Spirt Es 1993

Lecture Notes in Statistics 81
Edited by J. Berger, S. Fienberg, J. Gani, K. Krickeberg,

I. OIkin, and B. Singer
Peter Spirtes
Clark Glymour
Richard Scheines
Causation, Prediction,
and Search
Springer-Verlag
New York Berlin Heidelberg London Paris
Tokyo Hong Kong Barcelona Budapest
Peter Spirtes
Clark Glymour
Richard Scheines
Department of Philosophy
Carnegie Mellon University
Pittsburgh, P A 15213
Mathematics Subject Classification: 62-07
library of Congress Cataloging-in-Publication Data

Spirtes, Peter
Causation, prediction, and search I Peter Spirtes, Clarie Glymour
Richard Scheines
p. em. - (Lecture notes in statistics ; 81)
ISBN-13:978-1-4612-7650-0
1. Mathematical statistics I. Glymour, Clarie N. ll. Scheines,
Richard. In. Title. IV. Series: Lecture notes in statistics
(Springer-Verlag) : v. 81.
QA276.S65 1993
519.5--dc20 92-40263
Printed on acid-free paper.
© 1993 Springer-Verlag New Yorlc, Inc.

All rights reserved. This worie may not be translated or copied in whole orin part without the written
permission of the publisher (Springer-Verlag New Yorie, Inc., 175 Fifth Avenue, New Yorie, NY
10010, USA), except for brief excetpts in connection with reviews or scholarly analysis. Use in
connection with any foon of infonnation storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademaries, etc., in this publication, even if the
fonner are not especially identified, is not to be taken as a sign that such names, as understood by the
Trade Maries and Merchandise Maries Act, may accordingly be used freely by anyone.
Camera ready copy provided by the authors.
9 8 765 432 1
ISBN-13:978-1-4612-7650-0 e-ISBN-13:978-1-4612-2748-9
DOl: 10.1007/978-1-4612-2748-9
To my parents, Morris and Cecile Spirtes - P.S.
In memory of Lucille Lynch Schwartz Watkins Speede Tindall Preston - C. O.
To Martha, for her support and love - R.S.

It is with data affected by numerous causes that Statistics is
mainly concerned. Experiment seeks to disentangle a complex of
causes by removing all but one of them, or rather by
concentrating on the study of one and reducing the others, as far
as circumstances permit, to comparatively small residium.
Statistics, denied this resource, must accept for analysis data
subject to the influence of a host of causes, and must try to
discover from the data themselves which causes are the
important ones and how much of the observed effect is due to the
operation of each.
--G. U. Yule and M. G. Kendall 1950
The Theory of Estimation discusses the principles upon which

observational data may be used to estimate, or to throw light
upon the values of theoretical quantities, not known numerically,
which enter into our specification of the causal system
operating.
-- Sir Ronald Fisher, 1956
George Box has [almost] said "The only way to find out what will
happen when a complex system is disturbed is to disturb the
system, not merely to observe it passively." These words of
caution about "natural experiments" are uncomfortably strong.
Yet in today's world we see no alternative to accepting them as,
if anything, too weak.
--G. Mosteller and J. Tukey, 1977
Causal inference is one of the most important, most subtle, and

most neglected of all the problems of Statistics.
-- P. Dawid, 1979
Preface
This book is intended for anyone, regardless of discipline, who is interested in the use of
statistical methods to help obtain scientific explanations or to predict the outcomes of actions,
experiments or policies.
Much of G. Udny Yule's work illustrates a vision of statistics whose goal is to investigate
when and how causal influences may be reliably inferred, and their comparative strengths
estimated, from statistical samples. Yule's enterprise has been largely replaced by Ronald
Fisher's conception, in which there is a fundamental cleavage between experimental and non-
experimental inquiry, and statistics is largely unable to aid in causal inference without
randomized experimental trials. Every now and then members of the statistical community
express misgivings about this turn of events, and, in our view, rightly so. Our work represents
a return to something like Yule's conception of the enterprise of theoretical statistics and its
potential practical benefits.
If intellectual history in the 20th century had gone otherwise, there might have been a
discipline to which our work belongs. As it happens, there is not. We develop material that
belongs to statistics, to computer science, and to philosophy; the combination may not be
entirely satisfactory for specialists in any of these subjects. We hope it is nonetheless
satisfactory for its purpose. We are not statisticians by training or by association, and perhaps
for that reason we tend to look at issues differently, and, from the perspective common in the
discipline, no doubt oddly. We are struck by the fact that in the social and behavioral sciences,
epidemiology, economics, market research, engineering, and even applied physics, statistical
methods are routinely used to justify causal inferences from data not obtained from
randomized experiments, and sample statistics are used to predict the effects of policies,
manipulations or experiments. Without these uses the profession of statistics would be a far
smaller business. It may not strike many professional statisticians as particularly odd that the
discipline thriving from such uses assures its audience that they are unwarranted, but it strikes
us as very odd indeed. From our perspective outside the discipline, the most urgent questions
about the application of statistics to such ends concern the conditions under which causal
inferences and predictions of the effects of manipulations can and cannot reliably be made,
and the most urgent need is a principled, rigorous theory with which to address these
viii Causation, Prediction, and Search
problems. To judge from the testimony of their books, a good many statisticians think any
such theory is impossible. We think the common arguments against the possibility of inferring
causes from statistics outside of experimental trials are unsound, and radical separations of the
principles of experimental and observational study designs are unwise. Experimental and
observational design may not always permit the same inferences, but they are subject to
uniform principles.
The theory we develop follows necessarily from assumptions laid down in the statistical
community over the last fifteen years. The underlying structure of the theory is essentially
axiomatic. We will give two independent axioms on the relation between causal structures
and probability distributions and deduce from them features of causal relationships and
predictions that can and that cannot be reliably inferred from statistical constraints under a
variety of background assumptions. Versions of all of the axioms can be found in papers by
Lauritzen, Wermuth, Speed, Pearl, Rubin, Pratt, Schlaifer, and others. In most cases we will
develop the theory in terms of probability distributions that can be thought of loosely as
propensities that determine long run frequencies, but many of the probability distributions can
alternatively be understood as (normative) subjective degrees of belief, and we will
occasionally note Bayesian applications. From the axioms there follow a variety of theorems
concerning estimation, sampling, latent variable existence and structure, regression,
indistinguishability relations, experimental design, prediction, Simpson's paradox, and other
topics. Foremost among the "other topics" are the discovery that statistical methods
commonly used for causal inference are radically suboptimal, and that there exist
asymptotically reliable, computationally efficient search procedures that conjecture causal
relationships from the outcomes of statistical decisions made on the basis of sample data.
(The procedures we will describe require statistical decisions about the independence of
random variables; when we say such a procedure is "asymptotically reliable" we mean it
provides correct information if the outcome of each of the requisite statistical decisions is true
in the population under study.)
This much of the book is mathematics: where the axioms are accepted, so must the theorems
be, including the existence of search procedures. The procedures we describe are applicable
to both linear and discrete data and can be feasibly applied to a hundred or more variables so
long as the causal relations between the variables are sufficiently sparse and the sample
Preface ix
sufficiently large. These procedures have been implemented in a computer program,

TETRAD II, which at the time of writing is publicly available.!
The theorems concerning the existence and properties of reliable discovery procedures of
themselves tell us nothing about the reliabilities of the search procedures in the short run. The
methods we describe require an unpredictable sequence of statistical decisions, which we
have implemented as hypothesis tests. As is usual in such cases, in small samples the
conventional p values of the individual tests may not provide good estimates of type I error
probabilities for the search methods. We provide the results of extensive tests of various
procedures on simulated data using Monte Carlo methods, and these tests give considerable
evidence about reliability under the conditions of the simulations. The simulations illustrate
an easy method for estimating the probabilities of error for any of the search methods we
describe. The book also contains studies of one large pseudo-empirical data set--a body of
simulated data created by medical researchers to model emergency medicine diagnostic
indicators and their causes--and a great many empirical data sets, most of which have been
discussed by other authors in the context of specification searches.
A further aim of this work is to show that a proper understanding of the relationship between
causality and probability can help to clarify diverse topics in the statistical literature,
including the comparative power of experimentation versus observation, Simpson's paradox,
errors in regression models, retrospective versus prospective sampling, the perils of variable
selection, and other topics. There are a number of relevant topics we do not consider. They
include problems of estimation with discrete latent variables, optimizing statistical decisions,
many details of sampling designs, time series, and a full theory of "non-recursive" causal
structures--i.e., finite graphical representations of systems with feedback.
Causation, Prediction and Search is not intended to be a textbook, and it is not fitted out with
the associated paraphernalia. There are open problems but no exercises. In a textbook
everything ought to be presented as if it were complete and tidy, even if it isn't. We make no
such pretenses in this book, and the chapters are rich in unsolved problems and open
questions. Textbooks don't usually pause much to argue points of view; we pause quite a lot.
The various theorems in this book often have a graph theoretic character; many of them are
long, difficult case arguments of a kind quite unfamiliar in statistics. In order not to interrupt
! To anyone with a workstation with a UNIX operating system, a PASCAL compiler and a network connection.
A less flexible version of the program is available for IBM compatible 80-386 and 80-486 personal computers.
Write Richard Scheines at RS2L@andrew.cmu.edu.
x Causation, Prediction, and Search
the flow of the discussion we have placed all proofs but one in a chapter at the end of the
book. In the few cases where detailed proofs are available in the published literature, we have
simply referred the reader to them. Where proofs of important results have not been published
or are not readily available we have given the demonstrations in some detail.
The structure of the book is as follows. Chapter I concerns the motivation for the book in the
context of current statistical practice and advertises some of the results. Chapter 2 introduces
the mathematical ideas necessary to the investigation, and Chapter 3 gives the formal
framework a causal interpretation, lays down the axioms, notes circumstances in which they
are likely to fail, and provides a few fundamental theorems. The next two chapters work out
the consequences of two of the axioms for some fundamental issues in contexts in which it is
known, or assumed, that there are no unmeasured common causes affecting measured
variables. In Chapter 4 we give graphical characterizations of necessary and sufficient
conditions for causal hypotheses to be statistically indistinguishable from one another in each
of several senses. In Chapter 5 we criticize features of model specification procedures
commonly recommended in statistics, and we describe feasible algorithms that from
properties of population distributions extract correct information about causal structure,
assuming the axioms apply and that no unmeasured common causes are at work. The
algorithms are illustrated for a variety of empirical and simulated samples. Chapter 6 extends
the analysis of Chapter 5 to contexts in which it cannot be assumed that no unmeasured
common causes act on measured variables. From both a theoretical and practical perspective,
this chapter and the next form the center of the book, but they are especially difficult. Chapter
7 addresses the fundamental issue of predicting the effects of manipulations, policies, or
experiments. As an easy corollary, the chapter unifies directed graphical models with Donald
Rubin's "counterfactual" framework for analyzing prediction. Chapter 8 applies the results of
the preceding chapters to the subject of regression. We argue that even when standard
statistical assumptions are satisfied multiple regression is a defective and unreliable way to
assess causal influence even in the large sample limit, and various automated regression
model specification searches only make matters worse. We show that the algorithms of
Chapter 6 are more reliable in principle, and we compare the performances of these
algorithms against various multiple regression procedures on a variety of simulated and
empirical data sets. Chapter 9 considers the design of empirical studies in the light of the
results of earlier chapters, including issues of retrospective and prospective sampling, the
comparative power of experimental and observational designs, selection of variables, and the
design of ethical clinical trials. The chapter concludes with a look back at some aspects of the
dispute over smoking and lung cancer. Chapters 10 and 11 further consider the linear case,
and analyze algorithms for discovering or elaborating causal relations among measured and
Preface xi
unmeasured variables in linear systems. Chapter 12 is a brief consideration of a variety of

open questions. Proofs are given in Chapter 13.
We have tried to make this work self-contained, but it is admittedly and unavoidably difficult
The reader will be aided by a previous reading of Pearl (1988), Whittaker (1990) or
Neopolitan (1990) .
Acknowledgments
One source of the ideas in this book is in work we began ten years ago at the University of
Pittsburgh. We drew many ideas about causality, statistics and search from the psychometric,
economic and sociological literature, beginning with Charles Spearman's project at the tum of
the century and including the work of Herbert Simon, Hubert Blalock and Herbert Costner.
We obtained a new perspective on the enterprise from Judea Pearl's Probabilistic Reasoning
in Intelligent Systems, which appeared the next year. Although not principally concerned with
discovery, Pearl's book showed us how to connect conditional independence with causal
structure quite generally, and that connection proved essential to establishing general, relillble
discovery procedures. We have since profited from correspondence and conversation with
Pearl and with Dan Geiger and Thomas Verma, and from several of their papers. Pearl's work
drew on the papers of Wermuth (1980), Kiiveri and Speed (1982), Wermuth and Lauritzen
(1983), and Kiiveri, Speed and Carlin (1984), which in the early 1980s had already provided
the foundations for a rigorous study of causal inference. Paul Holland introduced one of us to
the Rubin framework some years ago, but we only recently realized it's logical connections
with directed graphical models. We were further helped by J. Whittaker's (1990) excellent
account of the properties of undirected graphical models.
We have learned a great deal from Gregory Cooper at the University of Pittsburgh who
provided us with data, comments, Bayesian algorithms and the picture and description of the
ALARM network which we consider in several places. Over the years we have learned useful
things from Kenneth Bollen. Chris Meek provided essential help in obtaining an important
theorem that derives various claims made by Rubin, Pratt and Schlaifer from axioms on
directed graphical models.
Steve Fienberg and several students from Carnegie Mellon's department of statistics joined
with us in a seminar on graphical models from which we learned a great deal. We are
indebted to him for his openness, intelligence and helpfulness in our research, and to
Elizabeth Slate for guiding us through several papers in the Rubin framework. We are obliged
to Nancy Cartwright for her courteous but salient criticism of the approach taken in our
previous book and continued here. Her comments prompted our work on parameters in
xiv Causation, Prediction, and Search
Chapter 4. We are indebted to Brian Skyrms for his interest and encouragement over many
years, and to Marek Druzdzel for helpful comments and encouragement. We have also been
helped by Linda Bouck, Ronald Christensen, Jan Callahan, David Papineau, John Earman,
Dan Hausman, Joe Hill, Michael Meyer, Teddy Seidenfeld, Dana Scott, Jay Kadane, Steven
Klepper, Herb Simon,Peter Slezak, Steve Sorensen, John Worrall and Andrea Woody. We are
indebted to Ernest Seneca for putting us in contact with Dr. Rick Linthurst, and we are
especially grateful to Dr. Linthurst for making his doctoral thesis available to us.
Our work has been supported by many institutions. They, and those who made decisions on
their behalf, deserve our thanks. They include Carnegie Mellon University, the National
Science Foundation programs in History and Philosophy of Science, in Economics, and in
Knowledge and Database Systems, the Office of Naval Research, the Navy Personnel
Research and Development Center, the John Simon Guggenheim Memorial Foundation,
Susan Chipman, Stanley Collyer, Helen Gigley, Peter Machamer, Steve Sorensen, Teddy
Seidenfeld and Ron Overmann. The Navy Personnel Research and Development Center
provided us the benefit of access to a number of challenging data analysis problems from
which we have learned a great deal.
Table of Contents
Preface ................................................................................................................................ vii
Acknowledgments .............................................................................................................. xiii
Notational Conventions ••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••• xxi
1. Introduction and Advertisement •••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••.••..•••.••••.••• 1

1.1 The Issue ............................................................................................................ 1
1.2 Advertisements .................................................................................................. 10
1.2.1 Bayes Networks from the Data ........................................................ ,.. 11
1.2.2 Structural Equation Models from the Data ......................................... 13
1.2.3 Selection of Regressors ....................................................................... 14
1.2.4 Causal Inference without Experiment ................................................ 17
1.2.5 The Structure of the Unobserved ........................................................ 19
1.3 Themes ............................................................................................................... 21
2. Formal Preliminaries ..................................................................................................... 25

2.1 Graphs ................................................................................................................ 25
2.2 Probability .......................................................................................................... 31
2.3 Graphs and Probability Distributions ................................................................ 32
2.3.1 Directed Acyclic Graphs ..................................................................... 32
2.3.2 Directed Independence Graphs ........................................................... 34
2.3.3 Faithfulness ......................................................................................... 35
2.3.4 d-separation ......................................................................................... 36
2.3.5 Linear Structures ................................................................................. 36
2.4 Undirected Independence Graphs ...................................................................... 37
2.5 Deterministic and Pseudo-Indeterministic Systems .......................................... 38
2.6 Background Notes ............................................................................................. 39
3. Causation and Prediction: Axioms and Explications .................................................41

3.1 Conditionals ....................................................................................................... 41
3.2 Causation ........................................................................................................... 42
3.2.1 Direct vs. Indirect Causation ............................................................. .42
3.2.2 Events and Variables .......................................................................... 43
3.2.3 Examples ............................................................................................. 45
3.2.4 Representing Causal Relations with Directed Graphs ........................ 47
xvi Causation, Prediction, and Search
3.3 Causality and Probability ................................................................................... 49

3.3.1 Deterministic Causal Structures ........................................................ .49
3.3.2 Pseudo-Indeterministic and Indetenninistic Causal Structures .......... 51
3.4 The Axioms ....................................................................................................... 53
3.4.1 The Causal Markov Condition ............................................................ 53
3.4.2 The Causal Minimality Condition ...................................................... 55
3.4.3 The Faithfulness Condition ................................................................. 56
3.5 Discussion of the Conditions ............................................................................. 57
3.5.1 The Causal Markov and Minimality Conditions ................................ 57
3.5.2 Faithfulness and Simpson's Paradox ................................................... 64
3.6 Bayesian Interpretations .................................................................................... 70
3.7 Consequences of The Axioms ........................................................................... 71
3.7.1 d-Separation ........................................................................................ 71
3.7.2 The Manipulation Theorem ................................................................ 75
3.8 Detenninism ...................................................................................................... 81
4. Statistical Indistinguishability ...................................................................................... 87

4.1 Strong Statistical Indistinguishability ................................................................ 88
4.2 Faithful Indistinguishability ............................................................................... 89
4.3 Weak Statistical Indistinguishability ................................................................. 90
4.4 Rigid Indistinguishability .................................................................................. 93
4.5 The Linear Case ................................................................................................. 94
4.6 Redefining Variables ......................................................................................... 99
5. Discovery Algorithms for CausaUy Sufficient Structures •••••••••••••••••••••••••••••••••••••••••. 103

5.1 Discovery Problems ........................................................................................... 103
5.2 Search Strategies in Statistics ............................................................................ 104
5.2.1 The Wrong Hypothesis Space ............................................................ 105
5.2.2 Computational and Statistical Limitations .......................................... 107
5.2.3 Generating a Single Hypothesis .......................................................... 108
5.2.4 Other Approaches ............................................................................... 109
5.2.5 Bayesian Methods ............................................................................... 109
5.3 The Wermuth-Lauritzen Algorithm ................................................................... 111
5.4 New ~gorithms ................................................................................................. 112
5.4.1 The SGS Algorithm ............................................................................ 114
5.4.2 The PC Algorithm ............................................................................... 116
Contents xvii
5.4.3 The IG (Independence Graph) Algorithm .......................................... 124

5.4.4 Variable Selection ............................................................................... 125
5.4.5 Incorporating Background Knowledge ............................................... 127
5.5 Statistical Decisions ........................................................................................... 128
5.6 Reliability and Probabilities of Error ................................................................. 130
5.7 Estimation .......................................................................................................... 132
5.8 Examples and Applications ............................................................................... 132
5.8.1 The Causes of Publishing Productivity ............................................... 133
5.8.2 Education and Fertility ....................................................................... 139
5.8.3 The Female Orgasm ............................................................................ 140
5.8.4 The American Occupational Structure ............................................... 142
5.8.5 The ALARM Network ........................................................................ 145
5.8.6 Virginity .............................................................................................. 147
5.8.7 The Leading Crowd ............................................................................ 147
5.8.8 Influences on College Plans ................................................................ 149
5.8.9 Abortion Opinions .............................................................................. 150
5.8.10 Simulation Tests with Random Graphs ........................................... 152
5.9 Conclusion ......................................................................................................... 161
5.10 Background Notes ........................................................................................... 162
6. Discovery Algorithms without Causal Sufficiency ..................................................... 163

6.1 Introduction ........................................................................................................ 163
6.2 The PC Algorithm and Latent Variables ........................................................... 165
6.3 Mistakes ............................................................................................................. 168
6.4 Inducing Paths ................................................................................................... 173
6.5 Inducing Path Graphs ........................................................................................ 174
6.6 Partially Oriented Inducing Path Graphs ........................................................... 177
6.7 Algorithms for Causal Inference with Latent Common Causes ........................ 181
6.8 Theorems on Detectable Causal Influence ........................................................ 190
6.9 Non-Independence Constraints ........................................................................... 191
6.10 Generalized Statistical Indistinguishability and Linearity ............................... 193
6.11 The Tetrad Representation Theorem ............................................................... 196
6.12 An Example: Math Marks and Causal Interpretation ...................................... 197
6.13 Background Notes ........................................................................................... 200
7. Prediction ........................................................................................................................ 201

7.1 Introduction ........................................................................................................ 201
7.2 Prediction Problems ........................................................................................... 202
xviii Causation, Prediction, and Search
7.3 Rubin-Holland-Pratt-Schlaifer Theory .............................................................. 203

7.4 Prediction with Causal Sufficiency ................................................................... 213
7.5 Prediction without Causal Sufficiency .............................................................. 216
7.6 Examples ............................................................................................................ 227
7.7 Conclusion ......................................................................................................... 237
7.8 Background Notes ............................................................................................ 237
8. Regression, Causation and Prediction ••••••••••••••••••••••••••••••••••••••••••.•••••••••••.•••••••••••••••••• 238

8.1 When Regression Fails to Measure Influence ................................................... 238
8.2 A Solution and Its Application .......................................................................... 242
8.2.1 Components of the Armed Forces Qualification Test ........................ 243
8.2.2 The Causes of Spartina Biomass ........................................................ 244
8.2.3 The Effects of Foreign Investment on Political Repression ............... 248
8.2.4 More Simulation Studies .................................................................... 250
8.3 Error Probabilities for Specification Searches ................................................... 252
8.4 Conclusion ......................................................................................................... 257
9. The Design of Empirical Studies ••••••••••••••...••••••.••••••••••.••••••••••••••••.•••••••••••.•.••••••••. _•.••• 259

9.1 Observational or Experimental Study? .............................................................. 259
9.2 Selecting Variables ............................................................................................ 271
9.3 Sampling ............................................................................................................ 272
9.4 Ethical Issues in Experimental Design .............................................................. 276
9.4.1 The Kadane/SedranskiSeidenfeld Design ........................................... 277
9.4.2 Causal Reasoning in the Experimental Design ................................... 280
9.4.3 Towards Ethical Trials ........................................................................ 286
9.5 An Example: Smoking and Lung Cancer .......................................................... 291
9.6 Appendix ............................................................................................................ 302
10. The Structure of the Unobserved ............................................................................... 306

10.1 Introduction ...................................................................................................... 306
10.2 An Outline of the Algorithm ............................................................................ 307
10.3 Finding Almost Pure Measurement Models .................................................... 310
10.3.1 Intra-Construct Foursomes ............................................................... 310
10.3.2 Cross-Construct Foursomes ............................................... ,.............. 311
10.4 Facts about the Unobserved Determined by the Observed .............................. 315
10.5 Unifying the Pieces .......................................................................................... 316
10.6 Simulation Tests .............................................................................................. 320
10.7 Conclusion ....................................................................................................... 322
Contents xix
11. Elaborating Linear Tbeories witb Unmeasured Variables •••••••••••••••••••••••••••••••••••••. 323

11.1 Introduction ...................................................................................................... 323
11.2 The Procedure ................................................................................................. 324
11.2.1 Scoring .............................................................................................. 324
11.2.2 Search ............................................................................................... 327
11.3 The LISREL and EQS Procedures ................................................................. 329
11.3.1 Input and Output ............................................................................... 329
11.3.2 Scoring .............................................................................................. 330
11.3.3 The LISREL VI Search .................................................................... 331
11.3.4 The EQS Search ................................................................................ 331
11.4 The Primary Study .......................................................................................... 332
11.4.1 The Design of Comparative Simulation Studies .............................. 332
11.4.2 Study Design ..................................................................................... 333
11.5 Results .............................................................................................................. 343
11.6 Reliability and Infonnativeness ....................................................................... 346
11. 7 Using LISREL and EQS as Adjuncts to Search .............................................. 349
11.8 Limitations of the TETRAD IT Elaboration Search ......................................... 351
11.9 Some Morals for Statistical Search .................................................................. 352
12. Open Problems ......................................................................,'•••••••••••••••••••••••••••••••••••••• 354

12.1 Feedback, Reciprocal Causation, and Cyclic Graphs ...................................... 354
12.1.1 Mason's Theorem .............................................................................. 355
12.1.2 Time Series and Cyclic Graphs ......................................................... 356
12.1.3 The Markov Condition, Factorizability and FaithfuIness ................. 359
12.1.4 Discovery Procedures ....................................................................... 360
12.2 Indistinguishability Relations .......................................................................... 361
12.3 Time series and Granger Causality .................................................................. 363
12.4 Model Specification and Parameter Estimation from the Same Data Base ..... 365
12.5 Conditional Independence Tests ...................................................................... 366
13. Proofs of Tbeorems ••••••••••••••••••••••••••••••••••••.••••.••••••••••••••••••••.••••••••••••••••••••••••••••••••••••••• 367

13.1 Theorem 2.1 ..................................................................................................... 367
13.2 Theorem 3.1 ..................................................................................................... 367
13.3 Theorem 3.2 ..................................................................................................... 374
13.4 Theorem 3.3 ..................................................................................................... 376
13.5 Theorem 3.4 ..................................................................................................... 385
13.6 Theorem 3.5 ..................................................................................................... 386
13.7 Theorem 3.6 (Manipulation Theorem) ............................................................ 395
xx Causation, Prediction, and Search
13.8 Theorem 3.7 ..................................................................................................... 398

13.9 Theorem 4.1 ..................................................................................................... 401
13.10 Theorem 4.2 ................................................................................................... 403
13.11 Theorem 4.3 ................................................................................................... 403
13.12 Theorem 4.4 ................................................................................................... 404
13.13 Theorem 4.5 ................................................................................................... 404
13.14 Theorem 4.6 ................................................................................................... 405
13.15 Theorem 5.1 ................................................................................................... 405
13.16 Theorem 6.1 ................................................................................................... 408
13.17 Theorem 6.2 ................................................................................................... 411
13.18 Theorem 6.3 ................................................................................................... 414
13.19 Theorem 6.4 ................................................................................................... 417
13.20 Theorem 6.5 ................................................................................................... 418
13.21 Theorem 6.6 ................................................................................................... 419
13.22 Theorem 6.7 ................................................................................................... 424
13.23 Theorem 6.8 ................................................................................................... 425
13.24 Theorem 6.9 ................................................................................................... 425
13.25 Theorem 6.10 (Tetrad Representation Theorem) .......................................... 426
13.26 Theorem 6.11 ................................................................................................. 460
13.27 Theorem 7.1 ................................................................................................... 460
13.28 Theorem 7.2 ................................................................................................... 462
13.29 Theorem 7.3 ................................................................................................... 463
13.30 Theorem 7.4 ................................................................................................... 470
13.31 Theorem 7.5 ................................................................................................... 471
13.32 Theorem 9.1 ................................................................................................... 472
13.33 Theorem 9.2 ................................................................................................... 472
13.34 Theorem 10.1 ................................................................................................. 473
13.35 Theorem 10.2 ................................................................................................. 476
13.36 Theorem 11.1 ................................................................................................. 479
Glossary .............................................................................................................................. 481
Bibliography ....................................................................................................................... 495
Index .................................................................................................................................... 517

Notational Conventions
Text
In the text, each technical term is written in boldface where it is defined.
Variables: capitalized, and in italics, e.g, X

Values of variables: lower case, and in italics, e.g. X = x
Sets: capitalized, and in boldface, e.g, V
Values of sets of variables: lower case, and in boldface, e.g. V = v
Members of X that are not members of Y: x\Y
Error variables: e, {;,e
Independence of X and Y: xliY
Independence of X and Y conditional on Z: xli YIZ
XuY: XY
Covariance of X and Y: COV(X,Y) or'}'xy
Correlation of X and Y : PXY
Sample correlation of X and Y :
Partial Correlation of X and Y,
controlling for all members of set Z: pxy.z
In all of the graphs that we consider, the vertices are random variables. Hence we use the
terms "variables in a graph" and "vertices in a graph" interchangeably.
Figures
Figure numbers occur centered just below a figure, starting at I within each chapter. Where
necessary, we distinguish between measured and unmeasured variables by boxing measured
variables and circling unmeasured variables (except for error terms). Variables beginning with
e, e, or {; are understood to be "error," or "disturbance" variables. For example, in the figure
below, X and Yare measured, T is not, and e is an error term.
xxii Causation, Prediction, and Search
Figure}
We will neither box nor circle variables in graphs in which no distinction need be made
between measured and unmeasured variables, e.g., figure 2.
X.l X
........ X ....- 3~X
2 4
Figure 2
For simplicity, we state and prove our results for probability distributions over discrete
random variables. However, under suitable integrability conditions, the results can be easily
generalized to continuous distributions that have density functions by replacing the discrete
variables by continuous variables, probability distributions by density functions, and
summations by integrals.
If a description of a set of variables is a function of a graph G and variables in G, then we

make G an optional argument to the function. For example, Parents(G,x) denotes the set of
variables that are parents of X in graph G; if the context makes clear which graph is being
referred to we will simply write Parents(X).
If a distribution is defined over a set of random variables 0 then we refer to the distribution as
P(O). An equation between distributions over random variables is understood to be true for
all values of the random variables for which all of the distributions in the equation are
*
defined. For example if X and Y each take the values 0 or 1 and P(X = 0) 0 and P(){ = 1) 0 *
then P(yIX) = P(y) means P(Y = OIX= 0) = P(y = 0), P(y = OIX= I) = P(Y = 0), P(y = llX= 0)
= P(Y = I), and P(Y = llX= 1) = P(Y = I).
Notational Conventions xxiii
L ' which has the following properties:

~
We sometimes use a special summation symbol,
(i) when sets of random variables are written beneath the special summation symbol, it is
understood that the summation is to be taken over sets of values of the random variables,
not the random variables themselves,
(ii) if a conditional probability distribution appears in the scope of such a summation
symbol, the summation is to be taken only over values of the random variables for which
the conditional probability distributions are defined,
(iii) if there are no values of the random variables under the special summation symbol for
which the conditional probability distributions in the scope of the symbol are defined,
then the summation is equal to zero.
For example, suppose that X, Y, and Z can each take on the values 0 or 1. Then if
P(Y=O;Z=O) *" 0
~
LP(XIY = O,Z = 0) = P(X = OIY = O,Z = 0) + P(X = IIY = O,Z = 0)

x
However, if p(Y=o,z=o) = 0, then P(X=OIY=<i,z=O) and P(X=lIY=O,z=o) are not defined, so
LP(XIY = O,Z = 0) = 0
X
We will adopt the following conventions for empty sets of variables. If Y =0 then
(i) P(XIY) means P(X).
(ii) fJXZ.Y means fJXZ.
(iii) A Jl DIY means A Jl D.
(iv) A Jl Y is always true.
Chapter 1
Introduction and Advertisement
1.1 The Issue
Statistics textbooks provide interesting examples of causal questions: Did halothane do more to
cause surgical deaths than ether? Was the lower admission rate of women to graduate programs
at the University of California caused by discrimination against women? Does smoking cause
cancer? Issues about determining causes surround many of the introductory and even advanced
topics in statistical pedagogy: experimental design, randomization, collinearity in multiple
regression, observational versus experimental studies, and so forth. But except for the standard
warnings that correlation is not causation, the textbooks include little if any systematic
discussion of the connection between causation and probability. The mathematics of probability
and statistical inference is explicit; but the connection between probability relations and causal
dependencies is almost completely tacit. The same applies to prediction, at least outside of
econometrics. The textbooks consider cases where policy interventions are at issue, but they tell
us nothing systematic about the connections between statistical analysis of observations or
experiments and predictions of the effects of policies, actions or manipulations.
Even more curious to an outsider, many statistics textbooks claim that the methods they
describe cannot be reliably used to infer causal dependencies from random uncontrolled
samples, or to predict the effects of manipulations, and they say or suggest that all possible
statistical methods are equally impotent for these purposes. If widely believed, these bold
claims would remove much of the market for the books that advance them, and at least some of
the interest in statistics as a subject Linear regression, for example, is taught as a means of
fitting a line to sample data and also as a means of predicting new values of a variable. Yet
many of the real applications of regression are to predict values of a variable when the
regressors are manipulated, that is, when action or policy forces some novel distribution on the
regressors. Every author of a regression textbook must know that were it not for such uses
2 Causation, Prediction, and Search
much of the audience for the book would be lost. Modestly, the texts announce that regression
cannot be used to infer causes from non-experimental samples, or to make predictions about
manipulated systems; immodestly, often the same books announce that no possible statistical
methods can be used to reliably infer causes or to make predictions about manipulated systems.
The arguments advanced against the possibility of reliable inference from statistical samples to
causal structure are remarkably uniform from book to book. Mosteller and Tukey (1977), for
example, provide an abundance of cases in which regression yields erroneous causal
inferences. Rawlings (1988) makes the chief example of his textbook on applied regression a
problem of causal inference pursued through several chapters, only to conclude that regression
is unavailing for the purpose. But without some demonstration that regression is optimally
reliable for such inference problems, these considerations only argue that rather than teaching
regression we should look for better methods. An even more common textbook argument
attempts to demonstrate that what regression cannot do for causal inference, nothing can. The
argument asserts the underdeterrnination of causal structure by statistical dependencies. Readers
are told (Younger, 1978; Rawlings, 1988) that a "relationship"--that is, some statistical
dependency--between sample values of variables X and Y may be observed
(i) when X causes Y,
(ii) when Y causes X,
(iii) when each causes the other,
(iv) when some third variable causes both,
(v) when the sample is notrepresentative, or
(vi) when the values of X and Y form time series.
Some such list is inevitably followed by the warning that "experimentation, and not the
existence of the statistical relationship, is necessary to establish a cause" (Younger, p. 176).
Consider the logical force of this argument:
It is true enough that we cannot distinguish among (i), (ii) (iii) and (iv) when we measure only
two variables, X and Y, but what is the proof that we cannot distinguish among these
alternative causal relations if additional variables are measured? We know that in other contexts
identifiability and estimation properties of parameters relating a pair of variables can be changed
if further variables are measured; the method of instrumental variables in econometrics (Bowden
and Turkington, 1984) is a famous illustration. May the same not be the case with identifying
causal structure? Where have the scores of textbooks so confident of the negative answer
hidden their proofs? Or consider (v). The appeal to unrepresentative samples is as red a herring
as can be. Virtually every statistical estimation procedure will be unreliable if the sampling
Introduction 3
procedures are inappropriate, and bad luck in the draw can lead any estimation procedure
astray. There is something disingenuous in applying to causal inference a demand for reliability
that would be dismissed out of hand as a requirement on statistical estimation. Only (vi) makes
a sound point. Yule (1926) noted that time series samples of variables that tend to increase or
decrease will be correlated. One might conclude that such correlations are due to remote
unmeasured common causes, but for unspecified reasons Yule rejected that hypothesis. Non-
stationary time series are special in several ways; they represent, for instance, a mixture of
distributions. One might reasonably wonder what conditions exclude such cases, and whether
there are specifiable general conditions on distributions, or distributions and causal structures,
that, when they obtain, make causal inference possible in principle; one might reasonably
wonder whether regression or other standard methods are optimal for such inferences. Such
questions are never raised in the textbooks. For comparison, what would one think of a
textbook on estimation that, knowing of circumstances in which an estimator is biased or
inconsistent, refuses to discuss circumstances in which it is neither, or that, noting that some
latent variable models are not identifiable, concludes that none are?
Pedagogy reflects accepted statistical theory, and with some important exceptions, statistical
theory tiptoes around issues of causal inference. Even so, many research issues in statistics are
fundamentally motivated by problems about reliable causal inference, although the motivation is
sometimes hidden by the details.
The statistical literature on collapsibility provides one illustration. Some years ago the National
Halothane Study (Bunker, Forrest, Mosteller and Vandam, 1969) found itself with more
relevant variables than could be analyzed by methods then available. At stake were a set of
causal questions about the effects and side-effects of alternative anesthetics. The problem
suggested an important theoretical question: when can the same conclusions about the existence
and strength of an influence of one variable, A, on another, B, be obtained by analyzing a
reduced set of variables, K (containing A and B), rather than a larger set of variables that
properly includes K? A little more exactly the question is this: when and how can the analysis
of a set K of variables, including A and B, reliably determine whether variable A causes B,
even though there may be common causes of A and B that are not in K; and when and how, in
such circumstances, can we reliably predict the effect on B of manipulating A? The question
isn't peculiar to the halothane study; the issue arises in nearly every non-experimental study that
tries to establish or assess a causal relationship. The question haunts non-experimental
epidemiology. A pessimistic answer was part of Fisher's (1959) criticism of epidemiological
arguments that smoking causes lung cancer.
Motivated in part by the problems of the halothane study, the influential book of Bishop,
Fienberg and Holland (1975) replaced the question of identifying and estimating causal
influence in the presence of latent variables with another question altogether: when are the log-
linear parameters for a model obtained by marginalizing out some variables the same as the
corresponding parameters of the larger, unmarginalized model?1 Other authors have produced
variants of this collapsibility question, sometimes outside the log-linear formalism (Asmussen
and Edwards, 1983; Whittaker, 1990). The various conceptions of collapsibility have led to
some nice mathematical work, but their bearing on the original issue is not clear; there is no
reason to think that when a log-linear statistical model for a causally related set of variables
collapses over a subset of the variables, the collapsed log-linear parameters correctly
characterize anything about the causal relations among the variables in the marginal set
"Model selection" or "specification search" is an area of research in which questions of causal

inference have been badly obscured. Statistical "models," whether log-linear models, structural
equation models, regression models, or whatever, often have two distinct roles. One role is to
restrict the class of possible probability distributions among a set of variables and to parametrize
the family of distributions that satisfy the restriction. Thus a log-linear model, for example, is
given by specifying that particular parameters vanish in a linear expansion of the logarithm of
the probability of any cell. The importance of hypothesis selection in this respect is that the
restrictions and the parametrization should aid one in understanding and efficiently estimating
the distribution. The other role such models may have is to inform prediction; in econometrics,
epidemiology, market research and elsewhere, the predictions are often about the effects of
actions or events that, if they were to occur, would alter the probability distribution. These are
causal claims; they do not follow from any estimate of an actual probability distribution, and
they depend on a further interpretation of the representations through which the restrictions of a
statistical model are expressed. Statistical hypotheses used with a causal interpretation would
seem either to be correct or to be incorrect, and the difference is important: smoking does cause
lung cancer; ether, not halothane, was the riskier anesthetic.
We naive outsiders might therefore expect that research on the selection of linear or logistic
regressors, the selection of log-linear parameters, the modification of structural equation
parameters, and so forth, would pursue model selection as a kind of estimation problem for
causal structure, analogous to the estimation of features of probability distributions. One might
expect theoretical work to investigate conditions under which procedures can be found that give
correct information about structure in the large sample limit, to characterize their error
1We are indebted to Steve Fienberg for this account of the genesis of collapsibility questions.
Introduction 5
probabilities in various cases, and to describe other computational and statistical properties of
such procedures. But there is little if any work of this sort in the statistical literature. Instead,
model selection and specification search have been treated quite differently from estimation.
While texts and papers sometimes note conditions under which particular search procedures--
stepwise regression, for example--may fail, search methods come without any kind of relevant
assurances about their asymptotic reliability. There are not even very many large simulation
studies of the reliabilities of common methods for specifying models.
Any number of other topics in statistics that concern the relation between statistical dependency,
on the one side, and causal dependency on the other, have been addressed without the guidance
of any theory of that relation. They include the rare discussions of statistically indistinguishable
structural equation models (Basmann, 1965; Stetzl, 1986; Joreskog, 1990), discussions of
Simpson's "paradox," and many issues of experimental design, including retrospective versus
prospective sampling, and randomization in experiments.
Why is so much of statistical application and so little of statistical theory concerned with causal
inference? One reason sometimes given for avoiding any attempt at a mathematical analysis
joining causality and probability is that the idea of causality involves a lot of metaphysical murk
which a mathematical science does well to avoid. Some people try to, give accounts of causation
entirely in terms of probability relations, while others try to characterize causation in terms of
counterfactual conditions. Who is to say which is right, or if the counterfactual characterizations
are even partly right, what exactly is meant? (We, certainly, have no definition of causation to
promote.) But surely, the thought continues, there cannot be any rigorous theory about what is
undefmed.
This explanation of the neglect of causation in statistical theory is unsatisfactory on two counts.
First, the absence of a "definition" of causation doesn't keep statisticians from talking
comfortably about causal inference in experimental contexts, and surely the very fact that one
variable causes another cannot always depend essentially on whether we discover the fact by
experimentation. Second, while the notion of causality is vague in many respects, and wrapped
in metaphysical disputes, so is the idea of probability. Every interpretation of probability
appeals to obscure counterfactual assumptions or to mysterious properties. The classical
definition appealed to "equipossible" cases. Limiting frequency interpretations must associate
real finite empirical sequences with imaginary infinite sequences. Subjective Bayesian
interpretations -rest on a particularly obscure psychological notion, belief, and require
assignments of degrees of belief that, in complex cases, no cognitively limited human being can
instantiate (see, for example, Fine, 1973). Few Bayesian models in the literature are
specifications of the actual beliefs of any person, a fact sometimes sidestepped by claims that
probabilities are degrees of belief of ideally rational agents, yet another class of non-existents.
The obscurity of the fundamental notion of probability has not prevented the idea from bearing
enormous fruit, and one important step in that process was to give a clear mathematical form to
the theory. That happened first through the analytic characterization of special probability
distributions and their properties, and later through the theory of measure, the Kolmogorov
axioms and their variants. Notions of causality have paralleled and motivated developments in
the theory of probability since the 17th century, and are undeniably entangled with probability
assessments in scientific practice. Why, therefore, should we not give notions of causal
dependence a clear mathematical form and make the relation of the causal and stochastic
formalisms explicit in a way that reflects scientific practice? There are two attempts in the
statistical literature to do that very thing, each valuable and neither sufficient of itself.
A mathematical representation of causal dependencies among a set of variables has been in the
statistical literature for most of this century. Sewell Wright (1934) used directed graphs to
represent causal structures. The vertices of the graph represent variables and a directed edge
from one variable to another represents the claim that the first variable has a direct influence on
the second, an influence not blocked by holding constant all other variables considered. Ever
since, directed graphs have occasionally been used in representations of regression models,
factor models, simultaneous equation models, time series models, and elsewhere. From a
mathematical point of view, directed graphs are implicit whenever a statistical model is specified
through a set of algebraic equations in which a single variable occurs on one side and is treated
as the effect of variables on the other side.
Of itself, the mathematical representation of causal influence by directed graphs is trivial.

