppl 3): S161–S184.
chronic inflammatory diseases: a population-
Cooper, W., Hernandez-Diaz, S., Arbogast, P.
based cohort study. BMJ Open 7 (7): e016788.
et al. (2006). Major congenital malformations
Huybrechts, K.F., Palmsten, K., Avorn, J. et al.
after first-trimester exposure to ACE
(2014). Antidepressant use in pregnancy and
inhibitors. N. Engl. J. Med. 354: 2443–2451.
the risk of cardiac defects. N. Engl. J. Med. 370
D’Onofrio, B.M., Class, Q.A., Rickert, M.E. et al.
(25): 2397–2407.
(2016). Translational epidemiologic
Huybrechts, K.F., Bröms, G., Christensen, L.B.
approaches to understanding the
et al. (2018 Feb 1). Association between
consequences of early-life exposures. Behav.
methylphenidate and amphetamine use in
Genet. 46 (3): 315–328.
pregnancy and risk of congenital
Expert Panel on Therapeutic Products for Infants
malformations: a cohort study from the
Children and Youth (2015). Chapter 3:
international pregnancy safety study
children are not Small adults: considering
consortium. JAMA Psychiatry. 75 (2): 167–175.
variation in drug response. In: Improving
Huybrechts, K.F., Hernandez-Diaz, S., Straub, L.
Medicines for Children in Canada, 49–88.
et al. (2018). Association of maternal first-
Fiks, A.G., Grundmeier, R.W., Steffes, J. et al.
trimester ondansetron use with cardiac
(2015). Comparative effectiveness research
malformations and oral clefts in offspring.
through a collaborative electronic reporting
JAMA 320 (23): 2429–2437.
consortium. Pediatrics 136 (1): e215–e224.
Huybrechts, K.F., Bateman, B.T., and Hernandez-
Forrest, C.B., Margolis, P.A., Bailey, L.C. et al.
Diaz, S. (2019). Use of real-world evidence from
(2014). PEDSnet: a national pediatric learning
healthcare utilization data to evaluate drug
health system. J. Am. Med. Inform. Assoc. 21
safety during pregnancy. Pharmacoepidemiol.
(4): 602–606.
Drug Saf. 28: 906–922. https://doi.org/10.1002/
Hales, C.M., Kit, B.K., Gu, Q., and Ogden, C.L.
pds.4789pmid: 31074570.
(2018). Trends in prescription medication use
Huybrechts, K.F., Hernandez-Diaz, S., and
among children and adolescents-United
Bateman, B. (2020 Jun 1). Contextualizing
States, 1999-2014. JAMA 319 (19): 2009–2020.
potential risks of medications in pregnancy for
Hernan, M.A., Hernandez-Diaz, S., Werler,
the newborn: the case of ondansetron. JAMA
M.M., and Mitchell, A.A. (2002). Causal
Pediatr. https://doi.org/10.1001/jamapediatrics.
knowledge as a prerequisite for confounding
2020.1325. Online ahead of print.
evaluation: an application to birth defects
Khoury, M.J., Flanders, W.D., James, L.M., and
epidemiology. Am. J. Epidemiol. 155 (2): 176.
Erickson, J.D. (1989). Human teratogens,
Hernan, M.A., Hernandez-Diaz, S., and Robins,
prenatal mortality, and selection bias. Am. J.
J.M. (2004). A structural approach to selection
Epidemiol. 130: 361–370.
bias. Epidemiology 15 (5): 615–625.
Kinlaw, A.C., Sturmer, T., Lund, J.L. et al. (2017). Slaughter, J.L., Reagan, P.B., Newman, T.B., and
Trends in antibiotic use by birth season and Klebanoff, M.A. (2017). Comparative
birth year. Pediatrics 140 (3): e20170441. effectiveness of nonsteroidal anti-
https://doi.org/10.1542/peds.2017-0441. inflammatory drug treatment vs no treatment
Lash, T.L., Schmidt, M., Jensen, A.O., and for patent ductus arteriosus in preterm
Engebjerg, M.C. (2010). Methods to apply infants. JAMA Pediatr. 171 (3): e164354.
probabilistic bias analysis to summary Sturkenboom, M.C., Verhamme, K.M., Nicolosi,
estimates of association. Pharmacoepidemiol. A. et al. (2008). Drug use in children: cohort
Drug Saf. 19 (6): 638–644. study in three European countries. BMJ 337:
Lash, T.L., Fox, M.P., MacLehose, R.F. et al. a2245.
(2014). Good practices for quantitative bias US Food and Drug Administration, US
analysis. Int. J. Epidemiol. 43 (6): 1969–1985. Department of Health and Human Services
Margulis, A.V., Mittleman, M.A., Glynn, R.J. (2014). Content and format of labeling for
et al. (2015). Effects of gestational age at human prescription drug and biological
enrollment in pregnancy exposure registries. products; requirements for pregnancy and
Pharmacoepidemiol. Drug Saf. 24 (4): 343–352. lactation labeling. Final rule. Fed. Regist. 79
Mitchell, A.A. (2003). Systematic identification (233): 72063–72103.
of drugs that cause birth defects – a new Werler, M.M., Louik, C., and Mitchell, A.A.
opportunity. N. Engl. J. Med. 349 (26): (2011). Case-control studies for identifying
2556–2559. novel teratogens. Am. J. Med. Genet. C: Semin.
Mitchell, A.A., Gilboa, S.M., Werler, M.M. et al. Med. Genet. 157C (3): 201–208.
(2011). Medication use during pregnancy, Woolfenden, S., Eapen, V., Axelsson, E. et al.
with particular focus on prescription drugs: (2016). Who is our cohort: recruitment,
1976–2008. Am. J. Obstet. Gynecol. 205: 51. representativeness, baseline risk and retention
e1–51.e8. in the “watch me grow” study? BMC Pediatr.
Poulton, A.S., Bui, Q., Melzer, E., and Evans, R. 16: 46.
(2016). Stimulant medication effects on
growth and bone age in children with
attention-deficit/hyperactivity disorder: a Risk Management
prospective cohort study. Int. Clin.
Psychopharmacol. 31 (2): 93–99. Directive (2001) 2001/83 EC, Article 1 (28b).
Rothman, K.J. (1990). No adjustments are European Medicines Agency and Heads of
needed for multiple comparisons. Medicines Agencies (2012) Guideline on Good
Epidemiology 1 (1): 43–46. Pharmacovigilance Practices, Module V – Risk
Rothman, K.J. (2014). Six persistent research Management Systems (published July 2012.
misconceptions. J. Gen. Intern. Med. 29 (7): http://www.ema.europa.eu/docsen_GB/
1060–1064. document_library/Scientific_
Schatz, M., Chambers, C.D., Jones, K.L. et al. guideline/2012/06/WC500129134.pdf
(2011). The safety of influenza immunizations Institute of Medicine (2000). To Err Is Human:
and treatment during pregnancy: the vaccines Building a Safer Health System (eds. L.T.
and medications in pregnancy surveillance Kohn, J.M. Corrigan and M.S. Donaldson).
system. Am. J. Obstet. Gynecol. 204: S64–S68. Institute of Medicine, Washington DC:
Scott, F.I., Horton, D.B., Mamtani, R. et al. National Academies Press.
(2016). Administration of antibiotics to National Coordinating Council for Medication
children before age 2 years increases risk for Error Reporting and Prevention website,
childhood obesity. Gastroenterology 151 (1): http://www. http://nccmerp.org/
120–129. e125. aboutMedErrors.html.
The Pharmacoepidemiology of Medication Errors 461
Ovation Pharmaceuticals, Inc. Advisory Chertow, G.M., Lee, J., Kuperman, G.J. et al.
Committee Briefing Document. Sabril® (2001). Guided medication dosing for
(vigabatrin) Tablet and Powder for Oral inpatients with renal insufficiency. JAMA 286:
Solution For Adjunctive Treatment of 2839–2844.
Refractory Complex Partial Seizures in Adults Cypress BW (1982) Drug Utilization in Office
(NDA 20-427), For Monotherapy Treatment of Visits to Primary Care Physicians: National
Infantile Spasms (NDA 22-006). Peripheral and Ambulatory Medical Care Survey, 1980.
Central Nervous System Advisory Committee Department of Health and Human Services
January 7–8, 2009. Available at http://www. publication (PHS) 82–1250. Public Health
http://fda.gov/downloads/Advisory Service.
Committees/Committees MeetingMaterials/ Falconnier, A.D., Haefeli, W.E., Schoenenberger,
Drugs/PeripheralandCentralNervous System R.A. et al. (2001). Drug dosage in patients with
DrugsAdvisoryCommittee/UCM153780.pdf renal failure optimized by immediate concurrent
US Department of Health and Human Services. feedback. J. Gen. Intern. Med. 16: 369–375.
Food and Drug Administration. Guidance. Forster, A.J., Murff, H.J., Peterson, J.F. et al.
Drug Safety Information–FDA’s (2003). The incidence and severity of
Communication to the Public. March 2007. adverse events affecting patients after
Available at http://www.fda.gov/downloads/ discharge from the hospital. Ann. Intern. Med.
Drugs GuidanceComplianceRegulatory 138: 161–167.
Information/Guidances/ UCM072281.pdf Gandhi, T.K., Weingart, S.N., Borus, J. et al.
US Food and Drug Administration. Guidance for (2003). Adverse drug events in ambulatory
Industry (2005) Development and Use of Risk care. N. Engl. J. Med. 348: 1556–1564.
Minimization Action Plans. Hazlet, T.K., Lee, T.A., Hansten, P.D., and Horn,
J.R. (2001). Performance of community
pharmacy drug interaction software. J. Am.
The Pharmacoepidemiology Pharm. Assoc. 41: 200–204.
of Medication Errors Hennessy, S., Bilker, W.B., Zhou, L. et al. (2003).
Retrospective drug utilization review,
Aronson, J.K. (2009). Medication errors: prescribing errors, and clinical outcomes.
definitions and classification. Br. J. Clin. JAMA 290: 1494–1499.
Pharmacol. 67: 599–604. Institute of Medicine (1999). To Err Is Human:
Bates, D.W., Boyle, D.L., Vander Vliet, M.B. et al. Building a Safer Health System (eds. L.T.
(1995). Relationship between medication Kohn, J.M. Corrigan and M.S. Donaldson).
errors and adverse drug events. J. Gen. Intern. Washington, DC: National Academy Press.
Med. 10: 199–205. Kaushal, R., Bates, D.W., Landrigan, C. et al.
Bates, D.W., Cullen, D., Laird, N. et al. (1995). (2001). Medication errors and adverse drug
Incidence of adverse drug events and events in pediatric inpatients. JAMA 285:
potential adverse drug events: implications for 2114–2120.
prevention. JAMA 274: 29–34. Kopp, B.J., Erstad, B.L., Allen, A.M. et al. (2006).
Bates, D.W., Leape, L.L., Cullen, D.J. et al. Medication errors and adverse drug events in
(1998). Effect of computerized physician order an intensive care unit: direct observation
entry and a team intervention on prevention approach for detection. Crit. Care Med. 34:
of serious medication errors. JAMA 280: 415–425.
1311–1316. Koppel, R., Metlay, J.P., Cohen, A. et al. (2005).
Berwick, D.M. (1989). Continuous improvement as Role of computerized physician order entry
an ideal in health care. N. Engl. J. Med. 320: systems in facilitating medication errors.
53–56. JAMA 293: 1197–1203.
Krähenbühl-Melcher, A., Schlienger, R., Evans, S.R. and Follmann, D. (2016). Using
Lampert, M. et al. (2007). Drug-related outcomes to analyze patients rather than
problems in hospitals: a review of the recent patients to analyze outcomes: a step toward
literature. Drug Saf. 30: 379–407. pragmatism in benefit: risk evaluation. Stat.
Kuperman, G.J., Gandhi, T.K., and Bates, D.W. Biopharm. Res. 8 (4): 386–393.
(2003). Effective drug-allergy checking: Evans, S.R. et al. (2015). Desirability of outcome
methodological and operational issues. J. ranking (DOOR) and response adjusted for
Biomed. Inform. 36: 70–79. duration of antibiotic risk (RADAR). Clin.
Montesi, G. and Lechi, A. (2009). Prevention of Infect. Dis. 61 (5): 800–806.
medication errors: detection and audit. Br. J. Evans, S.R., Knutsson, M., Amarenco, P. et al.
Clin. Pharmacol. 67: 651–655. (2020). Methodologies for pragmatic and
Peterson, J.F. and Bates, D.W. (2001). Preventable efficient assessment of benefits and harms:
medication errors: identifying and eliminating application to the SOCRATES trial. Clinical
serious drug interactions. J. Am. Pharm. Assoc. Trials 17 (6): 617–626.
41: 159–160. Evans, S.R. et al. (2020). Presenting risks and
Potylycki, M.J., Kimmel, S.R., Ritter, M. et al. benefits: helping the data monitoring
(2006). Nonputitive medication error committee do its job. Ann. Intern. Med. 172
reporting: 3-year findings from one hospital’s (2): 119–125.
primum non nocere initiative. J. Nurs. Adm. Guidance for Industry and Food and Drug
36: 370–376. Administration Staff: Factors to Consider When
Schiff, G.D., Klass, D., Peterson, J. et al. (2003). Making Benefit-Risk Determinations in Medical
Linking laboratory and pharmacy: Device Premarket Approval and De Novo
opportunities for reducing errors and Classifications. 2016, FDA CDRH.
improving care. Arch. Intern. Med. 163: Food and Drug Administration, Guidance for
893–900. Industry and Food and Drug Administration
Seidling, H.M., Storch, C.H., Bertsche, T. et al. Staff: Patient Preference Information –
(2009). Successful strategy to improve the Voluntary Submission, Review in Premarket
specificity of electronic statin-drug interaction Approval Applications, Humanitarian Device
alerts. Eur. J. Clin. Pharmacol. 65: 1149–1157. Exemption Applications, and De Novo Requests,
and Inclusion in Decision Summaries and
Device Labeling 2016, FDA CDRH.
Benefit-Risk Assessments Food and Drug Administration Benefit-Risk
of Medical Treatments Assessment in Drug Regulatory Decision-
Making. Draft PDUFA VI Implementation
Coplan, P.M. et al. (2011). Development of a Plan 2018 March 30, 2018; Available from:
framework for enhancing the transparency, https://www.fda.gov/downloads/ForIndustry/
reproducibility and communication of the UserFees/PrescriptionDrugUserFee/
benefit-risk balance of medicines. Clin. UCM602885.pdf.
Pharmacol. Ther. 89 (2): 312–315. Hallgreen, C.E. et al. (2014). Benefit-risk
European Medicines Agency, Guidance assessment in a post-market setting: a case
document on the content of the <Co-> study integrating real-life experience into
Rapporteur day <60><80> critical benefit-risk methodology. Pharmacoepidemiol.
assessment report. 2011 Rev. 10.16]; Available Drug Saf. 23 (9): 974–983.
from: http://www.ema.europa.eu/docs/en_GB/ Ho, M.P. et al. (2015). Incorporating patient-
document_library/Regulatory_and_procedural_ preference evidence into regulatory decision
guideline/2016/05/WC500206989.pdf. making. Surg. Endosc.
Benefit Risk Assessments of Medical Treatment 463
Lackey. L., Thompson, G., Eggers, S. (2021). FDA’s org/wp-content/uploads/2015/05/MDIC_

Benefit-Risk Framework for Human Drugs and PCBR_Framework_Web.pdf.
Biologics: Role in Benefit-Risk Assessment and Muhlbacher, A. and Johnson, F.R. (2016). Choice
Analysis of Use for Drug Approvals. Ther Innov experiments to quantify preferences for health
Regul Sci. 55 (1): 170–179. doi: 10.1007/ and healthcare: state of the practice. Appl.
s43441-020-00203-6. Epub 2020 Aug 10. Health Econ. Health Policy 14 (3):
Levitan, B.S. et al. (2011). Application of the BRAT 253–266.
framework to case studies: observations and Pignatti, F. et al. (2015). Structured frameworks to
insights. Clin. Pharmacol. Ther. 89 (2): 217–224. increase the transparency of the assessment of
MDUFA Performance Goals And Procedures, benefits and risks of medicines: current status
Fiscal Years 2018 Through 2022. and possible future directions. Clin. Pharmacol.
Medical Device Innovation Consortium, A Ther. 98 (5): 522–533.
framework for incorporating information on Soekhai, V., et al. Compendium of Methods For
patient preferences regarding benefit and risk Measuring Patient Preferences in Medical
intro regulatory assessments of new medical Treatment. in ISPOR 20th Annual European
technology. 2015; Available from: http://mdic. Congress. 2017. Glasgow, Scotland.
464
24
The Future of Pharmacoepidemiology

Brian L. Strom1, Stephen E. Kimmel2, and Sean Hennessy3
1
2
University of Florida College of Public Health and Health Professions & College of Medicine, Gainesville, FL, USA
3
University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA
We should all be concerned about the International Society for Pharmacoepidemiology

future because we will have to spend the (ISPE), established in 1991, has grown to approx
rest of our lives there. imately 1500 members from over 60 countries. It
Charles Franklin Kettering 1946 has developed a set of guidelines for Good
Epidemiologic Practices for Drug, Device, and
Vaccine Research in the United States in 1996,
Introduction and updated these guidelines most recently in
2016 as the ISPE Guidelines for Good
Speculating about the future is at least risky Pharmacoepidemiology Practices. Many
and possibly foolish. Nevertheless, the future national pharmacoepidemiologic societies have
of pharmacoepidemiology seems apparent in been formed as well. The journal, Clinical
many ways, judging from past trends and Pharmacology and Therapeutics, the major US
recent events. Interest in the field by the phar academic clinical pharmacology journal,
maceutical industry, government agencies, actively solicits pharmacoepidemiologic manu
new trainees, and the public continues to grow, scripts, as did the Journal of Clinical
as is realization of what pharmacoepidemiol Epidemiology. The major journal devoted to the
ogy can contribute. Indeed, international field, Pharmacoepidemiology and Drug Safety,
attention on drug safety remains high, impor ISPE’s official journal, is indexed on Medline
tant safety questions involving widely used and achieved an impact factor of 2.314 in 2018,
drugs continue to emerge, and questions con remarkably high for a niche field. Other journals
cerning the effectiveness of systems of drug have been formed to publish pharmacoepidemi
approval and drug safety monitoring remain. ology research. The number of individuals seek
As the functions of academia, industry, and ing to enter the field continues to increase, as is
government have become increasingly global, so their level of training. The number of programs
has the field of pharmacoepidemiology. The of study in pharmacoepidemiology is increasing
number of individuals attending the annual in schools of medicine, public health, and phar
International Conference on Pharmacoepide macy. While in the 1980s the single summer
miology has increased from approximately 50 in short course in pharmacoepidemiology at the
the early 1980s to over 1400 in 2017. The University of Minnesota was sometimes
canceled because of insufficient interest, later Thus, from the perspective of those in the
the University of Michigan School of Public field, the trends in pharmacoepidemiology are
Health summer course in pharmacoepidemiol remarkably positive, although many important
ogy attracted 10% of all students in the entire challenges remain. In this chapter, we will
summer program, and thereafter McGill briefly give our own view on the future of
University, Erasmus University Rotterdam, pharmacoepidemiology. Following the format
Utrecht University, and the Johns Hopkins of Part I of the book, we explore this future
Bloomberg School of Public Health all conduct from the perspectives of academia, the phar
summer short courses in pharmacoepidemiol maceutical industry, regulatory agencies, and
ogy. Several other short courses are given as well, then the law.
including by ISPE itself which has seen a mas
sive increase in pre-conference courses offered
over the years. Regulatory bodies around the
The View from Academia
world have expanded their internal pharma
coepidemiologic programs. The number of
Scientific Developments
pharmaceutical companies with their own phar
macoepidemiologic units has also increased, Methodologic Advances
along with their support for academic units and Methodologically, the array of approaches
their funding of external pharmacoepidemio available for performing pharmacoepidemio
logic studies. Requirements that a drug be shown logic studies will continue to grow. Each of
to be cost-effective (see Chapter 18) have been the methodologic issues discussed in Part III
added to many national health care systems, can be expected to be the subject of further
provincial health care systems, and managed research and development. The future is
care organizations, either to justify reimburse likely to see ever more advanced ways of per
ment or even to justify drug availability. Drug forming and analyzing epidemiologic studies
utilization review is being widely applied (see across all content areas, as the field of epide
“Evaluating and Improving Prescribing” in miology continues to expand and develop.
Chapter 23), and many hospitals are becoming Some of these new techniques will, of course,
mini-pharmacoepidemiologic practice and be particularly useful to investigators in
research laboratories. The US Congress has rec pharmacoepidemiology (see for example
ognized the importance of pharmacoepidemiol Chapter 22). The next few years will likely see
ogy, requiring FDA to build a new data resource, continued expanded use of propensity scores,
containing at least 100 million lives, for evaluat instrumental variables, the trend-in-trend
ing potential adverse effects of medical products, design, sensitivity analysis, and novel meth
and most recently passing the twenty-first ods to analyze time-varying exposures and
Century Cures Act, encouraging the wide use of confounders. In addition, we believe that we
“real-world evidence.” The latter has been will see increasing application of pharma
deemed to range from traditional pharmacoepi coepidemiologic insight in the conduct of
demiology data sources like claims and medical clinical trials, as well as increased use of the
record databases, and even ad hoc pharmacoepi randomized trial design to examine questions
demiology studies (see Part II), to novel data traditionally addressed by observational phar
sources like e-health tools, m-health tools, and macoepidemiology (see Chapter 17), espe
other wearable devices, as well as pragmatic tri cially given the controversies resulting from
als using traditional pharmacoepidemiology inconsistencies between nonrandomized
databases to collect outcomes (see Chapter 17). studies vs randomized trials, and given the
The future is likely to see a marked expansion of emerging field of comparative effectiveness
these novel, technology-driven approaches. research.
466 24 The Future of Pharmacoepidemiology
Drug regulators have enthusiastically the study of pharmacogenetics, exciting devel

embraced therapeutic risk management (see opments have occurred in the ability of research
“Risk Management” in chapter 23). Yet, this ers to identify biologic factors that predispose
field is still very much in its infancy, with an patients to adverse drug reactions (see
enormous amount of work needed to develop Chapter 15). However, relatively few of these
new methods to measure, communicate, and discoveries have yet been shown useful in
manage the risks and benefits associated with improving patient care, and new studies and
medication use. Rigorous studies (i.e. program methods must be pursued to determine the clini
evaluations) of the effectiveness of risk man cal utility of genetic testing. Pharmacogenetics
agement programs remain the exception rather has evolved from studies of measures of slow
than the rule. Development of this area will drug metabolism as a contributor to adverse
require considerable effort from pharmacoepi reactions to the study of molecular genetic
demiologists as well as those from other fields. markers. This has been aided by the develop
We may see developments in the processes ment of new, noninvasive methods to collect and
used to assess causality from individual case analyze biosamples, making population-based
reports (see Chapters 7 and 14). Data mining genetic studies feasible. We believe that clinical
approaches will be used increasingly in sponta measurement of biologic factors will ultimately
neous reporting databases to search for early complement existing approaches to tailoring
signals of adverse reactions. Hopefully, we will therapeutic approaches for individual patients.
see studies evaluating the utility of such However, it is unlikely that genotype will be the
approaches. The need for newer methods to only, or even the major, factor that determines
screen for potential adverse drug effects, such as the optimal drug or dose for a given patient.
those using health care claims or medical record Future years are likely to see much more of this
data and data from social media, is also clear. cross-fertilization between pharmacoepidemiol
We are likely to see increasing input from ogy and molecular biology, and newer forms of
pharmacoepidemiologists into policy ques “-omics” such as the microbiome. From a
tions about drug approval (see “The View from research perspective, we can easily envision
Regulatory Agencies” in Chapter 6), with new pharmacogenetic studies added to the process of
attention to applying pharmacoepidemiology evaluating potential adverse reactions. We also
in the study of the growing opiate epidemic. anticipate the availability of genotypic informa
We anticipate that emphasis will shift from tion for members of large patient cohorts for
studies evaluating whether a given drug is whom drug exposures and clinical outcomes are
associated with an increased risk of a given recorded electronically, and even for selected
event to those that examine as well patient-and patients from electronic data systems, such as
regimen-specific factors that affect risk (see those described in Part II of this book.
also Chapter 19 and “Risk Management” in
Chapter 23). Such studies are crucial because, New Content Areas of Interest
if risk factors for adverse reactions can be bet In addition, there are a number of new content
ter understood before a safety crisis occurs, or areas that are likely to be explored and devel
early in the course of a crisis, then the clinical oped more. Studies of drug utilization will con
use of the drug may be able to be repositioned, tinue and will continue to become more
avoiding the loss of useful drugs (see innovative (see “Studies of Drug Utilization” in
Chapter 15 as well as “Risk Management” Chapter 23). Particularly as the health care
and “Benefit-Risk Assessments of Medical industry becomes more sensitive to the possi
Treatments in Chapter 23.). bility of overutilization, underutilization, and
With recent developments in molecular biol inappropriate utilization of drugs, and the risks
ogy and bioinformatics, and their application to associated with each, one would expect to see
an increased frequency of and sophistication in by the development of comparative effective