Things only get interesting when some condition is given connecting the graphical structure
with restrictions on probabilities. Some such connection was already implicit in the use of linear
models in the social and behavioral sciences, for the system of equations and independence
assumptions about error variables always determined a directed graph, sometimes given
explicitly, and entailed constraints on correlations and partial correlations (Simon, 1954;
Blalock, 1961), which for normal distributions is the same as constraints on independence and
conditional independence. But social scientists did not articulate any general principle
connecting their graphs with their probability distributions, and the literature developed no
further than analyses of various particular cases (Blalock, 1971). Without using graphical
representations, a few philosophers of science, most notably Reichenbach (1956) and Suppes
(1970) attempted to give analyses of the very notion of causality in terms of statistical
dependencies (and, in Suppes' case, time order). An explicit, general mathematical connection
Introduction 7
between directed graphs and probability distributions was, however, not introduced until about
ten years ago.
Kiiveri and Speed (1982) related causal dependencies, represented as directed graphs withou:
cycles, to conditional independence constraints. They formulated several equivalent versions of
the idea that if Y is not a cause of X and X influences Y, if at all, only through an intermediary
set Z of direct causes of Y, so that if the variables in Z are held constant no variation in X will
produce a variation in Y, then X and Y are independent conditional on Z. Formal statements of
the idea are called Markov conditions, and we use one such formulation in this book. Kiiveri
and Speed showed that social scientists' claims about vanishing correlations and partial
correlations required by various linear models follow from the Markov Condition. They further
showed that any strictly positive probability density satisfying the Markov condition for a
directed acyclic graph admits a "factorization" determined by the structure of the graph: any joint
density satisfying the Markov condition for a directed graph must equal a product of terms, one
term for each variable, with each term giving the conditional probability of that variable on its
parent variables in the graph. Thus for discrete variables if the graph is
Figure 1
the joint distribution must satisfy
P(X,Y,Z) =P(X) P(YIX) P(ZIY).
The conditional independence constraints required by the Markov condition applied to the graph
are a consequence of the factorization. A widespread practice in applied statistics was thus given
an elegant formal foundation and connected with statistical work on factorization of
distributions that had previously been carried out by Wermuth (1980) and others.
Consequences of the Markov Condition and additional constraints on the connection between
directed graphs and distributions were subsequently developed by a number of authors
(Wermuth and Lauritzen, 1983; Pearl, 1988; Wermuth and Lauritzen, 1990).
There is another thread in statistics connecting probability and causality, a connection that
emphasizes relations between the notion of causation and the predictability of the effects of
manipulations. It arose from work on experimental design. As early as 1935, Neyman noted a
counterfactual aspect to conclusions drawn from an experimental study in which some units are
treated one way and some another. Often the conclusion of such studies is about what would
happen ifall units were treated the same way, and is therefore in a sense about a condition the
experiment did not examine, or about a distribution it did not actually sample. Donald Rubin
(1974, 1977, 1978, 1986), and following him several others (Holland, 1986; Pratt and
Schlaifer, 1988), have interpreted causal hypotheses as postulating a family of random
variables, some of which never have their values observed. In an experiment, random variables
whose values are never observed represent the value an outcome variable would have (but
doesn't in fact have) for a unit subjected to one treatment if that unit had instead been subjected
to another treatment condition of the experiment. Various assumptions may constrain the
relations between the contrary-to-fact random variables and random variables whose values are
observed; for example, it might be assumed that if a treatment were applied to one unit, the
treatment of that unit would have no effect on the outcome of applying the same treatment to
other units. Rubin has used the framework to argue for the importance of randomization in
experimental design and to give methods for estimating the effect of one variable on another by
means of trials in which treatment is determined by the value of a variable that covaries with the
outcome variable. Pratt and Schlaifer have given rules for predicting the invariance of
conditional probabilities under interventions or manipulations: when does P(y), in a population
in which a variable X is/orced to have a value x, equal P(YIX=x) in the population of
unmanipulated units? While intuitively correct, their rules were not explicitly derived from any
general principle.
From our perspective, the Rubin, Holland, Pratt and Schlaifer theory is essentially an account
of special cases in which the effects of an intervention or manipulation can be predicted. The
work makes no use of directed graphical methods, but it makes a great deal of use of
conditional independence relations. And that ties the Rubin framework to the Markov
Condition. For a causal structure the Markov condition provides a factorization of the
probability distribution in terms of the conditional probability of each variable X on its parents:
P(X1Vt ... VA;). We can think of a "direct manipulation" of X as an intervention that changes
P(XIVt ... VA;) to some other distribution P*(X1Vt ... Vk) but leaves the other conditional
probabilities in the original factorization unaltered. The result of a direct manipulation is thus a
new distribution with a new factorization, obtained by replacing P (XIV t ..• VA;) by
P*(XIVI .. .vA;). So for example, suppose the causal structure is X <- Z -> Y with factorization
P(XIZ)P(YlZ)P(Z). Let some intervention change the distribution of X to P*(XIZ). Then if the
intervention were applied to all units in the population the new joint distribution would be
P*(X,Y,Z) = P*(XIZ)P(YlZ)P(Z). In this example the effect on Y of such an intervention is
trivial to predict since the marginal distribution of Y is unchanged.
Introduction 9
With an appropriate definition of "manipulation" the Markov Condition entails a generalization

of the principle just illustrated, a result we call the Manipulation Theorem. Although it seems
never to have been formulated, instances of the theorem are implicit in some treatments of
shocks in econometrics and in many discussions of experimental design. Rubin's (1977) claims
about unbiased estimates of differential effects from alternative manipulations and Pratt and
Schlaifer's rules all follow immediately from the assumption that the interventions in question
are special cases of "manipUlations." As we will show in Chapter 7, their analysis applies in the
special case in which the intervention makes X statistically independent of its causes in the
original unmanipulated system.
The Markov Condition is not given by God; it can fail for various reasons we will discuss in the
course of this book. The reliability of inferences based upon the Condition is only guaranteed if
substantive assumptions obtain. But the Condition is weak enough that there is often reason to
think it applies. In most of the investigations in this book we combine the Markov condition
with a further condition that assumes that all conditional independence relations among variables
occur because of the Markov Condition applied to the graph of causal relations among the
variables. This assumption, which we call the Faithfulness Condition, can be thought of
formally as the claim that when a causal graph is associated with a probability distribution, the
Markov Condition applied to the graph characterizes all of the conditional independence
relations that hold in the distribution. Informally, the Faithfulness Condition can be thought of
as the assumption that conditional independence relations are due to causal structure rather that
to accidents of parameter values. Sometimes we investigate consequences of the special case
that assumes the variables are linearly related, and in that context it can be shown that a version
of the Faithfulness Condition holds almost always for a "natural" probability distribution over
the parameters.
The immediate value of the unification of the graphical and experimental design approaches to
statistics and causality is aesthetic: two distinct theories are unified in a simple way, and
principles that appear ad hoc are derived uniformly. But aesthetics is the least of it; the real value
of the unified theory lies in recasting a number of research topics so that fundamental issues can
be addressed more directly, and in providing new algorithmic techniques for practical inference
problems.
The Markov and Faithfulness Conditions, and their consequences, provide a framework for
analyzing many of the questions about causality that we think are fundamental to the application
of statistics: When are regression methods reliable for causal inference? Under what conditions
are they optimal? Are more reliable methods possible? Are various model selection procedures
reliable? Are there more reliable methods? When are two alternative causal structures statistically
indistinguishable? When can the presence of latent variables be detected from the statistics? Can
predictions reliably be made when the causal structure is not known? Can predictions reliably be
made when it is not known beforehand whether there may be latent variables acting to produce
measured statistical dependencies? How does the reliability of causal inference depend on
sampling procedure? Is retrospective sampling as useful for causal inference as prospective
sampling? What discriminations about causal structure can be made by experiment but not by
observation, or--incredible thought--vice-versa? Under what additional assumptions can what
features of latent structure be reliably inferred? The unification of the graphical and experimental
design frameworks provides results pertinent to all of these questions. Sometimes the results
are versions of statistical cornmon sense, but rather often they reverse received opinions. The
results we have obtained are in many respects incomplete, but they are complete enough to
demonstrate that issues of causation, prediction and search form an area of research that is
fundamental to many statistical problems, and to illustrate that within the unified theory there is
no need either to ignore the essential questions about reliability or to have recourse to ersatz
technical questions.
There are two sorts of practical consequences of the theory. The unified theory yields precise
conditions under which a variety of causal inferences cannot be reliably made, and the unified
theory yields algorithms for inferring causal structure from statistical decisions, algorithms that,
under appropriate assumptions, can be shown in a special but well defined sense to be
asymptotically reliable. With such proofs in hand, statistics can collaborate with computer
science, and reliable algorithms that are computationally feasible and that require feasible sets of
statistical decisions can be obtained. By implementing these algorithms in a computer, the
reliabilities of the procedures on small and medium samples, and their robustness to violations
of the assumptions, can be explored by simulation methods. Most of the algorithms we have
derived from the theory have been implemented in experimental versions of the TETRAD II
program, and the results of various simulation tests are described in the course of this book.
1.2 Advertisements
Each of the examples in this section is also treated in other chapters. Their function here is only
to tempt the reader to muster the patience to work through a long book with a lot of
mathematics. Although the examples are not themselves proofs of the reliability of anything,
Introduction 11
most of them illustrate procedures for which we will later give proofs of reliability assuming
only the Markov and Faithfulness Conditions and the capacity to make correct statistical
decisions.
1.2.1 Bayes Networks from the Data
The diagram shown below, called the ALARM network, was developed to model an emergency
medical system (Beinlich, et al. 1989). The variables are all discrete, taking 2, 3 or 4 distinct
values. In most instances a directed arrow indicates that one variable is regarded as a cause of
another. The physicians who built the network also assigned it a probability distribution: each
variable V is given a probability distribution conditional on each vector of values of the
variables having edges directed into V. One use of such networks is to compute the probabilities
of values of some variables for any new unit from measurements of values (for that unit) of
other variables in the network. Of course such networks also constitute statistical models of
discrete data, and as such they are alternatives to log-linear models, logistic regression models,
or undirected graphical independence models. The directed graph has 37 variables and 46
edges. Herskovits and Cooper (1990) used the diagram to generate simulated emergency
medicine statistics for 20,000 individuals.
For many statistical search procedures 37 might as well be infinity. Commonly recommended
procedures for searching for models of discrete data in popular formalisms are stopped dead at
ten or twelve variables. In contrast, Herskovits and Cooper describe a fast Bayesian procedure
that, given a prior linear ordering of the variables consistent with the causal directions, makes
only a single error in recovering the adjacencies in this graph from the discrete sample statistics.
We will describe a procedure that recovers almost all of the ALARM network--including both
the adjacencies and the directions of the edges--from the sample data. For example, there is an
implementation on an ordinary workstation2 that recovers most of the network in less than
fifteen seconds from data generated by treating the dependencies in the graph as linear. For
Herskovits and Cooper's discrete data generated from the same network, most of the structure
can be recovered by computer with comparable reliability in a few minutes. Maximum
likelihood estimates of the probabilities can be obtained directly. We emphasize that in these
studies the procedure was not given any prior information about the graph, and that in most
cases the computer determined the directions of the arrows.
2Decstation 3100.
The ALARM belief network
KEY:
1 - central venous pressure 20 - insufficient anesthesia or
analgesia
2 - pulmonary capillary wedge pressure 21 - pulmonary embolus
3 - history of left ventricular failure 22 - intubation status
4 - total peripheral resistance 23 - kinked ventilation tube
5 - blood pressure 24 - disconnected ventilation tube
6 - cardiac output 25 - left-ventricular end - diastolic
volume
7 - heart rate obtained from blood pressure 26 - stroke volume
monitor
8 - heart rate obtained from electrocardiogram 27 - catecholamine level
9 - heart rate obtained from oximeter 28 - error in heart rate reading due to
low cardiac output
10 - pulmonary artery pressure 29 - true heart rate
11 - arterial-blood oxygen saturation 30 - error in heart rate reading due to
electrocautery device
12 - fraction of oxygen in inspired gas
13 - ventilation pressure 31 - shunt
14 - carbon-dioxide content of expired gas 32 - pulmonary-artery oxygen
saturation
15 - minute volume, measured 33 - arterial carbon-dioxide content
16 - minute volume, calculated 34 - alveolar ventilation
17 - hypovolemia 35 - pulmonary ventilation
18 - left-ventricular failure 36 - ventilation measured at
endotracheal tube
19 - anaphylaxis 37 - minute ventilation measured at
the ventilator
Figure 2
Introduction 13
1.2.2 Structural Equation Models from the Data
Rodgers and Maranto (1989) were interested in why some academic psychologists publish
more than others. Is it abilities people already have before entering graduate school? Gender?
The quality of their graduate training? The habits of publishing formed in graduate school? The
quality of their fIrst job? Rodgers and Maranto organized a survey and collected data on each of
these variables, as well as others. Common sense determines the direction of most possible
causal connections among their variables. For example, if gender and some of the other
variables are causally connected, it cannot be because the other variables cause gender. Quality
of fIrst job, or publication rate after graduation cannot cause the quality of someone's graduate
program. Rodgers and Maranto adapted two theories from economics to form two sets of linear
equations, another from sociology, and still another from social psychology, estimating and
testing each separate system of equations. Each of these theories failed a goodness of fIt test.
Combining the dependencies postulated in these rejected models, and adding more
dependencies for good measure because they seemed plausible, Rodgers and Maranto then
obtain a new model which they estimated, tested, and presented as a graph with eleven direct
causal dependencies and associated standardized coefficients:
.13
CITES
Figure 3
GPQ =graduate program quality PUBS =publications since graduate school

QFJ =quality of fIrst job CITES =citation requency
PREPROD = publications in graduate school
Rodgers and Maranto prefaced their paper with the remark that their elaborate exercise was
necessary because causal dependencies cannot be inferred from the probabilities, but only from
"theory." But given the common sense time order of the variables, a computer running a search
algorithm will in a few seconds produce ten of the eleven hypothetical connections directly from
their data, leaving out only one of the three dependencies associated with the smallest estimated
linear coefficients.
Chapters 5 and 8 illustrate a variety of cases in which features of linear models that have been
justified at length on theoretical grounds are produced immediately from empirical covariances
by the procedures we describe. We also describe cases in which the algorithms produce
plausible alternative models that show various conclusions in the social scientific literature to be
unsupported by the data.
1.2.3 Selection of Regressors
Regression may be the method most commonly used in empirical studies to yield predictions
about the effects of policies, often from variables measured roughly simultaneously and often
without guarantees that no unmeasured factors affect the outcome variable and one or more
regressors. Regression is something of a disaster when in such roles, and it is difficult to justify
the prominence it is given in both pedagogy and arguments over policy.
Consider the following example, given by a set of linear equations with error terms and an
unmeasured variable T, and also by a directed graph indicating the causal dependencies
assumed for each unit in a population. The equations are
Y=aIXI +a2Xs +a3T+ey

X I = a~2 + as X4 + EI
X3 = a6X2 + a7T + E3
The directed graph, with the error terms not represented, is:
Introduction 15
Figure 4
For definiteness, assume that all exogenous variables, including the error terms, are
independent and distributed normally with zero mean and unit variance, and that none of the ai
vanish. Consider the multiple regression of Y on X l-XS in such a population. The regression
coefficient for X3 provides an unreliable (biased, inconsistent) estimate of the influence of X3
on Y, and this fact is often remarked in textbooks, usually in terms that do not explicitly
mention causation. Perhaps exactly because they avoid explicit discussions of causality,
however, the textbooks3 miss the more important fact that when a regressor X and an outcome
variable Y have an unmeasured common cause, the estimate of the influence on Y of every other
regressor that directly influences X or has a common unmeasured cause with X will likewise be
unreliable. In the example in figure 4, even though only XI and Xs are direct causes of Y, not
mediated by other variables, only the regression coefficient for X4 will be zero. For suitable
ranges of values of the exogenous variances and linear coefficients, the regression coefficient of
X2 may even have a larger absolute value than that for X 1. A similar phenomenon can occur if
one of the regressors is actually an effect, rather than a cause, of the outcome variable, a
circumstance that may not be uncommon in uncontrolled studies. As we will show with
simulation studies later in this book, the usual computerized model selection techniques for
regression typically fail in such cases; in fact, in simulation studies with large samples these
methods do worse than simply regressing on all variables and choosing the significant
regressors.
The unified theory yields an algorithm that under the statistical assumptions commonly made
when such models are used (linearity, homoscedasticity, etc.) asymptotically produces correct
information about the regressor structure. In the example just considered the procedure yields
the object in figure 5, which we call a partially oriented inducing path graph.:
30ur nonrandom search through about 30 textbooks on regression found no mention of the fact in any book.
Figure 5
The double-headed arrow indicates the presence of unmeasured common causes; the circles at
the ends of edges indicate that the algorithm cannot determine whether there should be
arrowheads at those ends; that is, whether the 0-> edge should be -> or <->. The bar over X2
indicates that not both of the small 0 marks on edges adjacent to X2 may be arrowheads. If
obtained in a context in which the Markov and Faithfulness Conditions are warranted, this
graph would allow us to conclude that Xl is a direct cause of Y, that X5 may be a direct cause
of Y, thatX3 does not cause Yand Y does not causeX3, thatX3 and Yare effects of a common
cause, and that no other X variables are direct causes of Y.
Consider two empirical examples, which are described in more detail in Chapter 8. In the first,
a sample of test results for more than 6,000 people includes values for each subject on seven
test scores and on a combined score, AFQT, which is an average of three of the seven test
scores as well as of other tests not included in the data. The dependency of AFQT on the other
recorded tests is therefore linear. The problem is to identify which of the seven tests are
components of AFQT. Linear multiple regression of AFQT on the other seven test scores gives
significant regression coefficients to all of them. In contrast, an algorithm we will derive from
the Markov and Faithfulness conditions correctly finds the three tests that are components of
AFQT. In this case regression probably fails because the variables are related by an intricate
structure of unmeasured common causes.
Rawlings (1988) describes a study (Linthurst, 1979) of 45 samples of Spartina grass from the
Cape Fear Estuary. Besides the biomass of Spartina, 14 other variables were measured which
might be thought to be relevant to growth. A linear multiple regression of biomass on the
fourteen other variables found only two nutrient variables, copper and potassium concentration,
to have significant coefficients, a result that is not plausible on biological grounds. With the
same data, the alternative algorithm finds that pH controls growth in the sample, a result
confirmed by experiment In this case multiple regression may fail because significance tests of
each regression coefficient must control for 13 other variables, which effectively reduces the
Introduction 17
sample size from 45 to 32 and results in tests with little power against alternatives in which the
partial correlations are not very large. In contrast the algorithms we use never require tests of
vanishing partial correlations that control for more than one other variable in this case .
1.2.4 Causal Inference without Experiment
In his criticism of the epidemiological literature on smoking and lung cancer, Fisher (1959)
emphasized that the correlation of two variables cannot distinguish a direct effect from an
unmeasured common cause; the complaint was echoed in Brownlee's (1965) review in the
Journal of the American Statistical Association of the first Surgeon General's Report on
Smoking and Health. Fisher and Brownlee were, of course, both correct, but consider the cases
illustrated in figure 6:
(ii)
Figure 6
Suppose B is unmeasured. Fisher's point is that non-experimental studies of A and C cannot

distinguish hypothesis (i) from hypothesis (ii), even if every unit in the population has the very
same causal structure, but experiment can. But there are some points Fisher passed by.
If variable U is measured and causes A, (or has common causes with A ), and U affects Conly
through A, as in the figures with an asterisk, then structures (i*), and (ii*) can be distinguished
without experimental controls. It is not necessary that these restrictions on U be known
independently of the data; all that is necessary is that they be true of the causal process
generating the distribution.
Brownlee emphasized the case in which smoking is a cause of cancer, and smoking and cancer
also have unmeasured common causes. As we will see in Chapter 9, that case presents
interesting complexities, but with modest prior knowledge it too can be distinguished from
structures (i) and (ii) by appropriate observations.
The common presumption is that non-experimental evidence cannot determine whether

measured statistical dependencies are due to unmeasured factors: one must decide on non-
statistical grounds that there are unmeasured common causes at work producing statistical
dependencies in a population. We can't help but think that in practice such decisions are often
influenced more by convenience and the customs of disciplines than by any real knowledge.
But under quite general conditions, whether for continuous linearly related variables or for
discrete variables, the presumption is false. There are cases in which detailed inferences about
causal structure can be made without any prior assumption as to whether or not unmeasured
factors are at work, and in some of the same cases reliable predictions can be made about the
effects of policies that directly manipulate certain variables while leaving the causal structure
otherwise unaltered. Consider, for example, the following entirely hypothetical causal structure
related to Measured breathing dysfunction, in which the variables in rectangles are measured,
and those in ovals (Environmental Pollution and Genotype) are not.
Figure 7
Introduction 19
Suppose the structure in figure 7 represents the true causal dependencies among these variables
in a popUlation from which samples are drawn according to a multinomial distribution. The
Markov and Faithfulness Conditions yield an algorithm that from the relations of conditional
independence and dependence among the measured variables yields the graph in figure 8.
14----t~Lung capacity
~~
Figure 8
A double-headed arrow indicates the presence of unmeasured common causes; the circles at the
ends of two of the edges indicate that the algorithm cannot determine whether there should be
arrowheads at those ends. In this hypothetical example the statistics suffice to drive the
conclusion that smoking indirectly affects lung capacity, and the statistics also answer most
other questions about the causal dependencies among the variables, without any prior
assumption of the existence or of the non-existence of unmeasured factors. Given the
information about the population contained in the diagram in figure 8, and given adequate data,
the effects of smoking cessation on cilia damage, heart disease, lung capacity and breathing
dysfunction could be reliably predicted. Variants of the algorithm apply to continuous, linearly
related variables.
1.2.5 The Structure of the Unobserved
Investigators who use psychometric or sociometric tests or questionnaires often have hundreds
of item responses which in themselves are of no interest What is of interest is the nature and
causal relations of the features of persons or systems that the items indicate. On substantive or
other grounds it may be believed that various items form clusters that have a common
unmeasured cause. The clusters may, of course, be heavily confounded: some of the items in the
same cluster may have other common causes besides those common to all members of the
cluster, and some members of a cluster may be affected by responses to other items, whether in
that cluster or some other. Even if simplifying assumptions such as linearity are made, it is often
thought to be utterly hopeless to extract from assumptions about the clusters and from the data
any reliable conclusions about the causal relations of the unmeasured causes of the respective
clusters. But provided the variables are to good approximation linearly related (or binary) and
provided the clusters are not too confounded, reliable information about the causal relations
among unmeasured variables can be obtained. If those causal relations are sufficiently sparse and
the modeling assumptions apply, the causal structure may be identified almost uniquely. The
procedures required have been fully automated. Consider the graph in figure 9, in which five
(Yb Y2, Y4, Ys, and Y13) of the 16 item responses are confounded.
Figure 9
Given as input only the correlations among the Y variables and the four clusters shown, the
procedure we will derive correctly reconstructs the graphical connections among the latent
variables (the "1)" variables) up to alternative orderings of the 1)1 - 1)2 connection, which are
statistically indistinguishable.
Introduction 21
1.3 Themes
A fundamental theme of this book is that there are systematic connections between causal
dependency and stochastic dependency. By making use of these connections, the limits to
reliable causal inference can be established theoretically. What can go wrong in causal inference
and how wrong it can go only become clear once we have a theoretical understanding of the
connection between causality and probability. One important class of limitations can be
established by proving characterizations of equivalence relations between graph theoretic
representations of alternative causal structures. In effect, one investigates the classification of
directed graphs up to various relations of statistical indistinguishability. From the other side,
one can consider characterizing the class of all graphs consistent with a given probability
distribution. And the limits of causal inference can be investigated by characterizing
circumstances in which the typical (in our case axiomatic) connections between causal and
statistical structure are violated, a study begun by Yule and Pearson at the beginning of the
century. Within these limits, we can investigate rigorously the reliabilities and computational
properties of procedures that search for causal structure from statistical properties of samples.
The asymptotic reliabilities of well-defined procedures can be detennined mathematically, while
short run reliabilities are estimated experimentally through simulation studies. The theory of
model specification search consequently takes on some of the features of the theory of
estimation.
Most current statistical modeling and search procedures can analyze only rather small models.
Popular search and estimation procedures for log-linear models, for example, are stopped for
realistic sample sizes at about a dozen variables or even fewer, depending on the number of
values the variables may assume. By considering efficiencies of search, it is possible to find
algorithms that (under explicit assumptions) will make reliable causal inferences for linear
systems with a hundred or more variables when the causal connections among the variables are
sparse. With discrete variables on sparse structures models can be reliably identified with fifty
or more variables given samples of a few thousand units.
On the subject of latent variables, statistical practice is nearly schizophrenic. Statistical models
are presented either with unmeasured variables or without them, but almost never with any
credible argument from statistical considerations for their inclusion or their exclusion. Various
model construction techniques proceed as if unmeasured common causes were the only possible
kind, or alternatively as if they were absent and entirely irrelevant Such assumptions need not
be left unargued: there exist asymptotically reliable methods to obtain information about the
presence or absence of unmeasured common causes, and about their causal relations.
Informative sufficient conditions exist for the presence of unmeasured common causes,
assuming the Markov and Faithfulness Conditions. These conditions in turn yield theorems
about causal conclusions and predictions that can be reliably drawn whether or not latent
variables are present. If linearity is assumed then there are more powerful theorems,
summarized in what we call the Tetrad Representation Theorem, that can be used to identify the
presence of unmeasured common causes.
A further theme of this book concerns the conditions sufficient (or necessary) for correctly
predicting the effects of a policy applied to a population that has been studied through a sample,
and whose causal structure is not known prior to investigation. Statistical writers such as
Rubin, Pratt, and Schlaifer have emphasized questions such as: When will the conditional
distribution of Yon Z when X is/orced to have a certain value equal the probability of Y
conditional on X, Z in an observational or experimental population? Their answers assume that
various counterfactual claims are known. In our terms, this amounts to knowing aspects of the
causal structure of the systems under study. A natural question that follows is: When can the
relevant causal knowledge necessary to answer the question above be obtained from sample
data? We will describe results that answer this question and give information from sample data
about when the conditional probability of Y on Z and X is invariant under a manipulation of X.
In our view, however, the fundamental issue about prediction is this: If the distribution 0/
Xb ... ,Xn is to be directly manipulated, when can the resulting distribution 0/ a set Yo/
variables conditional on a set Z be calculated from the distribution 0/ Y, Z, X b ... , Xn and
other variables in an observational or experimental population in which Xl .. X n were not so
manipulated/or each unit in the sample? The formal connections between probability and causal
structure determine an answer, and we will unravel part of it.
The rigorous investigation of the reliabilities of search algorithms leads directly to the
conclusion that commonly used statistical search procedures are sub-optimal for causal
inference. The criticism falls most heavily on regression. Better methods are available and easily
applied to many regression problems. The best-founded objection to automated model search
procedures, especially in linear and logistic regression, is that the procedures are asymptotically
unreliable against alternative causal hypotheses that are often consistent with prior knowledge.
If procedures give the right answers in the ideal case of perfect information about the population
distribution, one can look around for better tests and more computationally efficient algorithms.
But if, as in the case of regression and many other automated techniques, probability relations
Introduction 23
and causal relations are incorrectly matched, all of the statistical subtlety in the world won't
make for good inference.
Yet another theme of this book is the importance of causal reasoning in the design of empirical
studies. The truth of the adage that correlation is not causation may be an obstacle to thinking
through what can be determined about causal dependencies from statistical dependencies, and
under what conditions. A recent textbook (Christensen, 1990, p. 279) on log-linear methods
opens and closes the question with a remark that we think would be endorsed by the majority of
statisticians:
Causation is not something that can be established by data analysis. Establishing

causation requires logical arguments that go beyond the realm of numerical manipulation.
For example, a well designed randomized experiment can be the basis for conclusions of
causality but the analysis of an observational study yields information only on
correlations. When observational studies are used as a basis for causal inference the jump
from correlation to causation must be made on non statistical grounds.
The passage contrasts the information about causal structure in experimental studies with the
information about causal structure in observational studies, and the view it expresses is
perfectly standard. But is it correct? Once a formal understanding of the connection between
causal structure and probability is in place, questions about the comparative power of
experiment versus observation can be answered by a mathematical study of the causal
information that can be extracted from experimental and from observational designs. The
standard claims of the power of experiment and the impotence of observation tum out not to be
so much false as misleading. There prove to be systematic parallels between the kinds of causal
conclusions and predictions that can be drawn from experimental and observational data
respectively. An intricate (and not yet complete) theory is required to understand the
implications of the design of empirical studies for reliable causal inference and prediction.
The results we obtain about inference and prediction have interesting implications for a number
of controversial subjects concerning the design of experiments, including the problem of
designing "ethical" clinical trials in which subjects' preferences can influence the treatment they
receive. We will describe a design that, depending on various discoverable empirical facts,
permits patient preferences to have such a role without any loss in the power to obtain relevant
information from experimental outcomes.
Although the methods we use are non-Bayesian, the broad issues of the subject cut across
divisions between Bayesian and non-Bayesian statistics, and we will indicate some ways in
which related inference results could be obtained by Bayesian procedures. The mathematical
methods appropriate to the subject include more of graph theory and computation theory than
measure theory. The empirical methods make heavy use of computer simulations to give
evidence of reliability where analytic results are unavailable. Various pieces of methodological
folklore are contradicted by the results already available in the subject, while others are given a
new perspective and significance.
Whatever value the results in this book may have for practical scientific inference, together they
illustrate a systematic and comparatively neglected area of inquiry that investigates causation,
prediction and search. The subject is full of well-formed and almost well-formed open
questions about systems with feedback, necessary and sufficient conditions for certain kinds of
inferences, short-run reliabilities of procedures, the existence of optimal search procedures,
optimal statistical decisions, indistinguishability properties of models, trade-offs between
informativeness and computational feasibility, and more.
Chapter 2
Formal Preliminaries
This chapter introduces some mathematical concepts used throughout the book. The chapter is
meant to provide mathematically explicit definitions of the formal apparatus we use. It may be
skipped in a first reading and referred to as needed, although the reader should be warned that
for good reason we occasionally use nonstandard defmitions of standard notions in graph
theory. We assume the reader has some background in finite mathematics and statistics,
including correlation analysis, but otherwise this chapter contains all of the mathematical
concepts needed in this book. Some of the same mathematical objects defmed here are given
special interpretations in the next chapter, but here we treat everything entirely formally.
We consider a number of different kinds of graphs: directed graphs, undirected graphs,

inducing path graphs, partially oriented inducing path graphs, and patterns. These different
kinds of objects all contain a set of vertices and a set of edges. They differ in the kinds of edges
they contain. Despite these differences, many graphical concepts such as undirected path,
directed path, parent, etc., can be defmed uniformly for all of these different kinds of objects.
In order to provide this uniformity for the objects we need in our work, we modify the
customary definitions in the theory of graphs.
2.1 Graphs
The undirected graph shown in figure 1 contains only undirected edges (e.g. A - B).
Figure 1
A directed graph, shown in figure 2, contains only directed edges, (e.g. A -> B).
Figure 2
An inducing path graph, shown in figure 3, contains both directed edges (e.g. A -> B) and bi-
directed edges (e.g. B <-> C). (Inducing path graphs and their uses are explained in detail in
Chapter 6.)
Figure 3
A partially oriented inducing path graph, shown in figure 4, contains directed edges (e.g. B ->
F), bi-directed edges (e.g B <-> C), nondirected edges (e.g. E 0-0 D), and partially directed
Fonnal Preliminaries 27
edges (e.g. A 0-> B.). (partially oriented inducing path graphs and their uses are explained in
detail in Chapter 6.)
F E
Ao
r
.B • .C • oD
I
Figure 4
A pattern, shown in figure 5, contains undirected edges (e.g., A - B) and directed edges (e.g. A
-> E). (Patterns and their uses are explained in detail in Chapter 5.)
Figure 5
In the usual graph theoretic defmition, a graph is an ordered pair <V,E> where V is a set of
vertices, and E is a set of edges. The members of E are pairs of vertices (an ordered pair in a
directed graph and an unordered pair in an undirected graph). For example, the edge A -> B is
represented by the ordered pair <A,B>. In directed graphs the ordering of the pair of vertices
representing an edge in effect marks an arrowhead at one end of the edge. For our purposes we
need to represent a larger variety of marks attached to the ends of undirected edges. In general,
we allow that the end of an edge can be unmarked, or can be marked with an arrowhead, or can
be marked with an "0".
In order to specify completely the type of an edge, therefore, we need to specify the variables
and marks at each end. For example, the left end of "A 0-> B" can be represented as the
ordered pair [A, 0]1, and the right end can be represented as the ordered pair [B, >]. The first
member of the ordered pair is called an endpoint of an edge, e.g. in [A, 0] the endpoint is A.
The entire edge is a set of ordered pairs representing the endpoints, e.g. {[A, 0], [B, >]}. The
edge {[B, >],[A, oJ) is the same as {[A, o],[B, >J) since it doesn't matter which end of the
edge is listed first
Note that a directed edge such as A -> B has no mark at the A endpoint; we consider the mark at
the A endpoint to be empty, but when we write out the ordered pair we will use the notation EM
to stand for the empty mark, e.g. [A,EM].
More formally, we say a graph is an ordered triple <V,M,E> where V is a non-empty set of
vertices, M is a non-empty set of marks, and E is a set of sets of ordered pairs of the form
{[VI,MIl,[V2,M2]}, where VI and V2 are in V, VI '# V2, and MI and M2 are in M. Except in
our discussion of systems with feedback we will always assume that in any graph, any pair of
vertices VI and V2 occur in at most one set in E, or, in other words, that there there is at most
one edge between any two vertices. If G = <V,M,E> we say that G is over V.
For example, the directed graph of figure 2 can be represented as <{A,B,C,D,E}, {EM, >},
{{[A,EM],[B, >]), {[A,EM],[E, >]}, {[A,EM],[D, >]}, {[D,EM],[B, >]}, {[D,EM],[C, >]},
{[B,EM],[C, >J), {[E,EM],[C, >J) I>·
Each member {[V}, MI],[V2,M2]) of E is called an edge (e.g. {[A,EM],[B, >J) in figure 2.}
Each ordered pair [V}, Mil in an edge is called an edge-end (e.g. [A,EM] is an edge-end of
{[A,EM],[B, >J).} Each vertex VI in an edge {[VI, MIl,[V2, M2]) is called an endpoint of
the edge (e.g. A is an endpoint of {[A,EM],[B, >]).} VI and V2 are adjacent in G if and only
if there is an edge in E with endpoints VI and V2 (e.g. in figure 2, A and B are adjacent, but A
and C are not.)
An undirected graph is a graph in which the set of marks M = {EM}. A directed graph is
a graph in which the set of marks M = {EM, >} and for each edge in E, one edge-end has mark
EM and the other edge-end has mark ">".
An edge {<[A,EM],[B, >]} is a directed edge from A to B. (Note that in an undirected graph
there are no directed edges.) An edge {[A,MI],[B, >]} is into B. An edge {[A,EM],[B,M2]}
is out of A. If there is a directed edge from A to B then A is a parent of Band B is a child
lit is customary to represent the ordered pair A, B with angle brackets as <A, B>, but for endpoints of an edge
we use square brackets so that the angle brackets will not be misread as arrowheads.
Formal Preliminaries 29
(or daugbter) of B. We denote the set of all parents of vertices in Vas Parents(V) and the set
of all children of vertices in V as Cbildren(V). The in degree of a vertex V is equal to the
number of its parents; the outdegree is equal to the number of its children; and the degree is
equal to the number of vertices adjacent to V. (In a directed graph, the degree of a vertex is
equal to the sum of it's indegree and outdegree.) In figure 2, the parents of B are A and D, and
the child of B is C. Hence, B is of indegree 2, outdegree 1, and degree 3.
We will treat an undirected path in a graph as a sequence of vertices that are adjacent in the
graph. In other words for every pair X, Y adjacent on the path, there is an edge
{[X,Mll,[Y,M2]} in the graph. For example, in figure 2, the sequence <.A,B,C,D> is an
undirected path because each pair of variables adjacent in the sequence (A and B, B and C, and
C and D) have corresponding edges in the graph. The set of edges in a path consists of those
edges whose endpoints are adjacent in the sequence. In figure 2 the edges in path <.A,B,C,D>
are {[A,EM],[B, >]}, {[B,EM],[C, >]}, and {[C, >],[D,EM]}.
More formally, an undirected path between A and B in a graph G is a sequence of vertices

beginning with A and ending with B such that for every pair of vertices X and Y that are
adjacent in the sequence there is an edge {[X,MI],[Y,M2]} in G. An edge {[X,MI],[Y,M2]}
is in path U if and only if X and Yare adjacent to each other (in either order) in U. If an edge
between X and Y is in path U we also say that X and Y are adjacent on U. If the edge
containing X in an undirected path between X and Y is out of X then we say that the path is
out of X; similarly, if the edge containing X in a path between X and Y is into X then we say
that the path is into X. In order to simplify proofs we call a sequence that consists of a single
vertex an empty path. A path that contains no vertex more than once is acyclic; otherwise it
is cyclic. Two paths intersect iff they have a vertex in common; any such common vertex is a
point of intersection. If path U is <U I, ... ,Un> and path V is <Un, ... ,Vm>, then the
concatenation of U and V is <UI, ... ,Un,VI, ... ,V m> denoted by U &V. The
concatenation of U with an empty path is U, and the concatenation of an empty path with U is
U. Ordinarily when we use the term "path" we will mean acyclic path; in referring to cyclic
path we will always use the adjective.
A directed path from A to B in a graph G is a sequence of vertices beginning with A and

ending with B such that for every pair of vertices X, Y, adjacent in the sequence and occurring
in the sequence in that order, there is an edge {[X,EM],[Y, >]} in G. A is the source and B
the sink of the path. For example, in figure 2 <.A,B,C> is a directed path with source A and
sink C. In contrast, in figure 2 <.A,B,D> is an undirected path, but not a directed path because
Band D occur in the sequence in that order, but the edge {[B,EM],[D, >]} is not in G
(although {[D,EMJ,[B, >]} is in G.) Directed paths are therefore special cases of undirected
paths. For a directed edge e from U to V (U -> V), head(e) = V and tail (e) = U. A directed
acyclic graph is a directed graph that contains no directed cyclic paths.
A semi-directed path between A and B in a partially oriented inducing path graph 11: is an
undirected path U from A to B in which no edge contains an arrowhead pointing towards A
(i.e. there is no arrowhead at A on U, and if X and Yare adjacent on the path, and X is between
A and Yon the path, then there is no arrowhead at the X end of the edge between X and Y.) Of
course every directed path is semi-directed, but in graphs with "0" end marks there may be
semi-directed paths that are not directed.
A graph is complete if every pair of its vertices are adjacent Figure 6 illustrates a complete
undirected graph.
Figure 6
A graph is connected if there is an undirected path between any two vertices. Figures I - 6 are
connected, but figure 7 is not
Figure 7
A subgraph of <V,M,E> is any graph <V',M',E'> such that V' is included in V, M' is
included in M, and E' is included in E. Figure 7 is a subgraph of figure 1. The subgraph of
<V,M,E> over V', where V' is included in V, is the subgraph <V',M,E'> in which an edge
is in E' if and only if it is in E and has both endpoints in V'.
A clique in graph G is any subgraph of G that is complete. In figure 1, for example, the
subgraph G' =
<{A,B,D},{EM},{ {[A,EM],[B,EM]},{[B,EM],[D,EM]},{[A,EM],[D,EM]}}>
is a clique with vertices A, B and D. A clique in G whose vertex set is not properly contained in
any other clique in G is maximal. In figure 1, both G' and G" = <{A,B},{EM},
{ {[A,EM],[B,EM]} }>, are cliques, but G", unlike G' is not maximal because G" is properly
contained in G'.2
A triangle in a graph G is a complete subgraph of G with three vertices; in other words,

vertices X, Y and Z form a triangle if and only if X and Yare adjacent, Y and Z are adjacent and
X and Z are adjacent. In graph G a vertex V is a collider on undirected path U if and only
if there are two distinct edges on U containing V as an endpoint and both are into V. Otherwise
V is a noncollider on U. In graph G, vertex V is an unshielded collider on U if V is a
collider on U, Vis adjacent to distinct vertices VI and V2 on U, and VI and V2 are not adjacent
in G. An ancestor of a vertex V is any vertex W such that there is a directed path from W to
V. A descendant of a vertex V is any vertex W such that there is a directed path from V to W.
In figure 2, A, B, C, D, and E are all ancestors of C, although neither A nor C is a parent of C.
Similarly, C is a descendant of A, B, C, D, and E, although it is not a child of A or C. Since
every vertex V is the source of a directed (empty) path from V to V, each vertex is its own
descendant and its own ancestor, but not of course its own parent or its own child.
2.2 Probability
The vertices of the graphs we consider will always be random variables taking values in one of
the following: a copy of the real line; a copy of the nonnegative reals; an interval of integers.
2Some writers, especially in statistics. understand "clique" as we have defmed maximal clique.
By a joint distribution on the vertices of a graph we mean a countably additive probability

measure on the Cartesian product of these objects. We say that two random variables, X, Yare
independent when the joint density of (X,Y) is the product of the density of X and the
density of Y for all values of X and Y. We write this as X 11 Y. We generalize in the obvious
way when asserting that one set of variables is independent of another set of variables. When
we say a set of random variables is jointly independent we mean that any two disjoint
subsets of the set are independent of one another. We say that random variables X, Y are
independent conditional on Z (or given Z), when the density of X, Y given Z equals the
product of the density of X given Z and the density of Y given Z, for all values of X, Y, and
for all values z of Z for which the density of z is not equal to O. We generalize in the obvious
way for sets of random variables, X, Y, Z. If X is independent of Y given Z we write X 11
YIZ, and we say that the order of the conditional independence is equal to the number of
variables in Z.
In the discrete case, we say that a distribution over V is positive if and only if for all values v of
V, P(v) #' O. (In general, a distribution over V is positive if the density function is non-zero for
all v.) If V is included in V' and
P(V)= L,P(V')
V'W
we will say that P(V) is the marginal of P(V') over V.
2.3 Graphs and Probability Distributions
We will examine several different graphical representations of conditional independence

relations true in a distribution.
2.3.1 Directed Acyclic Graphs
A directed acyclic graph can be used to represent conditional independence relations in a

probability distribution.
For a given graph G and vertex W let Parents(W) be the set of parents of W, and
Descendants(W) be the set of descendants of W.
Markov Condition: A directed acyclic graph G over V and a probability distribution

P(V) satisfy the Markov condition if and only if for every W in V, W is independent of
V\(Descendants(W) u Parents(W» given Parents(W).
A~
~C --'D
B
Figure 8
(Recall that W is its own descendant.) In the tenninology of Pearl (1988) G is an I-map of P.
In Figure 8, the Markov Condition entails the following conditional independence relations,3
Al1B
Dl1 {A,B} I C
For all values of v of V for which f(v) * 0, the joint density function f(V) satisfying the
Markov Condition is given by
f(V) = TIf(VIParents(V»
VeV
where f(VI Parents(V» denotes the density of V conditional on the (possibly empty) set of
vertices that are parents of V. (See Kiiveri and Speed, 1982. Recall our notation convention that
if Parents(V) = 0, thenf(VIParents(V» = ftV).)
If a joint distribution over discrete variables satisfies the Markov Condition for figure 8 it can be
factored in the following way:
P(A,B,C,D) = P(A) P(B) P(CIA,B) P(D IC)
3We do not include trivial independence relations, e.g., C 11 010 which are true by definition.
for all values of A, B, C, D such that P(A,B,C,D) * O. In a directed acyclic graph G, vertices
of zero indegree are said to be exogenous. If G satisfies the Markov Condition for a
distribution P, then for every pair of exogenous variables Vi and V2, Vi 11 V2 in P.
The Minimality Condition says, intuitively, that each edge in the graph prevents some
conditional independence relation that would otherwise obtain.
Minimality Condition: If G is a directed acyclic graph over V and P a probability

distribution over V, <G, P> satisfies the Minimality Condition if and only if for every
proper subgraph H of G with vertex set V, <H ,P> does not satisfy the Markov
Condition.
Returning to the example of figure 8, a distribution P' which satisfies the Markov Condition,
but in which A is independent of {B,C, D} does not satisfy the Minimality Condition, because
P' also satisfies the Markov Condition for the subgraph in which the edge between A and Cis
removed. In the terminology of Pearl (1988) if a distribution P(V) satisifies the Markov and
Minimality conditions for a directed acyclic graph G, then G is a minimal I-map of P.
If a distribution P satisfies the Markov and Minirnality Conditions for directed acyclic graph G,
we will say that G represents P. For any directed acyclic graph G and for any probability
distribution P satisfying the Markov and Minimality Conditions, if variables A and B are
statistically dependent, then either:
(i) there is a directed path in G from A to B; or
(ii) there is a directed path in G from B to A: or
(iii) there is a variable C and directed paths in G from C to B and from C to A.
A trek between distinct vertices A and B is an unordered pair of directed paths between A and
B that have the same source, and intersect only at the source. The source of the pair of paths is
also called the source of the trek. Note that one of the paths in a trek may be an empty path.
2.3.2 Directed Independence Graphs
Directed independence graphs are another (almost equivalent) way of representing conditional
independence relations true of a probability distribution. Say that directed acyclic graph G is a
directed independence graph of P(V) (Whittaker 1990) for an ordering> of the vertices of
G if and only if A -> B occurs in G if and only if -(A 11 B I K(B», where K(B) is the set of all
vertices V such that V '# A and V> B.
Theorem 2.1: H P(V) is a positive distribution, then for any ordering of the variables
in V, P satisfies the Markov and Minimality conditions for the directed independence
graph of P(V) for that ordering.
H a distribution P is not positive, it is possible that the directed independence graph of P for a
given ordering of variables is a subgraph of a directed acyclic graph for which P satisfies the
Minimality and Markov conditions (Pearl, 1988).
2.3.3 Faithfulness
Given any graph, the Markov condition determines a set of independence relations. These
independence relations in turn may entail others, in the sense that every probability distribution
having the independence relations given by the Markov condition will also have these further
independence relations. In general, a probability distribution P on a graph G satisfying the
Markov condition may include other independence relations be~ides those entailed by the
Markov condition applied to the graph. For example, A and D might be independent in a
distribution satisfying the Markov Condition for the graph in figure 9, even though the graph
does not entail their independence.
Figure 9
In linear models such an independence can arise if the product of the partial regression
coefficients for D on C and C on A cancels the corresponding product of D on B and B on A.
H all and only the conditional independence relations true in P are entailed by the Markov
condition applied to G, we will say that P and G are faithful to one another. We will,
moreover, say that a distribution P is faithful provided there is some directed acyclic graph to
which it is faithful. In the terminology of Pearl (1988) if P and G are faithful to one another
then G is a perfect map of P and P is a DAG-Isomorph of G. If distribution P is faithful to

directed acyclic graph G, X and Yare dependent if and only if there is a trek between X and Y.
2.3.4 d-separation
Following Pearl (1988), we say that for a graph G, if X and Y are vertices in G, X"* Y, and W
is a set of vertices in G not containing X or Y, then X and Y are d-separated given WinG if
and only if there exists no undirected path U between X and Y, such that (i) every collider on U
has a descendent in W and (ii) no other vertex on U is in W. We say that if X"* Y, and X and
Y are not in W, then X and Y are d-connected given set W if and only if they are not d-
separated given W. If U, V, and W are disjoint sets of vertices in G and U and V are not
empty then we say that U and V are d-separated given W if and only if every pair <U, V> in the
cartesian product of U and V is d-separated given W. If U, V, and W are disjoint sets of
vertices in G and U and V are not empty then we say that U and V are d-connected given W if
and only if U and V are not d-separated given W. An illustration of d-connectedness is given in
the following directed acyclic graph (but note that the definition also applies to other sorts of
graphs such as inducing path graphs, as explained in Chapter 6):
Figure 10
X and Y are d-separated given the empty set

X and Yare d-connected given set {Sl, S2}
X and Yare d-separated given the set {Sl, S2, V}
2.3.5 Linear Structures
A directed acyclic graph G over V linearly represents a distribution P(V) if and only if there
exists a directed acyclic graph G' over V' and a distribution P"(V') such that
Fonna! Preliminaries 37
(i) V is included in V';

(ii) for each endogenous (that is, with positive indegree) variable X in V, there is a unique
variable ex in V,\V with zero indegree, positive variance, outdegree equal to one, and a
directed edge from ex to X;
(iii) G is the subgraph of G' over V;
(iv) each endogenous variable in G is a linear function of its parents in G';
(v) in P"(V') the correlation between any two exogenous variables in G' is zero;
(vi) P(V) is the marginal of P"(V') over V.
The members of V,\V are called error variables and we call G' the expanded graph.
Directed acyclic graph G linearly implies PAB.H = 0 if and only if PAB.H = 0 in all
distributions linearly represented by G. (We assume all partial correlations exist for the
distribution.) If G linearly represents P(V) we say that the pair <G,P(V» is a linear model with
directed acyclic graph G.
2.4 Undirected Independence Graphs
There is a well-known representation of statistical hypotheses about conditional independence

by undirected graphs. The two representations, by directed and by undirected graphs, are
closely related, but it is important not to confuse them.
An undirected independence graph G with a set of vertices V represents a probability

distribution P if and only if there is no undirected edge between A and B just when A and B are
conditionally independent given V\{A,B lin P. If an undirected independence graph G
represents a distribution P, A and B are independent conditional on some set C if and only if
every undirected path between A and B contains a member of C.
Suppose we consider a particular directed acyclic graph G and faithful probability distribution
P. Let U be the undirected graph of adjacencies underlying G; that is, U is the undirected
graph with the same vertex set as G and the same adjacencies as G. Suppose that I is the
undirected independence graph for the distribution P formed according to the definition just
given. Then I and U are not in general the same, but U is always a subgraph of I. I and U will
be the same if and only if G contains no unshielded colliders (Wermuth and Lauritzen 1983).
2.5 Deterministic and Pseudo-Indeterministic Systems
We will use the notion of a deterministic system in a technical sense: A joint probability
distribution P on a set V of random variables represented by a directed acyclic graph G is
deterministic if each of the vertices of G of non-zero indegree is a function of the vertices that
are its immediate parents in G; we will also say that G is a deterministic graph of P. By
"function" we mean that for each assignment of a unique value to each of the parent vertices,
there is a unique value of the dependent vertex.
Figure 11
Suppose that the graph/distribution pair represented in figure 11 is deterministic, but that ec and
ED are not measured. Were we to consider only the measured variables, i.e., A, B, C, and D,
we would fmd that no variable has its value uniquely determined by the values of the others,
although some of the variables are statistically dependent. The system looks indeterministic,
although EC and ED are "hidden" variables which make it deterministic when added.
Furthermore, it is not necessary to posit that two measured variables depend upon the same
hidden variable, nor is it necessary to posit any dependence among the "hidden" variables in
order to make the system deterministic. When a distribution represented by a directed acyclic
graph among measured variables is not deterministic, but is embeddable in this way in a
distribution represented by a directed acyclic graph that is, we say the distribution is pseudo-
indeterministic.
In contrast, consider figure 12. Again suppose that only A, B, C, and D, were measured. In
this case we could not make the system deterministic by adding hidden variables unless either
Fonnal Preliminaries 39
the hidden variables were associated or at least one hidden variable is adjacent to at least two of
the measured variables.
fff:z?
~ ~. ~
Figure 12
More formally, <G,P> is pseudo indeterministic, where P is a probability distribution over

V and G is a directed acyclic graph over V, if and only if G is not a deterministic graph of P
and there exists a distribution P' and a directed acyclic graph G' over a set of variables V' that
properly includes V such that
(i) G' is a deterministic graph of P';
(ii) G is the subgraph of G' over V;
(iii) no vertex in V is an ancestor of a vertex in V,\V;
(iv) no vertex in V,\V, is the source of a trek connecting two vertices in V;
(v) P is the marginal of P';
(vi) G represents P.
If we say that <P,G> is linear pseudo indeterministic we mean that <P,G> is pseudo-
indeterministic and in addition in G', each vertex in V'is a linear function of its parents. A
distribution linearly represented by a directed acyclic graph is pseudo-indeterministic.
(Analogous definitions apply to Boolean pseudo indeterministic pairs of graphs and
distributions, etc.)
2.6 Background Notes
Drawing from purely graph-theoretical work of Lauritzen, Speed and Vijayan (1978), and on
statistical work in log-linear models (Bishop, Fienberg and Holland, 1975), in 1980 Darroch,
Lauritzen and Speed introduced undirected graphical representations of log-linear hypotheses of
conditional independence. Based on Kiiveri's thesis work, Kiiveri and Speed (1982) introduced
versions of the Markov Condition, defined the notion of recursive causal model, obtained
maximum likelihood estimates for a multinomial distribution and provided a systematic survey
of applications with both discrete and continuous variables. Shortly after, Kiiveri, Speed and
Carlin (1984) further developed the formal foundations. Wermuth and Lauritzen (1983)
introduced the notion of a recursive diagram, or what we have called a directed independence
graph. The definitions of minimality and faithfulness are due to Pearl (1988).
Chapter 3
Causation and Prediction: Axioms and Explications
Views about the nature of causation divide very roughly into those that analyze causal influence
as some sort of probabilistic relation, those that analyze causal influence as some sort of
counterfactual relation (sometimes a counterfactual relation having to do with manipulations or
interventions), and those that prefer not to talk of causation at all. We advocate no defmition of
causation, but in this chapter we try to make our usage systematic, and to make explicit our
assumptions connecting causal structure with probability, counterfactuals and manipulations.
With suitable metaphysical gyrations the assumptions could be endorsed from any of these
points of view, perhaps including even the last.
3.1 Conditionals
Intelligent planning usually requires predicting the consequences of actions. Since actions
change the states of affairs, assessing the consequences of actions not yet taken requires
judging the truth or falsity of future conditional sentences--l/ X were to be the case, then Y
would be the case. Judging the effects of past practice or policy requires judging the truth or
falsity of counterfactual sentences--lfX Iuld been the case, then Y would have been the case.
Giving a detailed description of the conditions under which a future conditional or

counterfactual conditional is true is a well-known and difficult philosophical problem. Lewis
(1973) notes thatl/kangaroos Iuld no tails, they would topple over is true even though we can
imagine circumstances in which kangaroos use crutches. We mean that if things were pretty
much as they are--given the scarcity of crutches for kangaroos and the disinclination of
kangaroos to use crutches--if kangaroos had no tails they would topple over. But making this
intuition precise is not easy.
It is widely recognized that causal regularities entail counterfactual conditionals; indeed this is
often taken to be the feature that distinguishes a causal law from generalizations that are true, as
it were, by accident. All of the coins in your pocket are made of silver does not entail the
counterfactual If this penny were in your pocket then it would be made of silver. But the causal
law All collisions of electrons and positrons release energy does entail the counterfactual If this
electron were to collide with this positron then energy would be released.
The connection between causal regularities and the truth of future conditional and counterfactual
sentences makes the discovery of causal structure essential for intelligent planning in many
contexts. A linear equation relating the fatality rate in automobile accidents to car weight may be
true of a given population, but unless it describes a robust feature of the world it is useless for
predicting what would happen to the fatality rate if car weight was manipulated through
legislation. Even quite accurate parametric representations of the distribution of values in a
population may be useless for planning unless they also reflect the causal structure among the
variables.
3.2 Causation
We understand causation to be a relation between particular events: something happens and

causes something else to happen. Each cause is a particular event and each effect is a particular
event. An event A can have more than one cause, none of which alone suffice to produce A. An
event A can also be overdetermined: it can have more than one set of causes that suffice for A to
occur. We assume that causation is transitive, irreflexive, and antisymmetric. That is, i) if A is a
cause of B and B is a cause of C, then A is also a cause of C, ii) an event A cannot cause itself,
and iii) if A is a cause of B then B is not a cause of A.
3.2.1 Direct vs. Indirect Causation
The distinction between direct and indirect causes is relative to a set of events. If C is the event
of striking a match, and A is the event of the match catching on fIre, and no other events are
considered, then C is a direct cause of A. If, however, we added B: the sulfur on the match tip
achieved sufficient heat to combine with the oxygen, then we would no longer say that C
Axioms and Explications 43
directly caused A, but rather that C directly caused Band B directly caused A. Accordingly, we
say that B is a causal mediary between C and A if C causes Band B causes A.
Having fixed a context and a set of events, what is it for one event to be a direct cause of
another? The intuition is this: once the events that are direct causes of A occur, then whether A
occurs or not no longer has anything to do with whether the events that are indirect causes of A
occur. The direct causes screen off the indirect causes from the effect. If a child is exposed to
chicken pox at her daycare center, becomes infected with the virus, and later breaks out in a
rash, the infection screens off the event of exposure from the occurrence of the rash. Once she
is infected, whether she gets the rash has nothing to do with whether she was exposed to the
virus from her daycare or from her Saturday morning playgroup.
Suppose V is a set of events including C and A. C is a direct cause of A relative to V just in

case C is a member of some set C included in v\{A} such that (i) the events in C are causes of
A, (ii) the events in C, were they to occur, would cause A no matter whether the eveqts in
V\( {A} u C) were or were not to occur, and (iii) no proper subset of C satisfies (i) and (ii).
3.2.2 Events and Variables
In order for causation to be connected with probabilities that can be estimated empirically,
events must be sorted; some actual or possible events must be gathered together, declared to be
of a type, and distinguished from other actual or possible events perhaps gathered into other
types. The simplest classifications describe events as of a kind, e.g., solar eclipses, or declines
in the Dow-Jones Industrial Average, and pair each event, A, of a kind with the event, -,,4, the
non-occurrence of A. Such classifications permit us to speak intelligibly of variables as causes.
We do so through the introduction of Boolean variables that take events of a kind, or their
absences, as values. We say that Boolean variable C causes Boolean variable A if and
only if at least one member of a pair (C, -,C) causes at least one member of a pair (A, -,,4).
Ordinarily no one would bother with collecting events into a type and examining causal relations
among such variables unless the causal relations among events of the two types had some
generality--that is, lots of events of type A have events of type C as causes and lots of events of
type C have effects of type A, or none do.
Events can be aggregated into variables X and Y, such that some events of kind X cause some
events of kind Y and some events of kind Y cause some events of kind X. In such cases there
will be no unambiguous direction to the causal relation between the variables.We consider this
case in Chapter 12.
Some events are of a quantity taking a certain value, such as bringing a particular pot of water to
a temperature of 100 degrees centigrade. Scales of many kinds are associated with an array of
possible events in which a particular system takes on a scale value or takes on a value within a
set of scale values. We can also speak of the variables of such scales as causes and effects, at
least for particular systems over particular time intervals. For any particular system S we say
that scaled variable Q causes scaled variable R in S provided that there is a value (or set
of values) q for Q and a value (or set of values) r for R and a possible event in which Q taking
value q in S would cause an event in which R takes value r in S. In practice we usually form
scales only when we think the causal relations among val!les of different measures are not
confined to particular values or particular systems but are more general. We sometimes say that
the value r for R is caused by the value q for Q if the system taking on the value q for Q caused
it to take on the value r for R. If K is a collection of systems, we say that variable Q causes
variable R in K provided that for every system SinK, Q causes R in S.
If our notion of causation between variables were strictly applied, almost every natural variable
would count as a cause of almost every other natural variable, for no matter how remote two
variables, A and B, may be, there is usually some physically possible--even if very unlikely--
arrangement of systems such that variation in some values of A produces variation in some
values of B. (A dictator could, we suppose, arrange circumstances so that the number of
childbirths in Chicago is a function of the price of tea in China.) In practice, we always consider
a restricted range of variation of other variables in judging whether A causes B. Strictly,
therefore, our definitions of causal relations for variables should be relative to a set of possible
values for other variables, but we will ignore this formality and trust to context The notion of
direct cause generalizes from events to variables in obvious parallel to the definition of causal
dependence between variables: Variable C is a direct cause of variable A relative to V provided
(i) C is a member of a set C of variables included in V, (ii) there exists a set of values c for
variables in C and a value a for A such that were the variables in C to take on values c, they
would cause A to take on value a no matter what the values of other variables in V, and (iii) no
proper subset of C satisfies (i) and (ii). We say that a variable X is a common cause of
variables Y and Z if and only if X is a direct cause of Y relative to {X,Y,Z} and a direct cause
of Z relative to {X,Y,Z}. If there is a sequence of variables in V beginning with A and ending
with B such that for each pair of variables X and Y that are adjacent in the sequence in that order
X is a direct cause of Y relative to V, then we say that there is a causal chain from A to B
relative to V. A is an indirect cause of B relative to V if there is a causal chain from A to B
relative to V of length greater than 2. We make the following two fundamental assumptions
about causal relations: (i) if A is a cause of B then A is a direct cause or an indirect cause of B
relative to V; (ii) if A, B, and C are in V, and there exists a causal chain from A to B relative to
V that does not contain C, then for any set V' that contains A and B there is a causal chain from
A to B relative to V' that does not contain C. When a cause is unmeasured it is sometimes
called a latent variable. We say that two variables are causally connected in a system if one
of them is the cause of the other or if they have a common cause. A causal structure for a
population is an ordered pair <V,E> where V is a set of variables, and E is a set of ordered
pairs of V, where <X,Y> is in E if and only X is a direct cause of Y relative to V. We assume
that in the population A is a direct cause of B either for all units in the population or no units in
the population, unless explicitly noted otherwise. If it is obvious which population is intended
we do not explicitly mention it. If P(V) is a distribution over V in a population with causal
structure C = <V,E>, we say that C generated P(V). Two causal structures <V,E> and
<V',E'> are isomorphic if and only if there is a one-to-one function/from V onto V' such
that for any two members of A and B of V, <A,B> is in E if and only if <f(A),f(B» is in E'.
A set V of variables is causally sufficient for a population if and only if in the population
every common cause of any two or more variables in V is in V, or has the same value for all
units in the population.! We will often use the notion of causal sufficiency without explicitly
mentioning the population.
3.2.3 Examples
Simple digital logic circuit elements present concrete examples of causal structures. They are not
of much intrinsic interest to most people, but they have the virtue that given a description of
such a circuit element almost everyone can agree about which events pertaining to the circuit
cause which other events. In the element illustrated below, the variables X 1, X2 and X3 have
two values, I and 0, accordingly as there is or is not a current through the corresponding line,
and the semi-circle represents an "and" gate. Current flows from top to bottom. The value of the
variable X3 is thus a simple Boolean function of the values of X 1 and X2. If "." represents
Boolean multiplication,X3 =Xl·X2.
! Strictly, we require for causal sufficiency of V for a population that if X is not in V and is a common canse of
two or more variables in V. that the joint probability of all variables in V be the same on each value of X that
occurs in the population.
Figure 1
We understand the event of X I taking on value I and the event of X2 taking on value I each to
be a cause of the event in which X3 takes on the value 1. We say that the Boolean variables X I
and X2 are each causes of the Boolean variable X3.
The form of the causal structure does not depend on the sort of variables involved or the
particular class of functions among them. Isomorphic causal structures might be realized by a
system of linear dependencies of continuous variables. Thus consider three different variables
Xl, X2, andX3, that represent the voltage in a given line and therefore range over the positive
reals. Suppose we have a mechanism that outputs the sum of the voltage into it (figure 2).
Figure 2
In this case X3 = X I + X2, but the causal structure is isomorphic to the causal structure in
figure 1: X I and X2 each are causes of X3.
These examples suggest that the causal dependencies and the functional dependencies are
related; X3 is the effect of X I and X2, and X3 is a function of X I and X2. In systems in which
variables that are effects have their values uniquely determined by the values of all of the
variables that are their direct causes, functional dependence can be inferred from causal
dependence by expressing each variable or event as a function of its direct causes. The converse
does not hold: from the fact that an equation correctly describes a system one cannot infer that
the direct causal dependencies in the system are reflected in the functional dependencies in the
equation. For example, ifthe equation X3 = X I + X2 is true of a system then the equation X2 =
X3 - XI is equally true of that system, but if XI and X2 causeX3, then ordinarily X3 and XI do
not cause X2. 2
3.2.4 Representing Causal Relations with Directed Graphs
U sing the notion of a direct cause, it is trivial to represent causal structures with directed graphs:
Causal Representation Convention: A directed graph G = <V, E> represents a

causally sufficient causal structure C for a population of units when the vertices of q
denote the variables in C, and there is a directed edge from A to B in G if and only if A
is a direct cause of B relative to V.3
We call a directed acyclic graph that represents a causal structure a causal graph. Figure 3
below is a causal graph for the circuit devices shown in figures 1 and 2.
Figure 3
Consistently with our previous definition, if G is a causal graph and there is a vertex X in G
and a directed path from X to Y that does not contain Z, and a directed path from X to Z that
does not contain Y, we will say X is a common cause of Y and Z.
2Using the notion of identifiability, Simon (1953) proposed a general means to derive causal structure from a set
of equations describing a system; later in the same paper Simon also proposed an account of causation using
invariances under perturbations of linear coefficients.
3Since causation for variables is assumed to be transitive and irreflexive, the directed graph representing a causal
structure must be acyclic. Introducing cyclic directed graphs requires a systematic reinterpretation.
There are important limitations to the Causal Representation Convention. Suppose drugs A and
B both reduce symptoms C, but the effect of A withoutB is quite trivial, while the effect of B
alone is not. The directed graph representations we have considered in this chapter offer no
means to represent this interaction and to distinguish it from other circumstances in which A and
B alone each have an effect on C. The interaction is only represented through the probability
distribution associated with the graph. Consider another example, a simple switch. Suppose as
in figure 4 battery A has two states: charged and uncharged. Charge in battery A will cause bulb
C to light up provided the switch B is on, but not otherwise. If A and B are independent
random variables, then A and C are dependent conditional on B and on the empty set, and B
and C are dependent conditional on A and the empty set, and A and B are dependent conditional
on C. The directed acyclic graph representing the distribution over A, B, and C therefore looks
like the directed graph shown above. There is nothing wrong with this conclusion except that it
is not fully informative. The dependence of A and C arises entirely through the condition B = 1.
When B = 0, A and C are independent
B
A
Figure 4
Since in discrete data the conditional independence facts, if known, identify the switch
variables, a better representation would identify certain parents of a variable as switches. But a
general representation of this sort would often not be very easy to grasp.4 Recent work on
extending the directed acyclic graph representation to represent switches is described in Geiger
and Heckerman (1991).
4A better practical arrangement might be a query system that, besides inferring the causal graph or graphs,
responds to the user's questions about the effects of the manipulation of variables.
3.3 Causality and Probability
3.3.1 Deterministic Causal Structures
To good approximation the devices in figures 1 and 2 are deterministic, i.e., the effects are
deterministic functions of their direct causes. If each effect is a linear function of its direct
causes in the population, we say the system is a linear deterministic causal structure in
the population.
Variables in a causal graph that have zero indegree, i.e., no causal input, are said to be
exogenous. Xl and X2 are exogenous variables in the causal graph in figure 3. Variables that
are not exogenous are endogenous. In a deterministic causal structure values for the
exogenous variables determine unique values for the remaining variables.
Circuit Diagram Causal Graph
Figure 5
Consider the circuit element in figure 1 and its causal graph, both of which are shown in figure
5. Imagine an experiment to verify whether or not the device works as described. We would
assign values to the exogenous variables, i.e., decide whether to put current into Xl and X2,
and then read whether or not X3 has current. We can represent the experiment with the
following table.
Xl X2 X3
1 1 ?
1 0 ?
0 1 ?
0 0 ?
Suppose we were satisfied that the device usually worked as designed, but we wanted to know
how often and in what way it fails. For each of a number of trials, we could randomly assign
values to X I and X2, and then read whether or not X 3 has current. That is, we could assign a
probability to each state the set of exogenous variables could occupy. For example,
P(X I = 1, X2 = 1) = 0.2
P(XI = 1,X2 = 0) = 0.3
P(XI =0,X2 = 1) = 0.2
P(XI =0,X2=0) =0.3
Because this causal structure is deterministic (even though the exogenous variables are
random), a probability distribution over the exogenous variables determines a joint distribution
for the entire set of variables in the system. For this example the joint distribution over (Xl, X2,
X3) is:
P(XI = 1, X2 = 1, X]= 1) = 0.2

P(XI = 1, X2 = 1,X3 = 0) = 0.0
P(XI = 1, X2 = 0, X]= 1) = 0.0
P(X I = 1, X2 = 0, X3 = 0) = 0.3
P(XI = 0, X2 = 1, X]= 1) = 0.0
P(XI = 0, X2 = 1, X3 = 0) = 0.2
P(X I = 0, X2 = 0, X3 = 1) = 0.0
P(X I = 0, X2 = 0, X3 = 0) = 0.3
We say that this distribution is generated by the causal structure of figure 5.
We use this example not to investigate sampling schemes for circuits but rather to illustrate how
probability distributions are generated by deterministic causal devices. The only assumption we
make about the connection between deterministic causal structures and the probability
distributions they may generate involves the distributions we will allow over the exogenous
variables. We assume that the exogenous variables are jointly independent in a probability
distribution over the variables in a causally sufficient structure. This is in part a substantive
assumption--that statistical dependence is produced by causal connection--and in part a
convention about representation. If exogenous variables in a structure are not independent, we
expect that the causal graph is incomplete and there is some further causal mechanism, not
represented in the graph, responsible for the statistical dependence. Either some of the input
variables are causes of others (in which case we have equivocated, and the causal graph is not
actually the graph of the causal structure of the structure) or else some nonconstant common
causes of observed variables have not been included in the description of the structure.
3.3.2 Pseudo-Indeterministic and Indeterministic Causal Structures
In practice, the variables people measure are seldom deterministic functions of one another. We
call a causal structure over a set V of variables for a population in which some variable is not a
determinate function of its immediate causes in V an indeterministic causal structure for
the population. An indeterministic causal structure might be pseudo-indeterministic. That is, a
deterministic causal structure for which not all of the causes of variables in V are also members
of V may appear to be indeterministic, even though there is no genuine indeterminism if the set
of variables is enlarged by adding variables that are not common causes of variables in V. For
example, suppose again that the device shown in figures 1 and 5 governs the current in lineX3.
Suppose also that X2 is hidden from us so that only Xl and X3 occur in the causal structure we
investigate. We might still hypothesize that Xl is a cause of X3, thereby forming the causal
graph on the right side of figure 6.
\ X3
Actual Circuit Diagram Hypothesized Causal Graph
Figure 6
Assuming that the joint distribution P(X 1, X2, X 3) generated by the actual circuit device is the
same as the one given for figure 5 in section 3.3.1, the observed distribution P(XI. X3) is just
the marginal of P(X 1. X2, X3), namely:
P(XI = 1, X3 = 1) =0.2
P(XI = I,X3 = 0) = 0.3
P(XI =0,X3= 1) =0.0
P(XI =0, X3 = 0) =0.5
In the observed distribution X3 is clearly not a function of its immediate parent Xl and the
causal structure appears to be indeterministic. We say the structure is pseudo-indetenninistic.
More formally, causal structure C = <V,E> is pseudo-indeterministic for a population, if
and only if C is not a detenninistic causal structure for the population and there exists a causal
structure C' for the population over a set of variables V' that properly includes V such that
(i) C' is a detenninistic causal structure for the population;
(ii) If A and B are in V, then <A,B> is in E if and only if <A,B> is in E';
(iii) no variable in V is a cause of a variable in V'W;
(iv) no variable in V'W, is a common cause of two variable in V;
We say a structure is linear pseudo-indeterministic if all functional dependencies in C' are

linear. Structural equation models in the social sciences are usually assumed to be pseudo-
indeterministic causal structures. The error terms in such models are often interpreted as omitted
causes.
A genuinely indeterministic causal structure for a population over a set of variables V is an

indeterministic causal structure that is not pseudo-indeterministic. It is at least conceivable that
there are genuinely indeterministic structures, even genuinely indeterministic macroscopic
structures, whose variables have a causal structure. We will assume that the same relations
between conditional independence and causal structure that obtain for pseudo-indeterministic
structures hold as well for genuinely indeterministic causal relations, although as we will see
later, there appear to be quantum mechanical systems for which that assumption must be
carefully qualified. For a discussion of the case in which measured variables are exact functions
of other measured variables see section 3.8.
3.4 The Axioms
We consider three conditions connecting probabilities with causal graphs: The Causal Markov
Condition, the Causal Minimality Condition, and the Faithfulness Condition. These axioms are
not independent Consequences of various subsets of the conditions are investigated in the
course of this book. We will consider justifications and objections to the conditions in the next
section, but their importance--if not their truth--is evidenced 'by the fact that nearly every
statistical model with a causal significance we have come upon in the social scientific literature
satisfies all three: if the model were true, all three conditions would be met While it is easy
enough to construct models that violate the third of these conditions, Faithfulness, such models
rarely occur in contemporary practice, and when they do, the fact that they have properties that
are consequences of unfaithfulness is taken as an objection to them In Chapters 5 and 8 we will
consider published log-linear models, regression models, and structural equation models
satisfying the three conditions.
3.4.1 The Causal Markov Condition
The intuitions connecting causal graphs with the probability distributions they generate are
unified and generalized in one fundamental condition:
Causal Markov Condition: Let G be a causal graph with vertex set V and P be a
probability distribution over the vertices in V generated by the causal structure
represented by G. G and P satisfy the Causal Markov Condition if and only if for every
W in V, W is independent of V\(Descendants(W) v Parents(W» given
Parents(W).
When G describes causal dependencies in a population with variables distributed as P satisfying

the Causal Markov condition for G, we will sometimes say that P is generated by G. If V is
not causally sufficient and V is a proper subset of the variables in a causal graph G generating a
distribution P, we do not assume that the Causal Markov condition holds for the marginal over
VofP.
The factorization results described in Chapter 2 apply to the joint probability distribution for a
set V of variables in a population of systems with a causal structure satisfying the Causal
Markov Condition. If P(V I Parents (V» denotes the probability of V conditional on the
(possibly empty) set of vertices that are direct causes of V, then
P(V) = II P(VIParents(V»
VeV
for all values of V for which each P(VlParents(V» is defined.
Figure 7
For the graph in figure 7 direct application of the Markov Condition yields a list of
independence facts about the distribution generated by G.
Xl 11x2
x 2 11. {XI,X4}
X311x41 {XI,X2}
x411 {X2,X3} IXI
xsll. {XI,X2} I {X3,X4}
Other independence relations are entailed by these, for example
A discussion of axioms for conditional independence is found in Pearl (1988).
3.4.2 The Causal Minimality Condition
We will usually impose a further condition connecting probability with causality. The principle
says that each direct causal connection prevents some independence or conditional independence
relation that would otherwise obtain. For example, in the following causal graph G, C is a
direct cause of A .
C • A
~/
B
Figure 8
In a distribution P over {A,B,C} for which C 11 A, P satisfies the Markov condition even if
the edge between C and A is removed from the graph.
Causal Minimality Condition: Let G be a causal graph with vertex set V and P a
probability distribution on V generated by G. <G, P> satisfies the Causal Minimality
condition if and only if for every proper sub graph H of G with vertex set V, the pair
<H,P> does not satisfy the Causal Markov condition.
Since we will almost always give the graphs we consider a causal interpretation, we will in
most cases hereafter simply describe these two conditions as the Markov and Minimality
Conditions.
3.4.3 The Faithfulness Condition
Given a causal graph, the Markov condition determines a set of independence relations. These
independence relations in turn may entail others, in the sense that every probability distribution
having the independence relations given by the Markov condition will also have these further
independence relations. In general a probability distribution P on a causal graph G satisfying the
Markov condition may include other independence relations besides those entailed by the
Markov condition applied to the graph. If, however, that does not occur, and all and only the
independence relations of P are entailed by the Markov condition applied to G, we will say that
P and G are faithful to one another. We will, moreover, say that a distribution P is faithful
provided there is some directed acyclic graph to which it is faithful. So we consider a further
axiom:
Faithfulness Condition: Let G be a causal graph and P a probability distribution

generated by G. <G, P> satisfies the Faithfulness Condition if and only if every
conditional independence relation true in P is entailed by the Causal Markov Condition
applied to G.
Note that a distribution P is faithful to G if and only if it satisfies both the Markov and
Faithfulness Conditions. The Faithfulness and Markov Conditions entail Minimality, but
Minirnality and Markov do not entail Faithfulness. We will sometimes use the weaker axiom or
axioms and more often the stronger one. Faithfulness turns out to be important to discovering
causal structure, and it also turns out to be the "normal" relation between probability
distributions and causal structures.
3.5 Discussion of the Conditions
When and why should it be thought that probability and causality together satisfy these
conditions, and when can we expect the conditions to be violated? When should the values of
variables in a population be thought to be distributed in accordance with the conditions?
3.5.1 The Causal Markov and Minimality Conditions
If we consider probability distributions for the vertices of causal graphs of deterministic or

pseudo-indeterministic systems in which the exogenous variables are independently distributed,
then the Markov Condition must be satisfied. A proof is given in the last chapter. We conjecture
the Minimality Condition is true of all pseudo-indeterministic systems. The warrant for the
conditions lies in this fact, and in the history of human experience with systems that we can
largely control or manipulate. Electrical devices, mechanical devices, chemical devices all
satisfy the condition. Large areas of science and engineering--from auto mechanics to chemical
kinetics to digital circuit design--would be impossible without using the principles to diagnose
failures and infer mechanisms.
In an important class of cases the application of the Minimality and Markov Conditions may be
unclear. In 1903 G. Udny Yule concluded his fundamental paper on the theory of association of
attributes in statistics with a section "On the fallacies that may be caused by the mixing of
distinct records". (Yule uses lAB I CI to denote "the association between A and B in the universe
of C's" (p. 131»:
It follows from the preceding work that we cannot infer independence of a pair of
attributes within a sub-universe from the fact of independence within the universe at
large...The theorem is of considerable practical importance from its inverse application;
i.e. even if IABI have a sensible positive or negative value we cannot be sure that
nevertheless lAB I C I and lAB I 'Y I are not both zero. Some given attribute might, for
instance, be inherited neither in the male line nor the female line; yet a mixed record
might exhibit a considerable apparent inheritance. Suppose for instance that 50% of the
fathers and of the sons exhibit the attribute, but only 10% of the mothers and daughters.
Then if there be no inheritance in either line of descent the record must give
(approximately)
fathers with attribute and sons with attribute: 25%

fathers with attribute and sons without attribute: 25%
fathers without attribute and sons with attribute: 25%
fathers without attribute and sons without attribute: 25%
mothers with attribute and daughters with attribute: 1%

mothers with attribute and daughters without attribute: 9%
mothers without attribute and daughters with attribute: 9%
mothers without attribute and daughters without attribute 81 %
If these two records be mixed in equal proportions we get
parents with attribute and offspring with attribute 13%

parents with attribute and offspring without attribute 17%
parents without attribute and offspring with attribute 17%
parents without attribute and offspring without attribute 53%
Here 13/40 = 43 [and] 1/3% of the offspring of parents with the attribute possess the
attribute themselves, but only 30% of offspring in general, i.e. there is quite a large but
illusory inheritance created simply by the mixture of the two distinct records. A similar
illusory association, that is to sayan association to which the most obvious physical
meaning must not be assigned, may very probably occur in any other case in which
different records are pooled together or in which only one record is made of a lot of
heterogeneous material.
The fictitious association caused by mixing records finds its counterpart in the spurious
correlation to which the same process may give rise in the case of continuous variables,
a case to which attention was drawn and which was fully discussed by Professor
Pearson in a recent memoir. If two separate records, for each of which the correlation is
zero, be pooled together, a spurious correlation will necessarily be created unless the
mean of one of the variables, at least, be the same in the two cases.
Yule's example seems to present a problem for the Causal Markov condition. Let a mixture
over V be any population that consists of a combination of some finite number of
subpopulations Pi each having different joint distributions over the variables in V, with each
distribution satisfying the Causal Markov Condition for some graph. Consider a population that
is a mixture of structures <G,Pl> and <G,P2> where PI and P2 are distinct and satisfy the
Markov Condition for G. Let the proportions in the mixture be n:m.
Let P(X,Y,Z) = nPl (X,Y,Z) + mP2(X,Y,Z), with n + m = 1. A little algebra shows that
P(XYIZ) = P(XIZ)P(YIZ) if and only if
If n, m > 0 and in both distributions, X,Y are independent conditional on Z, that is PI (X,YIZ) =
Pl(XIZ)Pl(YIZ) and P2(X,YIZ) =P2(XIZ)P2(YIZ), then equation (1) reduces to
The old but still rather surprising conclusion is that when we mix probability distributions we
may find all possible conditional dependence relations. Thus, it seems, in many mixed
populations conditional independence and dependence will not be a reliable guide to causal
structure.
In the case of linear pseudo-indeterministic systems, when populations with two different
distributions each associated with a linear structure are mixed, vanishing correlations in each
separate distribution will not produce vanishing correlations in the mixed distribution, and
vanishing partial correlations in each separate distribution will not produce vanishing partial
correlations in the mixed distribution. It is easy to verify that for any mixture of two
distributions--based on linear structures or not--the covariance of two variables vanishes in the
mixture if and only if
klCOVl(Xy) + k2COV2(Xy) = klk2[.ulX J.l2Y + J.lIY J.l2X] +

kl(kl - l)J.lIXJ.lIY + k2(k2 - 1)J.l2X J.l2Y
where the proportion of population 1 to population 2 is n: m and kl =n/(n+m), k2 =m/(n+m),

and "J.li" denotes the mean in population i.
So the situation is that we can have population 1 with causal graph Gl and population 2 with
causal graph G2, and the joint population will have a distribution that does not satisfy the
Markov Condition for either graph. The question is whether such a mixed population violates
the Causal Markov Condition. When a cause 0/ membership in a subpopulation is rightly

regarded as a common cause o/the variables in V, the Causal Markov Condition is not violated
in a mixed popUlation; instead, we have a population of systems satisfying the Causal Markov
Condition but with a common cause (or causes) that may not have been measured. In some
cases the cause of membership in a subpopulation may act like a latent switch variable of the
kind considered in section 3.2.4; the distributions conditional on different values of the latent
variable determine probability relations that are faithful to distinct causal graphs. In Yule's
example, the missing common cause is gender. If, to take another example, we form a mixed
sample of lead and copper pennies, within each subpopulation density and electrical
conductivity will be independent, but in the mixed population they will be statistically
dependent. We should say that is because chemical composition is a common cause of density
and conductivity. In other cases the cause or causes of membership in relevant subpopulations
may seem like unnatural kinds, or may at least not be the sort of causes a scientist seeks. Thus
an important controversy (Caramazza, 1986) in contemporary cognitive neuropsychology
concerns the use of statistical results for samples of people selected by syndrome, for example
subjects with Broca's aphasia. One aim of studying such groups may be to discover if two or
more normal capacities have a common cause damaged in Broca's aphasics. Suppose in a
sample of Broca's aphasics a correlation is observed in scores on tests of two cognitive skills.
Should the psychologist conclude that the test performances have a common latent cause?
Perhaps, but the common cause need not be any functional capacity--darnaged or otherwise--
that causes both skills. Instead, the sample of Broca's aphasics might be a mixture of people
with different sorts of brain damage, and within each subgroup the skills in question might be
independently distributed. The common cause is only a variable representing membership in a
subpopulation.
There are contexts in which the statistics of mixtures do not reflect any variable for popUlation
membership. In linear models the correlations and partial correlations are determined by the
linear coefficients and the variances of exogenous variables. These parameters themselves may
be treated as random variables and the resulting population distribution is a (generally
uncountable) mixture of distributions. Statistical, but not causal, inference has been extensively
studied in such settings (Swamy, 1971). If X is a random variable, we denote the expected
value of X by E(X).
Theorem 3.1 Let M be a linear model with directed acyclic graph G and linear
coefficients aij. Let M' be a linear model with directed acyclic graph G, such that the
linear coefficients in M are random variables a'ij that are jointly independent of all other
random variables in M, and E(a'ij) =aij. Suppose the variances of the exogenous non-
coefficient random variables are the same in M and M'. Then PAB.C = 0 in M' if and
only if PAB.C = 0 in M.
Thus a population that is a mixture of linear pseudo indeterministic causally sufficient systems
with the same causal graph and with parameters independently distributed will satisfy the
Causal Markov Condition for that graph without any unmeasured common cause.
Professional philosophers have offered a spate of criticisms of consequences of the Causal