drug utilization review programs, which seek ness research. Other approaches to controlling
to improve care (see “Evaluating and Improving for confounding are similarly likely to become
Prescribing” in Chapter 23), potentially incor more common as they are further developed
porating techniques from molecular pharma (see Chapter 22). New analytic approaches, like
coepidemiology (see Chapter 15). machine learning, artificial intelligence, and
The US Joint Commission on Accreditation of cognitive computing, are also likely to make
Healthcare Organizations revolutionized US hos their way into pharmacoepidemiology studies.
pital pharmacoepidemiology through its stand We will also see more use of pharmacoepide
ards requiring adverse drug reaction surveillance miologic approaches prior to drug approval,
and drug use evaluation program in every hospi e.g. to understand the baseline rate of adverse
tal. Hospitals are also now experimenting with events that one can expect to see in patients
different methods of organizing their drug deliv who will eventually be treated with a new drug
ery systems to improve their use of drugs, e.g. use (see “The View from Industry” in Chapter 6).
of computerized clinical decision support and the Recent years have seen an explosion in the
addition of pharmacists to patient care teams (see worldwide use of herbal and other comple
“The Pharmacoepidemiology of Medication mentary and alternative medications, includ
Errors’ in Chapter 23). ing cannabis-based products. These are
Interest in the field of pharmacoeconomics, essentially pharmaceuticals sold without con
i.e. the application of the principles of health ventional standardization, and with no
economics to the study of drug effects, is con required premarketing testing of safety or effi
tinuing (see Chapter 18). Society is realizing cacy. In a sense, for these products, this is a
that the acquisition cost of drugs is often a very return to a pre-regulatory era. Therefore, it is
minor part of their economic impact, and that quite likely that the next few years will see an
their beneficial and harmful effects can be analogous set of safety concerns associated
vastly more important. Further, more govern with their use, and society will turn to pharma
ments and insurance programs are increasingly coepidemiologists to help evaluate the use and
requiring economic justification before permit effects of these products. Of course, if regula
ting reimbursement for a drug. As a result, the tory oversight is decreased in some countries,
number of studies exploring this is increasing. as has been suggested in the US, the same
As the methods of pharmacoeconomics could occur with traditional medications.
become increasingly sophisticated, and its Research interest in the entire topic of
applications clear, this could be expected to patient nonadherence with prescribed drug
continue to be a popular field of inquiry. regimens goes back to about 1960, but little
More non-experimental studies of beneficial fruitful research could be done because meth
drug effects, particularly of drug effectiveness, ods for ascertaining drug exposure in individ
can be expected, as the field becomes more ual ambulatory patients were grossly
aware that such studies are possible. This is unsatisfactory. This problem has been miti
being encouraged by the rapid increase in the gated greatly by advances in incorporating
use of propensity scores to adjust for measured time-stamping microcircuitry into pharmaceu
covariates, although investigators using this tical containers, which records the date and
method often place more confidence in that time each time that the container is opened.
technique than is warranted, some not recog Perhaps as a consequence of its inherent sim
nizing that its ability to control for confounding plicity and economy, electronic monitoring is
by indication remains dependent on one’s abil increasingly emerging as the de facto gold
ity to measure the true determinants of expo standard for compiling dosing histories of
sure (see Chapter 22). It is also being encouraged ambulatory patients, from which one can
e valuate the extent of adherence to the pre see the increased use of both very large obser
scribed drug regimen. Future years are likely vational studies and large simple trials after
to see a continuing increase in the use of this marketing, to study important clinical out
technique (see Chapter 21) in research, and comes (see Chapters 17 and 20).
perhaps in clinical practice. Perhaps equally In addition, with the growth of concerns
importantly, new methods of measuring about patient safety (see “The Pharmaco
adherence that do not rely on purchasing and epidemiology of Medication Errors” in Chapter
using alternative sources of drug dispensing, 23), there has been increasing attention to
such as smartphone-based measures of pill simultaneous use of pairs of drugs that have
taking, may expand our ability to measure been shown in pharmacokinetic studies (see
adherence in real-world epidemiology studies. Chapter 4) to cause increased or decreased drug
The next few years are also likely to see the levels. Yet, population studies informing the
increasing ability to target drug therapy to the clinical importance and pharmacologic aspects
proper patients. This will involve both increas of drug–drug interactions have only been per
ing use of statistical methods, and increasing formed in the past few years. The next few years
use of laboratory techniques from other bio are likely to see the emergence of more studies
logical sciences, as described above. Statistical to address such questions.
approaches will allow us to use predictive mod Finally, in the last few years, society has
eling to study, from a population perspective, increasingly turned to pharmacoepidemiology
who is most likely to derive benefit from a drug, for input into major policy decisions. For
and who is at greatest risk of an adverse out example, pharmacoepidemiology played a
come. Laboratory science will enable us to major role in the evaluations by the Institute of
measure individuals’ biomarkers to predict Medicine of the US National Academy of
responses to drug therapy (i.e. molecular sus Sciences of the Anthrax Vaccine (deciding
ceptibility). From the perspective of pre- whether the existing vaccine was safe to use
approval testing, these developments may and, thereby, whether the military vaccine pro
allow researchers to target specific patient- gram should be restarted) and the Smallpox
types for enrollment into their studies, those Vaccine program (deciding the shape of the
subjects most likely to succeed with a drug. program intended initially to vaccinate the
From a clinical perspective, it will enable health entire US population). Pharmacoepidemiology
care providers to incorporate biological factors is likely to contribute to future assessment of
in the individualization of choice of regimens. vaccine effectiveness and safety in the popula
The past few years have seen the increased tion (e.g. for the SARS-CoV-2 vaccine).
use of surrogate markers, presumed to repre
sent increased risk of rarer serious adverse Logistical Advances
effects when drugs are used in broader num Logistically, with the increased computeriza
bers of patients. These range from mild liver tion of data in society in general and within
function test abnormalities, used as predictors health care in particular, and the increased
of serious liver toxicity, to electrocardiographic emphasis on using electronic databases for
QTc prolongation as a marker of risk of suffer pharmacoepidemiology (see Part II), some
ing the arrhythmia torsades des pointes, which data resources will disappear (e.g. The Rhode
can lead to death. Indeed, some drugs have Island Drug Use Reporting System and the
been removed from the market, or from devel inpatient databases discussed in prior editions
opment, because of the presence of these sur of this book have disappeared, with new ones
rogate markers. Yet, the utility of these markers added) and a number of new computerized
as predictors of serious clinical outcomes is databases have emerged as major resources
poorly studied. The next few years are likely to for pharmacoepidemiologic research (e.g.
c ommercial insurance databases [Chapter 9], would be expected to be a useful source of

inpatient databases, the databases from information about dose-dependent adverse
Ontario and Denmark [Chapter 9]). The drug effects. Others will undoubtedly be devel
importance of these databases to pharmacoep oped as well.
idemiology is now clear: they enable research Of critical importance, there is increasing
ers to address, quickly and relatively concern about patient privacy in many coun
inexpensively, questions about drug effects in tries. The regulatory framework for human
different settings that require large sample research is actively changing, in the process,
sizes, with excellent quality data on drug expo e.g. Europe’s new data protection law. As dis
sures. Registries (Chapter 11) will also become cussed in Chapter 16, this is already beginning
increasingly important for pharmacoepidemi to make pharmacoepidemiologic research
ologic research. With the initiation of US more difficult, whether it is access to medical
Medicare Part D in 2006, which provides pre records in database studies, or access to a list of
scription drug coverage to US Medicare recipi possible cases with a disease to enroll in ad hoc
ents, the availability of this data resource is case–control studies. This will be an area of
advancing our ability to perform hypothesis great interest and rapid activity over the next
testing studies, as it is so large relative to other few years as electronic health records become
resources; nearly 27 million Medicare benefi much more commonplace, and one in which
ciaries were already subscribed to Part D cover the field of pharmacoepidemiology will need
age in 2009 (see Chapter 9). It has created an to remain very active, or risk considerable
enormous new data resource for pharmacoepi interference with its activities.
demiology, as well as increased interest from It is likely that new types of research oppor
the US government in what pharmacoepide tunities will emerge. For example, as the US
miology can do. The development of FDA’s finally implemented a drug benefit as part of
Sentinel Initiative has, similarly, provided a Medicare, its health program for the elderly,
vast new data resource, initially intended for US government drug expenditures suddenly
hypothesis generating, and more recently used increased by $49.5 billion in 2007. Outside the
for hypothesis strengthening and testing. US, as well, many different opportunities to
Nevertheless, even as the use of databases form databases are being developed. There is
increases, it is important to keep in mind the also an increased interest in the importance of
importance of studies that collect data de novo pharmacoepidemiology in the developing
(see Chapter 11). Each approach to pharma world. Many developing world countries spend
coepidemiology has its advantages and its dis a disproportionate amount of their health care
advantages, as described in Part II. No resources on drugs, yet these drugs are often
approach is ideal in all circumstances, and used inappropriately. There have been a num
often a number of complementary approaches ber of initiatives in response to this, including
are needed to answer any given research ques the World Health Organization’s development
tion (see Chapter 12). To address some of the of its list of “Essential Drugs.”
problems inherent in any database, we must
maintain the ability to perform ad hoc studies,
Funding
as well (see Chapter 11). Perhaps better, less
expensive, and complementary approaches to For a number of years, academic pharmacoep
ad hoc data collection in pharmacoepidemiol idemiology suffered from limited research
ogy will be developed. For example, a potential funding opportunities. In the early 1980s, the
approach that has not been widely used is the only available US funding for the field was an
network of regional and national poison control extramural funding program from FDA with a
centers. In particular, poison control centers total of $1 million/year. Industry interest and
support were similarly limited. With the are organized by organ system. Earlier in the
increasing interest in the field, this situation development of pharmacoepidemiology, the
appears to be changing rapidly. FDA has National Institute of General Medical Sciences
expanded its internal pharmacoepidemiology (NIGMS) provided most of the US government
program, and US National Institutes of Health support for our field. It remains, conceptually,
(NIH) is funding pharmacoepidemiologic perhaps the most appropriate source of such
studies as well. In the US, other funding now support, since it is the institute that is intended
comes from the Agency for Health Care to fund projects that are not specific to an
Research and Quality (AHRQ), and from the organ system, and it is the institute that funds
Patient-Centered Outcomes Research Institute clinical pharmacologic research. However,
(PCORI), created as part of the Affordable over the past few years there has been limited
Care Act. Much industry funding is available, funding from NIGMS for epidemiologic
as the perceived need for the field within research. A notable exception was the NIGMS-
industry grows (see below). This is likely to funded Pharmacogenetics Research Network
increase, especially as the FDA more often (PGRN), which has now been disbanded. In
requires industry to perform postmarketing the meantime, NIH funding continues to be
studies, and the legislative mandate for FDA to available if one tailors a project to fit an organ
pay more attention to “real world evidence.” system or in some other way fits the priorities
There is, of course, a risk associated with of one of the individual institutes, e.g. the
academic groups becoming too dependent on National Institute on Aging or the National
industry funding, both in terms of choice of Institute of Child Health and Human
study questions and credibility. Fortunately, in Development.
the US, AHRQ began to fund pharmacoepide
miologic research as well, as part of an initia
Personnel
tive in pharmaceutical outcomes research. In
particular, the AHRQ Centers for Education With the major increase in interest in the field
and Research on Therapeutics (CERTs) pro of pharmacoepidemiology, accompanied by an
gram provided federal support for ongoing increased number of funding opportunities, a
pharmacoepidemiologic activities (see “The major remaining problem, aggravated by the
View from Academia” in Chapter 6). While other trends, is one of inadequate personnel
still small relative to industry expenditures on resources. There is a desperate need for more
research, it was large relative to the US federal well-trained people in the field, with employ
funding previously available for pharmacoepi ment opportunities available in academia,
demiology. Similar programs have now been industry, and government agencies. Some
started in Europe and Canada. Unfortunately, early attempts were made to address this. The
the CERTs program has ended and periodi Burroughs Wellcome Foundation developed
cally, the future of AHRQ itself is in question. the Burroughs Wellcome Scholar Award in
Even the US NIH now funds pharmacoepi Pharmacoepidemiology, a faculty develop
demiologic projects more often. NIH is the ment award designed to bring new people into
logical major US source for such support, as it the field. This program, now discontinued, did
is the major funding source for most basic bio not provide an opportunity for fellowship
medical research in the US. Its funds are also training of entry-level individuals, but was
accessible to investigators outside the US, via designed for more experienced investigators.
the same application procedures. However, Unfortunately, it is no longer an active
NIH’s current organizational structure repre program.
sents an obstacle to pharmacoepidemiologic Outside of government, training opportuni
support. In general, the institutes within NIH ties are limited. In the US, the NIH is the major
source of support for scientific training but as In addition to being useful for exploring the
noted above, NIGMS, which funds training effects of their drugs, manufacturers are begin
programs in clinical pharmacology, now sup ning to realize that the field can contribute not
ports one program in pharmacoepidemiology, only to identifying problems, but also to docu
while the National Heart, Lung and Blood menting drug safety and developing and evalu
Institute supports another. The National ating risk management programs. An
Institute of Child Health and Human increasing number of manufacturers are
Development also has funded limited training mounting pharmacoepidemiologic studies
in pediatric pharmacoepidemiology. However, “prophylactically,” to have safety data available
pharmacoepidemiologic training is still too in advance of when crises may occur. Proper
dependent on non-federal sources of funds, practice would argue for postmarketing studies
especially at a time when such funding is for all newly marketed drugs used for chronic
becoming harder to obtain. There is a growing diseases, and all drugs expected to be either
number of institutions now capable of carry pharmacologically novel or sales blockbusters,
ing out such training, for example universities because of the unique risks that these situa
with faculty members interested in pharma tions present. Pharmacoepidemiology also can
coepidemiology, including those with clinical be used for measuring beneficial drug effects
research training programs supported by, for and even for marketing purposes, in the form
example, an NIH Clinical and Translational of descriptive market research and analyses of
Science Award and organ system-specific the effects of marketing efforts. Perhaps most
training grants. Young scientists interested in importantly for the industry’s financial bottom
undergoing training in pharmacoepidemiol line, pharmacoepidemiologic studies can be
ogy, however, can only do so if they happen to used to protect the major investment made in
qualify for support from such programs. No developing a new drug against false allegations
ongoing support is normally available from of adverse effects and protecting good drugs for
these programs for training in pharmacoepide a public that needs them. Further, even if a
miology per se. This was addressed in the past, drug is found to have a safety problem, the legal
primarily through the leadership and generos liability of the company may be diminished if
ity of some pharmaceutical companies. Much the company has, from the outset, been forth
more is needed, however. Fortunately, with the right in its efforts to learn about that drug’s
rapid rise in interest in comparative effective risks. Finally, as noted in Chapter 1 and above,
ness research, additional training support FDA now has new authority to require post
emerged from both NIH and AHRQ/PCORI, marketing pharmacoepidemiologic studies,
but this too is now in question going forward. and a new charge to focus on “real world evi
dence,” so one can expect to see many more
required of industry by regulators.
The View from Industry Industry is always interested in predictabil
ity. With that, there is increased interest in
It appears that the role of pharmacoepidemiol developing a formulaic approach to benefit–
ogy in industry is and will continue to be risk assessment (see “Benefit-Risk Assessments
expanding rapidly. All that was said above of Medical Treatments” in Chapter 23). The
about the future of pharmacoepidemiology sci next few years are likely to see considerable
entifically, as it relates to academia (see additional work in this area.
Chapter 6), obviously relates to industry, as In light of these advantages, most major
well (see again Chapter 6). The necessity of pharmaceutical firms have formed their own
pharmacoepidemiology for industry has pharmacoepidemiologic units. Of course, this
become apparent to many of those in industry. then means that industry confronts and, in
fact, aggravates the problem of an insufficient bodies. For example, in the US, pharmacoepi
number of well-trained personnel described demiology used to play an important role
above. Many pharmaceutical companies within the AHRQ, the Centers for Disease
increased their investment in external phar Control and Prevention, PCORI, and the NIH,
macoepidemiologic data resources, so that and there has been for over 40 years intermit
they will be available for research when crises tent debate about the wisdom of developing
arise. This has been declining, however. A risk an independent new Center for Drug
of the growth in the number of pharmacoepi Surveillance.
demiologic studies for industry is the genera As noted above, the use of therapeutic risk
tion of an increased number of false signals management approaches (see “Risk
about harmful drug effects. This is best Management” in Chapter 23) has been aggres
addressed by having adequately trained indi sively embraced by regulatory bodies around
viduals in the field, and by having personnel the world, and there has been considerable dis
and data resources available to address these cussion about benefit–risk assessments of
questions quickly, responsibly, and effectively, medical products (see “Benefit-Risk
when they are raised. Assessments of Medical Treatments” in
Chapter 23). This will continue to change regu
lation as more experience with it is gained.
he View from Regulatory
T There is considerable regulatory interest in
Agencies getting important new drugs onto the market
quickly, using mechanisms such as FDA’s ini
It appears that the role of pharmacoepidemiol tiatives on orphan drugs, expanded access pro
ogy in regulatory agencies is also expanding grams, compassionate use programs, fast track
(see Chapter 6). Again, all of what was said regulations, accelerated approval, priority
above about the future of pharmacoepidemiol review, breakthrough drug designation, and
ogy scientifically, as it relates to academia, obvi use of “real-world evidence,” and analogous
ously relates to regulatory agencies, as well. In initiatives elsewhere. On the other hand,
addition, there have been a large number of efforts like the Right to Try Act may compro
major drug crises, many described throughout mise the scientific rigor of the normal regula
this book. Many of these crises resulted in the tory approach. The future is likely to see
removal of the drugs from the market. The continued creative regulatory initiatives
need for and importance of pharmacoepidemi toward maintaining this balance.
ologic studies have become clear. Again, this There is also increased interest in encourag
can be expected to continue in the future. It has ing the use of generic drugs, reducing costs,
even been suggested that postmarketing phar and reducing regulatory obstacles to the avail
macoepidemiologic studies might replace some ability of generically equivalent drugs, once
premarketing Phase III studies in selected situ patents expire.
ations, as was done with zidovudine. As noted, Finally, there is an enormous increase in
regulatory agencies are being given increased attention to drug safety, e.g. driven by drug
authority to require such studies after market safety issues identified with COX-2 inhibitors
ing. Regulatory bodies are also expanding their and even traditional nonsteroidal anti-
pharmacoepidemiologic staffing and seeking inflammatory drugs, and then by the thiazoli
training in pharmacoepidemiology for those dinediones, used for treatment of diabetes. The
already employed by the agencies. net result has been major regulatory change,
We are also seeing increasing governmental and even new legislation. Between 2009 and
activity and interest in pharmacoepidemiol 2012, for example, FDA approved 110 new
ogy, outside the traditional realm of regulatory drugs and biologics for 120 indications, and
Key Point 473
only 13 of them did not have any postmarket against the loss of useful drugs. The past few
ing requirements. decades have demonstrated the utility of this
new field. They also have pointed out some of
its problems. With luck, the next few years will
The View from the Law see the utility accentuated and the problems
ameliorated.
Finally, the importance of pharmacoepidemi
ology to the law has also been increasing. The
potential financial risk to drug manufacturers Key Points
posed by lawsuits related to adverse drug
effects is very large. Some financial payments ●● The discipline of pharmacoepidemiology
have been enormous, and indeed put large has been growing and will likely continue to
multinational companies at risk. It is clear that grow within academia, industry, and
the interest in the field and the need for more government.
true experts in the field will, therefore, increase ●● Methodological advances are expected to con
accordingly. tinue in order to support pharmacoepidemio
logic studies as well as newer approaches
such as risk management programs and
Conclusion molecular pharmacoepidemiology.
●● Content areas such as pharmacoeconomics,
There are no really “safe” biologically active medication adherence, risk management,
drugs. There are only “safe” physicians. and intermediate surrogate markers will
Harold A. Kaminetzsky 1963 grow as interest and need for these foci
increases.
All drugs have adverse effects. ●● Both automated databases and de novo stud
Pharmacoepidemiology will never succeed in ies will continue to be important to the field
preventing them. It can only detect them, hope and will serve as important complements to
fully early, and thereby educate health care pro each other.
viders and the public, which will lead to better ●● Challenges faced by pharmacoepidemiology
medication use. Pharmacoepidemiology can include limited funding opportunities, regu
also lead to safer use of medications through a latory restrictions and privacy concerns sur
better understanding of the factors that alter rounding human research, limited training
the risk: benefit balance of medications. The opportunities, and inadequate personnel
net results of increased activity in pharmacoep resources.
idemiology will be better for industry and aca ●● All sectors responsible for the public health,
demia but, most importantly, for the public’s including academia, industry, and govern
health. The next drug disaster cannot be pre ment, must address the challenges facing
vented by pharmacoepidemiology. However, pharmacoepidemiology, and must support
pharmacoepidemiology can minimize its its continued development in order to maxi
adverse public health impact by detecting it mize the benefit and minimize the risk
early. At the same time, it can improve the use inherent in all medications and medical
of drugs that have a genuine role, protecting devices.
Further Reading
Andrews, E.B., Avorn, J., Bortnichak, E.A. et al. cytochrome P450 antibodies in aromatic
(1996). Guidelines for good epidemiology anticonvulsant-induced hypersensitivity
practices for drug, device, and vaccine reactions. J. Pharmacol. Exp. Ther. 263:
research in the United States. 360–367.
Pharmacoepidemiol. Drug Saf. 5: 333–338. Lunde, P.K.M. (1984). WHOs programme on
Bates, D.W., Leape, L.L., Cullen, D.J. et al. (1998). essential drugs. Background, implementation,
Effect of computerized physician order entry and present state and prospectives. Dan. Med. Bull.
a team intervention on prevention of serious 31 (suppl 1): 23–27.
medication errors. JAMA 280: 1311–1316. Pullar, T. and Feely, M. (1990). Problems of
Classen, D.C., Pestotnik, S.L., Evans, R.S., and compliance with drug treatment: new
Burke, J.P. (1991). Computerized surveillance solutions? Pharm. J. 245: 213–215.
of adverse drug events in hospital patients. Spielberg, S.P. (1992). Idiosyncratic drug
JAMA 266: 2847–2851. reactions: interaction of development and
Committee on Smallpox Vaccination Program genetics. Semin. Perinatol. 16: 58–62.
Implementation, Board on Health Promotion Strom, B.L. and Carson, J.L. (1990). Use of
and Disease Prevention (2005). The Smallpox automated databases for
Vaccination Program: Public Health in an Age pharmacoepidemiology research. Epidemiol.
of Terrorism (eds. A. Baciu, A.P. Anason, K. Rev. 12: 87–107.
Stratton and B. Strom). Washington, DC: The Strom, B.L. and Gibson, G.A. (1993). A
National Academies Press. systematic integrated approach to improving
Food and Drug Administration, eHealth drug prescribing in an acute care hospital: a
Foundation, and the Brookings Institute potential model for applied hospital
(2008) Sentinel Initiative: structure, function pharmacoepidemiology. Clin. Pharmacol.
and scope. Washington, D.C. December 16, Ther. 54: 126–133.
2008. http://www.fda.gov/oc/initiatives/ Strom, B.L., West, S.L., Sim, E., and Carson, J.L.
criticalpath/transcript121608.pdf (1989). The epidemiology of the acute flank
Howard, N.J. and Laing, R.O. (1991). Changes in pain syndrome from suprofen. Clin.
the World Health Organization essential drug Pharmacol. Ther. 46: 693–699.
list. Lancet 338: 743–745. Urquhart, J. (1997). The electronic medication
Institute of Medicine (2005). The Smallpox event monitor–lessons for pharmacotherapy.
Vaccination Program: Public Health in an Age Clin. Pharmacokinet. 32: 345–356.
of Terrorism. Washington, DC: National Woosley, R.L., Drayer, D.E., Reidenberg,
Academies Press. M.M. et al. (1978). Effect of acetylator
ISPE (2008). Guidelines for good phenotype on the rate at which
pharmacoepidemiology practices (GPP). procainamide induces antinuclear
Pharmacoepidemiol. Drug Saf. 17: 200–208. antibodies and the lupus syndrome. N.
Joellenbeck, L.M., Zwanziger, L.L., Durch, J.S., Engl. J. Med. 298: 1157–1159.
and Strom, B.L. (eds.) (2002). The Anthrax Young, F.E. (1988). The role of the FDA in the
Vaccine: Is It Safe? Does It Work? Washington, effort against AIDS. Public Health Rep. 103:
DC: National Academies Press. 242–245.
Kettering, C.F. (January 1946). Picturesque Yudkin, J.S. (1980). The economics of
Speech and Patter, Reader’s Digest, vol. 48, 102. pharmaceutical supply in Tanzania. Int. J.
Pleasantville, New York: The Reader’s Digest Health Serv. 10: 455–477.
Association. Yudkin, J.S. (1984). Use and misuse of drugs in
Leeder, J.S., Riley, R.J., Cook, V.A., and the third world. Dan. Med. Bull. 31 (suppl 1):
Spielberg, S.P. (1992). Human anti- 11–17.
475
Sample Size Tables
Table A1 Sample sizes for cohort studiesa.
Incidence Relative risk to be detected

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 970 717 2 788 497 6 306 290 29 429 320 37 837 603 10 510 431 3 153 120 1 634 946 1 051 034 756 742 583 904 394 133 211 445 142 727 61 134 22 318
0.00005 394 133 557 684 1 261 219 5 885 657 7 567 179 2 101 980 630 585 326 965 210 189 151 334 116 768 78 816 42 280 28 538 12 220 4458
0.0001 197 060 278 832 630 585 2 942 699 3 783 376 1 050 923 315 268 163 467 105 083 75 657 58 376 39 401 21 135 14 264 6106 2225
0.0005 39 401 55 751 126 078 588 332 756 333 210 078 63 015 32 669 20 999 15 117 11 662 7870 4219 2845 1215 439
0.001 19 694 27 865 63 015 294 037 377 953 104 973 31 483 16 320 10 488 7549 5823 3928 2104 1418 603 216
0.005 3928 5557 12 564 58 600 75 249 20 888 6257 3240 2080 1495 1152 775 412 276 114 37
0.01 1957 2769 6257 29 170 37 411 10 378 3104 1605 1028 738 568 381 201 133 53 15
0.05 381 538 1212 5627 7140 1969 582 297 188 133 101 65 32 19 4 –
0.10 184 259 582 2684 3357 918 266 133 82 57 42 26 10 4 – –
0.15 118 166 372 1703 2095 568 161 79 47 32 23 13 – – – –
0.20 85 120 266 1212 1465 393 109 52 30 19 13 6 – – – –
0.25 65 92 203 918 1086 287 77 35 19 12 7 – – – – –
0.30 52 73 161 722 834 217 56 24 12 6 – – – – – –
0.35 43 60 131 582 654 167 41 16 7 – – – – – – –
0.40 36 50 109 477 519 130 30 11 – – – – – – – –
0.45 30 42 91 395 414 101 21 6 – – – – – – – –
0.50 26 36 77 329 329 77 14 – – – – – – – – –
0.55 22 31 66 276 261 58 8 – – – – – – – – –
0.60 19 27 56 231 203 42 2 – – – – – – – – –
0.65 17 23 48 194 155 29 – – – – – – – – – –
0.70 15 20 41 161 113 17 – – – – – – – – – –
0.75 13 17 35 133 77 7 – – – – – – – – – –
0.80 11 15 30 109 46 – – – – – – – – – – –
0.85 10 13 25 87 18 – – – – – – – – – – –
0.90 8 11 21 68 – – – – – – – – – – – –
0.95 7 9 17 51 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: exposed ratio = 1 : 1. The sample size listed is the number of subjects needed in the exposed group. An equivalent number would be
included in the control group.
Table A2 Sample size for cohort studiesa.

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 529 057 2 153 636 4 825 616 22 279 822 28 149 090 7 764 537 2 302 889 1183 56 3 755 529 540 883 415 381 278 329 147 626 99 000 41 938 15 197
0.0001 152 896 215 349 482 527 2 227 804 2 814 625 776 367 230 258 118 337 151 093 108 167 83 068 55 659 29 520 19 795 8384 3036
0.0005 30 570 43 057 96 475 445 402 562 673 155 196 46 024 23 651 75 539 54 077 41 528 27 825 14 756 9895 4189 1516
0.001 15 280 21 521 48 218 222 602 281 179 77 550 22 994 11 815 15 095 10 805 8297 5558 2946 1974 834 300
0.005 3047 4292 9613 44 362 55 984 15 433 4571 2346 7540 5396 4143 2774 1469 984 414 148
0.01 1518 2138 4787 22 082 27 834 7668 2268 1163 1496 1069 820 548 288 192 79 26
0.05 295 415 927 4258 5315 1456 426 216 740 528 404 269 141 93 37 11
0.10 142 200 444 2030 2500 680 196 97 136 95 72 47 23 14 3 –
0.15 91 128 283 1287 1561 421 119 58 60 41 31 19 8 3 – –
0.20 66 92 203 916 1092 291 80 38 35 23 17 9 – – – –
0.25 50 70 155 693 811 214 57 26 22 14 10 4 – – – –
0.30 40 56 123 545 623 162 42 18 14 9 5 – – – – –
0.35 33 46 100 439 489 125 31 12 9 4 – – – – – –
0.40 27 38 82 359 388 97 22 8 5 – – – – – – –
0.45 23 32 69 297 310 76 16 – – – – – – – – –
0.50 20 27 58 248 248 58 11 – – – – – – – – –
0.55 17 23 49 207 196 44 5 – – – – – – – – –
0.60 15 20 42 173 154 32 – – – – – – – – – –
0.65 13 17 36 145 117 22 – – – – – – – – – –
0.70 11 15 31 120 86 13 – – – – – – – – – –
0.75 9 13 26 99 59 – – – – – – – – – – –
0.80 8 11 22 80 35 – – – – – – – – – – –
0.85 7 10 18 64 – – – – – – – – – – – –
0.90 6 8 15 49 – – – – – – – – – – – –
0.95 5 7 12 36 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: exposed ratio = 2 : 1. The sample size listed is the number of subjects needed in the exposed group. Double this number would be

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 369 471 1 930 847 4 322 614 19 888 657 24 913 372 6 843 626 2 014 756 1 029 014 653 418 465 696 356 275 237 254 124 571 83 030 34 793 12 510
0.00005 273 886 386 158 864 495 3 977 589 4 982 452 1 368 657 402 927 205 788 130 673 93 131 71 248 47 445 24 910 16 602 6955 2499
0.0001 136 938 193 072 432 230 1 988 706 2 491 087 684 286 201 449 102 885 65 330 46 560 35 619 23 719 12 452 8299 3476 1248
0.0005 27 380 38 603 86 418 397 599 497 995 136 790 40 266 20 563 13 055 9303 7117 4738 2486 1656 692 247
0.001 13 685 19 294 43 192 198 711 248 859 68 352 20 118 10 272 6521 4646 3554 2365 1240 825 344 122
0.005 2729 3847 8611 39 600 49 549 13 603 4000 2040 1294 921 703 467 244 161 66 21
0.01 1359 1916 4288 19 711 24 636 6759 1985 1011 640 455 347 230 119 78 31 9
0.05 264 372 830 3800 4705 1284 373 188 117 82 62 40 19 12 2 –
0.10 127 179 398 1811 2213 600 171 85 52 36 26 16 7 3 – –
0.15 81 114 254 1148 1383 372 104 50 30 20 14 8 – – – –
0.20 58 82 181 817 968 257 71 33 19 12 8 4 – – – –
0.25 45 63 138 618 719 189 50 23 13 7 4 – – – – –
0.30 36 50 109 485 552 143 37 16 8 4 – – – – – –
0.35 29 41 89 391 434 111 27 11 4 – – – – – – –
0.40 24 34 73 319 345 86 20 7 – – – – – – – –
0.45 20 28 61 264 275 67 14 – – – – – – – – –
0.50 17 24 52 220 220 52 9 – – – – – – – – –
0.55 15 21 44 184 175 39 – – – – – – – – – –
0.60 13 18 37 154 137 29 – – – – – – – – – –
0.65 11 15 32 128 105 19 – – – – – – – – – –
0.70 10 13 27 106 77 10 – – – – – – – – – –
0.75 8 11 23 87 53 – – – – – – – – – – –
0.80 7 10 19 71 31 – – – – – – – – – – –
0.85 6 8 16 56 – – – – – – – – – – – –
0.90 5 7 13 43 – – – – – – – – – – – –
0.95 4 6 11 31 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: exposed ratio = 3 : 1. The sample size listed is the number of subjects needed in the exposed group. Triple this number would be

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 285 566 1 815 876 4 068 209 18 690 665 23 293 643 6 381 472 1 869 238 950 463 601 217 427 061 325 766 215 895 112 429 74 554 30 945 11 048
0.00005 257 106 363 164 813 616 3 737 999 4 658 521 1 276 231 373 825 190 079 120 234 85 404 65 147 43 174 22 482 14 907 6186 2207
0.0001 128 548 181 575 406 791 1 868 916 2 329 131 638 076 186 899 95 031 60 111 42 697 32 569 21 583 11 238 7451 3091 1102
0.0005 25 702 36 304 81 332 373 649 465 619 127 552 37 358 18 993 12 013 8532 6507 4311 2244 1487 615 218
0.001 12 846 18 145 40 650 186 741 232 680 63 737 18 665 9488 6000 4261 3249 2152 1119 741 306 107
0.005 2562 3618 8104 37 214 46 329 12 684 3711 1884 1190 844 643 425 220 145 58 19
0.01 1276 1802 4035 18 523 23 035 6303 1842 934 589 417 318 209 107 70 27 8
0.05 248 349 781 3571 4399 1198 346 174 108 76 57 36 17 10 2 –
0.10 119 168 374 1702 2070 560 159 78 48 33 24 15 6 2 – –
0.15 76 107 238 1079 1294 347 97 47 28 19 13 7 – – – –
0.20 55 77 171 767 905 240 66 31 18 11 8 3 – – – –
0.25 42 59 130 580 672 177 47 21 12 7 4 – – – – –
0.30 33 47 103 456 517 134 34 15 7 – – – – – – –
0.35 27 38 83 366 406 103 25 10 3 – – – – – – –
0.40 23 32 69 300 323 81 18 6 – – – – – – – –
0.45 19 27 58 248 258 63 13 – – – – – – – – –
0.50 16 23 48 206 206 48 8 – – – – – – – – –
0.55 14 19 41 172 164 37 – – – – – – – – – –
0.60 12 16 35 144 128 27 – – – – – – – – – –
0.65 10 14 30 120 98 18 – – – – – – – – – –
0.70 9 12 25 99 72 7 – – – – – – – – – –
0.75 8 10 21 81 50 – – – – – – – – – – –
0.80 6 9 18 66 29 – – – – – – – – – – –
0.85 6 8 15 52 – – – – – – – – – – – –
0.90 5 6 12 39 – – – – – – – – – – – –
0.95 4 5 10 28 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: exposed ratio = 4 : 1. The sample size listed is the number of subjects needed in the exposed group. Quadruple this number would be

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 472 091 2 082 958 4 710 686 21 983 178 28 264 016 7 851 105 2 355 325 1 221 276 785 104 565 273 436 166 294 411 157 946 106 615 45 666 16 672
0.00005 294 411 416 580 942 108 4 396 481 5 652 548 1 570 142 471 036 244 238 157 008 113 044 87 224 58 875 31 583 21 318 9129 3330
0.0001 147 201 208 283 471 036 2 198 144 2 826 115 785 022 235 500 122 108 78 496 56 515 43 606 29 433 15 788 10 656 4562 1663
0.0005 29 433 41 645 94 178 439 474 564 968 156 925 47 071 24 404 15 686 11 292 8712 5879 3152 2126 908 329
0.001 14 711 20 816 47 071 219 641 282 325 78 413 23 518 12 191 7835 5639 4350 2935 1572 1060 451 162
0.005 2935 4152 9385 43 774 56 210 15 604 4675 2421 1554 1117 861 579 309 207 86 28
0.01 1463 2069 4675 21 790 27 946 7752 2319 1199 769 552 425 285 151 100 40 12
0.05 285 402 906 4204 5334 1471 435 222 141 100 76 49 24 15 3 –
0.10 138 194 435 2005 2508 686 200 100 62 43 32 20 8 4 – –
0.15 89 125 278 1273 1566 425 121 59 36 24 17 10 – – – –
0.20 64 90 200 906 1095 294 82 39 15 15 10 5 – – – –
0.25 49 69 152 686 812 215 58 27 10 9 6 – – – – –
0.30 40 55 121 540 623 163 42 19 6 5 – – – – – –
0.35 33 45 99 435 489 125 31 13 – – – – – – – –
0.40 27 38 82 357 388 97 23 8 – – – – – – – –
0.45 23 32 69 295 309 76 16 5 – – – – – – – –
0.50 20 27 58 247 247 58 11 – – – – – – – – –
0.55 17 24 50 207 195 44 7 – – – – – – – – –
0.60 15 20 42 173 152 32 2 – – – – – – – – –
0.65 13 18 36 145 116 22 – – – – – – – – – –
0.70 11 15 31 121 85 13 – – – – – – – – – –
0.75 10 13 27 100 58 6 – – – – – – – – – –
0.80 9 12 23 82 35 – – – – – – – – – – –
0.85 8 10 19 66 14 – – – – – – – – – – –
0.90 7 9 16 51 – – – – – – – – – – – –
0.95 6 8 14 38 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: exposed ratio = 1 : 1. The sample size listed is the number of subjects needed in the exposed group. An equivalent number would be

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1190 356 1663 432 3 680 447 16 792 779 20 878 641 5 726 194 1 683 582 859 799 546 209 389 547 298 242 198 909 104 767 69 986 29 458 10 630
0.00005 238 065 332 677 736 066 3 358 436 4 175 543 1 145 183 336 697 171 948 109 233 77 903 59 643 39 777 20 950 13 994 5889 2124
0.0001 119 028 166 332 368 018 1 679 143 2 087 655 572 556 168 336 85 967 54 611 38 947 29 818 19 886 10 473 6995 2943 1061
0.0005 23 799 33 257 73 580 335 708 417 346 114 455 33 648 17 182 10 914 7783 5958 3973 2091 1396 586 210
0.001 11 895 16 622 36 775 167 779 208 557 57 193 16 812 8584 5452 3887 2975 1983 1043 696 292 104
0.005 2372 3315 7332 33 436 41 526 11 382 3343 1705 1082 771 589 392 205 136 56 19
0.01 1182 1651 3651 16 643 20 647 5656 1659 845 536 381 291 193 100 66 26 8
0.05 230 321 707 3208 3944 1075 312 157 99 69 52 34 17 10 2 –
0.10 111 154 339 1529 1856 503 144 71 44 30 23 14 6 3 – –
0.15 71 99 216 969 1160 312 88 43 26 17 13 7 – – – –
0.20 51 71 155 689 812 216 60 28 17 11 8 4 – – – –
0.25 39 54 118 522 603 159 43 20 11 7 4 – – – – –
0.30 31 43 93 410 464 121 32 14 7 4 – – – – – –
0.35 26 35 76 330 365 93 23 10 4 – – – – – – –
0.40 21 29 63 270 290 73 17 6 – – – – – – – –
0.45 18 25 52 223 232 57 13 – – – – – – – – –
0.50 15 21 44 186 186 44 9 – – – – – – – – –
0.55 13 18 38 155 148 34 5 – – – – – – – – –
0.60 11 16 32 130 116 25 – – – – – – – – – –
0.65 10 13 27 108 89 18 – – – – – – – – – –
0.70 9 12 23 90 66 11 – – – – – – – – – –
0.75 8 10 20 74 46 – – – – – – – – – – –
0.80 7 9 17 60 28 – – – – – – – – – – –
0.85 6 7 14 47 – – – – – – – – – – – –
0.90 5 6 12 36 – – – – – – – – – – – –
0.95 4 5 9 26 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: exposed ratio = 2 : 1. The sample size listed is the number of subjects needed in the exposed group. Double this number would be

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 088 323 1516 254 3330 831 15 057 392 18 412 768 5 014 203 1456 566 736 622 464 207 328 848 250 342 165 451 85 870 56 861 23 565 8410
0.00005 217 658 303 242 666 145 3 011 370 3 682 391 1 002 792 291 297 147 315 92 835 65 764 50 064 33 087 17 171 11 370 4711 1681
0.0001 108 825 151 615 333 059 1 505 617 1 841 094 501 366 145 638 73 651 46 413 32 879 25 029 16 541 8584 5684 2355 839
0.0005 21 759 30 314 66 590 301 015 368 057 100 225 29 111 14 721 9276 6570 5001 3305 1714 1134 469 166
0.001 10 875 15 151 33 281 150 439 183 927 50 082 14 545 7354 4634 3282 2498 1650 855 566 233 82
0.005 2169 3021 6635 29 979 36 623 9968 2892 1461 920 651 495 326 168 111 45 15
0.01 1080 1505 3304 14 922 18 210 4954 1436 725 456 322 245 161 83 54 21 6
0.05 210 292 639 2876 3480 942 271 135 84 59 44 29 14 8 2 –
0.10 101 140 306 1370 1638 441 125 62 38 26 19 12 5 2 – –
0.15 65 90 195 868 1025 274 76 37 22 15 11 6 – – – –
0.20 46 64 139 617 718 190 52 25 14 9 6 3 – – – –
0.25 36 49 106 466 534 140 37 17 10 6 4 – – – – –
0.30 28 39 84 366 411 107 28 12 6 3 – – – – – –
0.35 23 32 68 294 323 83 21 9 4 – – – – – – –
0.40 19 26 56 240 257 65 15 6 – – – – – – – –
0.45 16 22 47 199 206 51 11 – – – – – – – – –
0.50 14 19 39 165 165 39 8 – – – – – – – – –
0.55 12 16 33 138 132 30 – – – – – – – – – –
0.60 10 14 28 115 104 23 – – – – – – – – – –
0.65 9 12 24 96 80 16 – – – – – – – – – –
0.70 8 10 20 79 60 9 – – – – – – – – – –
0.75 7 9 17 65 42 – – – – – – – – – – –
0.80 6 7 14 52 26 – – – – – – – – – – –
0.85 5 6 12 41 – – – – – – – – – – – –
0.90 4 5 10 31 – – – – – – – – – – – –
0.95 3 4 8 22 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: exposed ratio = 3 : 1. The sample size listed is the number of subjects needed in the exposed group. Triple this number would be

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 034 606 1 440 316 3 154 116 14 188 116 17 178 604 4 657 092 1 342 104 674 194 422 454 297 814 225 764 148 182 76 019 49 975 20 438 7223
0.00005 206 915 288 054 630 802 2 837 520 3 435 570 931 374 268 406 134 830 84 485 59 558 45 149 29 633 15 201 9993 4086 1443
0.0001 103 454 144 022 315 388 1418 696 1 717 691 465 659 134 194 67 410 42 238 29 776 22 572 14 815 7599 4995 2042 721
0.0005 20 685 28 795 63 057 283 636 343 387 93 087 26 824 13 473 8442 5950 4510 2960 1518 997 407 143
0.001 10 338 14 392 31 515 141 754 171 599 46 516 13 402 6731 4217 2972 2253 1478 757 497 203 71
0.005 2061 2870 6282 28 248 34 169 9259 2665 1338 837 590 446 292 149 98 39 13
0.01 1027 1429 3128 14 059 16 990 4601 1323 663 415 292 221 144 73 48 19 6
0.05 199 277 605 2709 3247 876 250 124 77 53 40 26 12 8 2 –
0.10 96 133 289 1290 1529 410 115 57 35 24 17 11 5 2 – –
0.15 61 85 184 817 957 255 71 34 20 14 10 6 – – – –
0.20 44 61 132 581 670 177 48 23 13 9 6 3 – – – –
0.25 34 47 100 439 499 130 35 16 9 5 3 – – – – –
0.30 27 37 79 344 384 99 26 11 6 – – – – – – –
0.35 22 30 64 277 302 77 19 8 3 – – – – – – –
0.40 18 25 53 226 241 60 14 5 – – – – – – – –
0.45 15 21 44 186 193 47 10 – – – – – – – – –
0.50 13 18 37 155 155 37 7 – – – – – – – – –
0.55 11 15 31 129 124 28 – – – – – – – – – –
0.60 9 13 26 108 97 21 – – – – – – – – – –
0.65 8 11 22 89 75 15 – – – – – – – – – –
0.70 7 9 19 74 56 7 – – – – – – – – – –
0.75 6 8 16 60 39 – – – – – – – – – – –
0.80 5 7 13 48 24 – – – – – – – – – – –
0.85 4 6 11 38 – – – – – – – – – – – –
0.90 4 5 9 28 – – – – – – – – – – – –
0.95 3 4 7 20 – – – – – – – – – – – –
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: exposed ratio = 4 : 1. The sample size listed is the number of subjects needed in the exposed group. Quadruple this number would be
Table A9 Sample sizes for case–control studiesa.
Prevalence Odds ratio to be detected

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 970 728 2 788 519 6 306 363 29 429 793 37 838 497 10 510 715 3 153 225 1 635 011 1 051 081 756 780 583 937 394 159 211 464 142 743 61 147 22 330
0.00005 394 143 557 705 1 261 292 5 886 130 7 568 072 2 102 264 630 690 327 029 210 236 151 372 116 801 78 842 42 300 28 555 12 234 4469
0.0001 197 070 278 853 630 659 2943 172 3 784 269 1 051 207 315 373 163 532 105 130 75 696 58 409 39 427 21 155 14 281 6120 2237
0.0005 39 412 55 772 126 151 588 806 757 227 210 362 63 120 32 734 21 046 15 155 11 695 7896 4238 2862 1228 451
0.001 19 704 27 887 63 088 294 510 378 847 105 257 31 588 16 384 10 535 7587 5856 3954 2124 1435 617 228
0.005 3939 5579 12 638 59 074 76 145 21 173 6363 3304 2127 1533 1184 801 432 293 128 49
0.01 1968 2790 6331 29 646 38 309 10 663 3210 1669 1076 777 601 407 221 150 67 27
0.05 391 560 1288 6111 8059 2261 690 363 237 172 135 93 52 37 18 9
0.10 195 281 659 3181 4302 1219 379 202 133 98 77 54 32 23 13 8
0.15 129 189 451 2215 3072 879 278 150 100 75 60 43 26 19 11 8
0.20 97 143 348 1741 2476 716 230 126 85 64 52 37 23 18 11 8
0.25 77 116 287 1465 2137 624 203 113 77 59 48 35 23 18 12 9
0.30 64 98 248 1289 1930 569 188 106 73 56 46 34 23 18 13 10
0.35 56 86 222 1174 1802 536 180 103 72 56 46 35 24 19 14 11
0.40 49 77 203 1097 1727 519 177 102 72 56 47 36 25 20 15 12
0.45 44 70 191 1048 1694 513 178 104 74 58 49 38 27 22 17 14
0.50 40 66 182 1023 1696 519 182 108 77 61 52 40 29 24 19 16
0.55 38 62 178 1019 1732 535 191 114 82 66 56 44 32 27 21 18
0.60 36 61 177 1035 1806 562 203 123 89 72 61 49 36 31 25 21
0.65 35 60 180 1077 1927 605 222 135 99 80 69 56 42 36 29 25
0.70 34 61 188 1149 2110 669 248 153 113 92 79 64 49 43 35 31
0.75 35 64 203 1268 2390 764 287 178 133 109 94 77 59 52 43 38
0.80 37 70 230 1465 2831 913 348 218 164 135 117 97 75 66 55 49
0.85 43 82 278 1811 3591 1168 451 285 216 179 156 129 101 90 75 68
0.90 54 108 379 2527 5143 1687 659 420 320 266 233 195 154 137 116 105
0.95 93 190 690 4717 9851 3257 1288 828 635 531 466 391 313 280 238 217
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: case ratio = 1 : 1. The sample size listed is the number of subjects needed in the case group. An equivalent number would be included
in the control group.