Markov Condition. Most of them appear to depend on omitting relevant latent variables. Wesley
Salmon (1984) claims that "[t]here is another, basically different, sort of common cause
situation" that cannot appropriately be characterized in terms of the Causal Markov Condition.
Salrilon calls this other causal relation an "interactive fork."
One putative example of an "interactive fork" is from Davis (1988):
Imagine a television set with a balky switch: it usually turns the set on, but not
always. When the set is on, it produces both sound and picture. Then the
probability of a picture given that the switch is on and given sound is greater
than the probability of a picture given just that the switch is on. (Davis 88,
p.156)
C =SwitchOn
B =SoundOn
A =ScreenOn
B A
Figure 9
So, P(BIC) < P(BIA & C).
Davis' example gives an inaccurate picture of the causal situation, which is better depicted as
follows:
C = Switch On
B = Sound On
A =ScreenOn
D = Circuit Closed
Figure 10
The state of the circuit, or some variable downstream from the switch event, makes A and B
independent.
Salmon's own illustration uses a slight variant of the following example from the game of pool
(where we replace his events by Boolean variables).
C is the description of causal conditions relevant to both A and B, but A and B are not
independent conditional on C.
A= I ballinL.
B=2ballinR.
C = Collision of any sort between
cue ball and either 1 or 2 ball.
Figure 11
Knowing C (that there was a collision) and A (that balll dropped into its pocket) tells us more
about whether B occurred (the 2 ball dropped into its pocket) than just knowing C. A and B are
not directly causally connected, and they are not independent conditional on C.
In Salmon's example, event C does not completely describe all of the common causes of A and
B. C tells us that there was a collision of some sort between the cue ball and the 1 or 2 balls, but
it does not tell us the nature of the collision. A informs us about the nature of the collision and
therefore tells us more about B. Were the prior event more informative--for example, were it to
specify the exact momentum of the cue ball on striking the two target balls--conditional
independence would be regained. The example simply reflects a familiar problem in real data
analysis that arises whenever some proxy variable is used in causal analysis or distinct values of
a variable are collapsed. In our view these examples give no reason to doubt the Causal Markov
Condition.
Elliott Sober (1987) argues that we routinely find correlations for which there are no common
causes, or for which residual correlations remain after conditioning on known common causes.
The correlation of bread prices in England and the sea level in Venice may have some common
causes (perhaps the industrial revolution), but not enough to account for all of the dependency.
His point seems to be Yule's: if we consider a series in which variable A increases with time
and a series in which variable B increases with time, then A and B will be correlated in the
population formed from all the units-at-times, even though A and B have no causal connection.
Any such combined population is obviously a mixture of populations given by the time values.
There is a more fundamental objection to the Causal Markov Condition, namely that there exist
non-deterministic causal systems for which, to the best of current knowledge, the condition is
false. Consider pair production: a quantum mechanical event produces two particles which
move off in different directions. Because of conservation laws, dynamical variables in the two
particles must be correlated; if one has a component of spin up, for example, the other must
have that spin component down. We can do experiments in which for pairs we measure either
of two different components of spin at two spatially separated sensors and compute the
correlations. Suppose there is some state S of the system at the moment the pair of particles is
produced such that, conditional on S, the dynamical variables of the two particles are
uncorrelated. J. S. Bell (1964) argued that on such an assumption there follows an inequality
constraining the correlations of the measured dynamical variables. While the assumptions
needed for the derivation are controversial, the empirical facts seem beyond doubt: Bell's
inequality is violated in certain quantum mechanical experiments. In such experiments the
correlated variables are associated with spatially remote subsystems, so unless principles
constraining causal processes to act "locally" that is, not instantaneously over a distance, are
abandoned, any statistical dependency is presumably rwt due to the effect of one sub-system on
the other or to a common cause. Thus unless the locality principles are abandoned, the Causal
Markov Condition appears to be false (Elby, 1992).
In our view the apparent failure of the Causal Markov Condition in some quantum mechanical
experiments is insufficient reason to abandon it in other contexts. We do not, for comparison,
abandon the use of classical physics when computing orbits simply because classical dynamics
is literally false. The Causal Markov Condition is used all the time in laboratory, medical and
engineering settings, where an unwanted or unexpected statistical depedency is prima facie
something to be accounted for. If we give up the Condition everywhere, then a statistical
dependency between treatment assignment and the value of an outcome variable will never
require a causal explanation and the central idea of experimental design will vanish. No weaker
principle seems generally plausible; if, for example, we were to say only that the causal parents
of Y make Y independent of more remote causes, then we would introduce a very odd
discontinuity: So long as X has the least influence on Y, X and Y are independent conditional
on the parents of X. But as soon as X has no influence on Y whatsoever, X and Y may be
statistically dependent conditional on the parents of Y.
The basis for the Causal Markov Condition is, first, that it is necessarily true of populations of
structurally alike pseudo-indeterministic systems whose exogenous variables are distributed
independently, and second, it is supported by almost all of our experience with systems that can
be put through repetitive processes and whose fundamental propensities can be tested. Any
persuasive case against the Condition would have to exhibit macroscopic systems for which it
fails and give some powerful reason why we should think the macroscopic natural and social
systems for which we wish causal explanations also fail to satisfy the condition. It seems to us
that no such case has been made.
3.5.2 Faithfulness and Simpson's Paradox
Faithfulness can be violated in cases that realize variants of Simpson's "paradox" as Simpson
originally presented it. We have already seen that both Yule and Pearson observed that two
variables may be independent in subpopulations but dependent in a combined population. In
1948, M. G. Kendall used an example in his Advanced Theory of Statistics illustrating the
reverse situation: two binary variables are independent but are dependent conditional on a third
variable. Kendall's case was given a twist in a paper by Simpson (1951) a few years later, who
thought his example introduced difficulties about the relation between causal dependencies and
contingency tables. Subsequently the phenomenon the example exhibits has been referred to as
"Simpson's paradox." Like examples have since become standard puzzlers in discussions of the
connection between causality and probability.
Kendall's example5 was as follows:
Consider the case in which a number of patients are treated for a disease and there is
noted the number of recoveries. Denoting A by recovery, -A by non-recovery, B by
treatment, -B by not-treatment6, suppose the frequencies are
B -B Totals
A 100 200 300
-A 50 100 150
Totals 150 300 450
Here (AB) = 100 = (A)(B)/N, so that the attributes are independent. So far as can be
seen, treatment exerts no effect on recovery. Denoting male sex by SM and female sex
by SF, suppose the frequencies among males and females are
Males
BSM -BSM Totals
ASM 80 100 180
-ASM 40 80 120
Totals 120 180 300
Females
BSF -B SF Totals
ASF 20 100 120
-ASF 10 20 30
Totals 30 120 150
In the male group we now have
5p. 319. Q is Yule's Q = (ad - bc)/(ad+bc) when the first row is a,b and the second c,d in a 2 X 2 table.
6(sic) Kendall means, of course, that the symbols denote the respective treatment and recovery states, not vice-
versa.
QAB.SM = 0.231
and in the female group
QAB.SF = - 0.429
Thus among the males treatment is positively associated with recovery, and among the
females negatively associated. The apparent independence in the two together is due to
canceling of these associations in the sub-populations.
Kendall's example is thus of a mixture of two distributions, one for males and one for females,
such that the positive association between two variables in one population is exactly canceled by
the negative association in the other. There is nothing paradoxical in that, and one may fmd
empirical examples for which the same structure is claimed. The mixed distribution will violate
the Faithfulness Condition, because it will exhibit a statistical independence relation that does
not follow from the Markov condition applied to the causal graph common to all units.
Kendall's explanation of his contingency table depends on the fact that in one population the
association of two variables is positive, and in the other negative. But what can be going on if
in both sub-populations the association is positive, and yet in the mixed population it vanishes?
That is exactly the question Simpson posed in 19517. Simpson gave the following table and
commentary:
Male Female
Untreated Treated Untreated Treated
Alive 4/52 8/52 2/52 12/52
Dead 3/52 5/52 3/52 15/52
This time ... there is a positive association between treatment and survival both among
males and among females; but if we combine the tables we ... find that there is no
association between treatment and survival in the combined population. What is the
7Pienberg (1977), citing Darroch, attributes the issue to Yule "since Yule discussed it in the final section of his
1903 paper on the theory of association of atbibutes."(p. 51.) But save for the first sentence of that section, Yule
actually discusses the reverse issue of mixtures, namely circumstances in which variables ate statistically
dependent in a population but independent in sub-populations.
"sensible" interpretation here? The treatment can hardly be rejected as valueless to the
race when it is beneficial when applied to males and to females.
The question is what causal dependencies can produce such a table, and that question is
properly known as "Simpson's paradox."8
In Simpson's example the variables G (male or female), T (treated or untreated) and S (survives
or does not) are given an interpretation that imposes tacit restrictions on causal structure. When
we read the example we naturally assume that gender G cannot be caused by treatment T or
survival S, but may cause them. As with Kendall's example, the distribution in Simpson's table
satisfies the Causal Markov Condition for a graph in which G causes T and S and T causes S.
Simpson's distribution is not, however, faithful to such a graph, because T and S are
independent in the distribution even though T is a parent of S in the graph.
Suppose for a moment that we ignore the interpretation that Simpson gave to the variable,S in his
example, which was, after all, entirely imaginary, and let ourselves consider causal structures
that would be excluded by that interpretation. To avoid substantive associations, we substitute
A for T, B for G and C for S and obtain graph (i) in figure 12. Distributions such as Simpson's
and Kendall's can also be realized by a graph in which A and C are not adjacent but each causes
B, as in graph (ii) in figure 12.9
With the substitution of variables just noted, Simpson's distribution is faithful to graph (ii) but
not to graph (i); moreover (ii) is the only graph faithful to the distribution.
8The subsequent literature has confused it with a number of other questions about how independence and
dependence relations in a population may be related to independence and dependence relations in sub-populations,
and the causal significance of such facts. The unfortunate aspect of collapsing these questions is that they have
distinct answers. A circumstance attributed to Simpson and now often called "Simpson's paradox," but
nonetheless distinct from the question Simpson actually posed, was described by Colin Blyth (1972):
It is possible to have simultaneously
(1) P(AIB) < P(AIB')

and
(2) P(AIBC) 2: P(AIB'C)
(3) P(AIBC12: P(AIB'C')
In fact, Simpson has equality in (1) and > in (2) and (3).
~e point is implicit in Blalock (1961) and no doubt other sources as well.
/'"
B B
A
/~.C A C
t t t t
(i) (ii)
Figure 12
Judea Pearl (1988) offers a Bayesian example that illustrates why, when a causal structure like
that in graph (ii) obtains, one should expect that A and C, though independent, are dependent
conditional on B: Whether or not your car starts depends on whether or not the battery is
charged and also on whether or not there is fuel in the tank, but these conditions are
independent of one another. Suppose you find that your car won't start, and you hold in that
case that there is some probability that the fuel tank is empty and some probability that the
battery is dead. Suppose next you find that the battery is not dead. Doesn't the probability that
the fuel tank is empty change when that information is added?
Were we to find that A and C are independent but dependent conditional on B, the Faithfulness
Condition requires that if any causal structure obtains, it is structure (ii). Still, structure (i) is
logically possible, and if the variables had the significance Simpson gives them we would of
course prefer it. But if prior knowledge does not require structure (i), what do we lose by
applying the Faithfulness Condition; what, in other words, do we lose by excluding causal
structures that are not faithful to the distribution?
In the linear case, the parameter values--values of the linear coefficients and exogenous
variances of a structure--form a real space, and the set of points in this space that create
vanishing partial correlations not implied by the Markov Condition have Lebesgue measure
zero.
Theorem 3.2: Let M be a linear model with directed acyclic graph G and n linear
coefficients aI, ... , an and k positive variances of exogenous variables VI, ... , Vk. Let
M«Ul, ... ,Un, Un+l, ... ,Un+k» be the distributions consistent with specifying values
n be the set of probability
<Ul, ... ,U n , un+t. ... ,Un+k> for at, ... , an and vt, ... vk. Let
measures P on the space 9t n+k of values of the parameters of M such that for every
subset V of 9t n+k having Lebesgue measure zero, P(V) = O. Let Q be the set of vectors
of coefficient and variance values such that for all q in Q every probability distribution
in with M(q) has a vanishing partial correlation that is not linearly implied by G. Then
for all P in n, P(Q) = O.
The theorem can be strengthened a little; it is not really necessary that the set of exogenous and
error variables be jointly independent--pairwise independence is sufficient. In the pseudo-
indeterministic case, faithfulness can be violated, if at all, only by very special choices of the
functional dependencies between variables. Consider a population of linear, pseudo-
indeterministic systems in which the exogenous variables are independently and normally
distributed. The conditional independence relations required by the Markov Condition will be
automatically fulfilled for every possible value of the linear coefficients--they are guaranteed just
by the way the device acts to compose linear functions. But conditional independence relations
that are not required by the Markov Condition--the sorts of conditional independence relations
that characterize distributions that are unfaithful to the causal structure of the devices--either
cannot be produced at all or can only be produced if the linear coefficients satisfy very strong
constraints.
The same moral applies to other classes of functions. While for discrete variables we have not
attempted a formal proof of a theorem analogous to 3.2, such a result should be expected on
intuitive grounds. The factorization formula for distributions satisfying the Markov Condition
for a graph provides a natural parametrization of the distributions. If an exogenous variable has
n values, it determines n-l parametric dimensions consisting of a copy of the open interval
(0,1). If an endogenous variable X has n values, a conditional probability P(XIParents(X)) in
the factorization determines another n-l parametric dimensions consisting of a copy of (0,1) for
each vector of values of the parents of X.. One expects that the set of probability values that
generate conditional independence relations not entailed by the factorization itself will be
measure zero in this parameter space.
3.6 Bayesian Interpretations
We have interpreted the conditions as about frequencies in populations in which all units have
the same causal structure. We wish to consider how the conditions can be given a Bayesian
interpretation in which the probabilities are subjective. Current subjectivist interpretations hold
that probability is an idealization of rational, subjective degree of belief. On a strict subjectivist
view there can be finite frequencies, but there is no such thing as objective probability. One
assumes the systems under study in the sciences are deterministic, and any appearance of
indeterminacy is due simply to ignorance. The likelihood structures of Bayesian statistical
models often look like ordinary un-Bayesian statistical models; Bayesians add a prior
probability distribution over the free parameters. For example, Bayesian linear models specify a
e
distribution over a parameter representing linear coefficients, variances, means, and so on.
The Bayesian model is thus a mixture of ordinary linear models, and the joint distribution over
the measured variables does not satisfy the conditions we have considered in this section.
Consider a study of systems with causal graph G. Suppose a Bayesian agent's degrees of belief
are represented by a density, f, satisfying the condition f(XIParents(G,X» =
h(Parents(G ,x); e), where e is a parameter whose values determine a density for X
conditional on its parents. Let the Bayesian agent also have a distribution over e. In such a case
we understand the Causal Markov and Causal Minimality conditions to constrain the agent's
degrees of belief conditional on e. The subjective joint distribution over the variables
conditional on e will satisfy the conditions, but typically the unconditional joint distribution
will not.
Suppose now that the agent entertains a set G of alternative possible causal structures, and
holds that in each structure Gin GftXIParents(G,x» = h(Parents(G,x); eG), as before.
Then we understand the Causal Markov and Causal Minimality conditions to constrain the
agent's degrees of belief conditional on eG, G.
So understood, the conditions are normative principles about "reasonable" degrees of belief. In
a later chapter we will consider in some detail a Bayesian proposal for clinical trials and argue
that the assumptions the proposal makes about the degrees of belief of scientific experts accords
with the Markov Condition.
3.7 Consequences of The Axioms
Consequences of the Causal Markov, Minimality and Faithfulness Conditions are developed
throughout this book,_ but some important connections between causal dependency and
statistical dependency should be noted here.
3.7.1 d-Separation
Given a causal graph G, the Markov Condition axiomatizes the set of independence and
conditional independence relations true of any distribution P faithful to G. But which
conditional independence relations follow from the Markov Condition for a given graph may
not be obvious. Suppose one wanted to know, for each pair of vertices X and Y and each set of
vertices Q not containing X and Y, whether or not X and Y are independent conditional gn Q,
i.e., all the atomic independence facts among sets of variables. Applying the Markov Condition
directly to G, that is, applying the definition for each vertex, does not in general suffice.
W Z v
~ /'-..Y /
X
Figure 13
For example, in a distribution faithful to the graph in figure 13, suppose we wanted to know
whether X and Y are independent conditional on the set Q = {Z}. Applying the Markov
Condition directly to figure 13, we obtain:
W 11 {Z,Y,V}
X 11 {y,v}1 {W,Z}
z11 {W,v}
Y 11 {W,X} I {V,Z}
V 11 {W,X,Z}
It is not obvious that these facts entail X 11 Y I {Z}. Pearl proposed a purely graphical
characterization--which he called d-separation--of conditional independence, and Geiger,
Pearl, and Verma (Geiger and Pearl 1989a; Verma 1987) proved that d-separation in fact
characterizes all and only the conditional independence relations that follow from satisfying the
Markov condition for a directed acyclic graph.
The definition of d-separation is sufficiently unintuitive that an analogy may be helpful.

Consider first the situation with unconditional independence. Think of an undirected path in a
graph as a pipe carrying causal flow, and each of the vertices are valves that are either active
(open), or inactive (closed). If a vertex is a collider, then causal flow cannot get through it, and
it is thus inactive. For example, in the causal graph at the top of figure 14, X and V are d-
separated by the empty set because Y is a collider on the only path between them .
.........W
...--Z~
~Y M'
. ./V
Figure 14
Conditioning on a node flip-flops its status. Whereas X and Y are not d-separated given the
empty set in the graph in figure 14, they are d-separated given {WI, {ZJ or {W,zJ. X and V
are d-separated given the empty set, but are not d-separated given {Y}. That conditioning on a
non-collider makes it inactive is similar to the intuition behind the Markov Condition. A non-
collider is either a common cause, e.g. Z, or part of a causal chain, e.g., W. Effects are made
independent when we condition on their common causes, and effects are made independent of
their remote causes when we condition on their more proximate ones. That conditioning on a
collider makes it active was noted in section 3.5.2 above.
Given a graph G, checking whether any two vertices X and Y are d-separated by a set Q and
thus independent conditional on Q in a distribution faithful to G appears straightforward. A path
is active if all of its vertices are active, i.e. if all colliders are in Q and all of its non-colliders are
not in Q. X and Yare d-separated given Z if no undirected path between X and Y is active
given Z. In figure 15, for example, X and Y are not d-separated given U, but they are d-
separated give {V,z}.
..,...-Z -----i.~ V
X
..........W
Y
./'
~U~
Figure 15
In the first case, conditioning on U activates the X -> U <-Y path, so X and Y are not d-
separated given U. The X <- W <- Z -> V <- Y path is inactive given U, but X and Yare not d-
separated given U as long as there is one undirected path between X and Y that is active given
U. In the second case, conditioning on {V,z} activates Von the X <- W <- Z -> V <- Y path,
but conditioning inactivates Z on this path and thus makes it inactive; the X -> U <- Y path is
also inactive given {V,z} because U is a collider on the path that is not in {V,Z}. Because all of
the undirected paths between X and Yare inactive given {V,z}, X and Yare d-separated given
{V,Z}.
Unfortunately, the full story is not quite so simple. Conditioning on a collider activates it, and
so does conditioning on any of its descendants. In the graph in figure 16, for example, X and Y
are not d-separated given W, because W is a descendent of U.
Figure 16
For a directed acyclic graph G, if X and Y are vertices in G, X * Y, and W is a set of vertices
in G not containing X or Y, then X and Y are d-separated given W in G if and only if there
exists no undirected path U between X and Y, such that (i) every collider on U has a descendent
in Wand (ii) no other vertex on U is in W. We say that if X;t Y, and X and Y are not in W,
then X and Yare d-connected given set W if and only if they are not d-separated with respect
to W. If U, V, and W are disjoint sets of vertices in G and U and V are not empty then we say
that U and V are d-separated given W if and only if every pair <U,V> in the cartesian
product of U and V is d-separated given W. If U, V, and W are disjoint sets of vertices in G
and U and V are not empty then we say that U and V are d-connected given W if and only if
U and V are not d-separated given W.
The essential results are the following:
Theorem 3.3: P(V) is faithful to directed acyclic graph G with vertex set V if and only
if for all disjoint sets of vertices X, Y, and Z, X and Y are independent conditional on Z
if and only if X and Yare d-separated given Z.
Theorem 3.4 provides a slightly more intuitive characterization of faithfulness, which motivates
algorithms developed in Chapter 5.
Theorem 3.4: If P(V) is faithful to some directed acyclic graph, then P(V) is faithful
to directed acyclic graph G with vertex set V if and only if
(i) for all vertices X, Y of G, X and Y are adjacent if and only if X and Y are
dependent conditional on every set of vertices of G that does not include X or Y; and
(ii) for all vertices X, Y, Z such that X is adjacent to Yand Y is adjacent to Z and X
and Z are not adjacent, X -> Y <- Z is a subgraph of G if and only if X, Z are
dependent conditional on every set containing Y but not X or Z.
The study of correlation is historically tied to the normal distribution, and for that distribution
vanishing partial correlations and conditional independence are equivalent. But the Markov and
Faithfulness Conditions tie vanishing correlation and partial correlation to graphical and causal
structure for linear systems, without any normality assumption. Thus for linear systems,
correlational structure is a guide to causal structure. We will say that a distribution P is linearly
faithful to a graph G if and only if for vertices A and B of G and all subset C of the vertices
of G, A and B are d-separated given C if and only PAB.C = O.
Theorem 3.5: If G is a directed acyclic graph with vertex set V, A and B are in V,
and H is included in V, then G linearly implies PAB." = 0 if and only A and B are d-
separated given H.
It follows that a distribution P is linearly faithful to a graph G if and only if for vertices A and B
of G and all subsets C of the vertices of G, A andB are d-separated given C if and only PAB.C
= O. Theorem 3.5 is the general principle behind all of the path analysis examples (Wright,
1934; Simon, 1954; Blalock, 1961; Heise, 1975) connecting causal structure in "recursive"
(i.e., acyclic) linear models with vanishing partial correlations.
In the chapters that follow we will frequently remark that some conditional independence or
conditional dependence relation, or vanishing or non-vanishing partial correlation, follows from
a causal structure, assuming the distribution is faithful. Conversely, we will often observe that
given certain conditional independence and dependence relations, or partial correlation facts, the
causal structure must have certain properties if the distribution is faithful. Whenever we make
such claims, we are using tacit corollaries of Theorems 3.3, 3.4 and 3.5.
3.7.2 The Manipulation Theorem
The fundamental aim of many empirical studies is to predict the effects of changes, whether the
changes come about naturally or are imposed by deliberate policy. How can an observed
distribution P be used to obtain reliable predictions of the effects of alternative policies that
would impose a new marginal distribution on some set of variables? The very idea of imposing
a policy that would directly change the distribution of some variable (e.g., drug use)
necessitates that the resulting distribution PMan will be different from P. P alone cannot be used
to predict PMan, but P and the causal structure can be.
Suppose that the Surgeon General is considering discouraging smoking, and he asks "What
would the distribution of Cancer be if no one in the U.S were allowed to smoke?" Let V =
{Drinking, Smoking, and Cancer}. For the purpose of illustration assume that in the actual
population in the U.S. the causal structure shown in figure 17 is correct.
Cancer
GUnman
Figure 17
Let us call the population actually sampled (or produced by sampling and some experimental
procedure) the unmanipulated population, and the hypothetical population for which
smoking is banned the manipulated population. Suppose that if the policy of banning
smoking were put into effect it would be completely effective, stopping everyone from
smoking, but would not affect the value of Drinking in the population. Then the causal graph
for the hypothetical manipulated population will be different than for the unmanipulated
population, and the distribution of Smoking is different in the two populations. The
manipulated causal graph is shown in figure 18.
G Man
Figure 18
The difference between the unmanipulated graph and the manipulated graph is that some
vertices that are parents of the manipulated variables in GUnman may not (depending upon the
precise form of the manipulation) be parents of manipulated variables in GMan and vice-versa.
How can we describe the change in the distribution of Smoking that will result from banning
smoking? One way is to note that the value of a variable that represents the policy of the federal
government is different in the two populations. So we could introduce another variable into the
causal graph, the Ban Smoking variable, which is a cause of Smoking. The full causal graph,
including the new variable representing smoking policy, is then shown in figure 19. In the
actual unmanipulated population the Ban Smoking variable is off, and in the hypothetical
population the Ban Smoking variable is on. In the actual population we measure
P(SmokinglBan Smoking = off); in the hypothetical popUlation that would be produced if
smoking were banned P(Smoking =0 lBan Smoking =on) = 1. For any subset X of V =
{Smoking, Drinking, Cancer} in the causal graph, let PUnman(Ban Smoking)(X) be P(XlBan
Smoking = off) and let PMan(Ban Smoking)(V) be P(VlBan Smoking = on) .
Ban
Smoking
G Comb
Figure 19
We can now ask if PUnman(Ban Smoking)(CancerISmoking) = PMan(Ban

Smoking)(CancerISmoking) (for those values of Smoking for which PMan(Ban
Smoking)(CancerISmoking) is defined, namely Smoking = 0)7 Clearly the answer is affirmative
exactly when Cancer and Ban Smoking are independent given Smoking; but if the distribution
is faithful this just reduces to the question of whether Cancer and Ban Smoking are d-separated
given Smoking, which they are not in this causal graph. Further, PUnman(Ban Smoking)(Cancer)
¢:PMan(Ban Snwking)(Cancer) because Cancer is not d-separated from Ban Smoking given the
empty set. But in contrast PUnman(Ban Smoking)(CancerISmoking,Drinking) = PMan(Ban
Smoking)(CancerISmoking,Drinking) (for those values of Smoking for which PMan(Ban
Smoking)(CancerISmoking,Drinking) is defined, namely Smoking =0), because Ban Smoking
and Cancer are d-separated by {Smoking, Drinking}. The importance of this invariance is that
we can predict the distribution of cancer if smoking is banned by considering the conditional
distribution of cancer given drinking in the observed subpopulation of non-smokers, and by
considering the distribution of drinking in the unmanipulated population.
Note that one of the inputs to our conclusion about PMan(Ban Smoldng)(Cancer) is that the ban on
smoking is completely successful and that it does not affect Drinking; this knowledge does not
come from the measurements that we have made on Smoking, Drinking and Cancer, but is
assumed to come from some other source. Of course, if the assumption is incorrect, there is no
guarantee that our calculation of PMan(Ban Snwking)(Cancer) will yield the correct result. If we
had instead considered a policy that does not effectively ban smoking, but intervenes to make
smoking less likely without affecting drinking, then the graph of the entire system including the
manipulation variable Ban Smoking, would be the same as in figure 19, and the graph GUnman
would be as in figure 17, but the manipulated graph GMan would look like figure 17 rather than
18. Intervention would not remove the influence of drinking on smoking.
The analysis of prediction for a system involves three distinct graphs: a causal graph GComb
which includes variables W representing manipulations, and a causal graph GUnman which is
the subgraph of GComb over a set of variables V not including the variables representing
manipulations, and a graph GMan over V which represents the causal relations among variables
in V that result from a manipulation. GMan may be a subgraph of GUnman if the manipulation
"breaks" causal dependencies in GUnman; otherwise GUnman and GMan will be the same graph.
Here are the formal definitions: IT G is a directed acyclic graph over a set of variables V u W,
and V n W = 0, then W is exogenous with respect to V in G if and only if there is no
directed edge from any member of V to any member of W. If GComb is a directed acyclic graph
over a set of variables V u W, and P(V u W) satisfies the Markov condition for GComb, then
changing the value of W from Wl to Wz is a manipulation of GComb with respect to V if and
only if W is exogenous with respect to V, and P(VIW = Wl) '# P(VIW = WZ).
We define PUnman(W)(V) = P(VIW = Wl), and PMan(W)(V) = P(VIW = wz), and similarly for
various marginal and conditional distributions formed from P(V).
We refer to GComb as the combined graph, and the subgraph of GComb over V as the
unmanipulated graph GUnman. (Note that while PUnman(W)(V) satisfies the Markov
Condition for GUnman, it may also satisfy the Markov Condition for a subgraph of GUnman.
This is because G Comb, and hence its subgraph GUnman, may contain edges that are needed to
represent the distribution of the manipulated subpopulation but not needed to represent the
distribution of the unmanipulated subpopulation.)
V is in ManipuIated(W) (that is, V is a variable directly influenced by one of the manipulation

variables) if and only if V is in Children(W) n V; we will also say that the variables in
Manipulated(W) have been directly manipulated. We will refer to the variables in W as
policy variables.
The manipulated graph, GMan is a subgraph of GUnman for which PMan(W)(V) satisfies the
Markov Condition and which differs from GUnman in at most the parents of members of
Manipulated(W). Exactly which subgraph GMan is depends upon the details of the
manipulation and what the causal graph of the subpopulation where W = wz is. For example,
if smoking is banned, then GMan contains no edge between income and smoking. On the other
hand, if taxes are raised on cigarettes, GMan does contain an edge between income and
smoking. We will prove (in Chapter 13) that given a manipulation as defined, there always
exists a subgraph of GUnman for which PMan(W)(V) satisfies the Markov Condition. All of our
theorems about manipulations hold for any GMan that is a subgraph of GUnman for which
PMan(W)(V) satisfies the Markov Condition, and which differs from GUnman in at most the
parents of members of Manipulated(W).
These definitions entail the Manipulation Theorem:
Theorem 3.6: (Manipulation Theorem): Given directed acyclic graph GComb over
vertex set V u Wand distribution P(V u W) that satisfies the Markov condition for
GComb, if changing the value of W from WI to Wz is a manipulation of GComb with
respect to V, GUnman is the unmanipulated graph, GMan is the manipulated graph, and
PUnman(W)(V) = II PUnman(W) (XIParents(Gunman ,X))

XEV
for all values of V for which the conditional distributions are defined, then
PMan(W) (V) =
II
PMan(W) (XI Parents(GMan , X)) X
X EManipulated(W)
II
PUnman(W) (XIParents(Gunman ,X))
X EV\Manipulated(W)
for all values of V for which each of the conditional distributions is defined.
The importance of this theorem is that if the causal structure and the direct effects of the
manipulation (i.e. PMan(W)(XIParents(X)) for each X in Manipulated(W) are known, then
the joint distribution can be estimated from the unmanipulated population.
The Manipulation Theorem is not applicable when a causal mechanism between a pair of
variables is reversible, in which case there can be two subpopulations in which the direction of
the causal relationship between a pair of variables is reversed. For example, the movement of a
motor of a car may cause the wheels to turn (as when the gas pedal is pressed), but also the
turning of the wheels can cause the motor to move (as when the car rolls downhill.)l0 An
intervention in a causal system which reverses the direction of some causal relationship is not a
manipulation in our technical sense because there is no one combined graph representing the
causal relations in the combined population. We are not suggesting any non-experimental
lOwe thank Marek Druzdel for suggesting this example. and pointing out the problem of reversible mechanisms
to us.
methods for determining whether a given mechanism is reversible. In some cases, such as
smoking and yellow fingers, it is obvious from background knowledge that the mechanism is
not reversible, because yellow fingers cannot cause smoking. In other cases, the relevant
background knowledge may not be available, in which case it is not known whether the
Manipulation Theorem is applicable.
Rubin (1977; 1978), and following him Pratt and Schlaifer (1988), have offered rules for when
conditional probabilities in an observed population of systems will equal conditional
probabilities for the same variables if the population is altered by a direct manipulation of some
variables for all population units. We will show in Chapter 7 that their various rules are direct
consequences of the special case of the Manipulation Theorem, illustrated in the discussion of
figures 17, 18 and 19, in which one variable is manipulated and the intervention makes that
variable independent of its causes in the unmanipulated graph.
Because the Manipulation Theorem is a consequence of the Markov condition, it requires no

separate justification. Although the Manipulation Theorem is abstract, it is just the general
formulation of inferences that are routine, if not always correct. When, for example, a
regression model is used to predict the effects of a policy that would force values on some of
the regressors, we have an application of the Manipulation Theorem. Of course the prediction
may be incorrect if the causal or statistical assumptions of the regression model are false, or if
the changes actually carried out do not satisfy the conditions for a manipulation. There are
striking examples of both sorts of failure. Application of the Manipulation Theorem may give
misleading predictions if the values of variables for each unit depend on the values of other
units and if that dependency is not represented in the causal graph. Some public policy debates
illustrate absurd violations of this requirement. Recently a research institute funded by
automobile insurers carried out a non-linear regression of the rate of fatalities of occupants of
various kinds of cars against car length, weight and other variables, finding unsurprisingly that
the smaller the car the higher the fatality rate. This statistical analysis was then used by others to
argue that proposed federal policies to downsize the American automobile fleet would increase
highway fatalities. But of course the fatality rate in cars of a given size depends on the
distribution of sizes of other cars in the fleet.
One can mistake which variables will be directly affected by a policy or intervention. Tacit
applications of the Manipulation Theorem in such cases can lead to disappointment. As we will
see in a later chapter, the literature on smoking, lung cancer and mortality provides vivid
examples of predictions that went wrong, arguably because of misjudgments as to which
variables would be directly manipulated by an intervention.
There is no reason why every intervention to deliberately alter the distribution of values of a set
of variables V among units in a population (or sample) must satisfy the conditions for a direct
manipulation of V and no others. But one of the chief aims in the design of experiments is to
see to it that experimental manipulations are in fact direct manipulations of the intended variables
and no others. The point of blind and double blind designs, for example, is exactly to obtain in
experiment a direct manipulation of only the treatment variables. The concern with chronic
wounding in drug trials with animals is essentially a worry that with respect to the outcome
variables of interest, outcome variables as well as pharmacological variables have been directly
manipulated. Typically, when we mistake the variables an intervention will directly manipulate,
predictions of the outcomes of intervention will fail.
Our discussion in this section has assumed that the causal structure of the system is fully
known. In Chapters 6 and 7 we will consider when and how the effects of interventions can be
predicted from an unmanipulated distribution, assuming the distribution is the marginal over the
measured variables of a distribution faithful to an unknown causal graph, and assuming the
intervention constitutes a direct manipulation in the sense we have defined here.
3.8 Determinism
Another way that the Faithfulness Condition can be violated is when there are deterministic
relationships between variables. In this section, we will give some rules for determining what
extra conditional independence relations are entailed by deterministic relationships among
variables.
We will say that a set of variables Z determines the set of variables A, when every variable in
A is a deterministic function of the variables in Z, and not every variable in A is a deterministic
function of any proper subset of Z. When there are deterministic relationships among variables
in a graph, there are conditional independencies that are entailed by the deterministic
relationships and the Markov condition that are not entailed by the Markov condition alone. For
example, if G is a directed acyclic graph over V, V contains Z and A, and Z determines A, then
A is independent of vv..Z u (A}) given Z. If Z is a proper subset of the parents of A then this
entails that A is independent of its other parents given Z, and also independent of its
descendants as well as its non-descendants given Z. But it could also be the case that the
members of Z are children of A, in which case given its children, A is independent of all other
variables including its parents. It is also possible that Z could contain non-parental ancestors of
A. Each of these cases entails conditional independence relations not entailed by the Markov
condition alone. For example, consider the graph in figure 20.
Figure 20
No conditional independence relations among A, B, or C are entailed by the Markov Condition

alone. However, if the grandparent A determines the grandchild C, then C 11 BIA. If the
parent B determines the child C then C 11 AlB. If the child C determines the parent B then B
11 AIC.
Hence d-separability relations do not capture all of the conditional independencies entailed by
the Markov condition and a set of deterministic relations. We will look for a graphical condition
which entails the conditional independence of variables given the Markov condition and a set of
detenninistic relations among variables.
Geiger has proposed a simple, provably complete rule for graphically determining the
conditional independencies entailed by the Markov and Minimality conditions and one kind of
deterministic relationship among variables. Following Geiger (1990), in a directed acyclic graph
G over V that includes A and Z, say that vertex A is a deterministic variable if it is a
deterministic function of its parents in G. (Note that if a variable A has no parents in G, but has
a constant value, then A is a deterministic variable.) A is functionally determined by Z if
and only if A is in Z, or A is a deterministic variable and all of its parents are functionally
determined by Z. If X, Y, and Z are three disjoint subsets of variables in V, X and Y are D-
separated given Z if and only if there is no undirected path U between any member of X and
any member of Y such that each collider has a descendant in Z and no other variable on U is
functionally determined by Z. Geiger has shown that X and Y are D-separated given Z if and
only if for every distribution that satisfies the Markov and Minirnality Conditions for G, and the
deterministic relations, X and Yare independent given Z. We will prove that Geiger's rule is
correct for a much wider class of deterministic relations; we do not know if it is complete for
this wider class of detenninistic relationships.
Suppose G is a directed acyclic graph over V. and Deterministic(V) is a set of ordered tuples
of variables in V. where for each tuple D in Deterministic(V). if D is <VI •...• Vn> then Vn is
a deterministic function of VI ..... Vn-1 and is not a deterministic function of any subset of VI
.... , Vn-1; we also say (V1, ... ,Vn-1l determines Vn. Note that Vn could be an ancestor in G
of members of V 1..... V n-1. Also, if A determines Band B determines A then
Deterministic(V) contains both <A,B> and <B,A>. We assume that Deterministic(V) is
complete in the sense that if it entails some deterministic relationships among variables. those
deterministic relations are in Deterministic(V). (For example. if A determines Band B
determines C. then A determines C.) Det(Z) is the set of variables determined by some subset
of Z. If a variable A has a constant value. then we say that it is determined by the empty set,
and is in Det(Z) for all Z.
Note that Deterministic(V) can entail dependencies between variables as well as

independencies. If Z determines A, and Z is a member of Z. then A is dependent on z\{Z}
given Z. (Other dependencies may be entailed by Deterministic(V) as well.) These
dependencies may conflict with independencies entailed by satisfying the Markov Condition for
a directed acyclic graph G. so not every Deterministic(V) is compatible with every directed
acyclic graph with vertex set V. If Deterministic(V) and directed acyclic graph G are
incompatible, Theorem 3.7 stated below is vacuously true. but obviously it would be desirable
to have a test for determining whether Deterministic(V) and G 'are compatible.
We will expand Geiger's concept of D-separability so that it is not limited to the kind of
deterministic relations that he considers. If G is a directed acyclic graph with vertex set V, Z is
a set of vertices not containing X or Y, X ¢ Y. then X and Y are D-separated given Z and
Deterministic(V) if and only if there is no undirected path U in G between X and Y such that
each collider on U has a descendant in Z. and no other vertex on U is in Det(Z); otherwise if X
¢ Y and X and Yare not in Z, then X and Y are D-connected given Z and
Deterministic(V). Similarly, if X. Y, and Z are disjoint sets of variables, and X and Y are
non-empty. then X and Yare D-separated given Z and Deterministic(V) if and only if each
pair <X .Y> in the Cartesian product of X and Y are D-separated given Z and
Deterministic(V); otherwise if X, Y. and Z are disjoint, and X and Y are non-empty. then X
and Y are D-connected given Z and Deterministic(V).
Theorem 3.7: If G is a directed acyclic graph over V, X, Y, and Z are disjoint

subsets of V. and P(V) satisfies the Markov condition for G and the deterministic
relations in Deterministic(V) then if X and Y are D-separated given Z and
Deterministic(V). X and Y are independent given Z in P.
For example, suppose G is the graph in figure 21, and Deterministic(V) = {<A,8>, <.B,C>,
<A,C>}.
Figure 21
Band C are D-separated given A and Deterministic(V), and A and C are D-separated given B
and Deterministic(V).
Suppose that G is still the graph in figure 21, but now Deterministic(V) = {<A,8>, <.B,A>,
<.B,C>, <C,8>, <A,C>, <C,A>}. In addition to the previous D-separability relations, now A
and B are D-separated given C and Deterrninistic(V) because C determines A.
In some cases, conditional independencies are entailed because a parent is determined by its
child. Consider the graph in figure 22, where De t e r min i s tic (V) =
{<Y,W,Z>,<Z,Y>,<Z,W>}. X and T are D-separated given Z and Deterministic(V) because
Z determines Y and W, and Y and W are non-colliders on the only undirected path between X
and T.
Figure 22
Finally, we note that it is possible that some non-parental ancestor X of Z determines Z, even
though X does not determine any of the parents of Z. Let G be the graph in figure 23 and
Deterministic(V) = {<x,z>}. Suppose X, R, and Z each have two values, and Y has four
values. Consider the following distribution (where we give the probability of each variable
conditional on its parents):
P(X = 0) =.2
P(R = 0) =.3
P(Y = OIX = O,R = 0) = 1
P(Y = llX = O,R = 1) = 1
P(Y = 21X = 1,R = 0) = 1

P(Y = 31X = 1,R = 1) = 1
P(Z = OIY= 0) = 1
P(Z = OIY = 1) = 1
P(Z = 11Y = 2) = 1
P(Z = lIY = 3) = 1
In effect Yencodes the values of both R and X, and Z decodes Y to match the value of X.
Figure 23
It follows that Y and Z are D-separated given X and Deterministic(V), and X and Z are D-
separated given Y and Deterministic(V).
The following example points up an interesting difference between the set of distributions that
satisfy the Markov condition for a given directed acyclic graph G, and the set of distributions
that satisfy the Markov condition and a set of deterministic relationships among the variables in
G. Suppose G is the graph shown in figure 24. For any directed acyclic graph, the set of
probability distributions that satisfy the Markov condition for the graph includes some
distributions that also satisfy the Minirnality Condition for the graph. Suppose however, that
Deterministic(V) = {<X ,Y> }. In this case, among the distributions that satisfy the Markov
Condition and the specified deterministic relations, there is no distribution that also satisfies the
Minimality Condition. All distributions that satisfy the Markov Condition and the specified
deterministic relation are faithful to the subgraph of figure 24 that does not contain the Z -> Y
edge. This suggests that to find all of the conditional independence relations entailed by
satisfying the Markov Condition for a directed acyclic graph G and a set of deterministic
relations, one would need to test for D-separability in various subgraphs G' of G with vertex
set V in which for each Y in V no subset of Parents(G',Y) determines Y.
X---"'Y~Z
Figure 24
We will not consider algorithms for constructing causal graphs when such deterministic
relations obtain, nor will we consider tests for deciding whether a set of variables X determines
a variable Y.
The ambiguous use of hypotheses to represent both causal and statistical constraints is nearly as
old as statistics. In modern form the use of the idea by Spearman (1904) early in the century
might be taken to mark the origins of statistical psychometrics. Directed graphs at once
representing both statistical hypotheses and causal claims were introduced by Sewell Wright
(1934) and have been used ever since, especially in connection with linear models. For a
number of particular graphs the connections between linear models and partial correlation
constraints were described by Simon (1954) and by Blalock (1961) for theories without
unmeasured common causes, and by Costner (1971) and Lazarsfe1d and Henry (1968) for
theories with latent variables, but no general characterization emerged. A distribution-free
connection of graphical structure for linear models with partial correlation was developed in
Glymour, Scheines, Spirtes and Kelly (1987), for first order partials only, but included cyclic
graphs. Geiger and Pearl (1989a) showed that for any directed acyclic graph there exists a
faithful distribution. The general characterization given here as Theorem 3.5 is due to Spirtes
(1989), but the connection between the Markov Condition, linearity and partial correlation
seems to have been understood already by Simon and Blalock and is explicit in Kiiveri and
Speed (1982). The Manipulability Theorem has been used tacitly innumerable times in
experimental design and in the analysis of shocks in econometrics but seems never to have
previously been explicitly formulated. A special case of it was first given in Spirtes, Glymour,
Scheines, Meek, Fienberg and Slate (1991). The Minimality Condition and the idea of d-
separability are due to Pearl (1988), and the proof that d-separability determines the
consequences of the Markov condition is due to Verma (1987), Pearl, and Geiger (1989a). A
result entailing theorem 3.4 was stated by Pearl, Geiger, and Verma (1990). Theorem 3.4 was
used as the basis for a causal inference algorithm in Spirtes, Glymour, and Scheines (1990c).
D-separability is described in Geiger (1990).
Chapter 4
Statistical Indistinguishability
Without experimental manipulations, the resolving power of any possible method for inferring
causal structure from statistical relationships is limited by statistical indistinguishability. If two
causal structures can equally account for the same statistics, then no statistics can distinguish
them. The notions of statistical indistinguishability for causal hypotheses vary with the
restrictions one imposes on the connections between directed graphs representing causal
structure and probabilities representing the associated joint distribution of the variables. If one
requires only that the Markov and Minimality Conditions be satisfied, then two causal graphs
will be indistinguishable if the same class of distributions satisfy those conditions for one of the
graphs as for the other. A different statistical indistinguishability relation is obtained if one
requires that distributions be faithful to graph structure; and still another is obtained if the
distributions must be consistent with a linear structure, and so on. For each case of interest, the
problem is to characterize the indistinguishability classes graph-theoretically, for only then will
one have a general understanding of the causal structures that cannot be distinguished under the
general assumptions connecting causal graphs and distributions.
There are a number of related considerations about the resolving power of any possible method
of causal inference from statistical properties. Given axioms about the connections between
graphs and distributions, what graph theoretic structure must two graphs share in order also to
share at least one probability distribution satisfying the axioms? When, for example, do two
distinct graphs admit one and the same distribution satisfying the Minirnality and Markov
Conditions? When do two distinct graphs admit one and the same distribution satisfying the
Minirnality and Markov Conditions for one and the Faithfulness and Markov Conditions for the
other? Reversing the question, for any given probability distribution that satisfies the Markov
and Minimality Conditions (or in addition the Faithfulness Condition) for some directed acyclic
graph, what is the set of all such graphs consistent with the distribution and these conditions?
Finally, there are relevant measure-theoretic questions. If procedures exist that will identify
causal structure under a more restrictive assumption such as Faithfulness, but not always under
weaker assumptions such as the Markov and Minimality Conditions, how likely are the cases in
which the procedures fail? Under various natural measures on sets of distributions, for
example, what is the measure of the set of distributions that satisfy the Minimality and Markov
Conditions for a graph but are not faithful to the graph?
These are fundamental questions about the limits of any possible inference procedure--whether
human or computerized--from non-experimental data to structure. We will provide answers for
many of these questions when the system of measured variables is causally sufficient. Statistical
indistinguishability is less well understood when graphs can contain variables representing
unmeasured common causes.
4.1 Strong Statistical Indistinguishability
Two directed acyclic graphs G, G' are strongly statistically indistinguishable (s.s.i) if
and only if they have the same vertex set V and every distribution P on V satisfying the
Minimality and Markov Conditions for G satisfies those conditions for G', and vice-versa.
That two structures are s.s.i. of course does not mean that the causal structures are one and the
same, or that the difference between them is undetectable by any means whatsoever. From the
correlation of two variables, X and Y, one cannot distinguish whether X causes Y, Y causes X
or there is a third common cause, Z. But these alternatives may be distinguished by experiment
or, as we will see, by other means.
Strong statistical indistinguishability is characterized by a simple relationship, namely that two

graphs have the same underlying undirected graph and the same collisions:
Theorem 4.1: Two directed acyclic graphs GI, G2, are strongly statistically
indistinguishable if and only if (i) they have the same vertex set V, (ii) vertices VI and
V2 are adjacent in GI if and only if they are adjacent in G2, and (iii) for every triple VI,
V2, V3 in V, the graph VI -> V2 <- V3 is a subgraph of GI if and only if it is a
subgraph of G2.
Given an arbitrary directed acyclic graph G, the graphs s.s.i. from G are exactly those that can
be obtained by any set of reversals of the directions of edges in G that preserves all collisions in
Statistical Indistinguishability 89
G. A decision as to whether or not two graphs are s.s.i. requires O(n3) computations, where n
is the number of vertices.
In figure I graphs Gl and Gz are s.s.i., but Gl and G3, and G2 and G3 are not s.s.i.
¥"D~
A B
~/
c
A .B A 4 B
'\..1 C
\c/
Figure 1
Note, however, if a set of variables V is totally ordered, as for example by a known time order,
and P(V) is positive, then there is a unique graph for which P(V) satisfies the Minirnality and
Markov conditions. (See Corollary 3 in Pearl 1988.)
4.2 Faithful Indistinguishability
Suppose we assume that all pairs <G, P> are faithful: all and only the conditional independence
relations true in P are a consequence of the Markov condition for G. We will say that two
directed acyclic graphs, G, G' are faithfully indistinguishable (f.i.) if and only if every
distribution faithful to G is faithful to G' and vice-versa. The problem is to characterize faithful
indistinguishability graphically.
Theorem 4.2: Two directed acyclic graphs G and H are faithfully indistinguishable if
and only if (i) they have the same vertex set, (ii) any two vertices are adjacent in G if
and only if they are adjacent in H, and (iii) any three vertices, X, Y, Z, such that X is
adjacent to Yand Y is adjacent to Z but X is not adjacent to ZinG or H, are oriented as
X -> Y <- ZinG if and only if they are so oriented in H.
The question of faithful indistinguishability for two graphs can be decided in O(n3) where n is
the number of vertices.
It is immediate from Theorems 4.1 and 4.2 that if two graphs are strongly statistically
indistinguishable they are faithfully indistinguishable, but not necessarily conversely. The
graphs G4 and Gs in figure 1 are not s.s.i. but they are f.i.
A class of f.i. graphs may be represented by a pattern. A pattern n is a mixed graph with
directed and undirected edges. A graph G is in the set of graphs represented by n if and
only if:
(i) G has the same adjacency relations as II;
(ii) if the edge between A and B is oriented A -> B in n, then it is oriented A -> BinG;
(iii) if Y is an unshielded collider on the path <X,y,z> in G then Y is an unshielded collider
on <X,Y,Z> in n.
For example, the set of all complete, acyclic directed graphs on three vertices forms a faithful
indistinguishability class that can be represented by a pattern consisting of the complete
undirected graph on the same vertex set. When the pattern of the faithful indistinguishability
class of a directed acyclic graph has no directed edges, and so is purely undirected, the
statistical hypothesis represented by the directed graph is equivalent to the statistical hypothesis
of the undirected independence graph corresponding to the pattern.
4.3 Weak Statistical Indistinguishahility
The indistinguishability relations characterized in the two previous sections ask for the graphs
that can accommodate the same class of probability distributions as a given graph. We can tum
the tables, at least partly, by starting with a particular probability distribution on a set of
variables and asking for the set of all directed acyclic graphs on those vertices that are consistent
with the given distributions. The answers characterize how much the probabilities and our
assumptions about the connection between probabilities and causes underdetermine the causal
structure. Assuming Markov and Minimality only, the equivalence of these two conditions
(under positivity) with the defining conditions for a directed independence graph provides an
(impractical) algorithm for generating the set of all graphs that satisfy the two conditions for a
given distribution P. For every ordering of the variables in P there is a directed acyclic graph G
compatible with that ordering (i.e. A precedes B in the ordering only if A is not a descendant of
B in G) satisfying the Minimality and Markov Conditions for P. It can be generated by
assuming the ordering and the conditional independence relations in P and applying the
definition of directed independence graph. An algorithm that does not assume positivity is given
by Pearl (1988). According to that algorithm let Ord be a total ordering of the variables, and
Predecessors(Ord,x) be the predecessors of X in the ordering Ord. For each variable X, let
the parents of X in G be a smallest subset R of Predecessors(Ord,x) such that X is
independent of Predecessors(Ord,x)\R given R in P. It follows that P satisfies the
Minimality and Markov Conditions for P.
The alternatives are more limited if we start with P and assume that any graph must be faithful
to P. In that case all of the graphs faithful to P form a faithful indistinguishability class, i.e., the
set of all graphs f.i. from anyone graph faithful to P. The next chapter presents a number of
algorithms that generate the faithful indistinguishability classes from properties of distributions.
Given axioms connecting causal graphs with probability distributions it makes sense to ask for
which pairs G, G' of graphs there exists some probability distribution satisfying the axioms for
both G and G'. Let us say that two graphs are weakly faithfully indistinguishable
(w.f.i.) if and only if there exists a probability distribution faithful to both of them. We say that
two graphs are weakly statistically indistinguishable (w.s.i.) if and only if there exists a
probability distribution meeting the Minirnality and Markov Conditions for both of them. Weak
faithful indistinguishability proves to be equivalent to faithful indistinguishability:
Theorem 4.3: Two directed acyclic graphs are faithfully indistinguishable if and only
if some distribution faithful to one is faithful to the other and conversely; i.e. they are
f.i. if and only if they are w.f.i.
This theorem tells us that faithfulness divides the set of probability distributions over a vertex
set into equivalence classes that exactly correspond to the equivalence classes of graphs induced
by faithful indistinguishability. It follows that if a distribution is faithful to some graph G then it
is faithful to all and only the graphs faithfully indistinguishable from G.
There is no reason to expect so nice a match in general. Suppose we assume only the
Minimality and Markov Conditions. Under what conditions will there exist a distribution P
satisfying those axioms for two distinct graphs, G, and G'? The answer is not: exactly when G
and G' are strongly statistically indistinguishable. The two graphs shown in figure 2 are not
s.s.i., but there exist distributions that satisfy the Minimality and Markov Conditions for both:
A /" C
Figure 2
The distributions in Simpson's "paradox" provide an example, as we have already seen in