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1529 065 2 153 652 4 825 672 22 280 178 28 149 758 7 764 749 2 302 966 1183 610 755 564 540 911 415 405 278 348 147 639 99 012 41 948 15 205
0.00005 305 811 430 731 965 148 4 456 162 5 630 233 1 553 041 460 628 236 743 151 128 108 194 83 091 55 678 29 534 19 807 8393 3044
0.0001 152 904 215 366 482 583 2 228 160 2 815 293 776 578 230 335 118 385 75 573 54 105 41 552 27 844 14 770 9906 4199 1524
0.0005 30 578 43 073 96 531 445 759 563 340 155 407 46 101 23 698 15 130 10 833 8321 5577 2960 1986 843 307
0.001 15 288 21 537 48 274 222 959 281 846 77 761 23 072 11 862 7574 5424 4167 2793 1483 996 424 155
0.005 3055 4308 9669 44 719 56 653 15 644 4649 2393 1530 1097 844 567 302 204 88 34
0.01 1526 2154 4843 22 440 28 505 7880 2346 1210 775 556 428 289 155 105 46 19
0.05 303 431 984 4623 6001 1674 506 264 171 124 97 66 37 26 13 7
0.10 150 216 503 2405 3207 904 279 148 97 71 56 39 23 17 9 6
0.15 100 145 343 1673 2292 653 205 111 74 55 44 31 19 14 8 6
0.20 74 110 265 1313 1849 533 170 93 63 47 38 28 17 13 8 6
0.25 59 89 218 1104 1597 465 151 84 57 44 35 26 17 13 9 6
0.30 49 75 188 971 1443 425 140 79 55 42 34 26 17 14 9 7
0.35 42 65 168 883 1349 401 135 77 54 42 34 26 18 14 10 8
0.40 37 58 154 825 1294 388 133 77 54 42 35 27 19 15 11 9
0.45 33 53 144 788 1270 385 133 78 56 44 37 28 20 17 13 10
0.50 31 50 137 768 1272 389 137 81 58 46 39 31 22 19 14 12
0.55 28 47 133 764 1301 402 144 86 62 50 42 33 24 21 16 14
0.60 27 45 133 775 1357 423 154 93 68 55 47 37 28 24 19 16
0.65 26 45 135 805 1449 456 168 103 76 61 52 42 32 28 22 19
0.70 26 45 140 859 1588 505 188 116 86 70 61 49 38 33 27 23
0.75 26 47 151 947 1799 577 218 136 102 84 72 59 46 40 33 29
0.80 28 51 170 1092 2133 690 265 166 125 104 90 74 58 51 42 38
0.85 31 60 205 1349 2708 884 343 218 165 137 120 100 78 70 58 53
0.90 39 78 279 1880 3881 1278 503 322 246 205 180 150 119 107 90 82
0.95 66 137 506 3505 7438 2472 984 635 489 410 360 303 243 218 186 169
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: case ratio = 2 : 1. The sample size listed is the number of subjects needed in the case group. Double this number would be included in
the control group.
Table A11 Sample size for case–control studiesa.

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 369 478 1930 861 4 322 663 19 888 975 24 913 964 6 843 813 2 014 824 1 029 056 653 448 465 720 356 295 237 271 124 583 83 040 34 800 12 517
0.00005 273 893 386 172 864 545 3 977 907 4 983 044 1 368 844 402 996 205 830 130 703 93 155 71 268 47 461 24 922 16 612 6963 2506
0.0001 136 945 193 086 432 280 1 989 023 2 491 679 684 473 201 517 102 927 65 360 46 584 35 640 23 735 12 464 8309 3483 1254
0.0005 27 387 38 617 86 468 397 917 498 587 136 977 40 334 20 604 13 086 9328 7137 4754 2498 1666 700 253
0.001 13 692 19 309 43 242 199 028 249 451 68 540 20 186 10 314 6551 4671 3574 2382 1252 836 352 128
0.005 2736 3862 8661 39 918 50 143 13 790 4068 2082 1324 945 724 484 256 171 73 28
0.01 1367 1931 4338 20 030 25 231 6947 2054 1053 671 480 368 246 131 88 39 16
0.05 271 387 881 4125 5313 1477 444 231 149 108 84 57 32 22 11 6
0.10 134 194 450 2145 2841 799 245 129 85 62 49 34 20 14 8 5
0.15 89 130 307 1491 2031 577 180 97 64 48 38 27 16 12 7 5
0.20 66 98 236 1171 1639 471 150 82 55 41 33 24 15 12 7 5
0.25 53 79 195 984 1417 412 133 74 50 38 31 23 15 12 8 6
0.30 44 67 168 865 1281 376 124 70 48 37 30 23 15 12 8 6
0.35 38 58 150 786 1197 355 119 68 47 37 30 23 16 13 9 7
0.40 33 52 137 734 1149 345 118 68 48 37 31 24 16 14 10 8
0.45 30 47 128 700 1128 342 119 69 49 39 32 25 18 15 11 9
0.50 27 44 122 682 1131 346 122 72 52 41 35 27 19 16 12 10
0.55 25 42 119 679 1156 357 128 76 55 44 38 30 22 18 14 12
0.60 24 40 118 689 1207 377 137 83 60 49 41 33 25 21 17 14
0.65 23 40 119 715 1289 406 150 91 67 55 47 38 28 24 20 17
0.70 23 40 124 762 1414 450 168 104 77 63 54 44 33 29 24 21
0.75 23 42 133 839 1602 515 195 121 91 75 65 53 41 36 29 26
0.80 24 45 150 968 1900 616 236 149 112 93 80 66 52 45 38 34
0.85 27 52 180 1194 2413 789 307 195 148 123 107 89 70 62 52 46
0.90 34 68 245 1664 3459 1142 450 288 220 184 161 134 107 95 80 72
0.95 57 119 444 3100 6632 2208 881 569 438 367 323 271 217 194 165 150
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: case ratio = 3 : 1. The sample size listed is the number of subjects needed in the case group. Triple this number would be included in the
control group.

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 285 573 1 815 890 4068 256 18 690 963 23 294 197 6 381 647 1 869 301 950 501 601 245 427 084 325 786 215 910 112 440 74 563 30 952 11 054
0.00005 257 112 363 178 813 662 3 738 297 4 659 075 1276 406 373 889 190 118 120 262 85 427 65 166 43 189 22 493 14 916 6193 2213
0.0001 128 555 181 589 406 838 1 869 214 2 329 685 638 251 186 963 95 070 60 139 42 720 32 588 21 599 11 249 7461 3098 1108
0.0005 25 709 36 318 81 379 373 947 466 173 127 727 37 422 19 032 12 041 8554 6526 4326 2255 1496 622 224
0.001 12 853 18 159 40 697 187 039 233 234 63 912 18 729 9527 6028 4284 3269 2167 1130 750 313 113
0.005 2568 3632 8151 37 513 46 884 12 860 3775 1923 1219 867 662 440 231 154 65 25
0.01 1283 1816 4082 18 823 23 592 6479 1906 973 618 440 337 224 118 79 34 14
0.05 255 363 829 3876 4969 1378 412 214 137 99 77 52 29 20 10 5
0.10 126 182 423 2015 2658 746 228 120 78 57 45 31 18 13 7 4
0.15 83 122 289 1401 1901 539 168 90 60 44 35 25 15 11 7 4
0.20 62 92 222 1099 1534 440 140 76 51 38 31 22 14 11 7 5
0.25 50 74 183 923 1326 385 125 69 47 36 29 21 14 11 7 5
0.30 41 63 158 812 1200 352 116 65 45 34 28 21 14 11 7 6
0.35 35 55 140 738 1122 333 111 63 44 34 28 21 14 12 8 6
0.40 31 49 128 688 1077 323 110 63 45 35 29 22 15 13 9 7
0.45 28 44 120 657 1058 320 111 65 46 36 30 23 17 14 10 8
0.50 25 41 114 640 1060 324 114 67 48 38 32 25 18 15 11 10
0.55 23 39 111 636 1084 335 120 72 52 41 35 28 20 17 13 11
0.60 22 38 110 645 1132 354 128 78 57 46 39 31 23 20 15 13
0.65 21 37 111 669 1209 381 140 86 63 51 44 35 26 23 18 16
0.70 21 37 116 713 1326 422 158 97 72 59 51 41 31 27 22 19
0.75 21 39 125 786 1504 483 183 114 85 70 61 50 38 33 27 24
0.80 22 42 140 905 1784 579 222 140 105 87 75 62 48 42 35 31
0.85 25 48 168 1117 2266 742 289 183 139 115 101 83 65 58 48 43
0.90 31 63 228 1556 3248 1073 423 271 207 173 151 126 100 89 75 67
0.95 52 110 412 2897 6229 2076 829 536 412 345 303 255 203 182 154 139
a
α = 0.05 (two-tailed), β = 0.10 (power = 90%), control: case ratio = 4 : 1. The sample size listed is the number of subjects needed in the case group. Quadruple this number would be included

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 472 099 2082 974 4 710 741 21 983 531 28 264 683 7 851 317 2355 404 1 221 324 785 139 565 302 436 191 294 430 157 960 106 627 45 676 16 681
0.00005 294 418 416 596 942 163 4 396 835 5 653 216 1 570 354 471 115 244 286 157 043 113 073 87 248 58 894 31 598 21 330 9139 3339
0.0001 147 208 208 299 471 091 2 198 497 2 826 782 785 234 235 579 122 156 78 531 56 544 43 631 29 452 15 803 10 668 4572 1671
0.0005 29 440 41 661 94 233 439 828 565 636 157 137 47 150 24 452 15 721 11 321 8736 5899 3166 2138 918 337
0.001 14 719 20 831 47 126 219 994 282 992 78 625 23 596 12 239 7870 5668 4375 2954 1587 1072 461 171
0.005 2943 4168 9441 44 128 56 879 15 816 4753 2469 1589 1146 885 599 323 219 96 37
0.01 1470 2085 4730 22 145 28 617 7966 2398 1248 804 581 449 305 165 113 50 20
0.05 293 419 962 4566 6020 1690 516 272 177 129 101 70 39 28 14 7
0.10 146 211 493 2377 3214 911 283 151 100 74 58 41 24 18 10 6
0.15 97 142 337 1655 2295 657 208 113 75 56 45 32 20 15 9 6
0.20 73 107 260 1301 1850 535 172 95 64 48 39 28 18 14 9 6
0.25 58 87 215 1095 1597 466 152 85 58 44 36 27 17 14 9 7
0.30 49 74 186 964 1442 425 141 80 55 42 35 26 18 14 10 8
0.35 42 65 166 877 1346 401 135 77 54 42 35 26 18 15 11 9
0.40 37 58 152 820 1291 388 133 77 54 42 35 27 19 16 12 10
0.45 33 53 143 784 1266 384 133 78 56 44 37 29 20 17 13 11
0.50 31 50 137 765 1267 388 137 81 58 46 39 31 22 19 15 12
0.55 29 47 133 761 1294 400 143 85 62 50 42 33 25 21 16 14
0.60 27 46 133 774 1350 421 152 92 67 54 46 37 28 24 19 16
0.65 26 45 135 805 1440 453 166 101 75 61 52 42 32 27 22 19
0.70 26 46 141 859 1577 500 186 115 85 69 60 49 37 32 26 23
0.75 27 48 152 948 1785 571 215 134 100 82 71 58 45 39 33 29
0.80 28 53 172 1095 2115 682 260 163 123 101 88 73 57 50 42 37
0.85 32 62 208 1353 2683 873 337 213 162 134 117 97 76 68 57 51
0.90 41 81 283 1888 3842 1260 493 314 240 200 175 146 116 103 87 79
0.95 70 142 516 3524 7359 2433 962 619 475 397 349 293 234 210 179 162
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: case ratio = 1 : 1. The sample size listed is the number of subjects needed in the case group. An equivalent number would be included

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1190 363 1663 444 3 680 489 16 793 046 20 879 138 5 726 351 1 683 639 859 834 546 235 389 568 298 260 198 923 104 777 69 995 29 465 10 635
0.00005 238 071 332 689 736 108 3 358 703 4 176 039 1 145 339 336 754 171 983 109 259 77 923 59 660 39 791 20 960 14 003 5896 2129
0.0001 119 034 166 344 368 060 1 679 410 2 088 152 572 713 168 393 86 001 54 637 38 967 29 835 19 899 10 483 7004 2950 1066
0.0005 23 805 33 269 73 622 335 976 417 842 114 612 33 705 17 216 10 939 7803 5975 3986 2101 1405 593 216
0.001 11 901 16 635 36 817 168 047 209 054 57 349 16 869 8618 5477 3907 2993 1997 1053 705 298 109
0.005 2378 3327 7374 33 704 42 024 11 540 3400 1740 1107 791 607 406 215 145 63 24
0.01 1188 1664 3693 16 911 21 146 5814 1717 880 561 402 308 207 211 75 33 14
0.05 236 333 750 3482 4455 1237 371 193 125 91 70 48 27 19 10 5
0.10 117 167 383 1810 2383 669 205 109 71 53 41 29 17 13 7 5
0.15 77 112 261 1258 1704 484 152 82 54 41 32 23 14 11 7 5
0.20 58 84 201 987 1376 396 126 69 47 35 28 21 13 10 7 5
0.25 46 68 166 829 1190 346 112 62 43 33 27 20 13 10 7 5
0.30 38 57 143 729 1076 316 105 59 41 32 26 20 13 11 7 6
0.35 33 50 127 662 1006 299 101 58 40 31 26 20 14 11 8 7
0.40 29 45 116 618 966 290 99 58 41 32 27 21 15 12 9 7
0.45 26 41 108 590 949 288 100 59 42 33 28 22 16 13 10 8
0.50 24 38 103 574 951 292 103 61 44 35 30 24 17 15 11 10
0.55 22 36 100 571 973 301 108 65 47 38 32 26 19 16 13 11
0.60 21 34 99 579 1016 318 116 70 52 42 36 29 22 19 15 13
0.65 20 34 101 601 1085 343 127 78 58 47 40 33 25 22 18 16
0.70 20 34 105 640 1190 380 143 89 66 54 47 38 29 26 21 19
0.75 20 35 112 705 1350 435 166 104 78 64 56 46 36 32 26 23
0.80 21 38 126 812 1601 520 201 127 96 80 70 58 45 40 34 30
0.85 23 44 152 1002 2034 667 261 167 127 106 93 77 61 55 46 42
0.90 29 58 205 1395 2916 965 383 246 189 158 139 117 94 84 71 65
0.95 48 100 371 2598 5592 1868 750 487 376 316 279 236 190 171 147 134
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: case ratio = 2 : 1. The sample size listed is the number of subjects needed in the case group. Double this number would be included in
the control group.

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 088 329 1516 265 3330 869 15 057 631 18 413 208 5 014 341 1456 616 736 652 464 229 328 865 250 357 165 463 85 879 56 868 23 570 8 415
0.00005 217 664 303 253 666 182 3 011 608 3 682 831 1002 930 291 347 147 345 92 856 65 782 50 079 33 098 17 180 11 377 4717 1 685
0.0001 108 831 151 626 333 096 1505 856 1 841 534 501 504 145 688 73 681 46 435 32 896 25 044 16 553 8592 5691 2360 844
0.0005 21 764 30 325 66 628 301 253 368 496 100 363 29 161 14 751 9298 6588 5016 3316 1723 1141 474 171
0.001 10 881 15 162 33 319 150 678 184 367 50 220 14 595 7384 4655 3299 2513 1662 864 573 239 87
0.005 2174 3032 6672 30 218 37 064 10 107 2943 1491 942 668 510 338 177 118 50 19
0.01 1086 1516 3342 15 161 18 652 5093 1486 755 478 340 259 173 91 61 27 11
0.05 215 303 678 3120 3932 1085 323 167 107 77 60 41 23 16 8 4
0.10 107 152 345 1620 2105 588 179 94 62 45 35 25 15 11 6 4
0.15 70 101 235 1125 1507 426 132 71 47 35 28 20 12 9 6 4
0.20 52 76 181 882 1218 349 111 60 41 31 25 18 11 9 6 4
0.25 42 62 149 741 1053 305 99 55 37 29 23 17 11 9 6 5
0.30 35 52 128 650 954 280 92 52 36 28 23 17 12 9 7 5
0.35 30 45 114 590 892 265 89 51 36 28 23 18 12 10 7 6
0.40 26 40 104 550 857 257 88 51 36 28 24 18 13 11 8 7
0.45 23 36 97 525 843 256 89 52 37 30 25 19 14 12 9 7
0.50 21 34 92 511 846 259 92 55 39 32 27 21 15 13 10 9
0.55 19 32 89 507 866 268 97 58 42 34 29 23 17 15 12 10
0.60 18 30 88 514 905 283 104 63 46 38 32 26 19 17 13 12
0.65 18 30 89 533 967 306 114 70 52 42 36 30 23 20 16 14
0.70 17 30 92 567 1061 339 128 80 60 49 42 35 27 23 19 17
0.75 17 31 99 624 1204 389 149 93 70 58 51 42 32 29 24 21
0.80 18 33 111 718 1429 466 181 115 87 72 63 52 41 37 31 28
0.85 20 38 132 884 1817 598 235 151 115 96 84 70 56 50 42 38
0.90 25 50 179 1230 2607 867 345 223 172 144 127 107 85 77 65 59
0.95 41 85 323 2288 5002 1678 678 442 342 288 255 215 174 157 134 123
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: case ratio = 3: 1. The sample size listed is the number of subjects needed in the case group. Triple this number would be included in the
control group.

in control
group 0.2 0.3 0.5 0.75 1.25 1.5 2.0 2.5 3.0 3.5 4.0 5.0 7.5 10.0 20.0 50.0
0.00001 1 034 611 1440 327 3 154 151 14 188 340 17 179 015 4 657 221 1 342 151 674 222 422 474 297 830 225 778 148 193 76 026 49 982 20 443 7227
0.00005 206 920 288 065 630 838 2 837 745 3 435 981 931 503 268 452 134 858 84 505 59 574 45 162 29 644 15 209 9999 4091 1447
0.0001 103 459 144 032 315 424 1418 920 1 718 102 465 788 134 240 67 438 42 259 29 792 22 585 14 825 7607 5002 2047 725
0.0005 20 690 28 806 63 092 283 861 343 799 93 216 26 870 13 501 8462 5966 4524 2970 1525 1003 412 147
0.001 10 344 14 403 31 551 141 978 172 011 46 645 13 449 6759 4237 2988 2266 1489 765 504 207 75
0.005 2067 2880 6318 28 473 34 581 9388 2712 1366 858 606 460 303 157 104 44 17
0.01 1032 1440 3164 14 285 17 404 4731 1370 691 435 308 234 155 81 54 23 10
0.05 205 288 641 2938 3670 1009 298 153 98 70 54 37 20 14 7 4
0.10 101 144 327 1525 1966 547 166 87 57 41 32 23 13 10 5 3
0.15 67 96 222 1059 1408 397 123 66 43 32 26 18 11 8 5 4
0.20 50 72 171 830 1138 325 103 56 38 28 23 17 10 8 5 4
0.25 39 58 140 696 985 285 92 51 35 26 21 16 10 8 6 4
0.30 33 49 121 611 892 261 86 48 33 26 21 16 11 9 6 5
0.35 28 42 107 554 836 248 83 47 33 26 21 16 11 9 7 5
0.40 24 38 97 517 803 241 82 48 34 26 22 17 12 10 7 6
0.45 22 34 91 492 790 240 83 49 35 28 23 18 13 11 8 7
0.50 20 32 86 479 793 243 86 51 37 30 25 20 14 12 9 8
0.55 18 30 83 475 812 252 91 55 40 32 27 22 16 14 11 9
0.60 17 28 82 481 849 266 97 59 44 35 30 24 18 16 13 11
0.65 16 28 83 498 908 288 107 66 49 40 34 28 21 18 15 13
0.70 16 28 86 530 997 319 121 75 56 46 40 33 25 22 18 16
0.75 16 29 92 583 1131 366 140 88 67 55 48 39 31 27 22 20
0.80 17 31 103 670 1343 439 171 108 82 68 60 50 39 35 29 26
0.85 18 35 123 826 1708 564 222 143 109 91 80 67 53 47 40 36
0.90 23 45 166 1148 2452 817 327 211 163 137 120 101 81 73 62 56
0.95 37 78 298 2133 4707 1583 641 419 325 273 242 205 165 149 127 116
a
α = 0.05 (two-tailed), β = 0.20 (power = 80%), control: case ratio = 4: 1. The sample size listed is the number of subjects needed in the case group. Quadruple this number would be included
Table A17 Tabular values of 95% confidence limit factors for estimates of a poisson-distributed variable.
Observed
number on Observed number Observed number
which estimate Lower limit Upper limit on which estimate Lower limit Upper limit on which estimate Lower limit Upper limit
is based (n) factor (L) factor (U) is based (n) factor (L) factor (U) is based (n) factor (L) factor (U)
1 0.0253 5.57 21 0.619 1.53 120 0.833 1.200

2 0.121 3.61 22 0.627 1.51 140 0.844 1.184
3 0.206 2.92 23 0.634 1.50 160 0.854 1.171
4 0.272 2.56 24 0.641 1.49 180 0.862 1.160
5 0.324 2.33 25 0.647 1.48 200 0.868 1.151
6 0.367 2.18 26 0.653 1.47 250 0.882 1.134
7 0.401 2.06 27 0.659 1.46 300 0.892 1.121
8 0.431 1.97 28 0.665 1.45 350 0.899 1.112
9 0.458 1.90 29 0.670 1.44 400 0.906 1.104
10 0.480 1.84 30 0.675 1.43 450 0.911 1.098
11 0.499 1.79 35 0.697 1.39 500 0.915 1.093
12 0.517 1.75 40 0.714 1.36 600 0.922 1.084
13 0.532 1.71 45 0.729 1.34 700 0.928 1.078
14 0.546 1.68 50 0.742 1.32 800 0.932 1.072
15 0.560 1.65 60 0.770 1.30 900 0.936 1.068
16 0.572 1.62 70 0.785 1.27 1000 0.939 1.064
17 0.583 1.60 80 0.798 1.25
18 0.593 1.58 90 0.809 1.24
19 0.602 1.56 100 0.818 1.22
20 0.611 1.54
493
Glossary
The accuracy of a measurement is the degree Studies of adverse effects examine case reports
to which the measurement approximates of adverse drug reactions, attempting to
the truth. judge subjectively whether the adverse
Ad hoc studies are studies that require events were indeed caused by the
primary data collection. antecedent drug exposure.
Active surveillance is surveillance carried out Adverseomics is the study of vaccine adverse
via a continuous, defined process in a reactions using immunogenomics and
specific population, using one of several systems biology approaches.
approaches. Active surveillance can be Agreement is the degree to which different
medical product-based, identifying adverse methods or sources of information give the
events in patients taking certain products; same answers. Agreement between two
setting-based, identifying adverse events in sources or methods does not imply that
certain healthcare settings where patients either is valid or reliable.
are likely to present for treatment (e.g. Analyses of secular trends examine trends in
emergency departments); or event-based, disease events over time and/or across
identifying adverse events likely to be different geographic locations, and correlate
associated with medical products (e.g. acute them with trends in putative exposures,
liver failure). such as rates of drug utilization. The unit of
Actual knowledge, in a legal sense, is defined observation is usually a subgroup of a
as literal awareness of a fact. Actual population, rather than individuals. Also
knowledge can be demonstrated by called ecological studies.
showing that the manufacturer was Analytic studies are studies with control
cognizant of reasonable information groups, such as case–control studies,
suggesting, for example, a particular risk. cohort studies, and randomized clinical
An adverse drug event, adverse drug experience, trials.
adverse event, or adverse experience is an Anticipated beneficial effects of drugs are
untoward outcome that occurs during or desirable effects that are presumed to be
following clinical use of a drug. It does not caused by the drug. They usually represent
necessarily have a causal relationship with the reason for prescribing or ingesting the
this treatment. It may or may not be drug.
preventable. Anticipated harmful effects of drugs are
An adverse drug reaction is an adverse drug unwanted effects that could have been
event that is judged to be caused by the predicted on the basis of existing
drug. knowledge.
494 Glossary
An association is when two events occur Changeability is the ability of an instrument to

together more often than one would expect measure a difference in score in patients
by chance. who have improved or deteriorated.
Autocorrelation is where any individual Channeling bias is a type of selection bias,
observation is to some extent a function of which occurs when a drug is claimed to be
the previous observation. safe and therefore is used in high-risk
Bias is any systematic (rather than random) patients who did not tolerate other drugs
error in a study. for that indication. It is sometimes used
Biological inference is the process of synonymously with confounding by
generalizing from a statement about an indication.
association seen in a population to a Clearance is the proportion of the apparent
causal statement about biological volume of distribution that is cleared of
relationships. drug in a specified time. Its units are
Case-cohort studies are studies that compare volume per time, such as liters per hour.
cases with a disease to a sample of subjects The total body clearance is the sum of
randomly selected from the parent cohort. clearances by different routes, e.g. renal,
Case–control studies are studies that compare hepatic, pulmonary.
cases with a disease to controls without the Clinical pharmacology is the study of the
disease, looking for differences in effects of drugs in humans.
antecedent exposures. Cohort studies are studies that identify defined
Case-crossover studies are studies that populations and follow them forward in
compare cases at the time of disease time, examining their frequencies (e.g.
occurrence to different time periods in the incidence rate, cumulative incidence) of
same individuals, looking for differences in disease. Cohort studies generally identify
antecedent exposures. and compare exposed patients to unexposed
Case reports are reports of the experience of patients or to patients who receive a
individual patients. As used in different exposure.
pharmacoepidemiology, a case report Combination-triggered drug–drug interaction
usually describes a patient who was is, in a potential drug–drug interaction, the
exposed to a drug and experienced a scenario in which both the object drug and
particular outcome, usually an adverse precipitant drugs are initiated
event. simultaneously.
Case series are reports of collections of Confidence interval can be conceptualized to
patients, all of whom have a common represent a range of values within which
exposure, examining what their clinical the true population value lies, with some
outcomes were. Alternatively, case series probability.
can be reports of patients who have a Confidentiality is the right of patients to limit
common disease, examining what their the transfer and disclosure of private
antecedent exposures were. No control information.
group is present. A confounding variable, or confounder, is a
An exposure causes a health event when it variable other than the risk factor and
truly increases the probability of that event outcome variable under study that is related
in some individuals. That is, there are at independently both to the risk factor and to
least some individuals who would the outcome. A confounder can artificially
experience the event given the exposure inflate or reduce the magnitude of
who would not experience the event absent association between and exposure and
the exposure. outcome.
Glossary 495
Confounding by indication can occur when the The defined daily dose (DDD) is the usual
underlying diagnosis or other clinical daily maintenance dose for a drug for its
features that affect the use of a certain drug main indication in adults.
are also related to the outcome under study. Descriptive studies are studies that do not have
Construct validity refers to the extent to which control groups, namely case reports, case
results from a given instrument are series, and analyses of secular trends. They
consistent with those from other measures are in contrast with analytic studies.
in a manner consistent with theoretical Detection bias is an error in the results of a
hypotheses. study due to a systematic difference
Constructive knowledge, from a legal between the study groups in the procedures
perspective, is knowledge that a person did used for ascertainment, diagnosis, or
not have, but could have acquired by the verification of disease.
exercise of reasonable care. Differential misclassification occurs when the
A cost is the consumption of a resource that degree of misclassification of one variable
could otherwise be used for another purpose. (e.g. drug usage) varies according to the
Cost–benefit analysis of medical care level of another variable (e.g. disease status).
compares the cost of a medical intervention The direct medical costs of medical care are
to its benefit. Both costs and benefits must the costs that are incurred in providing the
be measured in the same monetary units care.
(e.g. dollars). Direct non-medical costs are non-medical care
Cost-effectiveness analysis of medical care costs incurred because of an illness or the
compares the cost of a medical intervention need to seek medical care. They can include
to its effectiveness. Costs are expressed in the cost of transportation to the hospital or
monetary units, while effectiveness is physician’s office, the cost of special
determined independently and may be clothing needed because of the illness, and
measured in terms of any clinically the cost of hotel stays and special housing
meaningful unit. Cost-effectiveness (e.g. modification of the home to
analyses usually examine the additional accommodate the ill individual).
cost per unit of additional effectiveness. Discriminative instruments are those that
Cost-identification analysis enumerates the measure differences among people at a
costs involved in medical care, ignoring the single point in time.
outcomes that result from that care. Disease registries are registries characterized
Criterion validity refers to the ability of an by inclusion of subjects based on diagnosis
instrument to measure what it is supposed of a common disease or condition.
to measure, as judged by agreement with a A drug is any exogenously administered
reference (gold) standard. substance that exerts a physiologic effect.
Cross-sectional studies examine exposures and Drug–drug interaction is the phenomenon in
outcomes in populations at one point in which one or more drugs affects the
time; they have no time sense. pharmacokinetics and/or
Data mining is exploratory data analysis for pharmacodynamics of one or more other
hypothesis generation. As part of a drugs.
knowledge discovery process, data mining Drug utilization, as defined by the World
looks to uncover patterns or correlations in Health Organization (WHO), is the
the data set with no or limited “marketing, distribution, prescription and
presupposition, with the intent of more use of drugs in a society, with special
rigorous testing of any emerging hypothesis emphasis on the resulting medical, social,
tailored to the issue at hand. and economic consequences.”
496 Glossary
Drug utilization evaluation (DUE) programs A study of drug efficacy is a study of whether,
are ongoing structured systems designed to under ideal conditions, a drug has the ability
improve drug use by intervening when to bring about the effect intended when
inappropriate drug use is detected. See also prescribing it.
drug utilization review programs. A study of drug efficiency is a study of
Drug utilization evaluation studies are ad hoc whether a drug can bring about its desired
investigations that assess the effect at an acceptable cost.
appropriateness of drug use. They are Enriched or hybrid study designs draw upon
designed to detect and quantify the both primary and secondary data, with
frequency of drug use problems. some data collected de novo, specifically for
Drug utilization review programs are ongoing the purposes of the study and other
structured systems designed to improve study-specific data collected via
drug use by intervening when inappropriate probabilistic or deterministic linkage with
drug use is detected. other data sources, such as electronic health
Drug utilization review studies are ad hoc records, administrative claims and billing
investigations that assess the data, vital records and genetic information.
appropriateness of drug use. They are Epidemiology is the study of the distribution
designed to detect and quantify any drug and determinants of disease or health-
use problems. See also drug utilization related states in populations.
evaluation programs. Evaluative instruments are those designed to
Drug utilization studies are descriptive studies measure changes within individuals over
that quantify the use of a drug. Their time.
objective is to quantify the present state, the Experimental studies are studies in which the
developmental trends, and the time course investigator controls the therapy that is to
of drug usage at various levels of the health be received by each participant, generally
care system, whether national, regional, using that control to randomly allocate
local, or institutional. participants among the study groups.
Ecological studies examine trends in disease Face validity is a judgment about the validity
events over time or across different of an instrument, based on an intuitive
geographic locations and correlate them assessment of the extent to which an
with trends in putative exposures, such as instrument meets a number of criteria
rates of drug utilization. The unit of including applicability, clarity and
observation is a subgroup of a population, simplicity, likelihood of bias,
rather than individuals. See also analyses of comprehensiveness, and whether
secular trends. redundant items have been included.
Effect modification occurs when the Fixed costs are costs that are incurred
magnitude of effect of a drug in causing an regardless of the volume of activity.
outcome differs according to the levels of a General causation, from a legal perspective,
variable other than the drug or the addresses whether a product is capable of
outcome (e.g. sex, age group). Effect causing a particular injury in the
modification can be assessed on an population of patients like the plaintiff.
additive and/or multiplicative scale. See Generic quality-of-life instruments aim to cover
interaction. the complete spectrum of function,
A study of drug effectiveness is a study of disability, and distress of the patient, and
whether, in the usual clinical setting, a drug are applicable to a variety of populations.
in fact achieves the effect intended when Half-life (T1/2) is the time taken for the drug
prescribing it. concentration to decline by half. Half-life is
Glossary 497
a function of both the apparent volume of Hypothesis-testing studies are studies that
distribution and clearance of the drug. evaluate in detail hypotheses raised
Hawthorne Effect is when study subjects alter elsewhere.
their behavior simply because of their Inception cohort design is a cohort study that is
participation in a study, unrelated to the restricted to new users of the exposure(s) of
study procedures or intervention. interest.
Health profiles are single instruments that Incidence/prevalence bias, a type of selection
measure multiple different aspects of bias, may occur in studies when prevalent
quality-of-life. cases rather than new cases of a condition
Health-related quality-of-life is a multifactorial are selected for a study. A strong association
concept which, from the patient’s with prevalence may be related to the
perspective, represents the end-result of all duration of the disease rather than to its
the physiological, psychological, and social incidence, because prevalence is
influences of the disease and the proportional to both incidence and duration
therapeutic process. Health-related quality- of the disease.
of-life may be considered on different levels: The incidence rate of a disease is a measure of
overall assessment of well-being; several how frequently the disease occurs.
broad domains—physiological, functional, Specifically, it is the number of new cases of
psychological, social, and economic status; the disease which develop over a defined
and subcomponents of each domain—for time period in a defined population at risk,
example pain, sleep, activities of daily divided by the number of people in that
living, and sexual function within physical population at risk.
and functional domains. Indirect costs are costs that do not stem
A human research subject, as defined in US directly from transactions for goods or
regulation, is “a living individual, about services, but represent the loss of
whom an investigator (whether professional opportunities to use a valuable resource in
or student) conducting research obtains alternative ways. They include costs due to
either: 1) data through intervention or morbidity (e.g. time lost from work) and
interaction with the individual, or 2) mortality (e.g. premature death leading to
identifiable private information.” (Title 45 US removal from the work force).
Code of Federal Regulations Part 46.102 (f)). Information bias is an error in the results of a
Hybrid or enriched study designs draw upon study due to a systematic difference
both primary and secondary data, with between the study groups in the accuracy of
some data collected de novo, specifically for the measurements being made of their
the purposes of the study and other exposure or outcome.
study-specific data collected via Instrumental variable is a variable used to
probabilistic or deterministic linkage with adjust for confounding that meets certain
other data sources, such as electronic health specific criteria: it should affect treatment
records, administrative claims and billing or be associated with treatment choice by
data, vital records and genetic information. sharing a common cause; should be a factor
Hypothesis-generating studies are studies that that is as good as randomly assigned, so
give rise to new questions about drug effects that it is unrelated to patient characteristics;
to be explored further in subsequent and should not be related to the outcome
analytical studies. other than through its association with
Hypothesis-strengthening studies are studies treatment.
that reinforce, although do not provide Intangible costs are those of pain, suffering,
definitive evidence for, existing hypotheses. and grief.
498 Glossary
Interaction, see effect modification. repeated assignments to the experimental

Interrupted time-series designs include or control arms.
multiple observations of study populations Near misses are medication errors that have
before and after an intervention. high potential for causing harm but did not,
Knowledge, as used in court cases, can be either because they were intercepted prior
actual or constructive; see those terms. to reaching a patient or because the error
Large simple trials are randomized trials reached the patient who fortuitously did
characterized by large sample sizes, broad not have any observable untoward sequelae.
entry criteria consistent with the approved Negative control precipitant drug is, in a study
medication label, randomization based on of a potential drug–drug interaction, a drug
equipoise, minimal data requirements, that is used in similar clinical
objectively-measured endpoints, follow-up circumstances as the potential precipitant
that minimizes interventions or under study, yet by virtue of the control
interference with normal clinical practice, precipitant’s pharmacology is not believed
follow-up of all patients regardless of to interact with the study object
whether they discontinue randomized Negative control object drug is, in a study of a
medication; and intent-to-treat analysis. potential drug–drug interaction, a drug that
Medication errors are any error in the process is used for similar indications as the object
of prescribing, transcribing, dispensing, under study, but is not believed to interact
administering, or monitoring a drug, pharmacologically with the study precipitant
regardless of whether an injury occurred or Non-differential misclassification occurs when
the potential for injury was present. the misclassification of one variable does
Meta-analysis is a systematic, structured not vary by the level of another variable.
review of the literature and formal Non-differential misclassification usually
statistical analysis of a collection of analytic results in bias toward the null.
results for the purpose of integrating the Non-experimental studies are studies in which
findings. Meta-analysis is used to identify the investigator does not control the
sources of variation among study findings therapy, but observes and evaluates the
and, when appropriate, to provide an results of ongoing medical care. The study
overall measure of effect as a summary of designs that are used are those that do not
those findings. involve random allocation, such as case
Microbiome includes the microorganisms, reports, case series, analyses of secular
primarily bacteria in the gut, and their trends, case–control studies, and cohort
genes, harbored within each person. studies.
Misclassification bias is the error resulting Object drug is, in a drug–drug interaction, the
from classifying study subjects as exposed drug(s) whose pharmacokinetics or
when they truly are unexposed, or vice pharmacodynamics are affected by the
versa. Alternatively, misclassification bias other drug(s).
can result from classifying study subjects as Object-triggered drug–drug interaction is, in a
diseased when they truly are not diseased, study of a potential drug–drug interaction,
or vice versa. the scenario in which the object drug is
Molecular pharmacoepidemiology is the study started in a person already taking the
of the manner in which molecular precipitant drug.
biomarkers alter the clinical effects of Observational studies (or nonexperimental
medications. studies) are studies in which the
An N-of-1 RCT is a randomized controlled investigator does not control the therapy,
trial within an individual patient, using but observes and evaluates the results of
Glossary 499
ongoing medical care. The study designs of people. It is also the application of the
that are used are those that do not involve research methods of clinical epidemiology
randomization, such as case reports, case to the content area of clinical
series, analyses of secular trends, case– pharmacology, and the primary science
control studies, and cohort studies. underlying the public health practice of
The odds ratio is the odds of exposure in the drug safety surveillance.
diseased group divided by the odds of Pharmacogenetics is the study of genetic
exposure in the non-diseased group. When determinants of responses to drugs.
the underlying risk of disease is low (about Although it is sometimes used
10% or lower) it is an unbiased estimator of synonymously with pharmacogenomics, it
the relative risk. It is also an unbiased often refers to a candidate-gene approach as
estimate of the rate ratio in a nested or opposed to a genome-wide approach.
population-based case–control study in Pharmacogenomics is the study of genetic
which controls are selected at random from determinants of responses to drugs.
the population at risk of disease at the time Although it is sometimes used
that the case occurred. synonymously with pharmacogenetics, it
One-group, post-only study design consists of often refers to a genome-wide approach as
making only one observation on a single opposed to a candidate-gene approach.
group which has already been exposed to a A pharmacokinetic compartment is a
treatment. theoretical space into which drug molecules
An opportunity cost is the value of a resource’s are said to distribute, and is represented by
next best use, a use that is no longer a given linear component of the log-
possible once the resource has been used. concentration versus time curve. It is not an
A p-value is the probability that a difference as actual anatomic or physiologic space, but is
large as or larger than the one observed in sometimes thought of as a tissue or group
the study could have occurred purely by of tissues that have similar blood flow and
chance if no association truly existed. drug affinity.
Patient reported outcomes are any report of the Pharmacokinetics is the study of the
status of a patient’s health condition that relationship between the dose administered
comes directly from the patient, without of a drug and the concentration achieved in
interpretation of the patient’s response by a the blood, in the serum, or at the site of
clinician or anyone else. action. It includes the study of the processes
Pharmacodynamics is the study of the of drug absorption, distribution,
relationship between drug level and drug metabolism, and excretion.
effect. It involves the study of the response Pharmacovigilance is the identification and
of the target tissues in the body to a given evaluation of drug safety signals. More
concentration of drug. recently, some have also used the term as
Pharmacoeconomics is the study of how the synonymous with pharmacoepidemiology.
price of pharmaceutical products and their WHO defines pharmacovigilance as the
economic impact health and the health care science and activities relating to the
system. detection, assessment, understanding and
Pharmacogenetic epidemiology is the study of prevention of adverse effects or any other
the effects of genetic determinants of drug possible drug-related problems (WHO.
response on outcomes in large numbers of Safety monitoring of medicinal products.
people. The importance of pharmacovigilance.
Pharmacoepidemiology is the study of the use Geneva, World Health Organization, 2002).
of and the effects of drugs in large numbers Mann defines pharmacovigilance as “the
500 Glossary
study of the safety of marketed drugs under potent drug, although it is not necessarily
the practical conditions of clinical usage in more effective.
large communities” (Pharmacovigilance. Potential adverse drug events are medication
R.D. Mann and E.B. Andrews, eds. John errors that have high potential for causing
Wiley & Sons Ltd., Chichester, 2002). harm but did not, either because they were
Pharmacology is the study of the effects of intercepted prior to reaching a patient or
drugs in a living system. because the error reached the patient who
Pharmacotherapeutics is the application of the fortuitously did not have any observable
principles of clinical pharmacology to untoward sequelae.
rational prescribing, the conduct of clinical The power (statistical power) of a study is the
trials, and the assessment of outcomes probability of detecting a difference in the
during real-life clinical practice. study if a difference really exists (either
Pharmionics is the study of how patients use between study groups or between treatment
or misuse prescription drugs in ambulatory periods).
care. Pragmatic clinical trials typically fall
Population-based databases or studies refers to somewhere in between a typical
whether there is an identifiable population randomized trial and a simple and a large
(which is not necessarily based in simple trial, where the goal is to introduce
geography), all of whose medical care one or more pragmatic elements into the
would be included in that database, design but with substantial protocol-
regardless of the provider. This allows one required follow-up and testing outside of
to determine incidence rates of diseases, as usual care practice.
well as being more certain that one knows Precipitant drug is, in a drug–drug
of all medical care that any given patient interaction, the drug that affects the
receives. pharmacokinetics or pharmacodynamics of
Positive control precipitant drug is, in a study the other drug(s).
of a potential drug–drug interaction, a Precipitant-triggered drug–drug interaction is,
precipitant drug known to produce an in a study of a potential drug–drug
association with an outcome in patients interaction, the scenario in which the
receiving the object drug of interest. precipitant drug is started in a person
Postmarketing surveillance is the study of already taking the object drug.
drug use and drug effects after release onto Precision is the degree of absence of random
the market. This term is sometimes used error. Precise estimates have narrow
synonymously with “pharmacoepi confidence intervals.
demiology,” but the latter can be relevant Precision medicine has been defined by the
to premarketing studies, as well. National Institutes of Health (NIH) in the
Conversely, the term “postmarketing United States as an “approach to disease
surveillance” is sometimes felt to apply to prevention and treatment based on people’s
only those studies conducted after drug individual differences in environment,
marketing that systematically screen for genes and lifestyle.”
adverse drug effects. However, this is a Pre-post with comparison group design
more restricted use of the term than that includes a single observation both before
used in this book. and after treatment in a non-randomly
Potency refers to the amount of drug that is selected group exposed to a treatment (e.g.
required to elicit a given response. A more physicians receiving feedback on specific
potent drug requires a smaller milligram prescribing practices), as well as
quantity to exert the same response as a less simultaneous before and after observations
Glossary 501
of a similar (comparison) group not study; namely, patient outcomes have not
receiving treatment. yet occurred as of the outset of the study.
Prescribing errors refer to issues related to Proteomics is, within the context of
underuse, overuse, and misuse of pharmacoepidemiology, the study of how
prescribed drugs, all of which contribute to proteins are responsible for variability in
the suboptimal utilization of medication response.
pharmaceutical therapies. Protopathic bias is interpreting a factor to be a
The prevalence of a disease is a measurement result of an exposure when it is in fact a
of how common the disease is. Specifically, determinant of the exposure, and can occur
it is the number of existing cases of the when an early sign of the disease under
disease in a defined population at a given study led to the prescription of the drug
point in time or over a defined time period, under study.
divided by the number of people in that Publication bias occurs when publication of a
population. study’s results is related to the study’s
Prevalence study bias, a type of selection bias findings, such that study results are not
that may occur in studies when prevalent published or publication is delayed because
cases rather than new cases of a condition of the results.
are selected for a study. A strong association Qualitative drug utilization studies are studies
with prevalence may be related to the that assess the appropriateness of drug use.
duration of the disease rather than to its Quality-of-life is the description of aspects
incidence, because prevalence is (domains) of physical, social, and emotional
proportional to both incidence and duration health that are relevant and important to
of the disease. the patient.
Privacy, in the setting of research, refers to Quantitative drug utilization studies are
each individual’s right to be free from descriptive studies of frequency of drug use.
unwanted inspection of, or access to, Random allocation is the assignment of
personal information by unauthorized subjects who are enrolled in a study into
persons. study groups in a manner determined by
Procedure registries are registries characterized chance.
by inclusion of subjects based on receipt of Random error is error due to chance.
specific services, such as procedures, or Random selection is the selection of subjects
based on hospitalizations. into a study from among those eligible in a
Product registries are registries characterized manner determined by chance.
by inclusion of subjects based on use of a Randomized clinical trials are studies in which
specific product (drug or device) or related the investigator randomly assigns patients
products in a given therapeutic area. to different therapies, one of which may be
Propensity scores are an approach to a control therapy.
controlling for confounding that uses Recall bias is an error in the results of a study
mathematical modeling to predict exposure due to a systematic difference between the
based on observed variables, and uses the study groups in the accuracy or
predicted probability of exposure as the completeness of their memory of their past
basis for matching or adjustment. exposures or health events.
Prospective drug utilization review is designed Referral bias is error in the results of a study
to detect drug-therapy problems before an that occurs when the reasons for referring a
individual patient receives the drug. patient for medical care are related to the
Prospective studies are studies performed exposure status, e.g. when the use of the
simultaneously with the events under drug contributes to the diagnostic process.
502 Glossary
Registries are organized systems that use Code of Federal Regulations Part 46.102
observational study methods to collect (d))
uniform data (clinical and other) to A research subject is “a living individual, about
evaluate specified outcomes for a whom an investigator (whether professional
population defined by a particular disease, or student) conducting research obtains
condition, or exposure, and that serves one either: 1) data through intervention or
or more predetermined scientific, clinical, interaction with the individual, or 2)
or policy purposes. Registries can be identifiable private information” (US Code
thought of as both the process for collecting of Federal Regulations 46.102f).
data from which studies are derived, as well Responsiveness is an instrument’s ability to
as referring to the actual database. detect change.
Regression to the mean is the tendency for Retrospective drug utilization review compares
observations on populations selected on the past drug use against predetermined criteria
basis of an abnormality to approach to identify aberrant prescribing patterns or
normality on subsequent observations. patient-specific deviations from explicit
The relative rate is the ratio of the incidence criteria.
rate of an outcome in the exposed group to Retrospective studies are studies conducted
the incidence rate of the outcome in the after the events under study have occurred.
unexposed group. It is synonymous with Both exposure and outcome have already
the terms rate ratio and incidence rate ratio. occurred as of the outset of the study.
The relative risk is the ratio of the cumulative Risk is the cumulative probability that
incidence of an outcome in the exposed something will happen.
group to the cumulative incidence of the Risk evaluation and mitigation strategy
outcome in the unexposed group. It is (REMS) is a pharmacovigilance assessment
synonymous with the term cumulative plan in the United States, approved by
incidence ratio. regulators in advance of implementation, to
Reliability is the degree to which the results ensure that the benefits of a drug or
obtained by a measurement procedure can biological product outweigh its risks.
be replicated. The measurement of In the EU, pharmacovigilance legislation
reliability does not require a gold standard, explicitly requires the active monitoring of
since it assesses only the concordance the outcome of risk minimization activities
between two or more measures. contained in the risk management plan,
A reporting rate in a spontaneous reporting placing the obligation on manufacturers
system is the number of reported cases of and regulatory authorities for this activity.
an adverse event of interest divided by A judgment about safety is a personal and/or
some measure of the suspect drug’s social judgment about the degree to which
utilization, usually the number of a given risk is acceptable.
dispensed prescriptions. This is perhaps Safety signal is a concern about an excess of
better referred to as a rate of reported adverse events compared to what is
cases. expected to be associated with a product’s
Reproducibility is the ability of an instrument (drug or device) use.
to obtain more or less the same scores upon Service registries are registries characterized by
repeated measurements of patients who inclusion of subjects based on receipt of
have not changed. specific services, such as procedures, or
Research, as defined in US regulation, is any based on hospitalizations.
activity designed to “develop or contribute Sample distortion bias is another name for
to generalizable knowledge”. (Title 45 US selection bias.
Glossary 503
Scientific inference is the process of Signal is a hypothesis that calls for further
generalizing from a statement about a work to be performed to evaluate that
population, which is an association, to a hypothesis.
causal statement about scientific theory. Signal detection is the process of looking for or
Selection bias is error in a study that is due to identifying signals from any source.
systematic differences in characteristics Signal generation, sometimes referred to as
between those who are selected for the data mining, is an approach that uses
study and those who are not. statistical methods to identify a safety
Self-controlled designs are studies that include signal. No particular medical product
only persons who experienced the outcome, exposure or adverse outcome is
using each person as her/his own control, prespecified.
and include self-controlled case series and Signal refinement is a process by which an
case-crossover designs. identified safety signal is further evaluated
Self-controlled case series (SCCS) design is a to determine whether evidence exists to
self-controlled design that is analogous to support a relationship between the
the cohort design. It includes only exposure and the outcome.
individuals who experienced the outcome, Specific causation, from a legal perspective,
and examines the rate of the outcome addresses whether the product in question
during exposed vs. unexposed periods actually caused an alleged injury in the
within those individuals. individual plaintiff.
Sensibility is a judgment about the validity of Specific quality-of-life instruments are focused
an instrument, based on an intuitive on disease or treatment issues specifically
assessment of the extent to which an relevant to the question at hand.
instrument meets a number of criteria Specificity is the proportion of persons who
including applicability, clarity and truly do not have a characteristic, who are
simplicity, likelihood of bias, correctly classified by a diagnostic test as
comprehensiveness, and whether not having it.
redundant items have been included. Spontaneous reporting systems are maintained
Sensitivity is the proportion of persons who by regulatory bodies throughout the world
truly have a characteristic, who are and collect unsolicited clinical observations
correctly classified by a diagnostic test as that originate outside of a formal study.
having it. Statistical inference is the process of
Sensitivity analysis is a set of procedures in generalizing from a sample of study
which the results of a study are recalculated subjects to the entire population from
using alternate values for some of the which those subjects are theoretically
study’s variables, in order to test the drawn.
sensitivity of the conclusions to altered Statistical interaction, see effect modification.
specifications. A statistically significant difference is a
A serious adverse experience is any adverse difference between two study groups that
experience occurring at any dose that is unlikely to have occurred purely by
results in any of the following outcomes: chance.
death, a life-threatening adverse Steady state, within pharmacokinetics, is the
experience, inpatient hospitalization or situation when the amount of drug being
prolongation of existing hospitalization, a administered equals the amount of drug
persistent or significant disability/ being eliminated from the body.
incapacity, or congenital anomaly/birth Systematic error is any error in study results
defect. other than that due to random variation.
504 Glossary
The therapeutic ratio is the ratio of the drug An unexpected adverse experience means any
concentration that produces toxicity to the adverse experience that is not listed in the
concentration that produces the desired current labeling for the product. This
therapeutic effect. includes an event that may be
Therapeutics is the application of the symptomatically and pathophysiologically
principles of clinical pharmacology to related to an event listed in the labeling, but
rational prescribing, the conduct of clinical differs from the event because of greater
trials, and the assessment of outcomes severity or specificity.
during real-life clinical practice. Utility measures of quality-of-life are
Type A adverse reactions are those that are the measured holistically as a single number
result of an exaggerated but otherwise along a continuum, e.g. from death (0.0) to
predictable pharmacological effect of the full health (1.0). The key element of a utility
drug. They tend to be common and instrument is that it is preference-based.
dose-related. Validity is the degree to which an assessment
Type B adverse reactions are those that are (e.g. questionnaire or other instrument)
aberrant effects of the drug. They tend to be measures what it purports to measure.
uncommon, not dose-related, and Variable costs are costs that increase with
unpredictable. increasing volume of activity.
A type I statistical error is concluding there is Apparent volume of distribution (VD) is the
an association when in fact one does not apparent volume that a drug is distributed
exist, i.e. erroneously rejecting the null in after complete absorption. It is usually
hypothesis. calculated from the theoretical plasma
A type II statistical error is concluding there is concentration at a time when all of the drug
no association when in fact one does exist, was assumed to be present in the body and
i.e. erroneously accepting the null uniformly distributed. This is calculated
hypothesis. from back extrapolation to time zero of the
Unanticipated beneficial effects of drugs are plasma concentration time curve after
desirable effects that could not have been intravenous administration.
predicted on the basis of existing Vaccinovigilance is the identification and
knowledge. evaluation of adverse events following
Unanticipated harmful effects of drugs are immunizations.
unwanted effects that could not have been Voluntariness is the concept in research ethics,
predicted on the basis of existing that investigators must tell subjects that
knowledge. participation in the research study is
Uncontrolled studies refer to studies without a voluntary, and that subjects have the right
comparison group. to discontinue participation at any time.
505
Index
a new content areas of definition of 355