Chapter 3. We conjecture that if a distribution satisfies the Minimality and Markov Conditions
for two graphs G and G', then G and G' have the same edges and the same colliders, save that
triangles such as Gl in one graph may be replaced by collisions such as G2 in the other,
provided appropriate conditions are met by other edges. We don't know how to characterize the
"appropriate" conditions. There is, however, a related property of interest we can characterize.
No distribution that is faithful to graph Gl in figure 2 can be faithful to graph G2, but a
distribution that satisfies the Minimality and Markov Conditions for Gl can be faithful to graph
G2. Just when can this sort of thing happen? When, in other words, can the generalization of
Simpson's "paradox" arise? If probability distribution P satisfies the Minimality and Markov
Conditions for G, and P is faithful to graph H, what is the relation between G and H?
Theorem 4.4: If probability distribution P satisfies the Markov Condition for directed
acyclic graphs G and H, and P is faithful to H, then for all vertices X, Y, if X, Y are
adjacent in H they are adjacent in G.
Theorem 4.5: If probability distribution P satisfies the Markov and Minimality

Conditions for directed acyclic graphs G and H, and P is faithful to graph H, then (i) for
all X, Y, Z such that X -> Y <- Z is in H and X is not adjacent to Z in H, either X -> Y
<- ZinG or X, Z are adjacent in G and (ii) for every triple X, Y, Z of vertices such that
X -> Y <- Z is in G and X is not adjacent to ZinG, if X is adjacent to Yin H and Y is

adjacent to Z in H then X -> Y <- Z.
Corollary 4.1: If probability distribution P satisfies the Markov Condition for

directed acyclic graph G, P is faithful to directed acyclic graph H, and G and H agree on
an ordering of the variables (as, for example, by time) such that X -> Yonly if X < Yin
the order, then H is a subgraph of G.
4.4 Rigid Indistinguishability
In addition to the notions of strong, faithful and weak statistical indistinguishability, there is still
another. Suppose two directed acyclic graphs, G and G', are statistically indistinguishable in
some sense over a common set 0 of vertices. Then without experiment, no measurement of the
variables in 0 will reliably determine which of the graphs correctly describes the causal
structure that generated the data. It might be, however, that G and G' can be distinguished if
other variables besides those in G or G' are measured and stand in appropriate causal relations
to the variables in O. For example, the following simple graphs are both s.s.i. and f.i. (where
A and B are assumed to be measured and in 0.)
Figure 3
But if we also measure a variable C that is a cause of A or has a common cause with A and no
connection with B save possibly through A, then the two structures can be distinguished.
Figure 4
The graphs in figure 4 are not f.i. or s.s.i. It is equally easy to give examples of w.s.i.
structures that can be embedded in graphs that are not w.s.i. Which causally sufficient
structures can be distinguished by measuring extra variables? To answer the question we
require some further definitions.
Let Gl. G2 be two directed acyclic graphs with common vertex set O. Let H 1. H2 be directed
graphs having a common set U of vertices that includes 0 and such that
(i) the subgraph of HI over 0 is GI and the subgraph of H2 over 0 is~;
(ii) every directed edge in HI but not in GI is in H2 and every directed edge in H2 but not
in G2 is in HI.
We will say then that directed acyclic graphs Gl and G2 with common vertex set 0 have a
parallel embedding in HI and H2 over 0 and U. In figures 3 and 4, Gl and G2 have a
parallel embedding in HI and H2 over 0 = {A,B} and U = {A,B,C,D}. The question of
whether two s.s.i. structures can be distinguished by measuring further variables then becomes
the following: do the structures have parallel embeddings that are not s.s.i.? If no such
embedding exists we will say the structures Gland G2 are rigidly statistically
indistinguishable (r.s.i.).
Theorem 4.6: No two distinct s.s.i. directed acyclic graphs with the same vertex set
are rigidly statistically indistinguishable.
In other words, provided additional variables with the right causal structure exist and can be
measured, the causal structure among a causally sufficient collection of measured variables can
in principle be identified. The proof of Theorem 4.6 also demonstrates that a parallel result for
faithfully indistinguishable structures. We conjecture that an analog of Theorem 4.6 also holds
for weak: statistical indistinguishability assuming positivity.
4.5 The Linear Case
Parameter values can force conditional independencies or zero partial correlations that are not
linearly implied by a graph. The graphs in figure 2 (reproduced in figure 5 with error variables
explicitly included) illustrate the possibility: treat the vertices of the graphs as each attached to an
"error" variable, and let the graphs plus error variables determine a set of linear equations. (We
assume that any pair of exogenous variables, including the error terms, have zero covariance.)
The result is, up to specification of the joint distribution of the exogenous variables, a structural
equation model. A linear coefficient is attached to each directed edge. The correlation matrix,
and hence all partial correlations, is detennined by the linear coefficients and the variances of the
exogenous variables.
r"<
B
A
c
.c A C
t t t t
(i) (ii)
Figure 5
If in the structure on the left ab = -c, then A, C will be uncorrelated as in the model on the right
This sort of phenomenon--vanishing partial correlations produced by values of linear
coefficients rather than by graphical structure--is bound to mislead any attempt to infer causal
structure from correlations. When can it happen? We have already answered that question in the
previous chapter, when we considered the conditions under which linear faithfulness might fail.
In the linear case, the parameter values--values of the linear coefficients and exogenous
variances of a structure with a directed acyclic graph G--forrn a real space, and the set of points
in this space that create vanishing partial correlations not linearly implied by G have Lebesgue
measure zero.
Theorem 3.2: Let M be a linear model with directed acyclic graph G and n linear
coefficients al, ... , an and k positive variances of exogenous variables VI, •.• , vk. Let
M«UI, ... ,Un , Un+I, ... ,Un+k» be the distributions consistent with specifying values
<UI, ... ,U n• Un+I, ... ,Un+k> for at, ... , an and Vt, ... Vk. Let IT be the set of probability
measures P on the space 9\n+k of values of the parameters of M such that for every
subset V of 9\n+k having Lebesgue measure zero, P(V) = O. Let Q be the set of vectors
of coefficient and variance values such that for all q in Q every probability distribution
in with M(q) has a vanishing partial correlation that is not linearly implied by G. Then
for all P in IT, P(Q) = O.
Measure theoretic arguments of this sort are interesting but may not be entirely convincing. One
could, after all, argue that in the general linear model absence of causal connection is marked by
linear coefficients with the value zero, and thus form a set of measure zero, so by parity of
reasoning everything is causally connected to everything else. In a recent book Nancy
Cartwright (1989) objects that since in linear structures independence relations may be produced
by special values of the linear coefficients and variances as well as by the causal structure, it is
illegitimate to infer causal structure from such relations. In effect, she rejects any inference
procedure that is unable to distinguish the true causal structure from w.s.i. alternatives. Such a
position may be extreme, but it does serve to focus attention on two interesting questions: when
is it impossible for two structures to be w.s.i. but not f.i. or s.s.i., and are there special marks
or indicators that a distribution satisfies the Markov and Minimality conditions for two w.s.i.
but not s.s.i. or f.i. causal structures? The answers to these questions are essentially just
applications to the linear case of the theorems of the preceding sections.
We will assume, with Cartwright, that a time ordering of the variables is known. Pearl and
Verma (Pearl 1988) have proved that for a positive distribution P and a given ordering of
variables, there is only one directed acyclic graph for which P satisfies the Minimality and
Markov Conditions. It follows that for a positive distribution with a given correlation matrix
and a given ordering of a causally sufficient set of variables there is a unique directed acyclic
graph that linearly represents the distribution and is consistent with the ordering.
In some cases at least, the positivity of a distribution can be tested for. (For example, in a bi-
variate normal distribution the density function is everywhere non-zero if the correlation is not
equal to one.) It follows for those cases that for a given ordering of variables either there is a
unique directed acyclic graph for which P satisfies the Markov and Minimality Conditions, or it
is detectable that more than one such directed aCyclic graph exists. However, even if for a given
ordering of variables there is a unique directed acyclic graph for which P satisfies the Markov
and Minimality Conditions, algorithms for finding that graph are not feasible for large numbers
of variables, because of the number and order of the conditional independence relations that
they require be tested.
Suppose that we wrongly assume that a distribution is faithful to the causal graph that generated
it. Then Corollary 4.1 applies, which means, informally, that if faithfulness is assumed but not
true, then conditional independence relations or vanishing partial correlations due to special
parameter values can only produce erroneous causal inferences in which a true causal
connection is omitted; no other sorts of error may arise. We will consider when this
circumstance is revealed in the correlations.
Recall that a trek is an unordered pair of directed acyclic paths having a single comrnon vertex
that is the source of both paths (one of the paths in a pair may be the empty path defined in
Chapter 2). For standardized models, in which the mean of each variable is zero and non-error
variables have unit variance, the correlation of two variables is given by the sum over all treks
connecting X, Y of the product for each trek of the linear coefficients associated with the edges
in that trek (we call this quantity the trek sum). For example, in directed acyclic graph (i) in
figure 5, the trek sum between A and Cis ab + c. We will use standardized systems throughout
our examples in this section. The system of correlations determines all partial correlations of
every order through the following formula.
P _ PXY.Z -PXR.z X PrR.Z

XY.Zu{R} - ~ 2 ~ 2
1- PXR.Z x l - PYR.Z
Since the recursion relations give the same partial correlation between two variables on a set U
no matter in what sequence the partials on the members of U are taken, a vanishing partial
correlation corresponds to a system of equations in the coefficients of a standardized system.
Suppose now that special values of the linear parameters in a normal, standardized system G
produce vanishing partial correlations that are exactly those linearly implied only by some false
causal structure, say H. Then the parameter values must generate extra vanishing partial
correlations not linearly implied by G. Any partial correlation is a function just of the trek sums
counecting pairs of variables, and the trek sums in this case involve just the linear parameters in
G. Hence each additional vanishing partial correlation not linearly implied by G determines a
system of (non-linear) equations in the parameters of G that must be satisfied in order to
produce the coincidental vanishing partial correlation. (For example, in directed acyclic graph (i)
of figure 5, the correlation between A and C is 0 only if the single equation ab =-c is satisfied).
Now for some G and some H (a sub-graph of G), these systems of equations may have no
simultaneous solution. In that case there are no values for the parameters of G that will produce
partial correlations that are exactly those linearly implied by H. For other choices of G and a
subgraph H, it may be that the system of equations has a solution, but only solutions that allow
only a finite number of alternative values for one or more parameters and that require some error
variance to vanish. Such a solution must "give itself away" by special correlation constraints
that are not themselves vanishing partial correlation relations. Consider the following choices of
G and H, where in each pair G is on the left hand side and H is on the right hand side:
x X
y
1)y
Z
JXy=I,-1
(i)
• z/ '"
W X
y w
X
r~/ "'II'
z/
* (ii)
Z
,/
1~ y
X
X
l~
/y PxY
•
= 1,-1
Z Z
(iii)
Figure 6
In (i) and (iii), coefficients and variances can be chosen for the graph on the left hand side so
that it appears as though an edge does not occur, but only by making the coefficient labeled b
equal to either 1 or -1. Since the variables are standardized, this requires that the error term for
Y have zero variance and zero mean--i.e., it vanishes. Thus in order for the true graph to be the
one on the left hand side and the parameter values to produce vanishing partial correlations that
are exactly those linearly implied by the graph on the right hand side, variable Y must be a linear
function of variable X and only variable X. The same result obtains if the edges that are not
eliminated in the first and last examples are replaced by directed paths of any length. Clearly in
these cases special parameter values that create vanishing partial correlations not linearly implied
by the true graph will be revealed by the correlations. In (ii) the edge between variables X and Z
cannot be made to appear to be eliminated by any choice of parameter values for the true graph.
We conjecture that even without a prior time order, unless three edges fonn a triangle in G, if
parameter values of G determine exactly the collection of vanishing partial correlations linearly
implied by a graph H--whether or not H is a subgraph of G--then there are extra constraints on
the correlations not entailed by the vanishing partial correlations.
4.6 Redefining Variables
The indistinguishability results so far considered relate alternative graphs over the same set of
vertices. The vertices are interpreted as random variables whose values are subject to some
system of measurement. New random variables can always be defined from a given set, for
example by taking linear or Boolean combinations. For any specified apparatus of definitions,
and any axioms connecting graphs with distributions, questions about indistinguishability
classes arise parallel to those we have considered for fixed sets of variables. A distribution P
over variable set V may correspond to a graph G, and a distribution P' over variable set V' may
correspond to a different graph G' (with P' and V' obtained from P and V by defming new
variables, ignoring old ones, and marginalizing). The differences between G and G' may in
some cases be unimportant, and one may simply want to say that each graph correctly describes
causal relations among its respective set of variables. That is not so, however, when the original
variables are ordered by time, and redefinition of variables results in a distribution whose
corresponding graphs have later events causing earlier events. Consider the following pair of
graphs.
B B
A
/\ C
/\
(A - C) (A + C)
(i) (ii)
Figure 7
In directed acyclic graph (i), A and C are effects of B; suppose that B occurs prior to A and C.
By the procedure of definition and marginalization, a distribution faithful to graph (i) can be
transformed into a distribution faithful to graph (ii). First, standardize A and C to form variables
A' and C' with unit variance. Then consider the variables (A' - C') and (A'+ C'). Their
covariance is equal to the expected value of A'2 - C'2 which is zero. Simple algebra shows that
the partial correlation of (A' - C') and (A'+ C') given B does not vanish. The marginal of the
original distribution is therefore linearly faithful to (ii), and faithful to (ii) if the original
distribution is normal.
Note that the transformation just illustrated is unstable; if the variances of A' and C' are unequal
in the slightest, or if the transformation gives (xA' + zC') and (yA' + wC') for any values of x,
y, z, and w such that xy + wz + PA'C'(ZY + xw) * 0 then the marginal on the transformed
distribution will be faithful, not to (ii), but to all acyclic orientations of the complete graph on
the three variables, a hypothesis that is not inconsistent with the time order.
Viewed from another perspective, a transformation of variables that produces a "coincidental"

vanishing partial correlation is just another violation of the Faithfulness Condition. Consider the
linear model in figure 8.
r/\
B
A' C'
xC><:}
D E
Figure 8
Let A' = rB +EA', C' = sB +EC', D =xA'+ zC' + ED, and E =yA'+ wC' + EE. If the variables
are standardized, PDE is equal to xy + zw + rysz + rxsw = xy + zw + rs(yz + xw), which,
since rs =PA'C', is the formula of the previous paragraph. If PDE = 0, the Faithfulness
Condition is violated. Hence the conditions under which we obtain a linear transformation of A
and C that produces a "coincidental" zero correlation are identical to the conditions under which
the treks between A and C exactly cancel each other in a violation of the Faithfulness Condition.
We get the example offigure 7 when D =A' + C' (i.e. x = z = 1), and E = A' - C' (i.e. y =-w
= 1) where the variances and means of the error terms have been set to zero. Since the set of
parameter values that violate Faithfulness in this example has Lebesgue measure zero, so does
the set of linear transformations of A and C that produce a "coincidental" zero correlation.
The underdetermination of linear statistical models by values of measured variables has been
extensively discussed as the "identification problem," especially in econometrics (Fisher, 1966)
where the discussion has focused on the estimation of free parameters. The device of
"instrumental variables," widely used for linear models, is in the spirit of Theorem 4.6 on rigid
distinguishability, although instrumental variables are used to identify parameters in cyclic
graphs or in systems with latent variables. The possibility of "rewriting" a pure linear
regression model so that the outcome variable is treated as a cause seems to have been familiar
for a long while, and we do not know the original source of the observation, which was
brought to our attention by Judea Pearl.
Accounts of statistical indistinguishability in something like one or another of the senses

investigated in this chapter have been proposed by Basmann (1965), Stetzl (1986) and Lee
(1987). Basmann argued, in our terms, that for every simultaneous equation model with a
cyclic graph (i.e., "non-recursive") there exists a statistically indistinguishable model with an
acyclic graph. The result is a weak indistinguishability theorem (see Chapter 12). Stetzl and Lee
focus exclusively on linear structural equation models with free parameters for linear
coefficients and variances, and they defme equivalence in terms of maximum likelihood
estimates of the parameters and hence of the covariance matrix. No general graph theoretic
characterizations are provided, although interesting attempts were made in Lee's thesis.
The notion of a pattern and Theorem 4.2 are due to Verma and Pearl (1990b). We state some
results about indistinguishability relations for causally insufficient graphs in Chapter 6. A well-
known result due to Suppes and Zanotti (1981) asserts that every joint distribution P on a set X
of discrete variables is the marginal of some joint distribution P* on X u {T} satisfying the
Markov Condition for a graph G in which T is the common cause of all variables in X and there
are no other directed edges. The result can be viewed as a weak indistinguishability theorem
when causally insufficient structures are admitted. Except in special cases, p* cannot be faithful
to G.
Chapter 5
Discovery Algorithms
for
Causally Sufficient Structures
5.1 Discovery Problems
A discovery problem is composed of a set of alternative structures, one of which is the source
of data, but any of which, for all the investigator knows before the inquiry, could be the
structure from which the data are obtained. There is something to be found out about the actual
structure, whichever it is. It may be that we want to settle a particular hypothesis that is true in
some of the possible structures and false in others, or it may be that we want to know the
complete theory of a certain sort of phenomenon. In this book, and in much of the social
sciences and epidemiology, the alternative structures in a discovery problem are typically
directed acyclic graphs paired with joint probability distributions on their vertices. We usually
want to know something about the structure of the graph that represents causal influences, and
we may also want to know about the distribution of values of variables in the graph for a given
population.
A discovery problem also includes a characterization of a kind of evidence; for example, data
may be available for some of the variables but not others, and the data may include the actual
probability or conditional independence relations or, more realistically, simply the values of the
variables for random samples. Our theoretical discussions will usually consider discovery
problems in which the data include the true conditional independence relations among the
measured variables, but our examples and applications will always involve inferences from
statistical samples.
A method solves a discovery problem in the limit if as the sample size increases without bound
the method converges to the true answer to the question or to the true theory, whatever
(consistent with prior knowledge) the truth might be. A procedure for inferring causes does not
solve the problem posed if for some of the alternative possibilities it gives no answer or the
wrong answer, although it may solve another, easier problem that arises when some of the
alternative structures are excluded. Which causal discovery problems are solvable in the limit,
and by what methods, are determinate, mathematical questions. The metaphysical wrangling
lies entirely in motivating the problems, not in solving them. The remainder of this book is an
introduction to the study of these formal questions and to the practical applications of particular
answers.
5.2 Search Strategies in Statistics
The statistical literature is replete with procedures that use data to guide a search for some
restricted parametrization of alternative distributions. When the representation of the statistical
hypothesis is used to guide policy or practice, to predict what will happen if some of the
variables are manipulated or to retrodict what would have happened if some of the variables had
in the past been manipulated, then the statistical hypotheses are usually also causal hypotheses.
In that case the first question is whether the search procedures are any good at finding causal
structure.
Many of the search strategies proposed in the statistical literature are best-only beam searches,
beginning either with an arbitrary model, or with a complete (or almost complete) structure that
entails no constraints, or with a completely (or almost completely) constrained structure in
which all variables are independent. Statisticians sometimes refer to the latter procedure as
"forwards" search, and the former procedures as "backwards" search. Depending on which
order is followed, the procedures iteratively apply a fit measure of some kind to determine
which fixed parameter in the parametrization will most improve fit when freed--or which free
parameter should be fixed. They then reestimate the modified structure to determine if a
stopping criterion is satisfied. A "forward" procedure of this kind was proposed by Arthur
Dempster (1972) for covariance structures, and a "backward" procedure was proposed by his
student, Nanny Wermuth (1976), for both log-linear and linear systems whose distributions are
"multiplicative"--in our terms, satisfy the Markov condition for some directed acyclic graph.
Forward search algorithms using goodness of fit statistics were proposed for mutinormallinear
systems by Byron (1972) and by Sorbom (1975) and versions of them have been automated in
the LISREL (Joreskog and Sorbom 1984) and EQS (Bentler 1985) estimation packages. The
Discovery Algorithms for Causally Sufficient Structures 105
latter program also contains a backwards search procedure. Versions of the general strategy for
log-linear parametrizations are described by Bishop, Fienberg and Holland (1975), by Fienberg
(1977) by Aitkin (1979), by Christensen (1990) and many others. The same representations
and search strategies have been used in the systems science literature by Klir and Parviz (1976)
and others under the title of "reconstructability" analysis. Stepwise regression procedures in
logistic regression can be viewed as versions of the same strategies. The same strategies have
been applied to undirected and directed graph representations. They are illustrated for a variety
of examples by Whittaker (1990).
In each of these cases the general statistical search strategy is unsatisfactory if the goal is not
just to estimate the distribution but also to identify the causal structure or to predict the results of
manipulations of some of the variables. When used to these ends, these searches are inefficient
and unreliable for at least three reasons: (i) they often search a hypothesis space that excludes
many causal hypotheses and includes many hypotheses of no causal significance; (ii) the
specifications of distributions typically force the use of numerical procedures that for statistical
or computational reasons unnecessarily limit search; (iii) restrictions requiring the search to
output a single hypothesis entail that the search fails to output alternative hypotheses that may be
indistinguishable given the evidence. We will consider each of these points in more detail.
5.2.1 The Wrong Hypothesis Space
In searching for the correct causal hypothesis the space of alternatives should, insofar as
possible, include all causal hypotheses that have not been ruled out by background knowledge
and no hypotheses that do not have a causal interpretation. The log-linear formalism, introduced
by Birch in 1963, provides an important example of a search space poorly adapted to the goal of
finding correct causal hypotheses. For discrete data a more appropriate search space turns out to
be a sub-class of conjunctions of log-linear hypotheses.
The log-linear formalism provides a general framework for the analysis of contingency tables of
any dimension. In the discrete case we are concerned with variables that take a finite number of
values, whether ordered or not. For a system with four variables, for example, we will let i
range over the values of the first variable, j the second, k the third and I the fourth. In a
particular sample or population, Xijkl will then denote the number of units that have value i for
the first variable, value j for the second variable, k for the third and I for the fourth. We will
refer to a particular vector of values for the four (or other number of) variables as a "cell." In the
formalism the joint distribution over the cells is given by an equation for the logarithm of the
expected value of each cell, expressed as the sum of a number of parameters. For example, in
Birch's notation in which mijk denotes the expected number in cell i, j, k,
In(mijk) = U + Uli + U2j + U3k + U12ij + U13ik + U23jk + U123ijk
The various u's are arbitrary parameters with an associated set of indices; only seven of the U
terms can be independent for a system of three binary variables. The power of Birch's
parametrization lies in at least two features. First, associations in multi-dimensional contingency
tables that had long been studied in statistics can be represented as hypotheses that certain of the
parameters are zero. For example Bartlett's representation of the hypothesis of no "three factor
interaction" among three binary variables is given by the following relation among the cell
probabilities:
PllIP12W21W221 = P22W21IP12IP112
Birch shows that a generalization of this condition to variables of any finite number of
categories obtains if and only if various of the U terms are zero. Second, for each hypothesis
obtained by setting some of the U terms to zero, there exist iterative methods for obtaining
maximum likelihood estimates for a variety of sampling procedures.
Birch's results were extended by several researchers. A hypothesis in the log-linear

parametrization has come to be treated as a specification that particular U terms vanish. There are
direct maximum likelihood estimates of the expected cell counts for certain forms of such
specifications, and for other specifications iterative algorithms have been developed that
converge to the maximum likelihood estimates. Various formal motivations have been
developed for focusing on particular classes of log-linear parametrizations. Using his
information-based distance measure, for example, Kullback (1959) derived a class of log-linear
relations that could be obtained in the same way from a slightly different perspective, the
maximum entropy principle. Fienberg (1977) and others have urged restricting attention to
"hierarchical models" --log linear parametrizations in which if a U term with a set of indices is
put to zero so are all other U terms whose indices contain the first set. The motivation for the
restriction is that these parametrizations bear a formal analogy to analysis of variance, so that the
Ul term, for example, may be thought of as the variation from the grand mean due to the action
of the first variable.
To see the difficulties in representing causal structure in the log-linear formalism, consider the
most fundamental causal relation of the preceding chapters, namely any collider A -> B <- C in
which A and C are not adjacent. Such a structure corresponds (assuming faithfulness) to two
facts about conditional independence: first, A and C are independent conditional on some set of
variables that does not contain B; second, A and C are dependent conditional on every set that
does contain B but not A or C. In the very simplest case of this kind, in which A, B and C are
the only variables, A and B are independent, but dependent conditional on C. The hypothesis
that these relations obtain cannot be expressed in the log-linear formalism by vanishing u terms.
Birch himself observed that in a three variable system the hypothesis that in the marginal
distribution two of the variables are independent cannot be expressed by the vanishing of any
subset of parameters in the general log-linear expansion for the three variables. There are of
course log-linear hypotheses that are consistent with marginal independence hypotheses, but do
not entail them.
Another inappropriate search space is provided by the LISREL program. The LISREL
formalism--at least as intended by its authors, Joreskog and Sorbom--allows search for
structures corresponding to causal relations among measured variables when there are no
unmeasured common causes, but when the search includes structures with unmeasured
common causes, causal relations among measured variables are forbidden. Users have found
ways around these restrictions (Glymour, et aI., 1987; Bollen, 1989), rather to the
dissatisfaction of the authors of the program (Joreskog and Sorbom, 1990). LISREL owes
these peculiarities to its ancestry in factor analysis, which provides still another example of an
artificially contracted search space. Thurstone (1935) carefully and repeatedly emphasized that
his "factors" were not to be taken as real causes but only as a mathematical simplifications of the
measured correlations. Of course factors were immediately treated as hypothetical causes. But
so applied, Thurstone's methods exclude a priori any causal relations among measured
variables themselves, they exclude the possibility that measured variables are causes of
unmeasured variables, and they cannot determine causal structure--only correlations--among the
latent variables.
5.2.2 Computational and Statistical Limitations
Some searches examine only a small portion of the possible space of hypotheses because they
require computationally intensive iterative algorithms in order to test each hypothesis. For
example, the automatic model respecification procedure in LISREL re-estimates the entire model
every time it examines a new hypothesis. One consequence is that the slowness of the search
prohibits the procedure from examining large portions of the hypothesis space where the truth
may be hiding.
Another common problem is that many searches require the determination of conditional
independence relations that cannot be reliably tested. Many log-linear search procedures
implicitly require the estimation of probabilities conditional on a set of variables whose size
equals the total number of variables minus two, no matter what the true structure turns out to
be. Estimates of higher order conditional probabilities and tests of higher order conditional
independencies tend to be unreliable (especially with variables taking several discrete values)
because at reasonable sample sizes most cells corresponding to an array of values of the
variables will be empty or nearly empty. This disadvantage is not inherent in the log-linear
formalism. A recent algorithm proposed by Fung and Crawford (1990) for searching the set of
graphical models (the subset of the hierarchical log-linear models that can be represented by
undirected independence graphs) reduces the need for testing high order conditional
independencies. A version of the same problem arises for linear regression with a large number
of regressors and small sample size, since in tests of the hypothesis that a regression coefficient
vanishes, the sample size is effectively reduced by the number of other regressors, or the
degrees of freedom are altered, so that the test may have little power against reasonable
alternatives.
A related but equally fundamental difficulty is that searches for models of discrete data that use
some measure of fit requiring model estimation at each (or any) stage are subject to an
exponential increase in the number of cells that must be estimated as the number of variables
increases. IT, to take the simplest case, the variables are binary, then the number of cells for
which an expected value must be computed is 2n. When n = 50, say, the number of cells is
astronomical.
One might think: that these difficulties will beset any possible reliable search procedure. As we
will see in this chapter and the next, that is not the case.
5.2.3 Generating a Single Hypothesis
IT a kind of evidence is incapable of reliably distinguishing when one rather than another of
several alternative hypotheses is correct, then an adequate search procedure should reflect this
fact by outputting all of them. Producing only a single hypothesis in such circumstances
misleads the user, and denies her information that may be vital in making decisions.
An example of this sort of flaw is illustrated by the LISREL and EQS programs. Beginning
with a structure constructed from background knowledge, each of these programs searches for
causal models among linear structures using a best-only beam search. At each stage they free
the fixed parameter that is judged will most increase the fit of the model to the data. Since
freeing a number of different fixed parameters may result in the very same improvement in fit,
the programs employ an arbitrary tie-breaking procedure. The output of the search is a single
linear model and any alternative statistically indistinguishable models are ignored.
In a later chapter we will describe a large simulation study of the reliabilities of the statistical
search procedures implemented in the LISREL and EQS programs for linear models. Because
of the computational problems and arbitrary choices from among indistinguishable models at
various stages of search, we find that the procedures are of little value in discovering
dependencies in the structures from which the data are generated, even when the programs are
given most of the structure correctly to start with, including even correct linear coefficients and
variances. The study involves systems with unmeasured variables, but we expect that similar
results would be obtained in studies with causally sufficient systems.
5.2.4 Other Approaches
There are several exceptions to the generalization that statistical search strategies have been
confined to generate-and-test-best-only procedures. Edwards and Havranek (1987) describe a
form of procedure that tests models in sequence, under the assumption that if a model passes
the test so will any more general model and if a model fails the test so will any more restricted
model. Their proposal is to keep track of a bounding set of rejected hypotheses and a bounding
set of accepted hypotheses until all possible hypotheses (in some parametrization) are classified.
Apparently unknown to Edwards and Havranek, the same idea was earlier developed at length
in the artificial intelligence literature under the name of "version spaces" (Mitchell, 1977). For
the applications they have in mind, no analysis of complexity or reliability is available.
5.2.5 Bayesian Methods
The best known discussion of search problems in statistics from a Bayesian perspective is
Leamer's (1978). Leamer's book contains a number of interesting points, including a
consideration of what a Bayesian should do upon meeting a novel hypothesis, but it does not
contain a method for reliable search. Considering the use of regression methods in causal
inference, for example, Leamer subsequntly recommended analyzing separately the sets of
relevant regressors endorsed by any opinion, and giving separate Bayesian updates of
distributions of parameters for ~h of these sets of regressors. The problem of deciding which
variables actually influence an outcome of interest is effectively ignored.
A much more promising Bayesian approach to search has been developed by Cooper and
Herskovits (1991, 1992). At present, their procedure is restricted to discrete variables and
requires a total ordering such that no later variable can cause an earlier variable. Each directed
graph compatible with the order is assigned a prior probability. The joint distribution of the
variables assigns each vertex in the graph a distribution conditional on its parents, and these
conditional probabilities parametrize the distributions for each graph. Using Dirchelet priors, a
density function is imposed on the parameters for each graph. The data are used to update the
density function by Bayes' rule. The probability of a graph is then just the integral of the
density function over the distributions compatible with the graph. The probability of an edge is
the sum of the probabilities of all graphs that contain it. Cooper and Herskovits use a greedy
algorithm to construct the output graph in stages. For each vertex X in the graph, the algorithm
considers the effect of adding to the parent set of X each individual predecessor of X that is not
already a parent of X; it chooses the vertex whose addition to the parent set of X most increases
the posterior probability of the local structure consisting of X and its parents. Parents are added
to X in this fashion until there is no single vertex that can be added to the parent set of X that
will increase the posterior probability of the local structure. The program runs very well even on
quite large sets of variables provided the true graph is sparse, and on discrete data with a prior
ordering appears to determine adjacencies with remarkable accuracy. Its accuracy on dense
graphs is not known at this time.
The Bayesian approach developed by Cooper and Herskovits has the advantages that
appropriate prior degrees of belief can be used in search, that models are output with ratios of
posterior distributions consistent with the specified prior distribution and the data, and that
under appropriate assumptions! the method converges to the correct graph. Because the method
can calculate the ratio of the posterior probabilities of any pair of graphs, it is possible to make
inferences over multiple graphs weighted by the probability of the graph (although generally
some heuristic to consider only the most probable graphs must be used because of the sheer
1 In particular, when the method is idealized 10 give up the greedy algorithm. Because of the greedy algorithm,
we would expect the specific search procedure 10 be asymplOtically unreliable when there are two or more treks
between a pair of non-adjacent variables, say X and Y, that result in a close statistical association between those
variables. This is the circumstance in the case of the one edge the procedure erroneously introduces in the
ALARM network. In practice, such sttuctures may be sufficiently uncommon for the error 10 be tolerable., and
Cooper and his colleagues are investigating techniques 10 ameliorate the problem.
number of possibilities.) The method works with Dirchelet priors because the relevant integrals
are available analytically and posterior densities can therefore be rapidly evaluated without any
numerical analysis. In view of the combinatorics of graphs, any other application of the search
architecture must have the same feature. One problem is to extend the method to continuous
variables, which depends on finding a family of conjugate priors that can be rapidly updated for
parameters that describe graphs. Another, more fundamental, problem concerns whether the
requirement of a prior ordering of the variables can be relaxed while preserving computational
feasibility. Using a fixed ordering of the variables reduces the combinatorics enormously, but in
many applied cases any such ordering may be uncertain. Since the procedure is reasonably fast,
requiring about 15 minutes (on a Macintosh II) to analyze data from the ALARM network
described in Chapter I, Cooper and his colleagues are investigating procedures that use a
number of orderings and compare the posterior probabilities of the graphs obtained.
5.3 The Wermuth·Lauritzen Algorithm
In 1983 Wermuth and Lauritzen defined what they called a recursive diagram. A recursive
diagram is a directed acyclic graph G together with a total ordering of the vertices of the graph
such that VI -> V2 occurs only if VI < V2 in the ordering. In addition there is a probability
distribution P on the vertices such that Vi is a parent of Vk if and only if Vi < Vk and Vi and Vk
are dependent conditional on the set of all other variables previous to Vk in the ordering.
Following Whittaker (1990), we call such systems directed independence graphs.
We can view this definition as an algorithm for constructing causal graphs from conditional
independence relations and a time ordering of the variables. It has in fact been used in this way
by some authors (Whittaker, 1990). Given an ordering of the variables and a list of the
conditional independence relations, proceed through the variables in their time order, and for
each variable Vk to each variable Vi such that Vi < Vk apply the dependence test in the
definition, and add Vi -> Vk if the test is passed. The procedure will correctly recover the
directed graph from the order and the independence relations of a faithful distribution in which,
for discrete variables, every combination of variable values has positive probability. In a sense,
the discovery problem for causally sufficient faithful systems is solved. In practice, however,
the Wermuth-Lauritzen algorithm is not feasible save for very small variable sets. The
remaining issues are therefore these:
(i) how to remove the requirement that an ordering of the variables be known beforehand;
(ii) how to improve on the computational efficiency and statistical requirements of the
Wermuth-Lauritzen procedure;
(iii) how to remove the tacit restriction to causally sufficient systems of variables.
In this chapter we will address the first two of these problems. The problem of causal inference
when unmeasured common causes, or "latent variables," may be acting will be taken up in
Chapter 6.
5.4 New Algorithms
We will describe several algorithms for discovering causal structure (assuming causal
sufficiency); they eliminate the need for a prior ordering of the variables and all but two of them
improve computational efficiency and reduce the difficulty of statistical decisions in comparison
with the Wermuth-Lauritzen algorithm. Some of the improvements are dramatic, others less so.
Each of the search procedures described can also be used on discrete data to search for graphical
log-linear models. (For each triple of variables, if X -> Y <- Z occurs in the directed graph, and
X is not adjacent to Z, add an undirected edge between X and Z; then remove all arrowheads
from the graph. The result is an undirected independence graph.)
Under the following assumptions all of the algorithms presented in this section provably
recover features of graphs faithful to the population distribution:
(i) The set of observed variables is causally sufficient
(ii) Every unit in the population has the same causal relations among the variables.
(iii) The distribution of the observed variables is faithful to an acyclic directed graph of
the causal structure (in the discrete case) or linearly faithful to such a graph (in the linear
case).
(iv) The statistical decisions required by the algorithms are correct for the population.
The fourth requirement is unnecessarily strong, since the algorithms will in many cases succeed
even if some statistical decisions are in error. Nonetheless, this is a strong set of assumptions
that is often not met in practice, but it is no stronger than the assumptions that would be
required to warrant most of the particular statistical models with a causal interpretation found in
the medical, behavioral, and social scientific literature. In subsequent chapters we will examine
the consequences of weakening some of these assumptions.
In practice, the algorithms take as input either a covariance matrix or cell counts. Where d-
separation facts are needed by an algorithm, in the discrete case the procedure performs tests of
conditional independence and in the linear continuous case tests for vanishing partial
correlations. (Recall that if P is a discrete distribution faithful to a graph G, then A and B are d-
separated given a set of variables C if and only A and B are conditionally independent given C,
and if P is a distribution linearly faithful to a graph G, then A and B are d-separated given C if
and only if PAB.C = 0.) The algorithms construct the set of directed acyclic graphs that satisfy
the given set of d-separability relations, if any such graph exists. Since the results of either kind
of test are used only to determine the d-separation relations among the variables, we will speak
as if the input to the algorithms is simply the d-separation relations themselves. 2
Let us say that a graph G faithfully represents a list of d-separations L if and only if
all and only the d-separations in L are true of G. A list L of d-separations is faithful if
and only some acyclic directed graph faithfully represents L. In practice, even if a distribution is
faithful to the causal structure that generates it, sampling error or minor violations of the
assumptions of the statistical tests employed can lead to errors in judgment about the properties
of the population. The robustness of the procedures to erroneous specification of the
distribution family or to sampling variation can be investigated by Monte Carlo simulation
methods.
Each of the following algorithms can have as output either a class of directed acyclic graphs, or
else a single mixed object with both directed and undirected edges-- the pattern that represents a
class of graphs. Recall that pattern II represents a set of directed acyclic graphs. A graph G is in
the set of graphs represented by II if and only if:
(i) G has the same adjacency relations as II;
(ii) if the edge between A and B is oriented A -> B in II, then it is oriented A -> BinG;
(iii) if Y is an unshielded collider on the path <X,y,z> in G then Y is an unshielded collider
on <X,Y,Z> in II.
If any of the algorithms use as input a covariance matrix from a distribution linearly faithful to
G, or cell counts from a distribution faithful to G, we will say the input is data faithful to G.
All of the algorithms we will discuss in this section have the following correctness property:
2Indeed, any statistical consttaint can be used as input for the algorithms for any pairing of distributions with
graphs such that the constraint is satisfied in the distribution if and only if the corresponding d-separation relation
holds in the graph.
Tbeorem 5.1: If the input to any of the algorithms is data faithful to G, the output of
each of the algorithms is a pattern that represents G.
The algorithms do not, however, always provide a pattern that explicitly characterizes all of the
orientation information implicit in the d-separation facts; a pattern may be produced that is
consistent only with one orientation of an edge but does not explicitly contain the corresponding
arrowhead.
5.4.1 Tbe SGS Algoritbm
The correctness of the SGS algorithm (Spirtes, Glymour and Scheines, 1990c) follows from
Theorem 3.4:
Tbeorem 3.4: If P is faithful to some directed acyclic graph, then P is faithful to G if

and only if
(i) for all vertices, X, Y of G, X and Y are adjacent if and only if X and Y are
dependent conditional on every set of vertices of G that does not include X or Y; and
(ii) for all vertices X, Y, Z such that X is adjacent to Y ,~nd Y is adjacent to Z and X
and Z are not adjacent, X -> Y <- Z is a subgraph of G if and only if X, Z are
SGS Algoritbm:
A.) Form the complete undirected graph H on the vertex set V.

B.) For each pair of vertices A and B, if there exists a subset S ofv\{A,B} such that A
and Bared-separated given S, remove the edge between A and B from H.
C.) Let K be the undirected graph resulting from step B). For each triple of vertices A,
B, and C such that the pair A and B and the pair B and C are each adjacent in K (written
as A - B - C) but the pair A and C are not adjacent in K, orient A - B - C as A -> B <- C
if and only if there is no subset S of {B} u V\{A,C} that d-separates A and C.
D.) repeat
If A -> B, Band C are adjacent, A and C are not adjacent, and there is no
arrowhead at B, then orient B - Cas B -> C.
If there is a directed path from A to B, and an edge between A and B, then orient
A -B asA ->B.
until no more edges can be oriented.
5.4.1.1 Complexity
Reliability is one thing, efficiency another. Step B) of the SOS algorithm requires that for each
pair of variables adjacent in G we look at all possible subsets of the remaining variables, and
that, of course, is an exponential search. In the worst case that complexity is unavoidable if
reliability is to be maintained. Two variables can be dependent conditional on a set U but
independent on a superset or subset of U. Any procedure that in the worst case does not
examine the conditional independence relations of variables X, Yon all subsets of vertices not
containing that pair will fail--there will be some structure the procedure does not get correctly.
5.4.1.2 Stability of SGS
We need to consider whether an algorithm remains reasonably reliable when the data are
imperfect. We will use the notion of stability informally: If intuitively small errors of input
produce intuitively large errors of output, the algorithm is not stable. For the SOS algorithm, an
intuitively small error in input consists of a few d-separation relations that are falsely included
or falsely excluded from the input. An intuitively small error for Step B is a few undirected
edges erroneously included in or omitted from the output. An intuitively small error for Step C
is a few edges misoriented.
Step B) of the SOS algorithm is stable. If, for example, a single correct d-separation relation is
omitted from the input, the algorithm will nonetheless produce the correct undirected graph
unless there is no other set besides U on which X, Y are d-separated. Even in that case Step B
will make an error in postulating an X - Y connection, but no other errors. If X and Y are
adjacent in the true graph, but it is incorrectly judged that X and Yare d-separated given U, the
algorithm will fail to connect X and Ybut no other error will be made.
Step C) of the SOS algorithm is less stable. A small error in either component of the input,
either the undirected graph or the list of d-separation relations, can (and often will) produce
large errors in the output. That is because the edges that occur in collisions determine the
orientations of other edges in the graph, and if input errors lead the algorithm erroneously to
include or exclude a collision, the error may affect the orientations of many other edges in the
graph.
Suppose, for example, an edge connecting X, Z is erroneously omitted in the undirected graph
input to step C), and X - Y - Z correctly occurs in the input. Then if X and Z are not d-separated
116 Causation:, Prediction, and Search
by any subset of variables containing Y but not X, Z, the algorithm will mistakenly require a
collision at Y, and this requirement will ramify through orientations of other edges. Or, if the
true structure contains a collision at Y but X - Y is omitted in the input to step C), no unique
orientation will be given to Y - Z, and this uncertainty may ramify through the orientations of
other edges on paths including Z.
Instabilities may also arise in Step C) because of errors in the list of d-separation relations input,
even when the underlying undirected graph is correct. If in the input to C), X is adjacent to Y
and Y to Z but not X to Z and a d-separation relation between X and Z given S containing Y is
omitted from the input, no orientation error will result unless no other set containing Y d-
separates X and Z. But if in the true directed graph, the edges between X and Yand between Y
and Z collide at Y, and a d-separation relation involving X and Z and some set U containing Y
but not X or Z is erroneously included in the input, the algorithm will conclude that there is no
collision at Y, and this error may be ramified to other edges.
A little reflection on Step C) reveals that its output may not be a collection of directed acyclic
graphs if one of the four assumptions listed at the beginning of this section is violated. This is
not necessarily a defect of the algorithm. If the algorithm fmds that the edges X - Y - Z collide at
Y, and Y - Z - W collide at Z, it will create a pattern with an edge Y <-> Z. Double headed
edges can occur when the causal structure is not causally sufficient, or when there is an error in
input (as from sampling variation). They have a theoretical role in identifying the presence of
unmeasured common causes, an issue discussed further in the next chapter.
5.4.2 The PC Algorithm
In the worst case, the SOS algorithm requires a number of d-separation tests that increases
exponentially with the number of vertices, as must any algorithm based on conditional
independence relations or vanishing partial correlations. But the SOS algorithm is very
inefficient because for edges in the true graph the worst case is also the expected case. For any
undirected edge that is in the graph G, the number of d-separation tests that must be conducted
in stage B) of the algorithm is unaffected by the connectivity of the true graph, and therefore
even for sparse graphs the algorithm rapidly becomes infeasible as the number of vertices
increases. Besides problems of computational feasibility, the algorithm has problems of
reliability when applied to sample data. The determination of higher order conditional
independence relations from sample distributions is generally less reliable than is the
determination oflower order independence relations. With, say, 37 variables taking three values
each, to determine the conditional independence of two variables on the set of all remaining
variables requires considering the relations among the frequencies of 335 distinct states, only a
fraction of which will be instantiated even in very large samples.
We should like an algorithm that has the same input/output relations as the SOS procedure for
faithful distributions but which for sparse graphs does not require the testing of higher order
independence relations in the discrete case, and in any case requires testing as few d-separation
relations as possible. The following procedure (Spirtes, Glymour, and Scheines, 1991) starts
by forming the complete undirected graph, then "thins" that graph by removing edges with zero
order conditional independence relations, thins again with fIrst order conditional independence
relations, and so on. The set of variables conditioned on need only be a subset of the set of
variables adjacent to one or the other of the variables conditioned.
Let Adjacencies(C,A) be the set of vertices adjacent to A in directed acyclic graph C. (In the
algorithm, the graph C is continually updated, so Adjacencies(C,A) is constantly changing as
the algorithm progresses.)
PC Algorithm:
A.) Form the complete undirected graph C on the vertex set V.

B.)
n =0.
repeat
repeat
select an ordered pair of variables X and Y that are adjacent in C such
that Adjacencies(C,x)\{y} has cardinality greater than or equal to
n, and a subset S of Adjacencies( C ,x)\{ Y} of cardinality n, and if
X and Y are d-separated given S delete edge X - Y from C and
record S in Sepset(X,y) and Sepset(Y,x);
until all ordered pairs of adjacent variables X and Y such that
Adjacencies(C,x)\{Y} has cardinality greater than or equal to n and all
subsets S of Adjacencies(C,x)\{Y} of cardinality n have been tested for
d-separation;
n =n + 1;
until for each ordered pair of adjacent vertices X, Y, Adjacencies(C,K)\{Y} is
of cardinality less than n.
C.) For each triple of vertices X, Y, Z such that the pair X, Y and the pair Y, Z are each
adjacent in e but the pair X, Z are not adjacent in e, orient X - Y - Z as X -> Y <- Z if
and only if Y is not in Sepset(X,z).
D. repeat
If A -> B, B and e are adjacent, A and e are not adjacent, and there is no
arrowhead at B, then orient B - e as B -> e.
A -B asA ->B.
Figure I traces the operation of the first two parts of the PC algorithm:
~e"
A --"B
f~77j
True Graph Complete Undirected Graph
n=0 No zero order independencies
n= 1 First order independencies Resulting Adjacencies
A lle IB A llD IB
A--B
AllE IB e llD IB
n =2: Second order independencies Resulting Adjacencies
B llE I {e,D}
A--B
Figure 1
Although it does not in this case, stage B) of the algorithm may continue testing for some steps
after the set of adjacencies in the true directed graph has been identified. The undirected graph at
the bottom of figure I is now partially oriented in step C). The triples of variables with only two
adjacencies among them are:
A -B - C; A -B -D;
C -B -D; B - C -E;
B -D -E; C-E-D
E is not in Sepset(C,D) so C - E and E - D collide at E. None of the other triples form

colliders. The final pattern produced by the algorithm is shown in figure 2.
Figure 2
The pattern in figure 2 characterizes a faithful indistinguishability class. Every orientation of the
undirected edges in figure 2 is permissible that does not include a collision at B.
5.4.2.1 Complexity
The complexity of the algorithm for a graph G is bounded by the largest degree in G. Let k be
the maximal degree of any vertex and let nbe the number of vertices. Then in the worst case the
number of conditional independence tests requited by the algorithm is bounded by
2(n)±(n~l)
2 .=0 I
which is bounded by
n2(n _1)<-1
(k -I)!
This is a loose upper bound even in the worst case; it assumes that in the worst case for nand
k, no two variables are d-separated by a set of less than cardinality k, and for many values of n
and k we have been unable to find graphs with that property. While we have no formal expected
complexity analysis of the problem, the worst case is clearly rare, and the average number of
conditional independence tests required for graphs of maximal degree k is much smaller. In
practice it is possible to recover sparse graphs with as many as a hundred variables. Of course
the computational requirements increase exponentially with k.
The structure of the algorithm and the fact that it continues to test even after having found the
correct graph suggest a natural heuristic for very large variable sets whose causal connections
are expected to be sparse, namely to fix a bound on the order of conditional independence
relations that will be tested.
5.4.2.2 Stability of PC
In theory, the PC Algorithm is unstable in both steps B) and C) although in practice step B) has
proved to be much more reliable than step C).
If an edge is mistakenly removed from the true graph at an early stage of step B) of the
algorithm, then other edges which are not in the true graph may be included in the output.
Consider the following example.
/\
----1.~D
Figure 3
If the edge E - D is mistakenly removed from the initial complete graph then at a subsequent
stage of the search the edge B - D will not be removed, because E will no longer be in the
adjacency set for D, and B and D are dependent on every subset of A and C. The omission of
an edge can also lead to orientation errors. If an edge is mistakenly left in the graph and there
Discovery Algoritluns for Causally Sufficient Structures 121
are no additional errors in the list of d-separations in the input, the only further errors that result
are that some edges which theoretically could be oriented, will not be oriented.
Step C) of the algorithm is unstable for the same reasons as in step C) of the SGS algorithm.
The PC algorithm is faster than the SGS algorithm because it tests fewer d-separation relations.
Given a faithful list of d-separability relations, the two algorithms output the same set of pattern
graphs. But if the list of d-separability relations is not faithful, due to sampling error for
example, the two algorithms can output different pattern graphs. Consider the following
example.
A -~ •• c ---i.~E
Figure 4
According to this graph, A and E are d-separated from each other given any non-empty subset
of B, C, and D. If, after the A - C and E - C edges have been removed from the initial
undirected graph, the procedure incorrectly judges that A and E are not d-separated given any
non-empty subset of Band D, the PC algorithm will incorrectly include an edge between A and
E, because it only tests whether A and E are d-separated given subsets of the adjacencies of A
and E. On the other hand, because the SGS algorithm tests whether A and E are d-separated
given any subset of v\{A,B}' it would properly recognize that there is not an edge between A
and E because A and E are d-separated given C.
In contrast, if after the A - E and B - E edges are removed from the initial undirected graph, it is
mistakenly judged that A and B are d-separated given E, the SGS algorithm will mistakenly
remove the A -> B edge. If the A - E and B - E edges are removed fIrst, the PC algoritlun,
would correctly leave the A -> B edge in, because it would not test whether A and B are d-
separated given E.
Because the PC algorithm attempts to use "local" information to judge whether an edge exists or
not, it is not guaranteed to produce a graph that is in some sense "closest" to an unfaithful
distribution. Consider the following example.
Figure 5
In a distribution faithful to this graph, every variable is dependent on every other variable.
Suppose a test determines that A and B are independent conditional on some other variable,
either because of some coincidental parameter values, or because of sampling error. The PC
algorithm would then remove the A - B edge in order to satisfy that constraint. In doing so,
however, it would disconnect the graph. The resulting graph would entail that A and all of its
descendants to the left are independent of B and all of its descendants. So, in order to satisfy
one conditional independence constraint, the PC algorithm may produce a graph that violates a
great many independence constraints. In a number of data sets the correlations between two
variables do not vanish but the output pattern disconnects them. For greater reliability the
procedure should be supplemented with a repair algorithm, for which the Cooper and
Herskovits Bayesian procedure might suffice in the discrete case; alternatively, a variation of the
procedure described in Chapter 11 could be applied.
5.4.2.3 Tbe PC* Algoritbm
The PC algorithm is computationally efficient and asymptotically reliable, but on sample data
the procedure takes unnecessary risks. In determining whether to eliminate an undirected edge
between variables A and B, the procedure may test every subset of the adjacency set of A and of
the adjacency set of B. But the independence or dependence of A and B on many of these
subsets of variables may be entirely irrelevant to the causal relations between A and B. For a
distribution faithful to a directed acyclic graph, if variables A and B are independent conditional
given Parents(A) or given Parents(B) then they are independent given a subset of
Parents(A) or given a subset of Parents(B) consisting only of vertices lying on undirected
paths between A and B. It is sufficient, then, to test for the conditional independence of A and
B given subsets of variables adjacent to A and subsets of variables adjacent to B that are on
undirected paths between A and B. Call the modified algorithm PC*~
The PC and PC* algorithms yield the same output given a faithful list of conditional
independence relations or correlations as input, but they may differ given conditional
independence relations determined from tests on sample data. The PC* algorithm avoids one
kind of error made by the PC algorithm. If, however, at an early stage the PC* algorithm
mistakenly disconnects a path between X and Y it may then mistakenly leave the X - Yedge in
the undirected graph, while the PC algorithm, given the same data, might avoid that error.
Moreover, whatever increased reliability the PC* algorithm may have is bought at great cost,
since the algorithm must at each stage of step B) keep track of all of the undirected paths in the
graph it considers at that stage. The number of undirected paths is typically very large, and the
memory requirements of the PC* algorithm are not feasible save for relatively small numbers of
variables, in which case it may be the algorithm of choice. For large numbers of variables the
PC algorithm must be used instead, although if the true graph is sparse, the PC algorithm can
be used until the average degree of the undirected graph C is small, after which stage the PC*
algorithm may be used. Later in this chapter we will describe the performance of the two
algorithms on discrete data taken from Christensen (1990).
5.4.2.4 Speed-Up Heuristics for Ordering Tests
Step B of the PC algorithm selects some variable pair and some subset S of a given size to test
for d-separation. The faster edges are eliminated from the complete graph, the smaller the search
that has to be conducted at later stages of the algorithm, and the faster the algorithm runs.
Hence, it is best to select first for testing those variable pairs A and B and subsets S for which
A and B are most likely to be d-separated by S. We have considered three variants of the PC
algorithm that use different methods of selecting the order of tests.
Heuristic 1: Test the variable pairs and subsets S in lexicographic order. (We will call
this PC-I.)
Heuristic 2: First test those variables pairs that are least dependent3 in probability. The
conditioning subsets are selected by lexicographic order. (We will call this PC-2.)
Heuristic 3: For a given variable A, first test those variables B that are least
probabilistically dependent on A, conditional on those subsets of variables that are most
probabilistically dependent on A. (We will call this PC-3.)
The intuition behind heuristic 2 is that variables with the highest probabilistic dependence are
most likely to be adjacent in the true graph, and hence not ever eliminated from the graph being
constructed, while those with the smallest probabilistic dependence are least likely to be adjacent
in the true graph. Of course, no such relation strictly holds.
The intuition behind heuristic 3 is similar. A variable B that is not genuinely adjacent to a
variable A is d-separated from A given some subset of the variables that are adjacent to A or
given some subset of the variables that are adjacent to B in the true graph. Assuming that
variables with the highest probabilistic dependence upon A are most likely to be adjacent to A in
the true graph, this suggests testing whether A is d-separated from variables with a low
probabilistic dependence on A, conditional on variables with a high probabilistic dependence
upon A.
5.4.3 The IG (Independence Graph) Algorithm
Verma and Pearl (1990) have suggested a variation of the SGS algorithm. In their alternative,
the first step in searching for the directed acyclic graph is to construct the undirected
independence graph N, i.e. for each pair of variables A, B introduce an undirected edge
between them if they are dependent conditional on the set of all other variables. In the
undirected independence graph for a distribution faithful to a directed acyclic graph the parents
of any variable form a maximal complete subgraph--a clique. Again for each pair of variables A,
B adjacent in N, determine if A and B are d-separated given any subsets of variables in the
cliques in N containing A or B. If so A is not adjacent to Bin G. The complexity is thus a
function of the size of the largest clique in N.
31n the following heuristics, "high probabilistic dependence" means high partial correlation in the linear case, and
high G2 statistic in the discrete case.
Determining the cliques in a graph would appear to require unnecessary computation, and in
other than the worst case, testing for conditional independence of two variables conditional on
all members of the maximal clique of one or the other will involve a test of unnecessarily high
order. A better idea might be to blend the procedure with the PC algorithm: modify step A of the
PC algorithm by setting the initial graph in the PC procedure to the undirected independence
graph, rather than the complete undirected graph, and then proceed in the same way. We will
call this algorithm IG (independence graph.)
The efficiency of these algorithms obviously depends upon how easily the independence graph
can be constructed. The off-diagonal elements of the standardized inverse of the correlation
matrix are the negatives of the partial correlation coefficients between the corresponding
variables given the remaining variables (see e.g. Whittaker, 1990). Hence in the linear case, the
independence graph can be efficiently constructed by placing an edge between A and B if and
only if the entry in the standardized inverse correlation matrix is non-zero. In the discrete case,
Fung and Crawford (1990) have recently proposed a fast algorithm for constructing an
independence graph from discrete data. We have not tested their procedure as a preprocessor for
the PC algorithm.
5.4.4 Variable Selection
While prior knowledge of causal structure can sometimes make the results of the algorithms we
have described more informative on real samples, correct selection of variables is essential for
reliable inference, and for that algorithms (at least these algorithms) provide no help.
We can aggregate variables or we can aggregate distinct values of a variable. As in Salmon's

imaginary example discussed in Chapter 3, we sometimes measure a variable that is an
imprecise version of a more precise natural variable; we fail, in other words, to distinguish
values that have differing effects on other variables. Continuous variables are often deliberately
collapsed into a few discrete categories, sometimes because contingency table methods offer the
promise of statistical analysis free of the substantive assumptions that would otherwise be
required about the form of the functional dependencies--e.g., linear or otherwise--and
sometimes because some of the variables to be analyzed are necessarily discrete and there are
few methods available for problems with mixtures of discrete and continuous variables.
Sometimes, whether through ignorance or even deliberately, we may aggregate two or more
distinct variables with distinct causal structures into a single scale. What effects can aggregation
and collapse have on the reliability of causal inference?
We have already observed that if C is a cause of A and B and some proxy C' for C is used that
is not so precise as C and not perfectly correlated with C, it may be that A and B are statistically
dependent conditional on C. Examples of this sort appear whenever a theory postulates a cause
that is measured by proxies. Friedman (1957), for example, advocated a much discussed theory
in which consumption is caused by "permanent" income which can only be measured by
proxies; if Friedman's theory were true, regression of consumption on measured income would
provide a biased estimate of the regression coefficient of consumption on permanent income and
might leave unexplained correlations between consumption and other variables. Klepper (1988)
has shown how, in the linear nurmal case, such errors may be bounded.
Suppose we are given variables A, B, C such that A and B are independent conditional on C.
Let C = PROJ(C) where PROJ(C) is a projection mapping the set of n values of C to a set of m
< n values. If there exist values C}, C2 for C such that P(A,BIC=C}) "* P(A,BIC=C2) and
PROJ(C=q) = PROJ(C=C2), then A and B are not independent conditional on C'.
Independence relations can be made to appear rather than disappear by collapsing values of a
variable. Suppose that variable B, C are dependent. Let C' = PROJ(C) where PROJ(C) is a
projection mapping the set of n values of C to a set of m < n values. If there exists a value ct of
C such that P(C = C} I B) = P(C = q) and PROJ(ct) has a unique inverse and PROJ(q) =
PROJ(Cj) for all k, j not equal to 1, then Band C are independent.
Pearl (personal communication) has pointed out that a very simple sort of aggregation can
produce an unfaithful distribution. Suppose A causes Cl and B causes C2, and Cl and C2 are
each binary, and there is no other causal connection among the variables. So {A, Cd is
independent of the set {B, C2l, but A and Cl are dependent and so are Band C2. Introduce
variable C taking values 0, 1, 2, 3 coding the different value pairs for Cl and C2. Then the
actual causal structure among A, B and C is shown in figure 6.
Figure 6
But in the joint distribution A and B are independent conditional on C, and so the joint
distribution is not faithful to any causal structure whatsoever. In this case the unfaithfulness of
the distribution is due to the fact that it is the marginal of a distribution that is unfaithful because
of deterministic relationships among the variables: the independence of A and B given C
follows directly from the application of D-separability (see Chapter 3) to figure 6. This sort of
thing may sometimes happen in practice, but it could always be tested for and in principle
identified: Conditioning on A divides the values of C into two equivalence classes each
containing values of C with the same conditional probability, and conditioning on B divides the
values of C into a distinct pair of equivalence classes. Letting the equivalence classes induced
by A be values of one variable and the equivalence classes induced by B be values of another
variable recovers Cl and C2.
5.4.5 Incorporating Background Knowledge
A user of any of these algorithms may have a great deal of background knowledge--or at least
belief--that could constrain the search. This knowledge might be about the existence or non-
existence of certain edges in the graph, or it might be about the orientation of some of the edges,
or it might be about the time order of the variables. How can this background knowledge be
used by the algorithms?
The most common sort of reliable prior belief orders or partially orders the variables by time of
occurrence: either measurements of A were taken before measurements of B, or A and Bare
believed to be exact measures of events that are so ordered. Any of the algorithms of this
section can be easily modified to make two uses of such knowledge:
(i) In determining whether A and B are adjacent in the true graph by testing whether B is
independent of A conditional on some subset of the current adjacencies of A, do not test for
independence conditional on any set of variables that includes a variable that is later than A.
(u) If A and B are adjacent and B is later than A, orient the edge as A -> B.
In the examples we give throughout this book the algorithms have been so modified, and we
sometimes make use of common sense time order, always noting where such assumptions have
been made.
Prior belief about whether one variable directly influences another can also be incorporated in
these algorithms: if prior belief forbids an adjacency, for example, the algorithms need not
bother to test for that adjacency; if prior belief requires than there be a direct influence of one
variable on another, the corresponding directed edge is imposed and assumed in the orientation
procedures for other edges. These procedures assume that prior belief should override the
results of unconstrained search, a preference that may not always be judicious; they are
nonetheless inc'orporated in versions of the TETRAD II program with the PC algorithm.
5.5 Statistical Decisions
The algorithms we have described are completely modular, and can be applied given any
procedures for making the requisite statistical decisions about conditional independence or
vanishing partial correlations. The better the decisions the better the performance to be expected
from the algorithms. While tests of conditional independence relations form the most obvious
class of such decisions, any statistical constraints that give d-separability relations for graphical
structure will suffice. For example, in the linear normal case, vanishing partial correlation is
equivalent to conditional independence, and the statistical decisions required by the algorithms
could be provided by t-tests of the hypotheses that partial correlations vanish. But vanishing
partial correlation marks d-separability whether or not the distribution is normal, so long as
linearity and linear Faithfulness hold.4 Hence under these assumptions the test of any statistic
that vanishes when partial correlations vanish would suffice; one might, for example, use an F
test for the square of the semi-partial correlation coefficient, which equals the square of the t-test
for a corresponding regression coefficient (Edwards, 1976).
In the examples in this book we test whether PXY.C = 0 using Fisher's z:
PXY.C = population partial correlation of X and Y given C, and ICI equals the number of
variables in C. If X, Y, and C are normally distributed and rxy.c denotes the sample partial
correlation of X and Y given C, the distribution of z(pxy.c,n) - z(l'XY.c,n) is standard normal
(Anderson, 1984).
In the discrete case, for simplicity consider two variables. Recall that we view the count in a
particular cell, Xij, as the value of a random variable obtained from sampling N units from a
4 For causally sufficient structures, if a distribution P, obtained by imposing a linear distribution compatible
with a graph G, implies some vanishing partial correlation not linearly implied by G, is then P not faithful to
G? If P is not faithful to G, does P necessarily imply some vanishing partial correlation not linearly implied by
G? We don't know the answer to either question.
multinomial distribution. Let Xi+ denote the sum of the counts in all cells in which the first
variable has the value i, and similarly let x+j denote the sum of the counts in all cells in which
the second variable has the value j. On the hypothesis that the fIrst and second variables are
independent, the expected value of the random variable Xij is:
Analogously, we can compute the expected values of cells on any hypothesis of conditional
independence from appropriate marginals. For example, on the hypothesis that the first variable
is independent of the second conditional on the third, the expected value of the cell Xijk is
If there are more than three variables this formula applies to the expected value of the marginal
count of the i, j, k values of the first three variables, obtained by summing over all other
variables. The number of independent constraints that a conditional independence hypothesis
places on a distribution is an exponential function of the order of the conditional independence
relation and also depends on the number of distinct values each variable can assume.
Tests of such independence hypotheses have used--among others--two statistics:
X2 = I, (Observed - Expected)2
Expected
G2 =2I,(Observed)ln(Observed)
Expected
each asymptotically distributed as '1.2 with appropriate degrees of freedom. In the examples in
this book we calculate the degrees of freedom for a test of the independence of A and B
conditional on C in the following way. Let Cat(X) be a function which returns the number of
categories of the variable X, and n be the number of variables in C. Then the number of degrees
of freedom (dj) in the test is:
n
df =(Cat(A) -1) x (Cat(B) -1) x II Cat(C;)
i=1
We assume that there are no structural zeroes. As a heuristic, for each cell of the distribution
that has a zero entry, we reduce the number of degrees offreedom by one5.
Because the number of cells grows exponentially with the number of variables, it is easy to
construct cases with far more cells than there are data points. In that event most cells in the full
joint distribution will be empty, and even non-empty cells may have only small counts. Indeed,
it can readily happen that some of the marginal totals are zero and in these cases the number of
degrees of freedom must be reduced in the test. For reliable estimation and testing, Fienberg
recommends that the sample size be at least five times the number of cells whose expected
values are determined by the hypothesis under test.
For discrete data we fIll out the PC algorithm with tests for independence using G2 which in
simulations we have found more often leads to the correct graph than does X 2. In testing the
conditional independence of two variables given a set of other variables, if the sample size is
less than ten times the number of cells to be fitted we assume the variables are conditionally
dependent.
5.6 Reliability and Probabilities of Error
Most of the algorithms we have described require statistical decisions which, as we have just
noted, can be implemented in the form of hypothesis tests. But the parameters of the tests
cannot be given their ordinary significance. The usual comforts of a statistical test are the
significance level, which offers assurance as to the limiting frequency with which a true null
hypothesis would erroneously be failed by the test, and the power against an alternative, which
is a function of the limiting frequency with which a false null hypothesis would not be rejected
when a specified alternative hypothesis is true. Except in very large samples, neither the
significance level nor the power of tests used within the search algorithms to decide statistical
5An exact general rule for calculating the reduction of degrees of freedom given cells with zero entries seems not
to be known. See Bishop, Fienberg, and Holland (1975).
dependence measures the long run frequency of anything interesting about the search. What
does?
The error probabilities one might naturally want to know for a search procedure include:
1. Given that model M is true, what is the probability that the procedure will return a
conclusion inconsistent with M on sample size n?
2. Given that model M* is true, what is the probability that the procedure will return a
conclusion inconsistent with M* but consistent with M on sample size n?
3. Given that model M is true, for samples of size n what is the probability that a search
procedure will specify an adjacency not in M? What is the probability that a search
procedure will omit an adjacency in M? What is the probability that a search procedure will
add an arrowhead not in M to an edge that is in M? What is the probability that a search
procedure will omit an arrowhead in M? What are these probabilities for any particular
variable pair, A, B?
For large models, where we expect some errors of specification from most samples, questions
of kind 3 are the most important.
There is little hope of obtaining analytic answers to these questions. In repeated tests of
independence hypotheses in a sample, each using the same significance level, the probability
that some true hypothesis will be rejected is not given by the significance level; depending on
the number of hypotheses and the sample size, that probability may in fact be much higher than
the significance level, but in any case the probability of some erroneous decision depends on
which hypotheses are tested, and for all of the algorithms considered that in tum depends in a
complex way on the actual structure. Further, each of the algorithms can produce correct output
even though some required statistical decisions are made incorrectly. For example, suppose in
graph G, vertices A and B are not adjacent. Suppose in fact A and B are independent
conditional on C, on D, on C and D, and so on. If the hypothesis that A and B are independent
conditional on C is rejected in the search procedure, and the algorithm goes on to test whether A
and B are independent conditional on D, and decides in favor of the latter independence, then
despite the earlier error the procedure will correctly conclude that A and B are not adjacent.
For any particular M and M* estimates of the answers to questions 1,2, and 3 can be found
empirically by Monte Carlo methods. Simulation packages for linear normal models are now
common in commercial statistical packages, and the TETRAD II program contains a simulation
package for linear and for discrete variable models with a variety of distributions. For small
models it takes only a few minutes to generate a hundred or more samples and run the samples
through the search procedures. Most of the time required is in counting the outcomes, a process
that we have automated ad hoc for our simulations, and that can and should be automated in a
general way for testing the reliabilities of particular search outcomes.
5.7 Estimation
There are well known methods for obtaining maximum likelihood estimates subject to a causal
hypothesis under the assumption of normality, even with unmeasured variables, (Joreskog,
1981; Lohmoller, 1989». A variety of computerized estimation methods, including ordinary
and generalized least squares, are also available when the normality assumption is given up. In
the discrete case, for a positive multinomial distribution, the maximum likelihood estimates
(when they exist) for a cell subject to the independence constraints of the graph over a set of
variables V can be obtained by substituting the marginal frequencies for probabilities in the
factorization formula of Chapter 3 (Kiiveri and Speed, 1982).
P(V) = II P(VIParents(V»)
VEV
When there are unmeasured variables that act as common causes of measured variables, the
pattern obtained from the procedures we have described can have edges with arrows at each
end. In that rather common circumstance we do not know how to obtain a maximum likelihood
estimate for the joint distribution of discrete measured variables, but work on estimating log-
linear models with latent variables (Haberman, 1979) may serve as a guide.
5.8 Examples and Applications
We illustrate the algorithms for simulated and real data sets. With simulated data the examples
illustrate the properties of the algorithms on samples of realistic sizes. In the empirical cases we
Discovery Algorithms for Causally Sufficient Structmes 133
often do not know whether an algorithm produces the truth. But it is at the very least interesting
that in cases in which investigators have given some care to the treatment and explanation of
their data, the algorithm reproduces or nearly reproduces the published accounts of causal
relations. It is also interesting that in cases without these virtues the algorithm suggests quite
different explanations from those advocated in published reports.
Studies of regression models and alternatives produced by the PC algorithm and by another
procedure, the Fast Causal Inference (FCI) algorithm, are postponed until Chapter 8, after latent
variables and prediction have been considered in Chapters 6 and 7, respectively.
5.8.1 The Causes of Publishing Productivity
In the social sciences there is a great deal of talk about the importance of "theory" in
constructing causal explanations of bodies of data. Of course in explaining a data set one will
always eliminate causal graphs that contradict common sense or that violate the time order of
variables. But in addition, many practitioners require that every attempt to provide a causal
explanation of observational data in the social sciences proceed through the particulars of
principles in sociology, psychology, economics, political science, or whatever, and come
accompanied with a denial of the possibility of determining a correct explanation from the
statistical dependencies alone. In many of these cases the necessity of theory is badly
exaggerated. Indeed, for every "recursive" structural equation model in the entire scientific
literature, if the assumptions of the model are correct and no unmeasured common causes are
postulated, then if the distribution is faithful the statistical dependencies in the popUlation
uniquely determine the undirected graph underlying the directed graph of causal relations. And
in many cases the population statistics alone determine a direction of some, or even all, edges.
When the variables are linearly ordered by time, so that variable A can be a cause of variable B
only if A occurs later than B, the statistical dependencies and the time order determine a unique
directed graph assuming only that the distribution is positive and the Markov and Minimality
Conditions are satisfied. The efforts spent citing literature to justify specifications of causal
dependencies are not misplaced, but in many cases effort would be better directed towards
establishing the fundamental statistical assumptions, including the approximate homogeneity of
the units, the correctness of the sampling assumptions, and sometimes the linearity of
dependencies.
Here is a recent and rather vivid example. There is a considerable literatme on causes of
academic success, including publication and citation rates. A recent paper by Rodgers and
Maranto (1989) considers hypotheses about the causes of academic productivity drawn from
sociology, economics, and psychology, and produces a combined "theoretically based" model.
Their data were obtained in the following way: solicitations and questionnaires were sent to 932
members of the American Psychological Association who obtained doctoral degrees between
1966 and 1976 and were currently working academic psychologists. Equal numbers of male
and female psychologists were sampled, and after deleting respondents who did not have
degrees in psychology, did not take their ftrst job in psychology, etc. a sample of 86 men and
76 women was obtained.
The response items were clustered into groups. For example, the ABILITY group consisted of
measures of the mean ACT, NMSQT and selectivity scores of the subject's undergraduate
institution, together with membership in Phi Beta Kappa and undergraduate honors at
graduation. Graduate Program Quality (GPQ) consisted of the scholarly quality of department
faculty and program effectiveness using national rankings, the fraction of faculty with
publications between 1978 and 1980, and whether an editor of a journal was on the department
faculty. These response items were treated as indicators--i.e., as effects--of the unmeasured
variables GPQ, and ABILITY. Other measures were quality of ftrst job (QFJ), SEX, citation
rate (CITES) and publication rate (PUBS). In preliminary analyses they also used an aggregated
measure of productivity (PROD). The various hypotheses Rodgers and Maranto considered
were then treated as linear "structural equation models"6 They report the following correlations
among the cluster variables
ABILITY GPQ PREPROD QFJ SEX CITES PUBS

1.0
.62 1.0
.25 .09 1.0
.16 .28 .07 1.0
-.10 .00 .03 .10 1.0
.29 .25 .34 .37 .13 1.0
.18 .15 .19 .41 .43 .55 1.0
6It is not clear from the article how the correlations of the latent variables. GPQ and ABILITY, with other
variables such as publishing productivity and QFJ were obtained. They can be obtained. for example. by using
the factor structure as a regression model to calculate estimated factor scores for each subject, or by including the
covariances of the Iatents among the free parameters in a set of structural equations and letting a program such as
LISREL estimate their values. In general the results of these procedures will be different
There follows a very elaborate explanation of causal theories suggested by the pieces of
sociological, economic and psychological literature. Rodgers and Maranto estimate no fewer
than six different sets of structural equations and corresponding causal theories. The six
structures they consider are as follows:
'" 1
ABILITY ABILITY
G:{ ~F]
"
,/
GPQ QF]
PROD
HUMAN CAPITAL
,/ " PROD
,/
JOB KNOWLEDGE
ABILITY ABILITY
~ GPL
GPQ
.............QF]
~ / F ]'-. ;ITES
- - . PROD
PREPROD PUBS
SCREENING ACCUMULATIVE
ADVANTAGE
ABILITY SEX
/
GPQ
......... QF] GP~I\~
/,
PREPROD ~PROD I }7T S
MERIT EMPIRICAL MODEL
Figure 7
The labels on the graphs indicate simply the social scientific theory from which Rodgers and
Maranto derived the causal graph. For example, the "Human Capital" and the "Screening"
graphs were obtained from economic theory in the following way:
In the human capital model (Becker, 1964) education has a direct effect on
productivity because it conveys relevant knowledge. People invest in education
until its marginal cost (the extra expenses and foregone earnings for an
additional year of education) is equal to its marginal benefit (the increase in
lifetime earnings caused by another year of education). More able individuals are
more productive in both work and the acquisition of skills than their less able
counterparts. Thus, ability has a direct effect on productivity and an indirect
effect through education, because more able individuals gain more from school.
Work experience also increases productivity by providing on-the-job training.
The quality as well as the quantity of education is relevant to the human capital
framework.
The screening hypothesis implicitly views ability as the primary determinant of

productivity. Employers wish to hire the most productive applicants, but ability
is not directly observable. Individuals invest in education as a means of
signaling their ability to employers. The marginal cost of education is inversely
related to ability, inducing a positive correlation between ability and the level of
education. Therefore, by selecting applicants based on their education,
employers hire by ability (Spence, 1973). In this model, education does not
affect productivity directly, but only through its association with ability.
Variations in the quality of education are consistent with the screening model
(Wise, 1975).
The "empirical model" was obtained from a previous study that did not appeal to social theory.
None of the structural equation systems based on these models save the phenomena. But
combining all of the edges in the "theoretical" models, adding two more that seem plausible,
and then throwing out statistically insignificant (at .05) dependencies, leads Rodgers and
Maranto instead to propose a different causal structure which fits the data quite well:
Y\
ABILITY
GPQ .25
.14
.2~
SEX 1QF~ l·~:EPROD
.41 pr BS .22
.16 .4~
CITES
Figure 8
It would appear that the tour through "theory" was nearly useless, but Rodgers and Mantnto say
otherwise:
Causal models based solely on the pattern of observed correlations are highly
suspect Any data can be fitted by several alternative models. The construction
of the best-fit model was thus guided by theory-based expectations. By using
the two measures of productivity, PUBS and CITES, and the five causal
antecedents, we initially estimated a composite model with all of the paths
identified by the six theories. This model produced a large positive deviation
between the observed and predicted correlation of ABILITY with PREPROD,
suggesting that we omitted one or more important paths. Reexamination of our
initial interpretation of the six theories led us to conclude that two paths had been
overlooked. One such path is from ABILITY to PREPROD ...The other
previously unspecified path is from ABILITY to PUBS. These two paths were
added and all nonsignificant paths were deleted from the composite model to
arrive at the best-fit model.
If the Rodgers and Maranto theory were completely correct, the undirected graph underlying
their directed graph would be uniquely determined by the conditional independence relations,
and the orientation would be almost uniquely determined; only the directions of the GPQ ->
QFJ, ABILITY -> GPQ and ABILITY -> PREPROD edges could be changed, and only in a
way that does not create a new collision.
When the PC algorithm is applied to their correlations with the common sense time order using
a significance level of .1 for tests of zero partial correlations, the output is the graph on the left
side of figure 9, which we show alongside the Rodgers and Maranto model.
ABILITY ABILITY
/
GPQ
" \/
QFJ PREPROD
GPQ
" \ QFJ PREPROD
ill
SEX
irl
SEX
CITES CITES
PC Output Rodgers and Maranto Graph
Figure 9
All but one of the edges in the Rodgers and Maranto model is produced instantaneously from
the data and common sense knowledge of the domain--the time order of the variables. EQS
gives this model a X2 of 13.58 with 11 degrees of freedom and p = .257. If the search
procedure is repeated using .05 as the significance level, the program deletes the PREPROD ->
PUBS edge. When that model is estimated and tested with the EQS program we find that X2 is
19.2 with 12 degrees of freedom and ap value of .08, figures that should be taken as estimates
of fit rather than of the probability of error.
Any claim that social scientific theory--other than common sense--is required to find the
essentials of the Rodgers and Maranto model is clearly false. Nor do the preliminary results of
Rodgers and Maranto's search afford any reason for confidence in social scientific theory. In
contrast, we know a good deal about the reliability and limitations of the PC algorithm. The
entire study with TETRAD and EQS takes a few minutes. A slight variant of the model is
obtained using the SGS algorithm rather than the PC algorithm.
5.8.2 Education and Fertility
Rindfuss, Bumpass and St. John (1980) were interested in the mutual influence in married
women of education at time of marriage (ED) and age at which a first child is born (AGE). On
theoretical grounds they argue at length for the model on the left in figure 10, where the
regressors from top to bottom are as follows:
DADSO= father's occupation

RACE = race
NOSIB = absence of siblings
FARM = farm background
REGN= region of the United States
ADOLF = presence of two adults in the subject's childhood family
REL= religion
YC!G= cigarette smoking
FEC= whether the subject had a miscarriage.
Regressors are correlated. The sample size is 1766, and the covariances are given below.
DADSO RACE NOSIB FARM REGN ADOLF REL YCIG FEC ED AGE
456.676
-.9201 .089
-15.825 .1416 9.212
-3.2442 .0124 .3908 .2209
-1.3205 .0451 .2181 .0491 .2294
-.4631 .0174 -.0458 -.0055 .0132 .1498
.4768 -.0191 .0179 -.0295 -.0489 -.0085 .1772
-0.3143 .0031 .0291 .0096 .0018 .0089 -.0014 .1170
.2356 .0031 .0018 -.0045 -.0039 .0021 -.0003 .0009 .0888
18.66 -.1567 -2.349 -.2052 -.2385 -.1434 -.0119 -.1380 .0267 5.5696
16.213 -.2305 -1.4237 -.2262 -.3458 .1752 .1683 .1702 .2626 3.6580 16.6832
Apparently to their surprise, the investigators found on estimating coefficients that the AGE ->
ED parameter is zero_ Given the prior information that ED and AGE are not causes of the other
variables, the PC algorithm (using .the 05 significance level for tests) directly finds the model
on the right in figure 10, where connections among the regressors are not pictured.
Rindfuss, et al.theoretical TETRAD II model

model; AGE -> ED coefficient
not statistically significant
Figure 10
5.8.3 The Female Orgasm
Bentler and Peeler (1979) obtained data from 281 female university undergraduates regarding
personality and sexual response. They include the Eysenck Personality Inventory which
measured neuroticism (N) and extraversion (E); a heterosexual behavior inventory (HET), a
monosexual behavior inventory (MONO); a scale of negative attitudes towards masturbation
(ATM) and an inventory of subjective assessments of coital and masturbatory experiences.
Using factor analysis the investigators formed scales, thought to be unidimensional, from these
responses, including two scales (SCaR) and (SMOR) from the subjective assessments of coital
and masturbatory experiences.
The investigators were interested in two hypotheses: (1) subjective orgasm responses in
masturbation and coitus are due to distinct internal processes; (2) extraversion, neuroticism and
Discovery Algorithms for Causally Suffic~nt Structures 141
attitudes toward masturbation have no direct effect on orgasmic responsiveness, measured by

SCOR and SMOR, but influence that phenomenon only through the history of the individual's
sexual experience measured by HEr and MONO.
We will not discuss the formation of the scales in this case, since the only data presented are the
correlations of the scales and inventory scores, which are:
E N ATM HET MONO SCOR SMOR

1.0
-.132 1.0
.009 -.136 1.0
.22 -.166 .403 1.0
-.008 .008 .598 .282 1.0
.119 -.076 .264 .514 .176 1.0
.118 -.137 .368 .414 .336 .338 1.0
Bentler and Peeler offer two linear models to account for the correlations. The models and the
probability values for the associated asymptotic 1,2 are shown in figure 11.
SCOR SCOR
l/f
----I.~ ----I.~
!/ONO~
ATM SMOR SMOR
Model I Model 2
p < .001 p = .21
Figure II
Only the second model saves the phenomena. The authors write that
.. .it proved possible to develop a model of orgasmic responsiveness consistent

with the hypothesis that extraversion (e), neuroticism (n), and attitudes toward
masturbation (aoo) influence orgasmic responsiveness only through the effect
these variables have on heterosexual (het) and masturbatory (mono) experience.