absolute risk 300, 301 interest 465–466 future 365
absorption 49 academic detailing 77 implementation 355
academia 73–74, 80 academic medical center 78 initiation 355
academic medical center 78 accelerated approval 93 methodological problems,
assessment of quality and Acceptable Mortality Reporting addressed by research
outcomes of medication, (AMR) 183 357
in populations 76–77 acetaminophen 54, 429 persistence 355
consortia of academic active ingredient(s) 120 phases of 355, 356
medical center programs active surveillance methods taxonomy of 356
for research 78–79 122 techniques for measurements
drug approval process acute myocardial infarction drug concentrations 361
74–75 (AMI) 373 electronic drug monitors
economic assessment of additional risk minimization (EDMs) 361
medication‐related measures (aRMMs) medication diaries 360
issues 78 82, 87 non‐pill formulations
evaluation of drugs, in health adherence 361–362
care system 76 analysis issues in 362 pharmacy dispensing
funding 469–470 clinical trials, data in 362 data 358–360
future 79 comparative effectiveness pill counts 360
interventional studies, data in 362 primary adherence
pharmacoepidemiology prediction, for measurement 361
77–78 interventions 365 self‐reports 357–358
personnel 470–471 selecting adherence ad‐hoc case–control
policy analysis 77 intervals 362–363 studies 210, 212, 469
prescribing practices 75–76 statistical analysis 363–364 ad hoc survey studies
scientific developments time‐varying nature self‐reported diagnosis and
465–469 364–365 hospitalization data
logistical advances trajectory modeling from 230
468–469 364–365 self‐reported drug data
methodologic advances clinical problems, addressed from 228
465–466 by research 356–357 administration errors 440
506 Index
administrative claims data 414 national pharmacovigilance cost‐effectiveness analysis

adverse drug event (ADE) systems 121–123 314–315
436, 438 non‐uniform temporal trends cost identification analysis
adverse drug reactions in reporting 130–131 315
(ADRs) 3, 5, 6, 8, opportunity for public to sensitivity analysis 315
115–117, 121, 281, 282, report 129 analytic studies 31
391 postmarketing adverse event Anatomic Therapeutic
case series and reporting database, signal detect Chemical (ATC)
rates 131 from 124–126 classification system
description of 116–117 reporting ratios 127–128 159, 393, 395
future 132 report quality 129–130 ancestry informative markers
to medical journals 248 review of individual case 271
national and international safety reports of anchor‐based methods 329
postmarketing adverse 126–127 angiotensin‐converting enzyme
event databases 123–124 signal detection 128 2 (ACE2) 188
national pharmacovigilance spontaneous reporting angiotensin converting enzyme
systems 121–123 system 118–119 (ACE) inhibitor 261
non‐uniform temporal trends underreporting 130 anticoagulants 448
in reporting 130–131 adverse events following antidepressant sertraline
opportunity for public to immunizations (Zoloft) 102
report 129 (AEFIs) 403 antiepileptic drugs 350
postmarketing adverse event Affordable Care Act in 2010 antihypertensive
database, signal detect 154 medications 261
from 124–126 Agency for Healthcare Research antimicrobials 237
reporting ratios 127–128 and Quality (AHRQ) anti‐tumor necrosis factor
report quality 129–130 11, 470 therapy 341
review of individual case AIDS Clinical Trials Group apparent volume of
safety reports of (ACTG) adherence distribution 49
126–127 questionnaire 358 applications of
signal detection 128 allele, definition of 258 pharmacoepidemiology
spontaneous reporting allergic/idiosyncratic events benefit–risk (B‐R)
system 118–119 299 assessment 442–454
underreporting 130 allergies 441 advanced methods on
adverse event reporting allergy errors 440 horizon 451–452
systems 412–413 α‐adducin polymorphism 262 appropriate data
adverse events (AEs) 115–117, alpha (α) error 36 identification for 443
193, 248 Alzheimer’s disease 157, 349 BRAT benefit–risk
case series and reporting American Hospital Formulary framework 444
rates 131 Service (AHFS) 157 communicating 444
description of 116–117 aminophylline 4 decision context 445
future 132 analyses of secular trends 27 definition of 442
national and international analysis types, in framework customization
postmarketing adverse pharmacoeconomics 447
event databases cost–benefit analysis 312, identify and define
123–124 314 outcomes 445
Index 507
identify and summarize digital health and patient‐ regression toward

source data 446–447 generated health data mean 398–399
initiatives, guidances and 415 unit of analysis 399
partnerships 451 evidence‐informed practice medication errors 436
integrating benefits and and policy 415–416 clinical problems,
risks 443–444 international addressed by 438–441
patient perspective 443 infrastructure 416 definition of 436
structured approaches mandated post marketing future 442
to 444 studies 413 generalizability 442
systematic approach methodological framework, information bias 441
to 443 for implantable device patient safety concepts
value judgments and patient outcome evaluation 437–438
preferences 447–448 415 safety theory 436–437
weighting and patient methodological problems sample size issues 441–442
perspective 450–451 solved by 410–412 types of 439–441
drug utilization 389–391, outcomes associated pregnancy and children,
397–398 with 409 medications effect in
classification systems public–private 418, 426–427
395–396 partnerships 416 case–control studies 424
clinical problems, registries 413–414 clinical practice, evidence
addressed by research signal detection/outlier to 419
391–392 identification 413 clinical problems,
current data sources translational addressed by research
393–394 epidemiology 416 418–419
definition of 389 evaluating and improving confounding 421–422
future 396 prescribing 398 exposure ascertainment,
methodological problems, clinical problems, timing, and
addressed by research addressed by research misclassification
392–393 398 420–421
operational concepts 391 conceptual future 424–425
units of measurement framework 399–400 methodological problems,
394–395 empirical evidence on solved by research
epidemiologic studies of effectiveness of 419–423
implantable medical interventions 400 newer designs 424
devices 408–409, ethical and legal problems outcome definition and
417–418 hindering ascertainment 421
administrative claims implementation of population, definition
data 414 RCT 399 of 419–420
adverse event reporting future 402 prospective cohorts 423
systems 412–413 internal validity 398 registries 423
automated surveillance 413 methodological problems, retrospective cohorts and
clinical problems 409–410 addressed by research nested case–control
Coordinated Registry 398–399 studies within
Networks (CRNs) patient outcomes, detecting automated health care
414–415 effects on 399 databases 424
508 Index
applications of assessment of causality b

pharmacoepidemiology 404 barbiturates 51, 56
(cont’d) clinical problems, Bar Code Medication
sample size requirements addressed by research Administration (BCMA)
and challenges 420 403–404 database 183
selection bias 422–423 confounding and bias baseline, definition of 56
treatment responses and 406–407 Bayesian approaches 251–252,
patterns 419 epidemiologic studies 271, 341, 344, 347
unique biology and 407–408 Bayesian Confidence
epidemiology 418–419 future 408 Propagation Neural
risk management measurement accuracy Network (BCPNN) 125
activities 430 406 Bayesian “piece‐wise”
additional risk methodological problems, exponential model 341
minimization addressed by research Bayh–Dole Act 308
strategies 431–432 404–407 behavioral economics
applies to all sample size 406 310–311
medicines 429 signal detection 404, 407 beneficial effects 212
clinical problems, Arrhenius Equation 104 benefit pool 318
addressed by arrhythmias 7 Benefit–Risk Action Team
research 428–433 artifactual association 22 (BRAT) 444
complexities of medication artificial intelligence 409 benefit–risk (B‐R) assessment
use system 428 Asia, encounter databases in 442–454
definition of 427 159–160, 170–171 advanced methods on
evaluation of risk Health Insurance Review and horizon 451–452
minimization and Assessment (HIRA) appropriate data
mitigation Data, South Korean identification for 443
measures 432–433 160–161 BRAT benefit–risk
future 434–435 Taiwanese National Health framework 444
pharmacoepidemiology Insurance Research communicating 444
role in 433–435 Database (NHIRD) 160 decision context 445
proactive process 429 Asian Pharmacoepidemiology definition of 442
product’s lifecycle, Network (AsPEN) 160, framework customization 447
proceeds throughout 283 identify and define
429 ATP‐binding cassette outcomes 445
risk assessment 430 transporter proteins identify and summarize
risk communication 431 (ABCB) 261 source data 446–447
risk minimization atrial fibrillation 76, 445, 446, initiatives, guidances and
430–431 449 partnerships 451
risk mitigation 430–431 attention deficit hyperactivity integrating benefits and
scientifically driven disorder (ADHD) 165 risks 443–444
428–429 automated databases see patient perspective 443
sources of risk from Electronic databases structured approaches to
medical products 428 automated healthcare 444
vaccine safety, methodological databases 167 systematic approach to
issues 403, 405–406 automated surveillance 413 443
Index 509
value judgments and patient protections against conflicts cardiovascular events and
preferences 447–448 of interest, for drug rofecoxib, risk of 348
weighting and patient industry‐sponsored care‐gap 398
perspective 450–451 research 288–289 case‐by‐case causality
benzodiazepines 157 biomarkers 257, 264, 416, 468 assessment 248, 249
beta‐blockers 75, 278, 347 European cohort study, in case–cohort design 373–374
beta (β) error 36 type 2 diabetes 200 case–control studies 28, 136,
bias 22, 226, 368, 406–407 biopharmaceutical risk 198, 203, 209, 213, 262,
in abstraction of data 338 management 81 406, 422, 424
susceptibility of original evolution of 81–83 gene‐exposure interaction
studies to 336 Biosimilar User Fee Act analysis in 268
big data 280, 290 (BsUFA) 10 modification in molecular
bioavailability 49, 50 biotransformation reactions pharmacoepidemiologic
biobanks 197 51, 56 studies 266
bioethical issues 276, Bioventus Exogen® device sample size, calculations
291–292 registry 199 for 40–42, 44, 45,
clinical problems, by research body composition 54–55 485–491
East Asian Boston Collaborative Drug case‐crossover study 375–376
pharmacoepidemiologic Surveillance case definition centers 126
trends and ethics 282–283 Program 10, 278 case reports 26
emergence, changing bradykinin B2 receptor 261 case reports, causality
methods, and moral bretylium 65 assessment from 246
stakes of Brief Medication adverse drug reactions to
pharmacoepidemiology, Questionnaire 358 medical journals
in Twentieth Century British National Formulary 248
North America 276–280 (BNF) codes 182, 395 appropriate approach
European choosing 253–254
pharmacoepidemiologic c calibration 253
trends and ethics calendar time 370, 371 clinical practice and
280–282 calibration 253 prescription 248
future 289–291 Canada, encounter databases clinical problems, by
Good Pharmacoepidemiology in 156–158, 169–170 research 246
Practice (GPP) Canadian Institute for Health clinical trials 248
286–288 Information (CIHI) future 254–255
Good Pharmacovigilance 157, 158 hypothesis generation and
Practices 286–288 Canadian Network for research 248
methodologic problems, by Observational Drug from individual cases 249
research 283–284 Effect Studies expert judgment/global
conflicts of interest for (CNODES) 167 introspection 249
drug industry cancer registries 196 probabilistic approaches
research 286 cancer targeted therapies 264 251–253
ethical benefits, for data candidate gene association structured guidelines and
integrity 285–286 studies 269 algorithms 250–251
ethics of surveillance 285 carbamezapine 56 methodological problems, by
informed consent 284–285 cardiac dysfunction 56 research 248–249
510 Index
case reports, causality assessment Centers for Disease Control and unique biology and
from (cont’d) Prevention’s (CDC’s) epidemiology 418–419
paradigms of 246–247 Vaccine Safety Datalink Children’s Health Insurance
spontaneous reporting (VSD) 79 Program (CHIP) 154
247–248 Centers for Education and China Food and Drug
case series 26–27 Research on Therapeutics Administration (CFDA)
sample size, calculations (CERTs) 11, 470 14
for 41, 43 Centers for Medicare and cholecystectomy 232
case–time–control Medicaid Services chronic drug therapy 393
design 376–377 (CMS) 149 chronic obstructive pulmonary
causality assessment from case chemotherapy 429 disease (COPD) 373
reports 246, 404 children, medications effect churning 151
adverse drug reactions to in 418, 426–427 cimetidine 27
medical journals 248 case–control studies 424 ciprofloxacin 56
appropriate approach clinical practice, evidence circulatory failure 56
choosing 253–254 to 419 claims databases 137–139
calibration 253 clinical problems, addressed classic randomized controlled
clinical practice and by research 418–419 trials (RCTs) 296–298
prescription 248 confounding 421–422 control treatment, choice
clinical problems, by exposure ascertainment, of 297
research 246 timing, and data analysis 297
clinical trials 248 misclassification data collection 297
future 254–255 420–421 generalizability of
hypothesis generation and future 424–425 results 297–298
research 248 methodological problems, masking treatment
from individual cases 249 solved by research assignment 297
expert judgment/global 419–423 sample size 297
introspection 249 newer designs 424 clinical assessments 193
probabilistic approaches outcome definition and clinical care, economic
251–253 ascertainment 421 evaluation of 312
structured guidelines population, definition of clinical decision support
and algorithms 419–420 systems (CDSS) 437
250–251 prospective cohorts 423 clinical pharmacology 47, 257
methodological problems, by registries 423 basics of 48–49
research 248–249 retrospective cohorts and drug interactions 56
paradigms of 246–247 nested case–control model‐informed drug
spontaneous reporting studies within development 59–60
247–248 automated health care pharmacodynamics 56–59
causal pathway of disease databases 424 vs. pharmacoepidemiology
261–263 sample size requirements and 4–5, 48
Center for Drug Administration challenges 420 pharmacogenomics 59
in Singapore 14 selection bias 422–423 pharmacokinetics
Centers for Disease Control and treatment responses and absorption 49
Prevention (CDC) 407 patterns 419 definition of 49
Index 511
elimination 51–52 sample size, calculations comparable patients,

metabolism 49–51 for 35–40, 476–484 analysis of 379–380
volume of distribution commercial insurance propensity score
49, 50 databases 151, 152 analysis 377–378
special populations 52 Committee for Medicinal proxy adjustment 380
body composition 54–55 Products for Human Use sensitivity analyses
elderly 52–53 (CHMP) 15 381–382
gastrointestinal 54 Comparative effectiveness treatment choice via
hepatic 54 research (CER) instrumental variable
organ impairment 55–56 12, 13, 79 analysis, random aspects
pediatrics 53–55 comparative effectiveness in 380–381
pregnancy 55 studies, data in 362 unobserved patient
renal 53–54 comparator group, choice of characteristics and
clinical practice and 379 residual confounding
prescription 248 completeness of drug data 237 380
Clinical Practice Research compliance, definition of clinical problems, addressed
Datalink (CPRD) 181, 355 by research 368
183 computerized decision support efficient sampling designs
clinical trials 248, 287 systems (CDSS) 400 within cohort study 370
adherence, data in 362 Computerized Online Medicaid case–cohort
patient reported outcomes Analysis and design 373–374
in 323 Surveillance System multi‐time case–control
Clinical Trials Regulation (COMPASS®) 11 design 372–373
(CTR) 287 computerized physician order nested case–control design
clinician or site‐reported entry (CPOE) 437 372
outcomes (ClinROs) concordance/agreement 222 prevalent new‐user cohort
195–196 confidence intervals 29 designs 374–375
clioquinol 278 conflicts of interest for drug structures of Cohorts
clopidogrel 78 industry research 286 370–372
code of conduct 79 confounding 406–407, within‐subject designs
cohort study 28–30, 209 421–422 375–377
efficient sampling designs controlling. See Controlling methodological problems,
within 370 confounding addressed by research
case–cohort design 373–374 by indication 296 368–370
multi‐time case–control confounding variable 22, 23 cooperative population 302
design 372–373 contract law 102–104 Coordinated Registry Networks
nested case–control contract research 102, 103 (CRNs) 414–415
design 372 controlled distribution copy number variants (CNV)
prevalent new‐user cohort products 195 258
designs 374–375 controlling confounding 210, Corporate Data Warehouse
structures of Cohorts 368, 384, 385 (CDW) 183
370–372 analytic approaches correlation coefficient 224
within‐subject designs balancing patient corticosteroid 78
375–377 characteristics 377 cost and clinical outcomes 317
512 Index
cost–benefit analysis 312, 314 in randomized controlled diethylstilbestrol (DES) 278

cost‐effectiveness analysis trials (RCTs) 297, 300 differential opportunity for
314–315, 318 data elements 299 exposure 420
cost identification analysis Data Extraction and digital health 415
315 Longitudinal Time digoxin 278
cost types 315 Analysis (DELTA) 413 direct effect 57
direct medical costs 315 data mining 125, 407, 466 Directive and Regulation 82
direct nonmedical costs signals 131 direct medical costs 315
315–316 data on outpatient direct nonmedical costs
intangible costs 316 diagnoses 210 315–316
productivity costs 316 data quality direct‐to‐patient data collection
cough 261 Europe 183–184 approaches 197
Council for International US/VA 184 Discharge Abstract Database
Organizations of Data Safety Monitoring (DAD) 157, 158
Medical Sciences Boards 286 distributed data systems, special
(CIOMS) 334 days of therapy (DOT) 395 considerations with
COVID‐19 pandemic 161, defined daily dose (DDD) 393, 239
186, 404 394 distribution‐based
COX‐2 inhibitors 472 degrees of safety 67 methods 329
criteria for the causal nature delayed drug effect 214 disulfiram 51
of an association Delphi Method 249 dose–response effects 24, 339
23–26 dementia 165 dosing errors 439
cross‐sectional studies 31 de novo questionnaire dread diseases 68
cumulative meta‐analysis 346, development 233–235 drug approval process 74–75
348–349 deoxyribonucleic acid (DNA) in other countries 14–15
Current Procedural Terminology 258 in US 13–14
(CPT) 151, 239 Department of Veterans Affairs drug concentrations 361
cyclosporine 58 (DVA) 159 drug crises
CYP2D6 260, 261 Department of Veterans Affairs intellectual development of,
cytochrome P450 2C9 variant Healthcare System emerging from 10–13
(CYP2C9) 264 (VA) 153 legislative actions resulting
cytochrome P450 system 50 depression diagnosis 234 from 9–10
DerSimonian and Laird (DL) and resulting regulatory
d methodology 341 actions 7–8
dabigatran 77 descriptive studies 31 drug data, special considerations
DAS‐28 (disease activity score) design defect theory 96, 98 with 237
assessment 329 desirability of outcome ranking drug development
data access 148–149 (DOOR) 452 economics of 307–308
data analysis, in RCT 297 Developing Evidence to Inform planning 92–93
database population, Decisions about drug–disease associations 228
characteristics of 145 Effectiveness (DEcIDE) drug–drug interaction
database size 164–165 Network 11 exposures 440
data collection device variation 415 drug‐event association 247
for meta‐analysis 339–340 Diagnosis Related Group drug exposure 224
via questionnaire 233 (DRG) 138 drug formularies 152
Index 513
drug‐illness relations 279 duration–response health care systems and

drug‐induced birth defects relationship 24 populations 174, 181
425–426 Dutch PHARMO system 148 future 186–187
Drug Information Number dwell time 146 limitations
(DIN) 157 clinical data,
drug interactions 56 e incompleteness of 185
drug–laboratory errors 440 East Asian drug data, incompleteness
drug lag 7, 9 pharmacoepidemiologic of 185–186
drug regulation 14 trends and overview of 175
and product liability ethics 282–283 in Spain 188–189
litigation 101–102 ecological studies 27 strengths
drug regulators 466 economic data 317–320 access original health
drug‐resistant epilepsy 261 economics of drug records, ability to 184–185
drug safety during pregnancy, development 307–308 accuracy of drug
FDA categories of 55 efficacy 56, 76 information 184
drug statistics 390 electronic length of follow‐up 184
drug supply diary 358 databases 136–137 linkage to external
drug toxicity 63 applications 140 patient‐level data 185
drug use evaluation (DUE) claims and other population‐based
studies 390 administrative data 184
drug use results surveys databases 137–138 sample size 184
(DURSs) 82 electronic health record validity of clinical
drug utilization 389–391, databases 138 information 184
397–398, 400, 419, 465 future 140–141 US/VA
classification systems 395–396 strengths 138–139 data access for
clinical problems, addressed weakness 139–140 researchers 184
by research 391–392 electronic drug monitors databases 182
current data sources (EDM) 358, 361 data collection and
393–394 electronic encounter structure 183
definition of 389 databases 236–237 data quality 184
future 396 electronic healthcare health care systems and
methodological problems, databases 380 populations 181
addressed by electronic health record (EHR) Elements to Assure Safe Use
research 392–393 databases 138, 142, (ETASU) 87
operational concepts 391 174, 221, 222, 236, elimination 51–52
units of measurement 438 eltrombopag 195
394–395 for epidemiologic Empirical Bayes method 125
drug utilization research research 176–180 ENCePP. See European
(DUR) 390, 392, 393, Europe Network of Centres for
396 data access for Pharmacoepidemiology
drug utilization review researchers 184 and Pharmacovigilance
studies 390 databases 181–182 (ENCePP)
Drug Safety Research Unit data collection and encounter databases 142–145
(DSRU) model 288, structure 182–183 access 148–149
289 data quality 183–184 applications 163
514 Index
encounter databases (cont’d) US private insurance public–private

database size 164–165 databases 151–153 partnerships 416
deciding between Eosinophilia‐Myalgia registries 413–414
individual 164–166 syndrome 7 signal detection/outlier
non‐encounter data, ability epidemiologic studies of identification 413
to supplement implantable medical translational
with 165–166 devices 408–409, epidemiology 416
non‐standard encounter 417–418 epidemiologic study
data, availability of 165 administrative claims designs 30
target population 164 data 414 advantages 25
typical activities involved adverse event reporting analyses of secular trends 27
in studies 163–164 systems 412–413 case–control studies 28
validate outcomes, ability automated surveillance 413 case reports 26
to 165 clinical problems 409–410 case series 26–27
in Asia 159–160 long‐term safety and cohort studies 28–30
Health Insurance Review effectiveness 410 disadvantages 25
and Assessment (HIRA) real‐word setting, benefits randomized clinical
Data, South and harms profile in trials 30–31
Korean 160–161 410 epidemiology 81, 257,
Taiwanese National Health Coordinated Registry 418–419
Insurance Research Networks in drug safety
Database (NHIRD) 160 (CRNs) 414–415 evaluation 83–86
attributes of 145 digital health and patient‐ in evaluation of risk
in Canada 156–158 generated health mitigation
core data domains 143, 144 data 415 interventions 86–87
in Europe evidence‐informed practice vs. pharmacoepidemiology 5
Nordic prescription and policy 415–416 post‐approval, contributions
databases 158–159 international infrastructure of 85–86
SNIIRAM 159 416 risk management and,
future 166–167 mandated post marketing regulatory on 81
limitations 162–163 studies 413 epigenetics 258
non‐encounter data, linkage methodological framework, epilepsy 84
to 147–148 for implantable device equipoise 300
non‐population‐based outcome evaluation 415 Eritrean Pharmacovigilance
studies 143 methodological problems Centre 132
population and coverage solved by 410–412 ethics of surveillance 285
period 145–146 comparative studies 411 eudravigilance 115, 281
population‐based ensuring comparability of EU pharmacovigilance
studies 143 study groups 412 legislation 82
purpose of 143, 144 individual patient exposure EuroBioBank network 197
selection of 149, 150 assessment 411 EUROCARE1–5 199
services covered and data national population Europe
completeness 146–147 exposure assessment 411 electronic health record
strengths 161–162 sample size and real world (EHR) databases
in United States 149, 151 performance data access for researchers
US government 153–156 issues 412 184
Index 515
databases 181–182 exome sequencing 272 gene–drug interactions

data collection and expected beneficial effects 261–263
structure 182–183 212 generalizability 412, 442
data quality 183–184 exposure‐measurement error General Practice Research
health care systems and (EME) 225 Database (GPRD) 181,
populations 174, 181 exposure variation 369 394
encounter databases in external validity 412 general practitioners
Nordic prescription (GPs) 394
databases 158–159 f genes affect drug response
SNIIRAM 159 factor V Leiden mutation 263 260–263
European Cancer Registry failure‐to‐warn claims 98–101 genetic variability 258–259,
Based Study On Survival false discovery rate (FDR) 271 262
And Care Of Cancer FDA Adverse Event Reporting genome‐wide approach 259
Patients (EUROCARE) System (FAERS) 115 genome‐wide association
registry 199 FDA Reauthorization Act of studies (GWAS) 261,
European cohort study of 2017 (FDARA) 10 269
biomarkers in type 2 federated virtual database 143 genotype 258
diabetes 200 fixed‐effect models 340, 349 glucuronidation of
European Economic Area fluconazole 56 morphine 54
(EEA) 280 fluoxetine 7 Good Epidemiology Practice
European Medicines Agency follow‐up data collection 303 (GEP) guidelines 288
(EMA) 15, 75, 194, follow‐up‐time cohorts 371 Good Pharmacoepidemiology
280, 431 Food and Drug Administration Practice
European Medicines Evaluation Amendment Act (GPP) 286–288
Agency (EMEA) 15 (FDAAA) in 2007 9, Good Pharmacovigilance
European Medicines Regulatory 15, 82, 86, 280 Practices 82, 286–288
Network (EMRN) 280 1938 Food, Drug, and Cosmetic Good Research for
European Network of Act 6 Comparative
Centres for formal logic 250 Effectiveness (GRACE)
Pharmacoepidemiology frequency errors 440 principles 197
and Pharmacovigilance future of graphical methods for
(ENCePP) 78, 79, pharmacoepidemiology meta‐analysis 342
281–282, 287, 288, 337 464–465 group education methods 400
European Patent Convention academia Guillain‐Barre
(EPC) 105 funding 469–470 Syndrome 407
European pharmacoepidemiologic personnel 470–471 gut microbiome 258
trends and scientific developments
ethics 280–282 465–469 h
evaluation of drugs, in health industry 471–472 half‐life of elimination 52
care system 76 law 473 haplotypes 259
evidence‐based medicine regulatory agencies Hartung‐Knapp‐Sidik‐Jonkman
(EBM) 335, 365 472–473 (HKSJ) method 341
evidence‐informed practice and Healthcare Common Procedure
policy 415–416 g Coding System
excess risk 30 gag clauses 103 (HCPCS) 151
exclusion criteria 126, 339 Gamma Poisson Shrinker 125 healthcare data linkages 148
516 Index
health care provider (HCP) human immunodeficiency mandated post marketing

experience 192 virus 92 studies 413
Health Care System Research human insulin 7 methodological framework,
Network (HCSRN) 153 hybrid/enriched designs 195 for implantable device
health economics and health hydroxychloroquine 342, 343 outcome evaluation
care financing hypertension 78 415
behavioral economics hypoalbuminemia 55 methodological problems
310–311 hypothesis‐generating solved by 410–412
economic evaluation of studies 66–67, 211 comparative studies 411
pharmaceuticals hypothesis generation and ensuring comparability of
311–312 research 248 study groups 412
health insurance 308–309 hypothesis‐strengthening individual patient exposure
moral hazard 309–310 studies 211 assessment 411
health information technology hypothesis‐testing studies 66, national population
(IT) 415 212 exposure
health insurance 308–309 hysteresis 58 assessment 411
Health Insurance Portability sample size and real world
and Accountability Act i performance issues
(HIPAA) 151, 290 Ibuprofen 294, 295 412
Health Insurance Review and Illumina’s MiSeqDx 270 outcomes associated with 409
Assessment Service Imatinib 262, 264 public–private partnerships
(HIRA), South implantable medical devices, 416
Korean 160–161 epidemiologic studies registries 413–414
health maintenance of 408–409, 417–418 signal detection/outlier
organizations administrative claims data identification 413
(HMOs) 239 414 translational
healthy user effect 362 adverse event reporting epidemiology 416
heparin 278 systems 412–413 important confounders 213
hepatic biotransformation automated surveillance 413 incidence of disease 212
reactions 54 clinical problems 409–410 incident users, limitations of
hepatic dysfunction 55–56 long‐term safety and 379
heterogeneity 341–343 effectiveness 410 inclusion criteria 339
heterozygous 258 real‐word setting, benefits Indian Pharmacopoeia
high‐dimensional PS and harms profile in Commission (IPC) 123
approach 380 410 indication bias 296
high‐throughput genotyping Coordinated Registry indirect costs 316
technologies 264 Networks (CRNs) indirect effect 57
Hill coefficient 57 414–415 indirect‐response models
HMO Research Network 153 digital health and patient‐ 58
homogeneity 346 generated health individual cases, causality
homozygous 258 data 415 assessment from 249
hospital‐based case–control evidence‐informed practice expert judgment/global
studies 209 and policy 415–416 introspection 249
hospitalization databases 156, international infrastructure probabilistic approaches
157 416 251–253
Index 517
structured guidelines and instrument strength 381 International Prospective

algorithms 250–251 insurance, health 308–309 Register Of systematic
individual‐level dispensing intangible costs 316 reviews (PROSPERO)
databases 393 integrated database 143 339
individual patient data (IPD) integrated healthcare International Society for
338 delivery system Pharmacoepidemiology
individual patient exposure databases 153 (ISPE) 464
assessment 411 Integrated Primary Care interventional
industry 79, 87–90, 471–472 Information (IPCI) pharmacoepidemiology
biopharmaceutical risk 182 77–78
management, evolution intellectual property interviewer bias 22
of 81–83 law 103–105 intraclass correlation
epidemiology intensive monitoring 83 coefficient 225
in drug safety evaluation internal validity 398 intraocular pressure (IOP) 198
83–86 International Agranulocytosis Investigational New Drug (IND)
in evaluation of risk and Aplastic Anemia application 6
mitigation interventions Study (IAAAS) 373 in vitro testing 83
86–87 International Classification of in vivo testing 83
post‐approval, contributions Diseases, Clinical Iowa Drug Information System
of 85–86 Modification (ICD‐CM) (IDIS) 395
risk management and, system 151 IQVIA Disease Analyzer
regulatory on 81 International Classification of databases 181
information bias 441 Diseases, 9th edition IQVIA Medical Research Data
Information Component (ICD‐9) 182 (IMRD) 181
(IC) 131 International Classification of IQVIA Private Drug Plan
informed consent 279, Diseases‐Ninth Database 158
284–285 Revision‐Clinical isoniazid 56
Innovative Medicines Initiative Modification (ICD‐9‐ Italy’s Health Search
(IMI) 121, 451 CM) codes 138 Longitudinal Patient
inpatient drug use 210, 213 International Classification of Database 181
Institute for Clinical and Disease (ICD) system item response theory (IRT)
Economic Review 237, 395 models 322, 329–330
(ICER) 312 International Classification of bookmarking and scale
Institute for Quality and Primary Care (ICPC) judgement
Efficiency in Health 182 of 329–330
Care (IQWiG) 312 International Coalition of
Institute of Medicine Breast Registries k
(IOM) 78, 81, 82, 404 Associations (I‐COBRA) Kaiser Permanente 239, 279
Institutional Claims 155 416 Kaiser programs 209
institutional review board (IRB) International Consortium of kappa (k) coefficient 224, 225,
184 Vascular Registries 231
instrumental variable (IV) (ICVR) 416 Ketoconazole 51
analyses 369, 380–381 International Network for Korea Food and Drug
instrumental variable‐based Rational Use of Drugs Administration
methods 412 (INRUD) 394 (KFDA) 282
518 Index
l Manufacturer and user facility Medicare Benefits Schedule

large simple trial (LST) 298, device experience (MBS) 159
299 (maude) database 412 Medicare Prescription Drug,
conditions appropriate for manufacturing defect 98 Improvement, and
300 marketing authorization Modernization Act 11
design characteristics of 299 application (MAA) 82 medication adherence. See
feasible 301 Master Beneficiary Summary Adherence
logistics of conducting 302 Files (MBSFs) 155 medication class, influence
Learning Healthcare measurement error 225–226 of 228
System 78 adjusting measures of medication diaries 360
legal system 105–107 association for 227–228 medication errors 436
contract law 102–104 indices of 222 clinical problems, addressed
drug regulation and product medicaid 64, 137, 149, 153–154, by 438–441
liability litigation 162, 164, 208–210, 213, definition of 436
101–102 214, 280, 318 future 442
failure‐to‐warn claims Medicaid Statistical generalizability 442
98–101 Information System information bias 441
intellectual property law (MSIS) 154 patient safety
103–105 medical certainty 101 concepts 437–438
pharmacoepidemiologic medical condition type, safety theory 436–437
expertise and influence of 230 sample size issues
Daubert 101 medical device 441–442
product liability law 98 definition of 409 types of 439–441
tort law and product liability studies of 194–195 medication possession ratio
lawsuits 98 Medical Device Epidemiology (MPR) 359
legal theory of negligence 65 Network (MDEpiNet) Medicines Agency in
Likelihood Ratios (LRs) 252 414 Denmark 14
linkage disequilibrium Medical Devices Agency Medicines Agency in
(LD) 261 (MDA) 15 Norway 14
literature data 250 Medical Dictionary for Medicines and Healthcare
loss to follow‐up 211 Regulatory Activities Products Regulatory
low incidence outcome 212 (MedDRA) 124 Agency (MHRA) 15
low prevalence exposure 213 Medical Information Database Medicines Control Agency
NETwork (MCA) 15
m (MID‐NET®) 96 Medicines & Medical Devices
machine learning 365, 412 Medical Information for Risk Safety Authority in New
major adverse cardiac events Assessment Initiative Zealand 14
(MACE) 349 (MIHARI) 82, 282 Mediplus 181
major birth defects 421 medical knowledge 250 MedWatch 115
managed care organizations Medical Product Safety meta‐analysis 74, 334–335
(MCOs) 407 Network bias in abstraction of data
Mandated post marketing (MedSun) 413 338
studies 413 medicare 149, 153, 155, 196, case studies of applications
Mantel–haenszel technique 213, 280, 318, 469 cumulative
343, 349, 375 medicare beneficiaries 155 meta‐analysis 348–349
Index 519
network meta‐analysis and microbiome 257, 466 pharmacogenetics 259–260

simultaneous evaluation microsatellite instability‐high pharmacogenomics
of treatment therapies (MSI‐H) 264 259–260
349–350 minimal clinically important progression and clinical
regulators’ role 350 difference (MCID) 327, application 264
saving time and 329 type I error and
resources 346–348 Minimum Data Set (MDS) 166 replication 270–271
clinical problems, addressed Mini‐Sentinel 211 type II error 271
by research 335–336 mismatch repair deficient monitoring post‐approval
combinability of studies (dMMR) 264 safety 94–95
336–337 mixed treatment comparisons moral hazard 309–310
future 350–351 336 Morisky Medication Adherence
indirect and simultaneous mobile health (mHealth) 415 Questionnaire
comparison of model‐informed drug (MMAS‐8) 358
treatments, for specific development multi‐criteria decision analysis
conditions 346 (MIDD) 59–60 (MCDA) 85, 444
methodological problems, modified Schlesselman multidrug‐resistance (MDR)‐1
solved by research formula 38 gene 261
336–338 molecular genetic multi‐item Gamma Poisson
publication bias 337, markers 466 Shrinker 125
344–346 molecular multi‐time case–control
steps in performing 338 pharmacoepidemiology design 372–373
data collections 339–340 257–258 myocardial infarction 126, 347
establish inclusion/ clinical problems, by
exclusion criteria 339 research 260 n
formulate conclusions and genes affect drug N‐acetyl‐p‐enzoquinone imine
recommendations 344 response 260–263 (NAPQI) 51
literature search 338–339 interplay of various naloxone 26
purpose defining 338 mechanisms 263–264 national and international
statistical confounding by population postmarketing adverse
analyses 340–344 admixture 271 event databases
susceptibility of original drug response, identifying 123–124
studies to bias 336 additional genetic National Cardiovascular Data
metabolism 49–51 contributions to Registry (NCDR) 194,
metastatic breast cancer, 269–270 195
high‐dose chemotherapy future 271–273 National Death Index 303
plus autologous stem genetic variability 258–259 National Drug Code
cell transplantation interactions 270 (NDC) 151, 411
for 313 methodological problems, by National Health Insurance
metiamide 27 research 264–266 (NHI) program 160
“me too” drugs 66 confounding by population National Health Insurance
metronidazole 56 admixture 268–269 Research Database
mibefradil 85 interactions 266–267 (NHIRD), Taiwanese 160
Michaelis Menten constant type I error 267 National Health Insurance
49, 50 type II error 267–268 Service (NHIS) 160
520 Index
National Institute for Health number needed to treat patient population, influence
and Care Excellence (NNT) 449 of 229–230, 232
(NICE) 312 number of prescriptions 392 patient preference studies 448
national pharmacovigilance patient reported outcomes
systems 121–123 o (PROs) 322
national population exposure Observational Health Data barriers to measurement
assessment 411 Sciences and Informatics 327–328
negative predictive values (OHDSI) 351 characterization of 193
(NPV) 223, 238, 240 observational studies 297 clinical problems, addressed
nested case–control observational study designs 31 by research 326–328
studies 372 oculomucocutaneous in clinical trials 323
within automated health care syndrome 278 completion and results
databases 424 odds ratio (OR) 29, 226, 339, review 326
net clinical benefit (NCB) 340 discordance in perspectives
measures 450 one size fits all approach 408 between patients,
network meta‐analysis one‐step model. See Peto clinicians and
349–350 method researchers 328–329
next‐generation sequencing oophorectomy 232 future 330–331
(NGS) techniques 266, open‐ended medication to guide interventions
272 questions 229 327–328
N‐of‐1 trial 255, 327 Open Trials project 346 individualized assessment
non‐adherence 356, 363 organic acids 51, 56 and treatment plans 327
non‐encounter data, linkage organ impairment 55–56 measuring within person
to 147–148 orphan drugs 91–92 change 329–330
nonexperimental study orphan genotypes 273 methodologic problems,
designs 31 outcome assessment/ solved by research 328
non‐interventional/non‐ adjudication 192–193 to new therapeutic strategies
experimental 287 over‐the‐counter (OTC) development 325–326
non‐interventional studies drugs 144, 146, 158, to optimize treatment in
(NIS) 281 186, 228, 393, 429 clinical care 324
non‐pill formulations 361–362 in routine care 323, 325
non‐population‐based p selection of patient reported
studies 143 parenteral ketorolac 65 outcomes and
non‐reimbursable drugs 163 Parkinson disease 126 implementation into
nonsteroidal anti‐inflammatory parmacovigilance plan 427 practice 330
drugs (NSAIDs) 41, passive reporting systems selection, score interpretation
130, 167, 215, 221, 228, 412 and interventions
237, 239, 338 Patient Centered Outcomes 326–327
nordic prescription Research Institute patient safety concepts
databases 148, 158–159 (PCORI) 79, 328, 470 437–438
not‐for‐profit healthcare patient characteristics, patient variation 415
delivery systems 151 matching on 379–380 PCSK‐9 inhibitors 319
Novartis’s CAR‐T therapy 308 patient engagement 322, 327 P‐D‐E system 278
number needed to harm patient‐generated health data Pediatric Investigation Plans
(NNH) 449 196, 415 (PIP) 81
Index 521
pembrolizumab 264 behavioral economics in other countries 14–15