Consequently hypothesis 2 appears to be accepted (p.419).
The logic of the argument is not apparent. As the authors note "it must be remembered that other
modes (sic) could conceivably also be developed that would equally well describe the data." (p.
419). But if the data could equally well be described, for example, by a model in which ATM
has a direct effect on SCOR or on SMOR, there is no reason why hypothesis 2 should be
accepted. Using the PC algorithm, one readily finds such a model.
E .HET ----i.~ SCOR E ---I.~SCOR
N
/ N
/MONO~
ATM .SMOR ------------~.~SMOR
TETRAD pattern at .05 TETRAD pattern at .10

significance level significance level
p = .148
Figure 12
The model on the right of figure 12 has an asymptotic X2 value of 17 with 12 degrees of
freedom, with P(X2) = .148.
The PC algorithm finds a model that cannot be rejected on the basis of the data and that
postulates a direct effect of attitude toward masturbation on orgasmic experience during
masturbation, contrary to Bentler and Peeler.
5.8.4 The American Occupational Structure
Blau and Duncan's (1967) study of the American occupational structure has been praised by the
National Academy of Sciences as an exemplary piece of social research and criticized by one
statistician (Freedman 1983a) as an abuse of science. Using a sample of 20,700 subjects, Blau
and Duncan offered a preliminary theory of the role of education (ED), first job (J 1), father's
education (FE), and father's occupation (FO) in determining one's occupation (OCC) in 1962.
They present their theory in the following graph, in which the undirected edge represents an
unexplained correlation:
Figure 13
Blau and Duncan argue that the dependencies are linear. Their salient conclusions are that
father's education affects occupation and first job only through the father's occupation and the
subject's education.
Blau and Duncan's theory was criticized by Freedman as arbitrary, unjustified, and statistically
inadequate (Freedman 1983a). Indeed, if the theory is subjected to the asymptotic x2likelihood
ratio test of the EQS (Bentler 1985) or LISREL (Joreskog and Sorbom, 1984) programs the
model is decisively rejected (p < .001), and Freedman reports it is also rejected by a bootstrap
test.
If the conventional .05 significance level is used to test for vanishing partial correlations, given
a common sense ordering of the variables by time, from Blau and Duncan's covariances the PC
algorithm produces the following graph:
(~ acc
Figure 14
In this case every collider occurs in a triangle and there are no unshielded colliders. The data
therefore do not determine the directions of the causal connections, but the time order of course
determines the direction of each edge. We emphasize that the adjacencies are produced by the
program entirely from the data, without any prior constraints. The model shown passes the
same likelihood ratio test with p > .3.
The algorithm adds to Blau and Duncan's theory a direct connection between FE and ft. The
connection between FE and ft would only disappear if the significance level used to test for
vanishing partial correlations were .0002. To determine a collection of vanishing partial
correlations that are consistent with a directed edge from FE to OCC in 1962 one would have to
reject hypotheses of vanishing partial correlations at a significance level greater than .3. The
conditional independence relations found in the data at a significance level of .0001 are faithful
to Blau and Duncan's directed graph.
Freedman argues that in the American population we should expect that the influences among
these variables differ from family to family, and therefore that the assumption that all units in
the population have the same structural coefficients is unwarranted. A similar conclusion can be
reached in another way. We noted in Chapter 3 that if a population consists of a mixture of
subpopulations of linear systems with the same causal structure but different variances and
linear coefficients, then unless the coefficients are independently distributed or the mixture is in
special proportions, the population correlations will be different from those of any of the
subpopulations, and variables independent in each subpopulation may be correlated in the
whole. When subpopulations with distinct linear structures are mixed and these special
conditions do not obtain, the directed graph found from the correlations will typically be
complete. We see that in order to fit Blau and Duncan's data we need a graph that is only one
edge short of being complete.
The same moral is if anything more vivid in another linear model built from the same empirical
study by Duncan, Featherman and Duncan (1972). They developed the following model of
socioeconomic background and occupational achievement, where FE signifies father's
education, FO father's occupational status, SIB the number of the respondent's siblings, ED the
respondent's education, OCC the respondent's occupational status and INC the respondent's
income.
In this case the double headed arrows merely indicate a residual correlation. The model has four
degrees of freedom, and entirely fails the EQS likelihood ratio test (X 2 is 165). When the
correlation matrix is given to the TETRAD IT program along with an obvious time ordering of
the variables, the PC algorithm produces a complete graph.
FE ---"'''~INC
ace
!
SIB
Figure 15
5.8.5 The ALARM Network
Recall the ALARM network developed to simulate causal relations in emergency medicine:
Figure 16
The SGS and PC* algorithms will not run on a problem this large. We have applied the PC
algorithm to a linear version of the ALARM network. Using the same directed graph, linear
coefficients with values between .1 and .9 were randomly assigned to each directed edge in the
graph. Using a joint normal distribution on the variables of zero indegree, three sets of
simulated data were generated, each with a sample size of 2,000. The covariance matrix and
sample size were given to a version of the TETRAD II program with an implementation of the
PC-1 algorithm. This implementation takes as input a covariance matrix, and it outputs a
pattern. No information about the orientation of the variables was given to the program. Run on
a Decstation 3100, for each data set the program required less than fifteen seconds to return a
pattern. In each trial the output pattern omitted two edges in the ALARM network; in one of the
cases it also added one edge that was not present in the ALARM network.
In a related test, another ten samples were generated, each with 10,000 units. The results were
scored as follows: We call the pattern the PC algorithm would generate given the population
correlations the true pattern. We call the pattern the algorithm infers from the sample data the
output pattern. An edge existence error of commission (Co) occurs when any pair of
variables are adjacent in the output pattern but not in the true pattern. If an edge e between A and
B occurs in both the true and output patterns, there is an edge direction error of
commission when e has an arrowhead at A in the output pattern but not in the true pattern,
(and similarly for B.) Errors of omission (Om) are defined analogously in each case. The
results are tabulated as the average over the trial distributions of the ratio of the number of actual
errors to the number of possible errors of each kind. The results at sample size 10,000 are
summarized below:
#trials %Edge Existence Errors %Edge Direction Errors

Commission Omission Commission Omission
10 .06 4.1 17 3.5
For similar data from a similarly connected graph with 100 variables, for ten trials the PC-l
algorithm required an average of 134 seconds and the PC-3 algorithm required an average of 16
seconds.
Herskovits and Cooper (1990) generated discrete data for the ALARM network, using variables
with two, three and four values. Given their data, the TETRAD II program with the PC
algorithm reconstructs almost all of the undirected graph (it omitted two edges in one trial; and
in another also added one edge) and orients most edges correctly. In most orientation errors an
edge was oriented in both directions. Broken down by the same measures as were used for the
linear data from the same network, (with simulated data obtained from Herskovits and Cooper
at sample size 10,000) the results are:
trial %Edge Existence Errors %Edge Direction Errors

Commission Omission Commission Omission
1 0 4.3 27.1 10.0
2 0.2 4.3 5.0 10.4
5.S.6 Virginity
A retrospective study by Reiss, Banwart and Foreman (1975) considered the relationship
among a sample of undergraduate females between a number of attitudes, including attitude
toward premarital intercourse, use of a university contraceptive clinic, and virginity. Two
samples were obtained, one of women who had used the clinic and one of women who had not;
the samples did not differ significantly in relevant background variables such as age, education,
parental education, and so on. Fienberg gives the cross-classified data for three variables:
Attitude toward extramarital intercourse (E) (always wrong; not always wrong); virginity (V)
and use of the contraceptive clinic (C) (used; not used). All variables are binary. The PC and
SGS procedures immediately produces the following pattern:
E --V--C
Figure 17
which is consistent with any of the orientations of the edges that do not produce a collision at V.
One sensible interpretation is that attitude affects sexual behavior which causes clinic use.
Fienberg (1977) obtains the same result with log linear methods.
5.S.7 The Leading Crowd
Coleman (1964) describes a study in which 3398 schoolboys were interviewed twice. At each
interview each subject was asked to judge whether or not he was a member of the "leading
crowd" and whether his attitude toward the leading crowd was favorable or unfavorable. The
data have been reanalyzed by Goodman (1973a, b) and by Fienberg (1977). Using Fienberg's
notation, let A and B stand for the questions at the first interview and C and D stand for the
corresponding questions at the second interview. The data are given by Fienberg as follows:
Second Interview
MembershiD Attitude + + - -
+ - + -
Membershio Attitude
+ + 458 140 110 49
First + - 171 182 56 87
Interview - + 184 75 531 281
- - 85 97 338 554
Fienberg summarizes his conclusions after a log-linear analysis in the path diagram in figure 18.
He does not explain what interpretation is to be given to the double-headed arrow:
Figure 18
When the PC algorithm is told that C and D occur after A and B, with the usual.05 significance
level for tests the program produces the pattern in figure 19:
A .C
B
/! .D
Figure 19
Orienting the undirected edge in the PC model as a directed edge from A to B produces expected
values for the various cell counts that are almost identical with Fienberg's (p. 127) expected
counts'? Note, however, this is a nearly complete graph, which may indicate that the sample is
a mixture of different causal structures.
5.S.S Influences on College Plans
Sewell and Shah (1968) studied five variables from a sample of 10,318 Wisconsin high school
seniors. The variables and their values are:
SEX [male =0, female = 1]

IQ = Intelligence Quotient, [lowest =0, highest =2]
CP = college plans [yes = 0, no = 1]
PE = parental encouragement [low = 0, high = I]
SES =socioeconomic status [lowest = 0, highest = 3]
They offer the following causal hypothesis:
SES~
SEX • PE -----+ CP'
~/
Figure 20
The data were reanalyzed by Fienberg (1977), who attempted to give a causal interpretation
using log-linear models, but found a model that could not be given a graphical interpretation.
Given prior information that orders the variables by time as follows
SEX
2 IQ PE SES
3 CP
71be small differences are presumably attributable to round-off errors.

so that later variables cannot be specified to be causes of earlier variables, the output with the
PC algorithm is the structure:
SES~
-~.~ PE ----.. CP
K2~
Figure 21
The program cannot orient the edge between IQ and SES. It seems very unIik:ely that the child's
intelligence causes the family socioeconomic status, and the only sensible interpretation is that
SES causes IQ, or they have a common unmeasured cause. Choosing the former, we have a
directed graph whose joint distribution can be estimated directly from the sample. We find, for
example, that the maximum likelihood estimate of the probability that males have college plans
(CP) is .35, while the probability for females is .31. Judged by this sample the probability a
child with low IQ, no parental encouragement (PE) and low socioeconomic status (SES) plans
to go to college is .011; more distressing, the probability that a child otherwise in the same
conditions but with a high IQ plans to go to college is only .124.
5.S.9 Abortion Opinions
Christensen (1990) illustrates log-linear model selection and search procedures with a data set
whose variables are Race (R) [white, non-white], Sex (S), Age (A) [six categories] and
Opinion (0) on legalized abortion (supports, opposes, undecided). Forward selection
procedures require fitting 43 log-linear models. A backwards elimination method requires 22
fits; a method due to Aitkin requires 6 fits; another backwards method due to Wermuth requires
23 fits. None of these methods would work at all on large variable sets. Christensen suggests
that the "best" log-linear model is an undirected conditional independence graphical model
whose maximal cliques are [RSO] and [OA]. This is shown in figure 22.
R 0
J/A
Figure 22
Subsequently, Christensen proposes a recursive causal model (in the terminology of Kiiveri and
Speed, 1982) for the data. He suggests on substantive grounds a mixed graph
R .0
I/i
S A
Figure 23
and says "The undirected edge between R and S ... represents an ,interaction between R and S."
Figure 23 is not a causal model in the sense we have described. It can be interpreted as a pattern
representing the equivalence class of causal graphs whose members are the two orientations of
the R - S edge, but R and S in Christensen's data are very nearly independent.
This example is small enough to use the PC* algorithm, which with significance level .05 for
independence tests gives exactly figure 24. Assuming faithfulness, the statistical hypothesis of
figure 24 is inconsistent with the independence of {R,S} and A conditional on 0, required by
the log-linear model of figure 22.
R .0
S
/i A
Figure 24
At a slightly lower significance level (.01) R and 0 are judged independent, and the same
algorithm omits the R -> 0 connection. On this data with significance level .05 the PC
algorithm also produces the graph of figure 24 but with the R -> 0 connection omitted. The
difference in the outputs of the PC* and PC algorithms occur in the following way. Both
algorithms produce at an intermediate stage the undirected graph underlying figure 24. In that
undirected graph A does not lie on any undirected path between Rand O. For that reason, the
PC* algorithm never tests the conditional independence of Rand 0 on A, and leaves the R - 0
edge in. In contrast, the PC algorithms does test the conditional independence of Rand 0 on A,
with a positive result, and removes the R - 0 edge.
5.8.10 Simulation Tests with Random Graphs
In order to test the speed and the reliability of the algorithms discussed in this chapter, we have
tested the algorithms SOS, PC-I, PC-2, PC-3, and IG on a large number of simulated
examples. The graphs themselves, the linear parameters, and the samples were all pseudo-
randomly generated. This section describes the sample generation procedures for both linear
and discrete data and gives simulation results for the linear case. Simulation results with discrete
data are considered in the chapter on regression.
The average degree of the vertices in the graphs considered are 2, 3, 4, or 5; the number of
variables is 10 or 50; and the sample sizes are 100, 200, 500, 1000, 2000, and 5000. For each
combination of these parameters, 10 graphs were generated, and a single distribution obtained
faithful to each graph, and a single sample taken from each such distribution.
Because of its computational limitations, the SOS algorithm was tested only with graphs of 10
variables.
5.8.10.1 Sample Generation
All pseudo-random numbers were generated by the UNIX "random" utility. Each sample is
generated in three stages:
(i) The graph is pseudo-randomly generated.
(ii) The linear coefficients (in the linear case) or the conditional probabilities (in the
discrete case) are pseudo-randomly generated.
(iii) A sample for the model is pseudo-randomly generated.
Discovery Algoritluns for Causally Sufficient Structures 153
We will discuss each of these steps in more detail.
(i) The input to the random graph generator is an average degree and the number of variables.
The variables are ordered so that an edge can only go from a variable lower in the order to a
variable higher in the order, eliminating the possibility of cycles. Since some of the procedures
use a lexicographic ordering, variable names were then randomly scrambled so that no
systematic lexicographic relations obtained among variable pairs connected by edges. Each
variable pair is assigned a probability p equal to
average degree
number of variables - 1
For each variable pair a number is drawn from a uniform distribution over the interval 0 to 1.
The edge is placed in the graph if and only if the number drawn is less than or equal to p. 8
(ii) For simulated continuous distributions, an "error" variable was introduced for each
endogenous variable and values for the linear coefficients between .1 and .9 were generated
randomly for each edge in the graph. For the discrete case, a range of values of variables is
selected by hand, and for each variable taking n values, the unit interval is divided into n sub-
intervals by random choice of cut-off points. A distribution (e.g., uniform) is then imposed on
the unit interval.
(iii) In the discrete case for each such distribution produced, each sample unit is obtained by
generating, for each exogenous variable, a random number between 0 and 1.0 according to the
distribution and assigning the variable value according to the category into which the number
falls. Values for the endogenous variables were obtained by choosing a value randomly with
probability given by the conditional probabilities on the obtained values of the parents of the
variable. In the linear case, the exogenous variables--including the error terms--were generated
independently from a standard normal distribution, and values of endogenous variables were
computed as linear functions of their parents.
8We do not know whether this method of graph generation produces "reaIistic" graphs. One feature of some of
the graphs generated in this fashion that may not be desirable is the existence of isolated variables. An informal
examination showed topologies not unlike the A1arm network.
5.S.10.2 Results
As before, reliability has several dimensions. A procedure may err by omitting undirected edges
in the true graph or by including edges--directed or undirected--between vertices that are not
adjacent in the true graph. For an edge that is not in the true graph, there is no fact of the matter
about its orientation, but for edges that are in the true graph, a procedure may err by omitting an
arrowhead in the true graph or by including an arrowhead not in the true graph. We count errors
in the same way as in section 5.8.5.
Each of the procedures was run using a significance level of .05 on all trials. The five
procedures tested are not equally reliable or equally fast. The SGS algorithm is much the
slowest, but in several respects it proves reliable. The graphs on the following pages show the
results. Each point on the graph is a number, which represents the average degree of the
vertices in the directed graphs generating the data. We plot the run times and reliabilities of the
PC-l PC-2, PC-3, IG, and SGS algorithms against sample size for data from linear models
based on randomly generated graphs with 10 variables, and similarly the reliabilities of the first
four of these algorithms for linear models based on randomly generated graphs with 50
variables. In each case the results are plotted separately for graphs of degree 2, 3, 4, and 5.
;
s
&I .\
J
::l 31,\
j
:;: 3
!.,.
~
~\'"
t\ .~:
!.,. - ~\.4 _ :---- 5
!il \3 . . . . . --..............
'. . . . ~::=::::::: J
""
II •
. OCO 2000 3000
~s.z..(50~IitIW)prc.. ,
""'" SOOO ,OCO 2000 3000
""'"
SItnJMe &ze (500 ....fl lILbIH ~ ,...... pc.,
SOOO
Ii!
, _ 5- 5- - - '
&I ~ ...... - ..
J
&I
' 3,","_. - - S
i< ~ ~2- ' ,","_ , __ , _________ • ~
i i •
".,. !il • .,.
" !il
\~
•....1=--.
,OCO
SMnplt
2000 3000
Sin {SO vaII."... . nut)

""'"
pe,
SOOO ' OCO
Soam plt
2000 3000
Stu (SO vanabI • • 1onNJ1 PC 1

""'" SOOO
;
&I •~S
i<
::l
) .\,\
,~:~:----.
.3_
~
:t
!.,.
5
.~:
!...
;;
:: ~
II •
!il \3 .......... --............. •
"""' 2 _ ~
. OCO 2000 3000

""'" 5000 ' OCO 2000 3000
""'" SOOO
s.nc- Silt (SO 14~ 1inMr) PC-2 ~ Sa. (50 ....nlilMt. ~""'l PC 2
Ii! ~
, _ S_S - --5
..... -
.. '-3 __ 3
.--""""""'!l Ii!
Ii!
i<
i
~
\-·'"'"-2--· _________ • I ~
j
".,. !il 2 .,. !il
\\ tE::::I==:.-.. .
, OCO 2000 3000 .... SOOO ' 000 2000 3000 .... SOOO
Saonple Sin (50 ......... .,.." PC-2 Swnp6e Size (SO ...n.oln M,.." PC-2
V" ~. J
II
.•
\
.i
,. _s----s
0
,~,\
S
Y ~\S"
~
! .- -s
N
~.
0
!
"~=:=:::;= : \',,:---.
~
~
~ !:
0
• ......... ::::::::: l
0
""" 1000 2000 <000 .... ,000 >000

""" """ ....
Sample Sll.ISO ....nlllle• .,..,., PC:) SarnI)58 SIol. (s.o .... tl abl •• ,,...,) PC3
51
fS-'---'
3. . ;-4 _____ .
iI
Ii!
.i I
i
~
~
i
!:
2t
'.-.----.
_t ........... 3 - 3 3
J
~
~
!: ,~
1_'_1
1000 2000
""" <000 .... '000 2000
""" """ !OOO
~ SI,. j50 ....1'1..,.•• 'ne.) PCl SamPle s.)'. (SO vaflatiIH ~nHf) PC-:)
~ •~
Ii!
J :::
;
~
i
s ~
,\.",
.~:
J !:
!
'" ;;
0
II - I
'" ---..
. ,----
2\4- .. ---- 5
\'"
........... Z_~
0
1000 2000
"""
SWr'Ipa. SI1. (SO wan""'., .......11G
<000
""" 1000 2000
s.m.pIe SI,. (SO <N1abret.

"""I.,..,.,,)Q<000 5000
II
\~I~:~
Ii!
1"-'-'
i!
j J
~ s
J
~
i ,'::::...1-='1
"
~ !:
-".-'- -. ~ !l
-=!
'000 2000
~Sll.(50~I~)1G
""" """ 5000 '000 2000
"""
Swnplt Sltt (SO wwiabIH IinNf) IG
<000
"""
II
,
.'
J
i!
j . 34 ~"S-S---' _ _ _ _
\_,_i:~,
S
.
'l "
2~';_4_____
! ! \.........
~ N ~ Iii 2. . . . 2_, 3_:)_
-,,
!a~.--=I~·
u"'" 3000
""'" ....
J
'000
,... 3000 5000
SAm"" SIZ.(10YallabCH~I'INt'PC" ""'"

2000
s.m~Sl"npO ..." . . . Ii,...,) pe..'
§
lil s ~S s iI
j /~ 1
' /-:7:~:
0 i!
!
'l
i i,
t2-2
~
"
~ Iii ~
~;\!~;::::>!
• Iii
,... ,...
'000 3000
SMl:pleSize( 1 0~ linNtl pe.,