penicillin G 54 310–311 in US 13–14
peripheral edema 56 economic evaluation of definition of 3–4
Peto method 341, 343, 349 pharmaceuticals drug crises
Pharmaceutical Benefits 311–312 intellectual development
Scheme (PBS) 159 health insurance of, emerging from 10–13
Pharmaceuticals and Medical 308–309 legislative actions resulting
Devices Agency moral hazard 309–310 from 9–10
(PMDA) 82, 282 high‐dose chemotherapy plus and resulting regulatory
pharmaceuticals, economic autologous stem cell actions 7–8
evaluation of transplantation for early legislation 5–7
311–312 metastatic breast vs. epidemiology 5
Pharmaceuticals, Medical cancer 313 historical background 5–13
Devices and Other methodological problems, by legal prophylaxis 65–66
Therapeutics Products research 312–320 marketing 64–65
Act (PMD Act) 93 Novel SARS‐CoV‐2 Vaccine, potential contributions of
pharmacodynamic gene–drug economic evaluation of 15–18
interactions 261 314 general 17–18
pharmacodynamics 4, 47, perspective of analysis 316 new types of information
56–59 pharmacoenvironmentology 17
pharmacoeconomics 290 supplementary
analysis types pharmacoepidemiology information 16–17
cost–benefit analysis 312, available approaches to regulatory 63–64
314 studies 203–208 safety vs. risk 67
cost‐effectiveness analysis characteristics of research pharmacogenetics 259–260
314–315 questions and impact, Pharmacogenetics Research
cost identification analysis on data resources Network (PGRN) 470
315 206–207, 211–215 pharmacogenomics 59, 259–260
sensitivity analysis 315 comparative characteristics pharmacokinetic gene–drug
clinical problems, by of data resources interactions 260–261
research 307–312 204–205, 208–211 pharmacokinetics 4, 13, 47, 419
cost types 315 example of 215–216 absorption 49
direct medical costs 315 nonsteroidal anti‐ definition of 49
direct nonmedical costs inflammatory drugs elimination 51–52
315–316 (NSAIDs) 215 metabolism 49–51
intangible costs 316 benefits of study 62–63 volume of distribution 49,
productivity costs 316 clinical 50
economic data 317–320 hypothesis pharmacology 47
economic evaluation of generating 66–67 pharmacotherapy 48
clinical care 312 hypothesis testing 66 pharmacovigilance 104, 115
economics of drug vs. clinical pharmacology Pharmacovigilance and Risk
development 307–308 4–5, 48 Assessment Committee
future 320 costs of study 62 (PRAC) 82
health economics and health current drug approval Pharmacovigilance research
care financing process network (PVNet) 282
522 Index
Pharmacovigilance system post‐authorization safety opportunity for public to

master file (PSMF) 82 studies (PASS) 82, 287 report AEs/ADRs 129
Pharmacy‐based medical record post‐authorization study scope 129
linkage systems 208, (PAS) 282 signal detection 128
212, 214 posterior probability 252 post‐marketing surveillance
Pharmacy Benefit Management post‐marketing commitments (PMS) 82, 279
(PBM) database 183, (PMC) 82 pragmatic clinical trial
184 postmarketing safety reporting (PCT) 86
pharmacy dispensing data system 128, 129 pragmatic trials 298
358–360 postmarketing spontaneous prazosin 16, 27, 66
PharmaNet 157 pharmacovigilance Preferred Reporting Items for
“Phase IV” postmarketing reporting systems Systematic Reviews and
surveillance study 27 applications Meta‐analyses
phenobarbital 54 blindness and retinal (PRISMA) 336
phenytoin 51 disorder, associated with pregnancy, medications effect
philadelphia chromosome‐ clomifene citrate 132 in 418, 426–427
positive leukemias 262 case series and reporting case–control studies 424
PhRMA Benefit‐Risk Action rates 131 clinical practice, evidence
Team (BRAT) data mining signals 131 to 419
framework 85 signals from developing clinical problems, addressed
pill counts 360 countries 131–132 by research 418–419
pill fatigue 356 future 132 confounding 421–422
pioglitazone 379 individual case safety reports, exposure ascertainment,
piroxicam (Feldene) 65 review of 126–127 timing, and
P450 isoforms 55 limitations misclassification
Poisson‐distributed variable, non‐uniform temporal 420–421
confidence limit factors trends in reporting future 424–425
for estimation 492 130–131 methodological problems,
Poisson distribution 43 report quality 129–130 solved by
policy analysis 77 underreporting 130 research 419–423
polychlorinated biphenyls national and international newer designs 424
(PCBs) 101 postmarketing adverse outcome definition and
population and coverage event databases ascertainment 421
period 145–146 123–124 population, definition of
population‐based databases national pharmacovigilance 419–420
143, 161, 184, 209 systems 121–123 prospective cohorts 423
population, definition of overview of 117–118 registries 423
419–420 postmarketing adverse event retrospective cohorts and
population health 166 database, signal detect nested case–control
positive predictive value (PPV) from 124–126 studies within
147, 223, 238, 240 report characteristics automated health care
post‐approval monitoring of 119–121 databases 424
drugs 75 reporting ratios 127–128 sample size requirements and
post‐authorization efficacy social media 121 challenges 420
studies (PAES) 82 strengths 128 selection bias 422–423
Index 523
treatment responses and clinical assessments 193 quality‐adjusted life years

patterns 419 controlled distribution (QALYs) 315
unique biology and products 195 quantitative approaches 448
epidemiology 418–419 designs involving quantitative drug use
pre‐marketing device randomization 192 evaluation (DUE)
trials 410 medical devices, studies studies 390
prescribed daily dose (PDD) of 194–195 quantitative drug utilization
394–395 outcome assessment/ research (DUR) 390
prescribing practices 75–76 adjudication 192–193 quantitative strength 24
prescription‐based exposure patient‐reported outcomes, quasi‐cohort approach 372
sets 374 characterization of 193 quasi‐experimental
Prescription Card Services rare populations, studies studies 402
(PCS) 280 of 193–194 Question Appraisal
Prescription Drug User Fee Act vaccine safety 194 System 233
(PDUFA) 9, 81, 323 strengths 197 questionnaire design
Prescription Event primary non‐adherence 76 best practices for 233
Monitoring 210–213, primary pulmonary influence of 228–229,
215, 216 hypertension (PPH) 288 232–233
prevalent new‐user cohort prior probability 252 Question Understanding Aid
designs 374–375 probabilistic linkage (QUAID) 233
preventable risks 428 methods 148 quinidine 56
primary adherence productivity costs 316
measurement 361 product liability law 98 r
primary data collection 420, product liability litigation, and random‐effects models 340,
425 drug regulation 101–102 341, 350
applications 198–199 propensity score analysis 374, random error 23
definition of 192 377–378 randomized controlled trials
hybrid/enriched proportional reporting rate ratio (RCTs) 30–31, 83, 286,
designs 195 (PRR) 407 294, 338, 398
limitations 197–198 proportion of days covered classic 296–298
methods of (PDC) 359 control treatment, choice
biobanks/specimen PROs. see Patient reported of 297
banks 197 outcomes (PROs) data analysis 297
clinician or site‐reported prospective studies 193 data collection 297
outcomes (ClinROs) prospective study 32 generalizability of results
195–196 provider variation 415 297–298
patient‐generated data proxy adjustment 380 masking treatment
196 publication bias 337, 344–346 assignment 297
quality of, guidelines on public–private partnerships 416 sample size 297
197 p‐values 29, 30 clinical problems, by
registries as means of data research 294–296
collection 196 q complete
site‐based data collection qualitative approaches 448 follow‐up 302–303
195 qualitative drug utilization cooperative population 302
research questions 192–195 studies 391 data collection 300
524 Index
randomized controlled trials reactive approaches 94–95 relative size of databases 208
(RCTs) (cont’d) reactogenicity AEs 194 relative speed 208
data elements 299 Real World Data and Evidence reliability 222
data monitoring/interim (RWD/RWE) 85 quantitative measurement
analyses 304 real‐world evidence 96, 465, of 224–225
ethical and legal problems 471 reminder‐fatigue 400
hindering reasonable certainty 101, 300 renal dosing errors 440–441
implementation of 399 receiver operating characteristic renal dysfunction 55
follow‐up data collection 303 (ROC) curves 329 renal function 53
future 304–305 refill gap method 359 renin–angiotensin–aldosterone
important research registries 413–414, 423 system (RAAS)
question 300 as means of data inhibitors 188
large simple trials 298, 299 collection 196 reporting rate 127–128
conditions appropriate for registry‐based case–control reporting ratios 127–128
300 studies 208, 212 report quality 129–130
design characteristics of registry‐based studies 192 representativeness, definition
299 Registry in Glaucoma of 208
feasible 301 Outcomes Research Research Data Assistance
logistics of (RiGOR) study 198 Center (ResDAC) 153
conducting 302 regression models 412 research questions, in primary
limitations of 298 regulatory agencies 90–91, 97, data collection
methodological problems, 431, 472–473 192–195
solved by advancing science of clinical assessments 193
research 296–298 pharmacoepidemiology controlled distribution
objective and easily measured 96 products 195
outcomes 301–302 impact of regulatory designs involving
power and actions, assessment randomization 192
confounding 300–301 of 95–96 medical devices, studies
power/sample size 299 medicine, patterns of 95 of 194–195
primary analysis 303 monitoring post‐approval outcome assessment/
simple hypothesis 301 safety 94–95 adjudication 192–193
simple treatments 301 need for medicines 91 patient‐reported outcomes,
subgroup analyses 303–304 orphan drugs 91–92 characterization of 193
treatment and outcome, no planning drug development rare populations, studies
interaction programs 92–93 of 193–194
between 301 planning for post‐approval vaccine safety 194
uncertainty must exist 300 studies 94 residual confounding 380
Randomized Evaluation of pre‐approval review of response shift 330
Long‐Term clinical safety data retrospective cohorts, within
Anticoagulation Therapy 93–94 automated health care
(RE‐LY) 445 reimbursement mechanism databases 424
rare events, analysis 77, 146 retrospective study 31–32
of 343–344 reinforcements 399 rhesus‐human rotavirus
rare populations, studies relative cost 208 reassortant‐tetravalent
of 193–194 relative risk 29, 301, 341 vaccine (RRV‐TV) 405
Index 525
Rhode Island Drug Use risk minimization 430–431 in randomized controlled

Reporting System 468 evaluation of 432–433 trials (RCTs) 297, 299
ribonucleic acid (RNA) 258 plan 427 and real world performance
rifampin 56 strategies 431–432 issues 412
Right to Try Act 472 Risk Minimization Action requirements and
risk assessment 81, 430 Plans (RiskMAPs) challenges 420
risk‐averse decision‐making 81 sanger sequencing 270
approach 311 risk mitigation 81, 430–431 SARS‐CoV‐2 vaccine 468
risk communication 431 evaluation of 86–87, economic evaluation of 314
risk difference 340, 349 432–433 scale judgement
RISK Evaluation and Mitigation Risk Mitigation Evaluation method 329–330
Strategies (REMS) 86, Strategies (REMS) 82 scientific inference 21
87, 431 risk ratio 340 secondary non‐adherence 76
risk interval 406, 408 risk tolerance 67–70 selection bias 227, 422–423
risk management acceptability of risks, factors self‐controlled designs 406
activities 430 affecting 67 self‐management support 325
additional risk minimization adverse outcomes, features self‐payment model 309
strategies 431–432 of 67–68 self‐reported adherence
applies to all medicines 429 exposure, characteristics measures 357–358
clinical problems, addressed of 68–69 semi‐quantitative approaches
by research 428–433 perceptions of evaluator 448
complexities of medication 69–70 sensitivity analysis 227, 315,
use system 428 “ROBIS” tool 336 344, 381–382, 422
definition of 427 rofecoxib 101 sensitivity measures 222, 223
and epidemiology, regulatory root cause analysis 436 Sentinel system 79, 211
on 81 rosiglitazone 103, 350, 379 shock 56
evaluation of risk “rule of three,” 43 83 short‐term drug effects 140
minimization and signal detection 128
mitigation measures s silicone breast implants 7
432–433 safety theory 436–437 simple hypothesis 301
future 434–435 safety vs. risk 67 simulation‐based approach 77
pharmacoepidemiology role safety specification 427 single nucleotide
in 433–435 SAKIGAKE review polymorphism (SNP)
proactive process 429 program 92 59, 258, 269
product’s lifecycle, proceeds salicylate 51–52 site‐based data collection 195
throughout 429 salk vaccine trial 298 size of database 145
risk assessment 430 sample size 43–45, 184, 406 Smallpox Vaccine program
risk communication 431 calculations for 468
risk minimization 430–431 case–control small study effects 344
risk mitigation 430–431 studies 40–42, social media 121
scientifically driven 428–429 44, 45, 485–491 Social Security Administration
sources of risk from medical case series 41, 43 Death Master File 183
products 428 cohort studies 35–40, Sofosbuvir 318
Risk Management Plan 476–484 spain, electronic health record
(RMP) 81, 434 issues 441–442 database in 188–189
526 Index
specificity 24 subacute myelo‐optic‐ time‐conditional propensity

measures 222, 223 neuropathy (SMON) 7 scores (TCPS) 374
specimen banks 197 substantial bias 302 time period covered by
spontaneous reporting (self‐ substantial evidence 6 database 146
reported) of medication substantially uncertain 300 tort law 98
errors 438 sudden infant death syndrome total costs 392
spontaneous reporting system (SIDS) 407 total hip replacement
118–119, 209, 214 Summary of Product system 415
standardized incidence ratio Characteristics (SmPC) toxic shock 68
(SIR) 372 431 trajectory modeling 364–365
statins 119 Surveillance, Epidemiology, Transformed Medicaid
statistical analysis 363–364 and End Results (SEER) Statistical Information
statistical inference 21 196 System (T‐MSIS) 154
statistically significant 399 Survey Quality Predictor (SQP) Transformed Medicaid
statistical method 329, 233 Statistical Information
340–341 Swedish–Danish study 166 System Analytic Files
Stevens–Johnson syndrome Swedish Prescribed Drug (TAF) 154
128, 235 Register 393 translational
stochastic multicriteria Swiss cheese model 436 epidemiology 416
acceptability analysis system 1 decision processes 311 trastuzumab 262, 264
(SMAA) 85 system 2 decision treatment anniversary
structured association processes 311 method 359
approach 271 systematic errors 237 trials, definition of 334
studies of adverse effects 5 Systematized Nomenclature of triazolam 7
studies of adverse events 5 Medicine (SNOMED) 395 trim‐and‐fill method 344
study design troglitazone 66
concepts 25 t 21st Century Cures Act 10, 75
criteria for the causal nature Taiwanese National Health type A reactions 4
of an association 23–26 Insurance Research type B reactions 4
cross‐sectional studies 31 Database (NHIRD) 160 type I error 36, 267
epidemiologic study designs 30 target population 164 and replication 270–271
advantages 25 target drug therapy 468 type II error 36, 267–268, 271
analyses of secular trends teratogenesis 421
27 teratogenic medications 432 u
case–control studies 28 thalidomide‐associated UK Clinical Practice Research
case reports 26 phocomelia 277 Database (CPRD) 138
case series 26–27 thalidomide children 277 UK electronic health record
cohort studies 28–30 thalidomide disaster 6 databases 208
disadvantages 25 The Health Improvement UK General Practice Research
randomized clinical trials Network (THIN) 138, Database (GPRD) 167
30–31 181 unavoidable risks 428
errors types 22–23 Theory of Planned unexpected beneficial effects
scientific method 20–22 Behavior 399 212
study populations 164 therapeutic window 47 Unique Device Identification
study subjects 20 thiazolidinediones 66 (UDI) 409
Index 527
unit cost 392 methodological issues 403, methodological problems,

United States 405–406 solved by research 222
electronic health record assessment of causality 404 adjusting measures of
(EHR) databases clinical problems, association, for
data access for researchers addressed by research measurement error
184 403–404 227–228
databases 182 confounding and bias indices of measurement
data collection and 406–407 error 222
structure 183 epidemiologic studies measurement error
data quality 184 407–408 225–226
health care systems and future 408 medical condition type,
populations 181 measurement influence of 230
encounter databases in 149, accuracy 406 medication class, influence
151 methodological problems, of 228
US government 153–156 addressed by patient population,
US private insurance research 404–407 influence of 229–230,
databases 151–153 sample size 406 232
drug approval in 13–14 signal detection 404, 407 questionnaire design,
unit of analysis 399 primary data collection 194 influence of 228–229,
units of measurement validity of drug and diagnosis 232–233
394–395 data reliability, quantitative
Upper Limit Factor (U) 43 best practices 239–242 measurement of
US Commercial Insurance clinical problems, by 224–225
Databases 168–169 research 221–222 self‐reported diagnosis and
user‐generated content 121 comparator selection on hospitalization data,
US Food and Drug validation studies, from ad hoc survey
Administration (FDA) influence of 235–236 studies 230
6, 73, 115, 345, 346, 350 conducting validation studies self‐reported drug data
US Food and Drug to assess self‐reported from ad hoc survey
Administration Center data 235 studies 228
for Drug Evaluation and de novo questionnaire timing of diagnosis and
Research (FDA development 233–235 emotional effects on
CDER) 199 diagnosis and hospitalization patient, influence of
230–232
US government 153–156 data, special
validity, quantitative
US healthcare environment considerations
measurement of
166 with 237–239
222–224
US Medicaid Analytic Extracts distributed data systems,
questionnaire design, best
(MAX) data 146 special considerations
practices for 233
US private insurance with 239
validity of exposure data
databases 151–153 drug data, special
209
considerations with
validity of outcome data 210
v 237
“Valley of Death,” 308
Vaccine Safety Datalink from electronic encounter
valproic acid 56
(VSD) 405, 407, 408 databases 236–237
value health 166
vaccine safety future 242
528 Index
VA Vital Status File 183 VigiLyze 124 within‐subject designs

venous thromboembolism 28, Virtual Research Data Center 375–377
209 (VRDC) 153
ventricular tachycardia 375, volume data 392 x
376 volume of distribution 49, 50 Xeljanz® 81
Veterans Affairs (VA) 174 volume‐outcome
Veterans Health Administration relationship 410 y
(VHA) 181 Yale University Open Data
Veterans Health Information w Access (YODA)
Systems Technology and warfarin 265, 356 Project 338
Architecture (VistA) warning defect 98
system 183 water‐soluble drugs 56 z
VigiBase 124, 131 whole‐genome sequencing zero‐order kinetics 50
VigiFlow 124 272 zidovudine 16, 64
WILEY END USER LICENSE AGREEMENT
Go to www.wiley.com/go/eula to access Wiley’s ebook EULA.

Brian L. Strom, Stephen E. Kimmel, Sean Hennessy - Textbook of Pharmacoepidemiology-Wiley-Blackwell (2021)

Uploaded by

Copyright:

Available Formats

Brian L. Strom, Stephen E. Kimmel, Sean Hennessy - Textbook of Pharmacoepidemiology-Wiley-Blackwell (2021)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Brian L. Strom, Stephen E. Kimmel, Sean Hennessy - Textbook of Pharmacoepidemiology-Wiley-Blackwell (2021)

Uploaded by

Copyright:

Available Formats

Textbook of Pharmacoepidemiology

Stephen E. Kimmel, MD, MSCE

Sean Hennessy, PharmD, PhD

Limit of Liability/Disclaimer of Warranty

Library of Congress Cataloging-­in-­Publication Data

Cover Design: Wiley

Set in 9.5/12.5pt STIXTwoText by Straive, Pondicherry, India

Part I Introduction to Pharmacoepidemiology 1

2 Study Designs Available for Pharmacoepidemiologic Studies 20

Analyses of Secular Trends 27

3 Sample Size Considerations for Pharmacoepidemiologic Studies 35

4 Basic Principles of Clinical Pharmacology Relevant to Pharmacoepidemiologic Studies 47

5 When Should One Perform Pharmacoepidemiologic Studies? 62

Part II Sources of Pharmacoepidemiology Data 113

7 Postmarketing Spontaneous Pharmacovigilance Reporting Systems 115

8 Overview of Electronic Databases in Pharmacoepidemiology 136

9 Encounter Databases 142

10 Electronic Health Record Databases 174

11 Primary Data Collection for Pharmacoepidemiology 192

12 How Should One Perform Pharmacoepidemiologic Studies?

Part III Special Issues in Pharmacoepidemiology Methodology 219

13 Validity of Drug and Diagnosis Data in Pharmacoepidemiology 221

Quantitative Measurement of Validity 222

14 Assessing Causality from Case Reports 246

15 Molecular Pharmacoepidemiology 257

16 Bioethical Issues in Pharmacoepidemiologic Research 276

17 The Use of Randomized Controlled Trials for Pharmacoepidemiology 294

18 Pharmacoeconomics: Economic Evaluation of Pharmaceuticals 307

19 Patient Engagement and Patient Reported Outcomes 322

0005160718.INDD 13 09-13-2021 15:12:51

­ ethodologic Problems to be Solved by Pharmacoepidemiologic Research 328

20 The Use of Meta-analysis in Pharmacoepidemiology 334

21 Studies of Medication Adherence 355

22 Advanced Approaches to Controlling Confounding in Pharmacoepidemiologic Studies 368

­ linical Problems to be Addressed by Pharmacoepidemiologic Research 368

Part IV Special Applications and the Future of Pharmacoepidemiology 387

23 Special Applications of Pharmacoepidemiology 389

Examples of Currently Available Solutions 423

24 The Future of Pharmacoepidemiology 464

Appendix A — Sample Size Tables 475

Trisha Acri Laura E. Bothwell

Jeffrey S. Barrett Robert T. Chen

Rachael L. DiSantostefano Center for Pharmacoepidemiology Research

Scott Evans Sean Hennessy

Robert Gross Tamar Lasky

Julie Lauffenburger Allen A. Mitchell

Annika Richterich Brian L. Strom

Mary Elizabeth Ritchey Samy Suissa

strategies are now required by regulatory bodies recipients, introduced to pharmacoepidemiol-

Textbook of Pharmacoepidemiology attempts Data” and includes important chapters about

B reactions may be some other idiosyncratic epidemiology, as well. Epidemiology is also

g­ astrointestinal bleeding, postoperative astemizole, cisapride, droperidol,

carry out independent assessments while ­ otential Contributions

pharmacoepidemiologic studies but not not typically included in studies conducted

this is an important assurance. One cannot ●● The history of pharmacoepidemiology is a

Study Designs Available for Pharmacoepidemiologic Studies

­Introduction principles of clinical pharmacology, the con-

Study Sample quantitate the probability that the observed

Table 2.4 Advantages and disadvantages of epidemiologic study designs.

Study Design Advantages Disadvantages

Library of Congress Cataloging-in-Publication Data

13 Validity of Drug and Diagnosis Data in Pharmacoepidemiology 221

14 Assessing Causality from Case Reports 246

15 Molecular Pharmacoepidemiology 257

16 Bioethical Issues in Pharmacoepidemiologic Research 276

17 The Use of Randomized Controlled Trials for Pharmacoepidemiology 294

18 Pharmacoeconomics: Economic Evaluation of Pharmaceuticals 307

19 Patient Engagement and Patient Reported Outcomes 322

ethodologic Problems to be Solved by Pharmacoepidemiologic Research 328

20 The Use of Meta-analysis in Pharmacoepidemiology 334

21 Studies of Medication Adherence 355

22 Advanced Approaches to Controlling Confounding in Pharmacoepidemiologic Studies 368

linical Problems to be Addressed by Pharmacoepidemiologic Research 368

23 Special Applications of Pharmacoepidemiology 389

g astrointestinal bleeding, postoperative astemizole, cisapride, droperidol,

carry out independent assessments while otential Contributions

Introduction principles of clinical pharmacology, the con-

Discussion viewed as a subset of cohort studies, a type of

Introduction can examine the resulting confidence intervals

Abnormal liver 0.01 0.05 (2-tailed) 0.1 2 1 3104 3104

where R, α, β, Z1-α, and Z1-β are as above, p is pR

Ibuprofen 0.01 0.05 (2-tailed) 0.1 2 1 3210 3210

Introduction Clinical pharmacology is an important bridg-

c learance with age because of declining renal Pediatrics

emonstrating the increased transport of phe-

and calcium channel blocking effects on the in the implementation of model-informed

Introduction case, decision-making involves weighing the

Table 5.1 Reasons to perform i mportance to different organizations and indi-

iseases kill the equivalent of three jumbo jet

Conclusion Source: data derived from O’Brien (1986), Silverberg

The View from Industry Regulatory and Industry Focus on Risk

r ecommendations for improvements to risk harmacovigilance legislation introduced a

he View from Regulatory

onobvious. These criteria ensure that patents

t emporal relationship of drug usage to the (https://web-radr.eu/) to develop new techni-

i ndividuals using the drug in a specified time Strengths

Introduction databases,” as potential data sources for phar-