""'" 5000 ' 000
..
~s.t "O ~ ~ """pc-,
3000
""'" 5000
iI
,
i! .'
J. I 3.~",5-5--_' _ _ _ _
j ~ 2~,
4
. _.
\3 . . . . _____ .
! !
2_,
~ ~
2 ........... 3 -
t:;!/~:;;:::~
N
Iii 3_
'000
,... 3000
""'" 5000 '000 2000 3000 <COO 5000
SM\pIe Siz. (10 ....,..,.... II.,.., PC2 SAmplt Slz.. (t O~ ~nNtl PC2
!!
i
<:
i!
l~i~:
~ .~
: I
lil
i!
!
'l
i '.
~{:=:7
I
3
~~~;~!=
'l
"
~ Iii
2
~
Iii
'000
s.m~
2000 3000
Siz. tl O~ inMr) prc..Z

""'" 5000 '000 2000 3000
~sa.(10""';--'InNt,PC-2
""'" 5000
.AI
i!
II •
j . 1
~
~\
3•• ,5~5_ S _ _ _ _ _ _ _ _
; 'i 2,3,4
2 . __ .. _____ 5
! !
~ N
' '2
~;~j~~
1
,
~
Ii: \ ........ ,---;-
' ........ ' - ,
"
...,., ...,.,
1000
Sample
""'" 3DIlO
Sil. (1 0 'l;tlriitlbl" kIMfl PC-3

5000 '000 2000 JOOO
S&moIe $lIt [1 0 'qII.tJI... ~ nnr l pc.;,
5000
1"; _ __ •
'",,- .........---' S
S! S, "-..... _ _ _ 41
i!
j .5/ i
~ s
t;/,/2~3
i
'i '".
~
i
~
'i ij~ __ , _
'
~ Ii: ~
2 Ii:
5~i==!
'000 2000 3000 ...,., 5000 1000 2000 3DIlO ...,., 5000
Sample 511.1'0 Yaliabl'f$1Ifttar) PC-3 ~..". SIze [10 \lltinabl~ ~I'MW' P<>3
il ,
I'~\.
--.
s
j . i
~ " '____
~
',3,.-.
5 __
-,- -,
!
~
~
~ N
~ Ii:
\ 3_.
2 3 -.... .. ..
' ........ ,- , ,
'-::::' • ..:= I~!
•000 2000 3DIlO 1000 5000 .. '000 2000 JOOO .000 5000
s.mpr. SIt. (10 vlNble:l 1.nNI) SGS SIiInCHS&nC'Ov..n.tH.tll"lNt) SCS
~ 5
J ~
!l il
\~~~;-
j ~
Il
i -,
- '=========:.
~
I;;-l='~:
'........ 5
~
~ ~
~ Ii: ~
~ ~,_, _ _ _ _ 2 _ _ _ _ _ _ _ _ Ii:
, 2
'000 2000 JDOO 1000 5000 1000 2000 3DIlO 1000 5000
~SIz.(10~IinMl1SGS ~. SQ.fl0W~ 1nMt) SGS

iI
,
'--
iii
,',
l 1'\
J.
v-:~:~.
3,4 ,5 -
05 4__ S _ S __________
2. . .2_.. .
~3
.,\ 3__
; 'I
!
~ N
!
~
Ii: ' - 3
4
~: = ~:: ;,~~ a
' 000 2000 3000
""'" 5000 . 000 2000 3000
""" ""'"
s.m"", Slr. 4' O ,"","~ iIoMI) IG ~ Slzt p OVWI.ab1 H hn• • , IG
!!
3
l-l?- ----------,.
i! iI
~~.~,~
l
I 'I '", ~: S iii
i i
_2_:.--2- ---=: " '"
, 3
"
~ Ii: ~"""'_3_ ~
Ii: 3~ ~
I
.000 2000 3000 5000 H>OO 2000 3000 ...,.
Sampll Slzt [ 10wriatHllinea1 kl
""" s.m ~ St.t. [10 vanllblH .....;t.tl to
SOOO
The following qualitative conclusions can be drawn.
The rates of arrow and edge omission decrease dramatically with sample size up to about
sample size 1000; after that the decreases are much more gradual.
The rates of arrow and edge commission vary much less dramatically with sample size than do
the rates of arrow and edge omission.
As the average degree of the variables increases, the average error rates increase in a very
roughly linear fashion, but the PC-2 algorithm tends to be less reliable than the other algorithms
with respect to edge omissions when the average degree of the graph is high.
The PC-I, PC-3, 10, and SOS algorithms have compensating virtues and disadvantages. None
of the procedures are reliable on all dimensions when the graphs are not sparse. One reliable
dimension is the addition of edges: If two vertices are not adjacent in the true graph, there is
very little chance they will be mistakenly output by any of these four procedures, no matter
what the average degree of the graph and no matter what the sample size.
In contrast, at high average degree and low sample sizes the output of each of the procedures
tends to omit over 50% of the edges in the true graph. At large sample sizes and low average
degree only a few percent of the true edges are omitted, but with high average degree the
percentage of edges omitted even at large sample sizes is significant. For example, at sample
size 5000 and average degree 5, PC-l omits over 30% of the edges in the true graph.
Arrow commission errors are much more common than edge commission errors. If an arrow
does not occur in a graph, there is a considerable probability for any of the procedures that the
arrow will be output, unless the sample size is large and the true graph is of low degree. For 10
variables with average degree about 2 and sample sizes of 1,000 or more, the SOS and 10
algorithms are quite reliable, with errors of commission for arrows around 6%. Under the same
conditions the error rates for the PC-l and PC-2 algorithms run about 20%. In the case of the
SOS algorithm, these relations are reversed for the question of arrow omission--if an arrow
occurs in the true graph, what is the chance that the procedure will fail to include the arrow in its
output? The answer is about 8% for PC-l and PC-2 and about 20% for the SOS procedure.
The 10 algorithm, while much less reliable for arrow omissions at low sample sizes, is only
slightly more unreliable at high sample sizes.
The return time of the PC-3 algorithm is dramatically smaller than the other algorithms. It's run
time also does not increase as sharply with average degree, but the procedure does produce
many more edge commission errors as the average degree increases.
The results suggest that the programs can reasonably be used in various ways according to the
size of the problem, the questions one wants answered, and the character of the output.
Roughly the same conclusions about reliability can be expected for the discrete case, but with
lower absolute reliabilities. For larger number of variables the same patterns should hold, save
that the SOS algorithm cannot be run at all.
More research needs to be done on local "repairs" to the graphs generated by these procedures,
especially for edge omission errors and arrow commission errors. In order for the method to
converge to the correct decisions with probability 1, the significance level used in making
decisions should decrease as the sample sizes increase, and the use of higher significance levels
(e.g., .2 at samples sizes less than 100, and .1 at sample sizes between 100 and 300) may
improve performance at small sample sizes.
5.9 Conclusion
This chapter describes several algorithms that can reliably recover sparse causal structures even
for quite large numbers of variables, and illustrates their application. The algorithms have each
been implemented using tests for conditional independence in the discrete case and for
vanishing partial correlations in the linear case. We make no claim that these uses of tests to
decide relevant probability relations is optimal, but any improvements in the statistical decision
methods can be prefixed to the algorithms. With the exception of the PC* and SOS algorithms,
the procedures described are feasible for large numbers of variables so long as the true causal
graphs are sparse.
The algorithms we have described scarcely exhaust the possibilities, and a number of very
simple alternative procedures should work reasonably well, at least for finding adjacency
relations in the causal graph.
The idea of discovery problems is already contained in the notion of an estimation problem, and
the requirement that an estimator be consistent is essentially a demand that it solve a particular
kind of discovery problem. An extension of the idea to general non-statistical settings was
proposed by Putnam (1965) and independently by Gold (1965) and has subsequently been
extensively developed in the literature of computer science, mathematical linguistics and logic
(Osherson, Stob and Weinstein, 1986).
A more or less systematic search procedure for causal/statistical hypotheses can be found in the
writings of Spearman (1904) and his students early in this century. A Bayesian version of
stepwise search was proposed by Harold Jeffreys (1957). Thurstone's (1935) factor analysis
inaugurated a form of algorithmic search separated from any precise discovery problem:
Thurstone did not view factor analysis as anything more than a device for finding
simplifications of the data, and a similar view has been expressed in many subsequent
proposals for statistical search. The vast statistical literature on search has focused almost
exclusively on optimizing fitting functions.
The SGS algorithm was proposed by Glymour and Spirtes in 1989, and appeared in Spirtes,
Giymour and Scheines (1990c). Verma and Pearl (1990b) subsequently proposed a more
efficient version that examines cliques. A version of the PC algorithm was developed by Spirtes
and Glymour (1990). The version presented here contains an improvement suggested by Pearl
and Verma in the efficiency of step C) of the algorithm. Bayesian discovery procedures have
been studied in Herskovits' thesis (1992).
The maximum likelihood estimation procedure for "recursive causal models" was developed in
Kiiveri's (1982) doctoral thesis. The mathematical properties of the structures are further
described in Kiiveri, Speed and Carlin (1984).
Chapter 6
Discovery Algorithms without Causal Sufficiency
6.1 Introduction
The preceding chapter complied with a common statistical fantasy, namely that in typical data
sets it is known that no part of the statistical dependencies among measured variables are due to
unmeasured common causes. We almost always fail to measure all of the causes of variables we
do measure, and we often fail to measure variables that are causes of two or more measured
variables. Any examination of collections of social science data gives the striking impression
that variables in one study often seem relevant to those in other studies. Record keeping
practices sometimes force econometricians to ignore variables in studies of one economy
thought to have a causal role in studies of other economies (Klein, 1961). In many studies in
psychometrics, social psychology and econometrics, the real variables of interest are
unmeasured or measured only by proxies or "indicators." In epidemiological studies that claim
to show that exposure to a risk factor causes disease, a burden of the argument is to show that
the statistical association is not due to some common cause of risk factor and disease; since not
everything imaginably relevant can be measured, the argument is radically incomplete unless a
case can be made that unmeasured variables do not "confound" the association. If, as we
believe, no reliable empirical study can proceed without considering whether relevant variables
are unmeasured, then few published uncontrolled empirical studies are reliable.
In both experimental and non-experimental studies the unrecognized presence of unmeasured

variables can lead to erroneous conclusions about the causal relations among the variables that
are measured, and to erroneous predictions of the effects of policies that manipulate some of
these variables. Until reliable, data-based methods are used to identify the presence or absence
of unmeasured common causes, most causal inferences from observational data can be no more
than guesswork at best and pseudo-science at worst. Are such methods possible? That question
surely ought to be among the most important theoretical issues in statistics.
Statistical methods for detecting unmeasured common causes, or "confounding" in the

terminology epidemiologists prefer, has been chiefly developed in psychometrics, where criteria
for the existence and numbers of common causes have been sought since the tum of the century
for special statistical models. The results include a literature on linear systems that contain
criteria (e.g., Charles Spearman's (1904) vanishing tetrad differences) for latent variables that
proved, however, to be neither necessary nor sufficient even assuming linearity. Criteria for
two latent common causes were introduced by Kelley (1928), and related criteria are used in
factor analysis, but they are not correct unless it is assumed that all statistical dependencies are
due to unmeasured common causes. For problems in which the measured variables are discrete
and their values a stochastic function of an unobserved continuous vector parameter 8, a
number of criteria have been developed for the dimensionality of 8 (Holland and Rosenbaum,
1986). Suppes and Zanotti (1981) showed that for discrete variables there always exists a
formal latent variable model in which all measured variables are effects of an unmeasured
common cause and all pairs of measured variables are independent conditional on the latent
variable. Their argument assumes the model must satisfy only the Markov Condition; the result
does not hold if it is required that the distributions be faithful.
Among epidemiologists, (Breslow and Day, 1980; Kleinbaum, Kupper and Morgenstern,
1982) the criteria introduced by the Surgeon General's report on Smoking and Health (1964)
are sometimes still advocated as a means for deciding whether a statistical dependency between
exposure to risk factor A and disease B is "causal," apparently meaning that A causes B and A
and B have no common causes. The criteria include (i) increase in response with dosage; (ii)
that the statistical dependency between a risk factor and disease be specific to particular disease
subgroups and to particular conditions of risk exposure; (iii) that the statistical association be
strong; (iv) that exposure to a risk factor precede the period of increased risk; (v) lack of
alternative explanations.
Even in causally sufficient systems, where all common causes of measured variables are
themselves measured, such criteria do not separate causes from correlated variables. They fail
even to come to grips with the problem of unmeasured "confounders." Criterion (v) is an
evasion; the problem in uncontrolled studies is exactly that there are too many alternative
explanations of the data. Criterion (iv) is a banality which is of no use at all in deciding whether
there are measured or unmeasured common causes at work. Criterion (iii) is defended on the
grounds that "If an observed association is not causal, but simply the reflection of a causal
association between some other factor and disease, then this latter factor must be more strongly
related to disease (in terms of relative risk) than is the former factor," (Breslow and Day, 1980).
Discovery Algorithms without Causal Sufficiency 165
But the inference is incorrect: if there are two or more common causes, measured or not, none
of them need be more strongly related to the disease than is the putative measured cause; and if
A causes B and A and B also have a common cause, the latter need not be more strongly
associated with B than is A. On behalf of Breslow and Day one might appeal to simplicity
against all hypotheses of multiple common causes, but that would be an implausible claim in
medical science, where multiple causal mechanisms abound. Nothing about the first two criteria
separates the situation in which A and B have common causes from circumstances in which
they do not.
In this chapter we present a more or less systematic account of how the presence of unmeasured
common causes can mislead an investigator about causal relationships among measured
variables, and of how the presence of unmeasured common causes can be detected. We deal
with these questions separately for the general case and for the case in which all structures are
linear. But the central aim of this chapter is to show how, assuming the Markov and
Faithfulness conditions, in principle reliable causal inferences can be made from appropriate
sample data without any prior knowledge as to whether the system of measured variables is
causally sufficient.
6.2 The PC Algorithm and Latent Variables
A natural idea is that a slight modification of the PC algorithm will give correct information
about causal structure even when unmeasured variables may be present. Suppose that P' is a
distribution over V that is faithful to a causal graph, and P is the marginal of P' over 0,
properly included in V. We will refer to the members of 0 as measured or observed variables.
As we have already seen, if there are unmeasured common causes, the output of the PC
algorithm can include bi-directed edges of the form A <-> B. We could interpret a bi-directed
edge between A and B to mean that there is an unmeasured cause C that directly causes A and B
relative to O. We modify the algorithm by using a "0" on the end of an arrow to indicate that it
is not known whether an arrowhead should occur in that place. We use a "*" as a metasymbol
to stand for any of the three kinds of endmarks that an arrow can have: EM (empty mark), ">",
or "0".
Modified PC Algorithm:
A.) Fonn the complete undirected graph C on the vertex set V.
B.)
n =0.
repeat
repeat
select an ordered pair of variables X and Y that are adjacent in C such
that Adjacencies(C,x)\{Y} has cardinality greater than or equal to
n, and a subset S of Adjacencies( C ,x)\{ Y} of cardinality n, and if
X and Y are d-separated given S delete edge X - Y from C and
record S in Sepset(X,Y) and Sepset(Y,x)
until all ordered pairs of adjacent variables X and Y such that
Adjacencies(C,x)\{Y} has cardinality greater than or equal to n and all
subsets of Adjacencies(C,x)\{y} of cardinality n have been tested for d-
separation.
n=n+ 1.
until for each ordered pair of adjacent vertices X, Y, Adjacencies(C,X")\{Y} is
of cardinality less than n.
C.) Let F be the graph resulting from step B). If X and Y are adjacent in F, orient the
edge between X and Y as X 0-0 Y.
D.) For each triple of vertices X, Y, Z such that the pair X, Yand the pair Y, Z are each
adjacent in F but the pair X, Z are not adjacent in F, orient X *-* Y *-* Z as X *-> Y <-*
Z if and only if Y is not in Sepset(X,z).
E.) repeat
If A *-> B, B *-* C, A and B are not adjacent, and there is no arrowhead at Bon
B *-* C, then orient B *-* Cas B -> C.
the edge as A *-> B.
(When we say orient X *-* Y as X *-> Y we mean leave the same mark on the X end of the
edge and put an arrowhead at the Yend of the edge.)
The result of this modification applied to the examples of the previous chapter is perfectly
sensible. For example, in figure 1 we show both the model obtained from the Rodgers and
Maranto data at significance level.l by the PC algorithm and the model that would be obtained
by the modified PC algorithm from a distribution faithful to the the graph in the PC output. (In
each case with the known time order of the variables imposed as a constraint.)
ABILITY
CITES CITES
PC Output Modified PC Output
Figure 1
The output of the Modified PC Algorithm indicates that GPQ and ABILITY, for example, may
be connected by an unmeasured common cause, but that PUBS is a direct cause of CITES,
unconfounded by a common cause. Where a single vertex has "0" symbols for two or more
edges connecting it with vertices that are not adjacent, a special restriction applies. ABILITY,
for example has an edge to GPQ and to PREPROD, each with an "0" at the ABILITY end, and
GPQ and PREPROD are not adjacent to one another. In that case the two "0" symbols cannot
both be arrowheads. There cannot be an unmeasured cause of ABILITY and GPQ and an
unmeasured cause of ABILITY and PREPROD, because if there were, GPQ and PREPROD
would be dependent conditional on ABILITY, and the modified pattern entails instead that they
are independent.
In many cases--perhaps most practical cases--in which the sampled distribution is the marginal
of a distribution faithful to a graph with unmeasured variables, this simple modification of the
PC algorithm gives a correct answer if the required statistical decisions are correctly made.
6.3 Mistakes
Unfortunately, this straightforward modification of the PC algorithm is not correct in general.

An imaginary example will show why.
Everyone is familiar with a simple mistake occasioned by failing to recognize an unmeasured

common cause of two variables, X, Y, where X is known to precede Y. The mistake is to think
that X causes Y, and so to predict that a manipulation of X will change the distribution of Y.
But there are more interesting cases that are seldom noticed, cases in which omitting a common
cause of X and Y might lead one to think, erroneously, that some third variable Z directly
causes Y. Consider an imaginary case:
A chemist has the following problem. According to received theory, which he very
much doubts, chemicals A and B combine in a low yield mechanism to form chemical D
through an intermediate C. Our chemist thinks there is another mechanism in which A
and B combine to form D without the intermediate C. He wishes to do an experiment to
establish the existence of the alternative mechanism. He can readily obtain reagents A
and B, but available samples may be contaminated with varying quantities of D and
other impurities. He can measure the concentration of the unstable alleged intermediate
C photometrically, and he can measure the equilibrium concentration of D by standard
methods. He can manipulate the concentrations of A and B, but he has no means to
manipulate the concentration of C.
The chemist decides on the following experimental design. For each of ten different
values of the concentration of A and B, a hundred trials will be run in which the
reagents are mixed, the concentration of C is monitored, and the equilibrium
concentration of D is measured. Then the chemist will calculate the partial correlation of
A with D conditional on C, and likewise the partial correlation of B with D conditional
on C. If there is an alternative mechanism by which A and B produce D without C, the
chemist reasons, then there should be a positive correlation of A with D and of B with D
in the samples in which the concentration of C is all the same; and if there is no such
alternative mechanism, then when the concentration of C is controlled for, the
concentrations of A, B on the one hand, and D on the other, should have zero
correlation.
The chemist finds that the equilibrium concentrations of A, B on the one hand and of D
on the other hand are correlated when C is controlled for--as they should be if A and B
react to produce D directly--and he announces that he has established an alternative
mechanism.
Alas within the year his theory is disproved. Using the same reagents, another chemist
performs a similar experiment in which, however, a masking agent reacts with the
intermediate C preventing it from producing D. The second chemist finds no correlation
in his experiment between the concentrations of A and B and the concentration of D .
What went wrong with the fust chemist's procedure?
By substituting a statistical control for the manipulation of C the chemist has run afoul of the
fact that the marginal probability distribution with unmeasured variables can give the appearance
of a spurious direct connection between two variables. The chemist's picture of the mechanism
is given in graph Gl, and that is one way in which the observed results can be produced.
Unfortunately, they can also be produced by the mechanism in graph G2, which is what
happened in the chemist's case: impurities (F) in the reagents are causes of both C and D:
Figure 2
The general point is that a theoretical variable F acting on two measured variables C and D can
produce statistical dependencies that suggest causal relations between A and D and between B
and D that do not exist. For faithful distributions, if we use the SGS or PC algorithms, a
structure such as G2 will produce a directed edge from A to D in the output
We can see the same point more analytically as follows: In a directed acyclic graph G over a set
of variables V, if A and D are adjacent in G, then A and D are not d-separated given any subset
of v\{A,D}. Hence under the assumption of causal sufficiency, either A is a direct cause of D
or D is a direct cause of A relative to V if and only if A and D are independent conditional on no
subset of v\{A,D}. However, if 0 is not causally sufficient, it is not the case that if A and D
are independent conditional on every subset of O\{A,D} that either A is a direct cause of D
relative to 0, or D is a direct cause of A relative to 0, or there is some latent variable F that is a
common cause of both A and D.
This is illustrated by G2 in figure 2, where V = {A,B,C,D,F} and 0 = {A,C,D}. 0 is not

causally sufficient because F is a cause of both C and D which are in 0, but F itself is not in O.
A and D are not d-separated given any subset of O\{A,D}' so in any marginal of a distribution
faithful to G, A and D are not independent conditional on any subset of O\{A,D}, and the
modified PC algorithm would leave an edge between A and D. Yet A is not a direct cause of D
relative to 0, D is not a direct cause of A relative to 0, and there is no latent common cause of
A and D. The directed acyclic graph Gl shown in figure 2, in which A is a direct cause of D,
and in which there is a path from A to D that does not go through C, has the same set of d-
separation relations over {A,C,D} as does graph G2. Hence, given faithful distributions, they
cannot be distinguished by their conditional independence relations alone.
A further fundamental problem with the simple modification of the PC algorithm described
above is that if we allow bi-directed edges in the graphs constructed by the PC algorithm, it is
no longer the case that if A and B are d-separated given some subset of 0, then they are d-
separated given a subset of Adjacencies(A) or Adjacencies(B). Consider the graph in figure
3, where Tl and T2 are assumed to be unmeasured.
Figure 3
Among the measured variables, Parents(A) = {D} and Parents(E) = {B}, but A and E are
not d-separated given any subset of {B} or any subset of {D}; the only sets that d-separate
them are sets that contain F, C, or H. The Modified PC algorithm would correctly fmd that C,
F, and H are not adjacent to A or E. It would then fail to test whether A and E are d-separated
given any subset containing C. Hence it would fail to fmd that A and Eared-separated given
{B,C,D} and would erroneously leave A and E adjacent. This means that it is not possible to
determine which edges to remove from the graph by examining only local features (i.e. the
adjacencies) of the graph constructed at a given stage of the algorithm. Similarly, once bi-
directed edges are allowed in the output of the PC algorithm, it is not possible to extract all of
the information about the orientation of edges by examining local features (i.e. pairs of edges
sharing a common endpoint) of the graph constructed at a given stage of the algorithm.
Because of these problems, for full generality we must make major changes to the PC algorithm
and in the interpretation of the output. We will show that there is a procedure, which we
optimistically call the Fast Causal Inference. (FCI) algorithm, that is feasible in large variable
sets provided the true graph is sparse and there are not many bidirected edges chained together.
The algorithm gives asymptotically correct information about causal structure when latent
variables may be acting, assuming the measured distribution is the marginal of a distribution
satisfying the Markov and Faithfulness conditions for the true graph. The FCI algorithm avoids
the mistakes of the modified PC algorithm, and in some cases provides more information.
For example, with a marginal distribution over the boxed variables from the imaginary structure
in figure 4, the modified PC algorithm gives the correct output shown in the first diagram in
figure 5, whereas the FCI algorithm produces the correct and much more informative result in
the second diagram in figure 5:
Figure 4
In figure 5, the double headed arrows indicate the presence of unmeasured common causes, and
as in the modified PC algorithm the edges of the form 0-> indicate that the algorithm cannot
determine whether the circle at one end of the edge should be an arrowhead. Notice that the
adjacencies among the set of variables {Cilia damage, Heart disease, Lung capacity, Measured
breathing dysfunction} form a complete graph, but even so the edges can be completely
oriented by the FCI algorithm.
The derivation of the FCI algorithm requires a variety of new graphical concepts and a rather
intricate theory. We introduce Verma and Pearl's notions of an inducing path and an inducing
path graph, and show that these objects provide information about causal structure. Then we
consider algorithms that infer a class of inducing path graphs from the data.
L..-____-.-....I
,.-----0 !Heartdisease! 01----""""
Modified PC Output
Income
----I.~!Heart disease!
L...---r.......,--r---l . . . ....t---.L;-....;;;........;_....;...J
FCIOutput
Income
Figure 5
6.4 Inducing Paths
Given a directed acyclic graph G over a set of variables V, and 0 a subset of V, Verma and
Pearl (1991) have characterized the conditions under which two variables in 0 are not d-
separated given any subset of O\{A,B}. If G is a directed acyclic graph over a set of variables
V, 0 is a subset of V containing A and B, and A*" B, then an undirected path U between A and
B is an inducing path relative to 0 if and only if every member of 0 on U except for the
endpoints is a collider on U, and every collider on U is an ancestor of either A or B. We will
sometimes refer to members of 0 as observed variables.
Figure 6: Grapb G3
For example, in graph G3, the path U = <A,B,C,D,E,F> is an inducing path over 0 =
{A,B,D,F} because each collider on U (B and D) is an ancestor of one of the endpoints, and
each variable on U that is in 0 (except for the endpoints of U) is a collider on U. Similarly, U is
an inducing path over 0 = {A,B,F}. However, U is not an inducing path over 0 =
{A,B,C,D,F} because C is in 0, but C is not a collider on U.
Tbeorem 6.1: If G is a directed acyclic graph with vertex set V, and 0 is a subset of
V containing A and B, then A and B are not d-separated by any subset Z of O\{A,B} if
and only if there is an inducing path over the subset 0 between A and B.
It follows from Theorem 6.1 and the fact that U is an inducing path over 0 = {A,B,D,F} that A
and F are d-connected given every subset of {B,D}. Because in graph G3 there is no inducing
path between A and F over 0 = {A,B,C,D,F} it follows that A and F are d-separated given
some subset of {B,C,D} (in this case, {B,C,D} itself.)
6.5 Inducing Path Graphs
The inducing paths relative to 0 in a graph G over V can be represented in the following
structure described (but not named) in Verma and Pearl (1990b). G' is an inducing patb
grapb over 0 for directed acyclic graph G if and only if 0 is a subset of the vertices in
G, there is an edge between variables A and B with an arrowhead at A if and only if A and B
are in 0, and there is an inducing path in G between A and B relative to 0 that is into A. (Using
the notation of Chapter 2, the set of marks in an inducing path graph is {>, EM}.) In an
inducing path graph, there are two kinds of edges: A -> B entails that every inducing path over
o between A and B is out of A and into B, and A <-> B entails that there is an inducing path
over 0 that is into A and into B. This latter kind of edge can only occur when there is a latent
common cause of A and B.
Figures 7 through 9 depict the inducing path graphs of G3 over 0 = {A,B,D,E,F}, 0 =

{A,B,D,F} and 0 = {A,B,F} respectively. Note that in G3 <B,D> is an inducing path between
Band Dover 0 = {A,B,D,E,F} that is out of D. However, in the inducing path graph the edge
between Band D has an arrowhead at D because there is another inducing path <B,C,D> over
0= {A,B,D,E,F} that is into D. There is no edge between A and F in the inducing path graph
over 0 = {A,B,D,E,F}, but there is an edge between A and F in the inducing path graphs over
0= {A,B,D,F} and 0 = {A,B,F}.
Figure 7: Inducing Path Graph of G3 Over {A,B,D,E,F}
Figure 8: Inducing Path Graph of G3 Over {A,B,D,F}
Figure 9: Inducing Path Graph of G3 Over {A,B,F}

We can extend without modification the concept of d-separability to inducing path graphs if the
only kinds of edges that can occur on a directed path are edges with one arrowhead, and
undirected paths may contain edges with either single or double arrowheads. If G is a directed
acyclic graph, G' is the inducing path graph for Gover 0, and X, Y, and S are disjoint sets of
variables included in 0, then X and Y are d-separated given SinG' if and only they are d-
separated given SinG.
Double-headed arrows make for a very important difference between d-separability relations in
an inducing path graph and in a directed acyclic graph. In a directed acyclic graph over 0, if A
and Bared-separated given any subset of O\{A,B} then A and B are d-separated given either
Parents(A) or Parents(B). This is not true in inducing path graphs. For example, in inducing
path graph G4, which is the inducing path graph of figure 3 over °
= {A,B,C,D,E,F,H},
Parents(A) = {D} and Parents(E) = {B}, but A and E are not d-separated given any subset
of {B} or any subset of {D}; all of the sets that d-separateA and E contain C, H, or F.
C~
Figure 10: Inducing Path Graph G4
There is, however, a kind of set of vertices in inducing path graphs that, so far as d-separability
is concerned, behaves much like the parent sets in directed acyclic graphs.
If G' is an inducing path graph over ° and A", B, let V E D-SEP(A,B) if and only if A", V
and there is an undirected path U between A and V such that every vertex on U is an ancestor of
A or B and (except for the endpoints) is a collider on U.
Theorem 6.2: In an inducing path graph G' over 0, where A and B are in 0, if A is
not an ancestor of B, and A and B are not adjacent then A and Bared-separated given
D.SEP(A,B).
In an inducing path graph either A is not an ancestor of B or B is not an ancestor of A. Thus we

can determine whether A and B are adjacent in an inducing path graph without determining
whether A and B are dependent conditional on all subsets of 0.
If 0 is not a causally sufficient set of variables, then although we can infer the existence of an
inducing path between A and B if A and B are dependent conditional on every subset of
o\{ A,B}, we cannot infer that either A is a direct cause of B relative to 0, or that B is a direct
cause of A relative to 0, or that there is a latent common cause of A and B. Nevertheless, the
existence of an inducing path between A and B over 0 does contain information about the
causal relationships between A and B, as the following lemma shows.
Lemma 6.1.4: If G is a directed acyclic graph over V, 0 is a subset of V that contains

A and B, and G contains an inducing path over 0 between A and B that is out of A, and
A and B are in 0, then there is a directed path from A to BinG.
It follows from Lemma 6.1.4 that if 0 is a subset of V and we can determine that there is an
inducing path between A and B over 0 that is out of A, then we can infer that A is a (possibly
indirect) cause of B. Hence, if we can infer properties of the inducing path graph over 0 from
the distribution over 0, we can draw inferences about the causal relationships among variables,
regardless of what variables we have failed to measure. In the next section we describe
algorithms for inferring properties of the inducing path graph over 0 from the distribution over
o.
6.6 Partially Oriented Inducing Path Graphs
A partially oriented inducing path grapb can contain several sorts of edges: A -> B, A 0-
> B,A 0-0 B, or A <-> B. We use "*" as a metasymbol to represent any of the three kinds of
ends (EM (the empty mark), ">", or "0"); the "*" symbol itself does not appear in a partially
oriented inducing path graph. (We also use "*" as a metasymbol to represent the two kinds of
ends (EM or ">") that can occur in an inducing path graph.)
A partially oriented inducing path graph dor directed acyclic graph G with inducing path graph
G' over 0 is intended to represent the adjacencies in G', and some of the orientations of the
edges in G' that are common to all inducing path graphs with the same d-connection relations as
G'. If G' is an inducing path graph over 0, Equiv(G') is the set of inducing path graphs over
the same vertices with the same d-connections as G. Every inducing path graph in Equiv(G')
shares the same set of adjacencies. We use the following definition:
1t is a partially oriented inducing path graph of directed acyclic graph G with

inducing path graph G' over 0 if and only if
(i). if there is any edge between A and B in 1&, it is one of the following kinds:
A -> B, B -> A, A 0-> B, B 0-> A, A 0-0 B, or A <-> B;
(ii). 1& and G' have the same vertices;
(iii). 1& and G' have the same adjacencies;
(iv). if A *-> B is in 1&, then A -> B or A <-> B is in every inducing path graph in
Equiv(G');
(v). if A -> B is in 1&, then A -> B is in every inducing path graph in Equiv(G');
(vi). if A *-* B *-* C is in 1&, then the edges between A and B, and Band C do not collide
at B in any inducing path graph in Equiv(G').
(Strictly speaking a partially oriented inducing path graph is not a graph as we have defined it
because of the extra structure added by the underlining.) Note that an edge A *-0 B does not
constrain the edge between A and B either to be into or to be out of B in any subset of
Equiv(G'). The adjacencies in a partially oriented inducing path graph 1& for G can be
constructed by making A and B adjacent in 1& if and only if A and Bare d-connected given every
subset of O\{ A,B }.
Once the adjacencies have been determined, it is trivial to construct an uninformative partially
oriented inducing path graph 1& for G. Simply orient each edge A *-* B as A 0-0 B. Of course
this particular partially oriented inducing path graph 1& for G is very uninformative about what
features of the orientation of G' are common to all inducing path graphs in Equiv(G'). For
example, figure 11 shows again the imaginary graph of causes of measured breathing
dysfunction. Figure 12 shows an uninformative partially oriented inducing path graphs of graph
G5 over 0 = {Cilia damage, Smoking, Heart disease, Lung capacity, Measured breathing
dysfunction, Income, Parents' smoking habits}.
Smoking
Income
Figure II: Graph Gs
----~o IHeart disease I()O----oO ILung capacity I
Figure 12: Uninformative Partially Oriented Inducing Path Graph of Gs Over 0

Let us say that B is a definite non-collider on undirected path U if and only if either B is an
endpoint of U, or there exist vertices A and C such that U contains one of the subpaths A <- B
*-* C, A *-* B -> C, or A *-!JL!-* C. In a maximally informative partially oriented
inducing path graph 11: for G with inducing path graph G',
(i) an edge A *-0 B appears only if the edge between A and B is into B in some
members of Equiv(G'), and out of B in other members of Equiv(G'.), and
(ii) for every pair of edges between A and D, and Band C, either the edges collide atB,
or they are defmite non-colliders at B, unless the edges collide in some members of
Equiv(G) and not in others.
Such a maximally informative partially oriented inducing path graph 11: for G could be oriented
by the simple but inefficient algorithm of constructing every possible inducing path graph with
the same adjacencies as G', throwing out the ones that do not have the same d-connection
relations as G', and keeping track of which orientation features are common to all members of
Equiv(G'). Of course, this is completely computationally infeasible. Figure 13 shows the
maximally oriented partially oriented inducing path graph of graph GS over 0 = {Cilia damage,
Smoking, Heart disease, Lung capacity, Measured breathing dysfunction, Income, Parents'
smoking habits}.
Income Parents'smoking
habits
Figure 13: Maximally Informative Partially Oriented Inducing Path Graph of
Gs Over 0
Our goal is to state algorithms that construct a partially oriented inducing path graph for a
directed acyclic graph G containing as much orientation information as is consistent with
computational feasibility. The algorithm we propose is divided into two main parts. First, the
adjacencies in the partially oriented inducing path graph are determined. Then the edges are
oriented in so far as possible.
6.7 Algorithms for Causal Inference with Latent Common Causes
In order to state the algorithm, a few more defInition are needed. In a partially oriented inducing
path graph 1C:
(i). A is a parent of B if and only if A -> B in 1C.
(ii). B is a collider along path <A,B,C> if and only if A *-> B <-* C in 1C.
(iii). An edge between B and A is into A if and only if A <-* B in 1C.
(iv). An edge between B and A is out of A if and only if A -> Bin 1C.
(v). In a partially oriented inducing path graph 'It, U is a definite discriminating

path for B if and only if U is an undirected path between X and Y containing B, B '#
X, B '# Y, every vertex on U except for B and the endpoints is a collider or a defInite
non-collider on U, and
(i) if V and V' are adjacent on U, and V' is between V and Bon U, then V *-> V' on U,
(ii) if V is between X and B on U and V is a collider on U then V -> Yin 1C, else V <-* Y
in 1C,
(iii) if V is between Y and Bon U and V is a collider on U then V -> X in 1C, else V <-*
X in 1C.
(iv) X and Yare not adjacent in 1C.
Figure 14 illustrates the concept of a defInite discriminating path.
Figure 14: <.E,F,G,A,C,B> is a definite discriminating path for C

In practice, the Causal Inference Algorithm and the Fast Causal Inference Algorithm (described
later in this section) take as input either a covariance matrix or cell counts. Where d-separation
facts are needed by the algorithms, the procedure perfonns tests of conditional independence (in
the discrete case) or of vanishing partial correlations (in the linear, continuous case. (Recall that
if P is a discrete distribution faithful to a graph G, then A and B are d-separated given a set of
variables C if and only A and B are conditionally independent given C, and if P is a
distribution linearly faithful to a graph G, then A and B are d-separated given C if and only if
PAB.C = 0.) Both algorithms construct a partially oriented inducing path graph of some directed
acyclic graph G, where G contains both measured and unmeasured variables.
Causal Inference Algorithm!
A). Form the complete undirected graph Q on the vertex set V.
B). If A and B are d-separated given any subset S of V, remove the edge between A
and B, and record S in Sepset(A,B) and Sepset(B,A).
C). Let F be the graph resulting from step B). Orient each edge as 0-0. For each triple
of vertices A, B, C such that the pair A, B and the pair B, C are each adjacent in F but
the pair A, C are not adjacent in F, orient A *-* B *-* C as A *-> B <-* C if and only if
B is not in Sepset(A,C), and orient A *-* B *-* C as A *-!..JL.!-* C if and only if B is
in Sepset(A,C).
D). repeat
If there is a directed path from A to B, and an edge A *-* B, orient A *-* B as

A *->B,
else if B is a collider along <A,B,C> in 1C, B is adjacent to D, and A and C are
not d-connected given D, then orient B *-* D as B <-* D,
else if U is a deftnite discriminating path between A and B for M in 1C, and P and
R are adjacent to M on U, and P - M - R is a triangle, then
if M is in Sepset(A,B) then M is marked as a non-collider on subpath P *-!
M..!.-* R
else P *-* M *-* R is oriented as P *-> M <- * R.
else if P *-~-* R then orient as P *-> M -> R.
If the CI or FCI algorithms use as input a covariance matrix from the marginal over 0 of a
distribution linearly faithful to G, or cell counts from the marginal over 0 of a distribution
faithful to G, we will say the input is data over 0 that is faithful to G.
Theorem 6.3: If the input to the CI algorithm is data over 0 that is faithful to G, the
output is a partially oriented inducing path graph of Gover O.
If data over 0 = {Cilia damage, Smoking, Heart disease, Lung capacity, Measured breathing
dysfunction, Income, Parents' smoking habits} that is faithful to the graph in ftgure 11 is input
!We thank Thomas Venna (personal communication) for pointing out an error in the original fonnuIation of the
CI algorithm.
to the CI algorithm, the output is the maximally informative partially oriented inducing path
graph over 0 shown in figure 13.
Unfortunately, the Causal Inference (CI) algorithm as stated is not practical for large numbers
of variables because of the way the adjacencies are constructed. While it is theoretically correct
to remove an edge between A and B from the complete graph if and only if A and B are d-
separated given some subset of O\{A,B}' this is impractical for two reasons. First, there are
too many subsets of O\{A,B} on which to test the conditional independence of A and B.
Second, for discrete distributions, unless the sample sizes are enormous there are no reliable
tests of independence of two variables conditional on a large set of other variables.
In order to determine that a given pair of vertices, such as X and Y are not adjacent in the
inducing path graph, we have to find thatX and Yare d-separated given some subset of O\{X,
Y}. Of course, if X and Y are adjacent in the inducing path graph, they are d-connected given
every subset of O\{X,Y}. We would like to be able to determine that X and Yare d-connected
given every subset of O\{X,Y} without actually examining every subset of O\{X ,Y}.
In a directed acyclic graph over a causally sufficient set V, by using the PC algorithm we are
able to reduce the order and number of d-separation tests performed because of the following
fact: if X and Yare d-separated by any subset of v\{X,Y}, then they are d-separated either by
Parents(X) or Parents(y). While the PC algorithm is constructing the graph it does not know
which variables are in Parents(X) or in Parents(y), but as the algorithm progresses it is able
to determine that some variables are definitely not in Parents(X) or Parents(Y) because they
are definitely not adjacent to X or Y. This reduces the number and the order of the d-separation
tests that the PC algorithm performs (as compared to the SOS algorithm).
In contrast, an inducing path graph over 0 it is not the case that if X and Y are d-separated
given some subset of O\{X,Y}, then X and Y are d-separated given either Parents(X) or
given Parents(Y). However, if X and Y are d-separated given some subset of O\{X,Y} , then
X and Y are d-separated given either D-Sep(X) or given D-Sep(Y). If we know that some
variable V is not in D-Sep(X) and not in D-Sep(Y), we do not need to test whether X and Y
are d-separated by any set containing V. Once again, we do not know which variables are in D-
Sep(X) or D-Sep(Y) until we have constructed the graph. But there is an algorithm that can
determine that some variables are not in D-Sep(X) or D-Sep(Y) as the algorithm progresses.
Let G be the directed acyclic graph of figure 3 (reproduced below in figure 15.) Let G' be the
inducing path graph of Gover 0 = {A,B,C,D,E,F ,H}. A and E are not d-separated given any
subset of the variables adjacent to A or adjacent to D (in both cases (B,D}). Because A and E
are not adjacent in the inducing path graph of A and E, they are d-separated given some subset
of O\{A,E}. Hence they are d-separated by either D-Sep(A,E) (equal to (B,D,F}) or by D-
Sep(E,A» (equal to (B,D,H}). (In this case A and E are d-separated by both D-Sep(A,E) and
by D-Sep(E,A).) The problem is: how can we know to test whether A and E are d-separated
given {B,D,H} or {B,D,F} without testing whether A and E are d-separated given every
subset of O\{A,E}?
A variable V is in D-Sep(A,E) in G' if and only if V '* A and there is an undirected path
between A and Von which every vertex except the endpoints is a collider, and each vertex is an
ancestor of A or E. If we could find some method of determining that a variable V does not lie
on such a path, then we would not have to test whether A and E were d-separated given any set
containing V (unless of course V was in D-Sep(E,A).) We will illustrate the strategy on G. At
any given stage of the algorithm we will call the graph constructed thus far 1r.
GraphG
Inducing Path Graph G'
Figure 15
The FCI algorithm determines which edges to remove from the complete graph in three stages.
The first stage is just like the first stage of the PC Algorithm. We intialize Jr to the complete
undirected graph, and then we remove an edge between X and Y if they are d-separated given
subsets of vertices adjacent to X or Y in Jr. This will eliminate many, but perhaps not all of the
edges that are not in the inducing path graph. When this operation is performed on data faithful
to the graph in figure 15, the result is the graph in figure 16.
Figure 16
Note that A and E are still adjacent at this stage of the procedure because the algorithm, having
correctly determined that A is not adjacent to F or H or C, and that E is not adjacent to F or H or
C, never tested whether A and E are d-separated by any subset of variables containing F, H, or
C.
Second, we orient edges by determining whether they collide or not, just as in the PC
algorithm. The graph at this stage of the algorithm is show in figure 17.
Figure 17
Figure 17 is essentially the graph constructed by the PC algorithm given data faithful to the
graph in figure 15, after steps A), B), and C) have been performed.
We can now determine that some vertices are definitely not in D-Sep(A,E) or in D-Sep(E,A);
it is not necessary to every test whether A and E are d-separated given any subset of O\{A,E}
that contains these vertices in order to find the correct adjacencies. At this stage of the
algorithm, a necessary condition for a vertex V to be in D-Sep(A,E) in G' is that in 'H there is
an undirected path U between A and V in which each vertex except for the endpoints is either a
collider, or has its orientation hidden because it is in a triangle. Thus C and H are defmitely not
in D-Sep(A,E) and C and F are definitely not in D-Sep(E,A). All of the vertices that we have
not definitely determined are not in D-Sep(A,E) in G' we place in Possible-D-Sep(A,E),
and similarly for Possible-D-Sep(E,A). In this case, Possible-D-Sep(A,E) is {B,F,D}
and Possible-D-Sep(E,A) is {B,D,H}. We now know that if A and E are d-separated given
any subset of O\{A,E} then they are d-separated given some subset of Possible-D-Sep(A,E)
or some subset of Possible-D-Sep(E,A). In this case we find that A and E are d-separated
given a subset of Possible-D-Sep(A,E) (in this case the entire set) and hence remove the
edge between A and E.
Once we have obtained the correct set of adjacencies, we unorient all of the edges, and then
proceed to re-orient them exactly as we did in the Causal Inference Algorithm. The resulting
output is shown in figure 18.
FCIOutput
Figure 18
For a given partially constructed partially oriented inducing path graph 'H, Possible-D-
SEP(A,B) is defined as follows: If A ~ B, V is in Possible-D-Sep(A,B) in 'H if and only if V
~ A , and there is an undirected path U between A and V in 'H such that for every subpath
<X,Y,2> of U either Y is a collider on the subpath, or Y is not a definite non-collider and on
U, and X, Y, and Z form a triangle in 'H.
Using this definition of Possible-D-Sep(A,E), we can prove that every vertex not in
Possible-D-Sep(A,E) in 'H is not in D-Sep(A,E) in G'. However, it may be possible to
determine from 'H that some members that we are including in Possible-D-Sep(A,E) are not
in D-Sep(A,E) in G'. There is clearly a trade-off between reducing the size of Possible-D-
Sep(A,E) (so that the number and order of tests of d-separability performed by the algorithm is
reduced) and performing the extra work required to reduce the size of the set, while ensuring
that it is still a superset of D-Sep(A,E) in G'. We do not know what the optimal balance is. If
G is sparse (Le. each vertex is not adjacent to a large number of other vertices in G), then the
algorithm does not need to determine whether A and B are d-separated given C for any
C containing a large number of variables.
Fast Causal Inference Algorithm
A). Form the complete undirected graph Q on the vertex set V.
B). n =O.
repeat
repeat
select an ordered pair of variables X and Y that are adjacent in Q such that
Adjacencies(Q,x)\{y} has cardinality greater than or equal to n, and a subset S
of Adjacencies(Q,x)\{ Y} of cardinality n, and if X and Y are d-separated given S
delete the edge between X and Y from Q, and record S in Sepset(X,y) and
Sepset(Y,x)
until all ordered variable pairs of adjacent variables X and Y such that
Adjacencies(Q,K)\{Y} has cardinality greater than or equal to n and all subsets S of
Adjacencies(Q,x)\{y} of cardinality n have been tested for d-separation;
n = n + 1;
until for each ordered pair of adjacent vertices X, Y, Adjacencies(Q ,x)\{ Y} is of
cardinality less than n.
C). Let F' be the undirected graph resulting from step B). Orient each edge as 0-0. For
each triple of vertices A, B, C such that the pair A, B and the pair B, C are each adjacent
in F' but the pair A, C are not adjacent in F', orient A *-* B *-* C as A *-> B <-* C if
and only if B is not in Sepset(A,C).
D). For each pair of variables A and B adjacent in F', if A and B are d-separated given
any subset S of Possible-D-SEP(A,B)\{A,B} or any subset S of Possible-D-
SEP(B,A)\{A,B} in F remove the edge between A and B, and record S in
Sepset(A,B) and Sepset(B,A).
The algorithm then re-orients an edge between any pair of variables X and Y as X 0-0 Y, and
proceeds to re-orient the edges in the same way as steps C) and D) of the Causal Inference
algorithm.
Theorem 6.4: If the input to the FCI algorithm is data over 0 that is faithful to G, the
output is a partially oriented inducing path graph of Gover O.
The Fast Causal Inference Algorithm (FCI) always produces a partially oriented inducing path
graph for a graph G given correct statistical decisions from the marginal over the measured
variables of a distribution faithful to G. We do not know whether the algorithm is complete, i.e.
whether it in every case produces a maximally informative partially oriented inducing path
graph.
As with the CI algorithm, if the input to the FCI algorithm is data faithful to the graph of figure
11, the output is the maximally informative partially oriented inducing path graph of figure 13.
Two directed acyclic graphs G and G' that have the same FCI partially oriented inducing path
graph over 0 have the same d-connection relations involving just members of O.
Corollary 6.4.1: If G is a directed acylic graph over V, G' is a directed acyclic graph
over V', and 0 is a subset of V and of V', then G and G' have the same d-separation
relations among only the variables in 0 if and only if they have the same FCI partially
oriented inducing path graph over O.
Given a directed acyclic graph G, it is possible to determine what d-separation relations

involving just members of 0 are true of G from the FCI partially oriented inducing path graph
of Gover O. In a partially oriented inducing path graph 11:, if X '" Y, and X and Y are not in Z,
then an undirected path U between X and Y definitely d-connects X and Y given Z if and
only if every collider on U has a descendant in Z, every definite non-collider on U is not in Z,
and every other vertex on U is not in Z but has a descendant in Z. In a partially oriented
inducing path graph 11:, if X, Y, and Z are disjoint sets of variables, then X is definitely d-
connected to Y given Z if and only if some member of X is d-connected to some member of Y
given Z.
Corollary 6.4.2: If G is a directed acylic graph over V, 0 is a subset of V, 11: is the

FCI partially oriented inducing path graph of Gover 0, and X, Y, and Z are disjoint
subsets of 0, then X is d-connected to Y given ZinG if and only if X is definitely d-
connected to Y given Z in 11:.
These corollaries are proved in Spirtes and Verma (1992).

6.8 Theorems on Detectable Causal Influence
In this section we show that a number of different kinds of causal inferences can be drawn from
a partially oriented inducing path graph.
Theorem 6.5: If 1r is a partially oriented inducing path graph of directed acyclic graph
Gover 0, and there is a directed path U from A to B in 1r, then there is a directed path
from A toB in G.
If G is a directed acyclic graph over V, and 0 is included in V, if the input to the CI algorithm
is data faithful to Gover 0, then we call the output of the CI algorithm the CI partially
oriented inducing path graph of Gover O. We adopt a similar terminology for the FCI
algorithm. A semi-directed path from A to B in partially oriented inducing path graph 1r is
an undirected path U from A to B in which no edge contains an arrowhead pointing towards A,
that is, there is no arrowhead at A on U, and if X and Y are adjacent on the path, and X is
between A and Y on the path, then there is no arrowhead at the X end of the edge between X
and Y.
Theorem 6.6: If 1r is the CI partially oriented inducing path graph of directed acyclic
graph Gover 0, and there is no semi-directed path from A to Bin 1r, then there is no
directed path from A to BinG.
Recall that a trek between distinct variables A and B is either a directed path from A to B, a
directed path from B to A, or a pair of directed paths from a vertex C to A and B respectively
that intersect only at C. The following theorem states a sufficient condition for when the edges
in a partially oriented inducing path graph indicate a trek in the graph that contains no measured
vertices except for the endpoints.
Theorem 6.7: If 1r is a partially oriented inducing path graph of directed acyclic graph
Gover 0, A and B are adjacent in 1r, and there is no undirected path between A and B
in 1r except for the edge between A and B, then in G there is a trek between A and B that
contains no variables in 0 other than A or B.
Tbeorem 6.8: If 1C is the CI partially oriented inducing path graph of directed acyclic
graph Gover 0, and every semi-directed path from A to B contains some member of C
in 1C, then every directed path from A to BinG contains some member of C.
Tbeorem 6.9: If 1C is a partially oriented inducing path graph of directed acyclic graph
Gover 0, and A <-> B in 1C, then there is a latent common cause of A and BinG.
Parallel results holds for the FCI algorithm.
To illustrate the application of these theorems, condsider the maximally informative partially
oriented inducing path graph in figure 13 of the causal structure of Gs. Applying Theorem 6.5
we infer that Smoking causes Cilia damage, Lung capacity and Measured breathing
dysfunction. Applying Theorem 6.6, we infer that Smoking does not cause Heart disease or
Income or Parents' Smoking Habits. It is impossible to determine from the conditional
independence relations among the measured variables whether Income causes Smoking, or
there is a common cause of Smoking and Income. The statistics among the measured variables
determine that Cilia damage and Heart disease have a latent common cause, Cilia damage does
not cause Heart disease, and Heart disease does not cause Cilia damage.
We note here a topic that will be more fully explored in the next chapter. In the example from
figure 11, in order to infer that smoking causes breathing dysfunction, it is necessary to
measure two causes of Smoking (whose collision at Smoking orients the edge from Smoking to
Cilia damage.) In general, this suggests that in the design of studies intended to determine if
there is a causal path from variable A to variable B, it is useful to measure not only variables
that might mediate the connection between A and B, but also to measure possible causes of A.
6.9 Non-Independence Constraints
The Markov and Faithfulness conditions applied to a causally insufficient graph may entail
constraints on the marginal distribution of measured variables that are not conditional
independence relations, and hence are not used in the FCI algorithm. Consider, the example in
figure 19, due to Thomas Verma (Verma and Pearl, 1991).
Figure 19
Assume T is unmeasured. Then a joint distribution faithful to the entire graph must satisfy the
constraint that the quantity
-+
L.P(BIA)P(DIB,C,A)
B
is a function only of the values of C and D.
-+ -+ -+
L.P(BIA)P(DIB,C,A) = L.L.P(BIA)P(DIB,C,A,T)P(TIB,C,A) =
B T B
-+ -+
L.P(DIC,T)L.P(BIA)p(TIB,A)
T B
(because A, B are independent of D given {C, T} and C is independent of T given (A, BD.
Hence
-+ -+ -+
L.P(DIC,T)L.P(BIA)P(TIB,A) = L.P(DIC,T)P(TIA) =
T B T
-+
L.P(DIC,T)P(T) = g(C,D)
T
(because T and A are independent).

This constraint is not entailed if a directed edge from A to D is added to the graph. The moral is
that there is further marginal structure not in the form of conditional independence relations that
could in principle be used to help identify latent structure. We will see a similar point when we
tum to linear models in the next section.
6.10 Generalized Statistical Indistinguishability and Linearity
Suppose that for whatever reasons an investigation were to be confined to linear structures and
to probability distributions that are consistent with the assumption that each random variable is a
linear function of its parents and of unmeasured factors. The effect of restrictions such as
linearity is to make distinguishable causal structures that would otherwise be indistinguishable.
That happens because the restriction, whatever it is, together with the conditional dependence
and independence relations required by the Markov, Minimality or Faithfulness Conditions,
entails additional constraints on the measured variables. These additional constraints may not be
in the form of conditional independence relations. In the linear case they typically are not.
Consider for example the two structures shown below, where the X variables are measured and
the T variables are unmeasured.
Figure 20
These structures each imply that in the marginal distribution over the measured variables every
pair of variables is dependent conditional on every other set of measured variables. In each case
the maximally informative partially oriented inducing path graph on the X variables is a
complete undirected graph. By examining conditional independence relations among these
variables, one could not tell which structure obtains. But if linearity is required, then it is easy
to tell which structure obtains. For under the linearity assumption, the second structure entails
all three of the following constraints on the correlations of the measured variables, while the
fIrst structure entails only the fIrst of these constraints (where we denote the correlation between
X 1 and X2 as P12 in order to avoid subscripts with subscripts.) :
P13P24 - P14P23 = 0
P12P34 - P14P23 = 0
P13P24 - P12P34 = 0
Early in this century Charles Spearman (1928) called constraints of these sorts vanishing
tetrad differences, and we will use his terminology.
Characterizing statistical indistinguishability under the linearity restriction thus presents an

entirely new problem, and one for which we will offer no general solution. It is not true, for
example, that conditional independence relations and vanishing tetrad differences jointly
determine the faithful indistinguishability classes of linear structures with unmeasured variables.
For example, each of the following linear structures entails that a single tetrad difference
vanishes in the marginal distribution over A, B, C and D, and has a partially oriented inducing
path graph for these variables consisting of a complete undirected graph:
ft-
(i)
[t] ~
(ii)
Figure 21
But the two graphs are not faithfully indistinguishable over the class of linear structures.
Structure (ii) permits distributions consistent with linearity in which the correlation of A and B
is positive, the correlation of Band C is positive and the correlation of A and C is negative.
Structure (i) admits no distributions consistent with linearity whose marginals satisfy this
condition.
Structures (i) and (ii) are not typical of the linear causal structures with unmeasured variables
one finds in the social science literature. For practical purposes, the examination of vanishing
tetrad constraints provides a powerful means to distinguish between alternative causal
structures, even in structures that are only partially linear. Tests for hypotheses of vanishing
tetrad differences were introduced by Wishart in the 1920s assuming normal variates, and
asymptotically distribution free tests have been described by Bollen (1989).
Algorithms that take advantage of vanishing tetrad differences will be described and illustrated
later in this book. In order to take that advantage, we need to be able to determine
algorithmically when a structure with or without unmeasured common causes entails a particular
vanishing tetrad difference among the measured variables. This question leads to an important
theorem.
6.11 The Tetrad Representation Theorem
We wish to characterize entirely in graph theoretic terms a necessary and sufficient condition for
a distribution on the vertices of an arbitrary directed acyclic graph G to linearly imply a
vanishing tetrad difference, that is the tetrad difference vanishes in all of the distributions
linearly represented by G. We will call a distribution linearly represented by some directed
acyclic graph G a linear model. (A slightly more formal definition is given in Chapter 13.) A
linear model is uniquely determined by the directed acyclic graph G that represents it, and linear
coefficients and the independent marginal distributions on the variables (including error terms)
of zero indegree.
First some terminology: Given a trek T(J,J) between vertices J and J, /(T(J,J» denotes the
directed path in T(J,J) from the source of T(J,J) to J and J(T(J,.I) denotes the directed path in
T(J,.I) from the source of T(J,.I) to J. (Recall that one of the directed paths in a trek maybe an
empty path.) T(J,.I) denotes the set of all treks between J and J.
In a directed acyclic graph G, if for all T(K,L) in T(K,L) and all T(J,J) in T(J,J),L(T(K,L»
and J(T(J,.I) intersect at a vertex Q, then Q is an LJ(T(J,.I),T(K,L» choke point. Similarly, if
for all T(K,L) in T(K,L) and all T(J,J) in T(J,J), L(T(K,L» and all J(T(J,J» intersect at a
vertex Q, and for all T(J,L) in T(J,L) and all T(J,K) in T(J,K), L(T(J,L» and J(T(J,K» also
intersect at Q, then Q is an LJ(T(J,J),T(K,L),T(J,L),T(J,K» choke point. Also see the
definition of trek.
The fundamental theorem for vanishing tetrad differences in linear models is this:
Tetrad Representation Theorem 6.10: In a directed acyclic graph G, there exists

an LJ(T(J,J),T(K ,L),T(J,L),T(J,K» or an JK(T(J,J),T(K,L),T(J ,L),T(J,K» choke
point if and only if G linearly implies PIJPKL - PILPJK = O.
A consequence of Theorem 6.10 is

Theorem 6.11: An acyclic graph G linearly implies PIJPKL - P1LPJK = 0 only if either
PIJ or PKL = 0, and PIL or PJK = 0, or there is a (possibly empty) set Q of random
variables in G such that PIJ.Q = PKL.Q = PIL.Q = PJK.Q = o.
Theorem 6.10 provides a fast algorithm for calculating the vanishing tetrad differences linearly
implied by any directed acyclic graph. Theorem 6.11 provides a means to determine when
unmeasured common causes are acting in linear structures. In later chapters we describe some
of the implications of these facts for investigating the structure of causal relations among
unmeasured variables.
6.12 An Example: Math Marks and Causal Interpretation
In several places in his recent text on graphical models in statistics, Whittaker (1990) discusses
a data set from Mardia, Kent and Bibby (1979) concerning the grades of 88 students on
examinations in five mathematical subjects: mechanics, vectors, algebra, analysis and statistics.
The example illustrates one of the uses of the Tetrad Representation Theorem, and provides
occasion to comment on some important differences of interpretation between our methods and
those Whittaker describes. The variance/covariance matrix for the data is as follows:
Mechanics Vectors Algebra Analysis Statistics

302.29
125.78 170.88
100.43 84.19 111.60
105.07 93.60 110.84 217.88
116.07 97.89 120.49 153.77 294.37
When given these data, the PC algorithm immediately determines the following pattern:
Figure 22
Whittaker obtains the same graph under a different interpretation. Recall that an undirected
independence graph is any pair <G,P> where G is an undirected graph and P is a distribution
such that vertices, X, Yin G are not adjacent if they are independent conditional on the set of all
other vertices of G; or to state the contrapositive: if X, Yare dependent conditional on the set of
all other vertices of G, then X, Yare adjacent in G. Undirected independence graphs hide much
of the causal structure, and sometimes many of the independence relations. Thus if variables X
and Z are causes of variable Y but X and Z are statistically independent and have no causal
relations whatsoever, the undirected independence graph has an edge between X and Z. In
effect, the independence graph fails to represent the conditional independence relations that hold
among proper subsets of a set of variables.
Every undirected pattern graph obtained from a faithful distribution (or sample) is a subgraph of
the undirected independence graph obtained from that distribution. In the case at hand the two
graphs are the same, but they need not be in general.
Whittaker claims that identifying the undirected independence graph is important for four
reasons: (i) it reduces the complex five dimensional object into two simpler three dimensional
objects--the two maximal cliques in the graph; (ii) it groups the variables into two sets; (iii) it
highlights Algebra as the one crucial examination in analyzing the interrelationship between
different subjects in exam performance; (iv) it asserts that Algebra and Analysis alone will be
sufficient to predict Statistics and that Algebra and Vectors will be sufficient to predict
Mechanics; but that all four marks are needed to predict Algebra (p. 6)
The second reason seems simply a consequence of the first, and the first seems of little
consequence: the cognitive burden of noting that there are five variables is not very great There
is a long tradition in statistics of introducing representations on grounds that they simplify the
data and in practice treating the objects of such reductions as causes. That is, for example, the
history of factor analysis after Thurstone. But as with factor analysis, causal conclusions drawn
from independence graphs would be unreliable. The third reason seems too vague to be worth
much trouble. The assertion given in the fourth reason is sound, but only if "predict" is
understood in all cases to have nothing to do with predicting the values of variables when they
are deliberately altered, as by coaching. We suspect statistical analyses of such educational data
are apt to be given a causal significance, and for such purposes directed graphical models better
represent the hypotheses.
Applying Theorem 6.11, the vanishing tetrad test for latent variables, we find that there are four
vanishing tetrad differences that cannot be explained by vanishing partial correlations among the
measured variables. This suggests that it is entailed by vanishing partial correlations involving
latent variables, and thus suggests the introduction of latent variables. A natural idea in view of
the mathematical structure of the subjects tested is that Algebra is an indicator of Algebraic
knowledge, which is a factor in the Knowledge of vector algebra measured by Vector and
Mechanics and is also a factor in Knowledge of real analysis that affects Analysis and Statistics.
The explanation of the data then looks like this:
Figure 23
The arrows without notation attached to them indicate other sources of variation. Assuming a
faithful distribution and linearity, this graph does not entail the vanishing fIrst order partial
correlations among the measured variables that the data suggest. But if the variance in Algebra
due to factors other than algebraic knowledge is sufficiently small, a linear distribution faithful
to this graph will to good approximation give exactly those vanishing partial correlations.
This structure (assuming linearity) entails eight vanishing tetrad differences, all of which the
TETRAD II program identifies and tests and cannot reject (p > .7). The model itself, when
treated as the null hypotheses in a likelihood ratio test, yields a p value of about .9, roughly the
value Whittaker reports for the undirected graphical independence model.
In a series of papers (Pearl and Verma 1990, 1991, Verma and Pearl 1990a, 1990b, 1991)
Verma and Pearl describe an "Inductive Causation" algorithm that outputs a structure that they
call a pattern (or sometimes a "completed hybrid graph") of a directed acyclic graph G over a set
of variables O. The most complete description of their theory appears in Verma and Pearl
(1990b). The key ideas of an inducing path, an inducing path graph, and the proof of (what we
call) Theorem 6.1 all appear in this paper. Unfortunately, the two main claims about the output
of the Inductive Causation Algorithm made in the paper, given in their lemma A2 and their
Theorem 2, are false (see Spirtes, 1992).
Early versions of the Inductive Causation Algorithm did not distinguish between A -> B and A
0-> B, and hence could not be used to infer that A causes B as in Theorem 6.5. This distinction
was introduced (in a different notation) in order to prove a version of Theorem 6.5 and
Theorem 6.6 in Spirtes and Glymour (1990); Verma and Pearl incorporated it in a subsequent
version of the Inductive Causation Algorithm. The Inductive Causation Algorithm does not use
definite discriminating paths to orient edges, and hence in some cases gives less orientation
information than the FCI procedure. The output of the Inductive Causation Algorithm has no
notation distinguishing between edges in triangles that definitely do not collide and merely
unoriented edges. Like the CI algorithm, the Inductive Causation Algorithm cannot be applied
to large numbers of variables because testing the independence of some pairs of variables
conditional on every subset of O\{A,B} is required.
The vanishing tetrad difference was used as the principle technique in model specification by
Spearman and his followers. A brief account of their methods is given in Glymour, Scheines,
Spirtes and Kelly (1987). Spearman's inference to common causes from vanishing tetrad
differences was challenged by Godfrey Thomson in a series of papers between 1916 and 1935.
In our terms, Thomson's models all violated linear faithfulness.
Chapter 7
Prediction
7.1 Introduction
The fundamental aim of many empirical studies is to predict the effects of changes, whether the
changes come about naturally or are imposed by deliberate policy: Will the reduction of sources
of environmental lead increase the intelligence of children in exposed regions? Will increased
taxation of cigarettes decrease lung cancer? How large will these effects be? What will be the
differential yield if a field is planted with one species of wheat rather than another~ or the
difference in number of polio cases per capita if all children are vaccinated against polio as
against if none are; or the difference in recidivism rates if parolees are given $600 per month for
six months as against if they are given nothing; or the reduction of lung cancer deaths in middle
aged smokers if they are given help in quitting cigarette smoking; or the decline in gasoline
consumption if an additional dollar tax per gallon is imposed?
One point of experimental designs of the sort found in randomized trials is to attempt to create
samples that, from a statistical point of view, are from the very distributions that would result if
the corresponding treatments were made general policy and applied everywhere. For such
experiments under such assumptions, the problems of statistical inference are conventional,
which is not to say they are easy, and the prediction of policy outcomes is not problematic in
principle. But in empirical studies in the social sciences, in epidemiology, in economics, and in
many other areas, we do not know or cannot reasonably assume that the observed sample is
from the very distribution that would result if a policy were adopted. Implementing a policy
may change relevant variables in ways not represented in the observed sample. The inference
task is to move from a sample obtained from a distribution corresponding to passive
observation or quasi-experimental manipulation, to conclusions about the distribution that
would result if a policy were imposed. In our view one of the most fundamental questions of
statistical inference is when, if ever, such inferences are possible, and, if ever they are possible,
by what means. The answer, according to Mosteller and Tukey, is "never." We will see
whether that answer withstands analysis.
7.2 Prediction Problems
The possibilities of prediction may be analyzed in a number of different sorts of circumstances,

including at least the following:
Case 1: We know the causal graph, which variables will be directly manipUlated, and what the
direct manipulation will do to those variables. We want to predict the distribution of variables
that will not be directly manipulated. More formally, we know the set X of variables being
directly manipulated, P(XIParents(X» in the manipulated distribution, and that Parents(X)
in the manipulated population is a subset of Parents (X) in the unrnanipulated population. That
is essentially the circumstance that Rubin, Holland and Pratt and Schlaifer address, and in that
case the causal graph and the Manipulation Theorem specify a relevant formula for calculating
the manipulated distribution in terms of marginal conditional probabilities from the
unmanipulated distribution. The latter can be estimated from samples; we can find the
distribution of Y (or of Y conditional on Z) under direct manipulation of X by taking the
appropriate marginal of the calculated manipulated distribution.
Case 2: We know the set X of variables being directly manipulated, P(XIParents(X» in the
manipulated distribution, that Parents(X) in the manipulated population is a subset of
Parents(X) in the unmanipulated population, and that the measured variables are causally
sufficient; unlike case 1, we do not know the causal graph. The causal graph must be
conjectured from sample data. In this case the sample and the PC (or some other) algorithm
determine a pattern representing a class of directed graphs, and properties of that class
determine whether the distribution of Y following a direct manipulation of X can be predicted.
Case 3. The difficult, interesting and realistic case arises when we know the set X of variables
being directly manipulated, we know P(XIParents(X» in the manipulated popUlation, and that
Parents(X) in the manipulated population is a subset of Parents(X) in the unmanipulated
population, but prior knowledge and the sample leave open the possibility that there may be
unmeasured common causes of the measured variables. If observational studies were treated
without unsupported pre-conceptions, surely that would be the typical circumstance. It is
Prediction 203
chiefly because of this case that Mosteller and Tukey concluded that prediction from
uncontrolled observations is not possible. One way of viewing the fundamental problem of
predicting the distribution of Y or conditional distribution of Y on Z upon a direct manipulation
of X can be formulated this way: find conditions sufficient for prediction, and conditions
necessary for prediction, given only a partially oriented inducing path graph and conditional
independence facts true in the marginal (over the observed variables) of the unmanipulated
distribution. Show how to calculate features of the predicted distribution from the observed
distribution. The ultimate aim of this chapter is to provide a partial solution to this problem.
We will take up these cases in turn. Case 1 is easy but we take time with it because of the
connection with Rubin's theory. Case 2 is dealt with very briefly. In our view Case 3 describes
the more typical and theoretically most interesting inference problems. The reader is wamed that
even when the proofs are postponed the issue is intricate and difficult.
7.3 Rubin-Holland-Pratt-Scblaifer Tbeoryl
Rubin's framework has a simple and appealing intuition. In experimental or observational

studies we sample from a population. Each unit in the popUlation, whether a child or a national
economy or a sample of a chemical, has a collection of properties. Among the properties of the
units in the population, some are dispositional--they are propensities of a system to give a
response to a treatment. A glass vase, for example, may be fragile, meaning that it has a
disposition to break if struck sharply. A dispositional property isn't exhibited unless the
appropriate treatment is applied--fragile vases don't break unless they are struck. Similarly, in a
population of children, for each reading program each child has a disposition to produce a
certain post-test score (or range of test scores) if exposed to that reading program. In
experimental studies when we give different treatments to different units, we are attempting to
estimate dispositional properties of units (or their averages, or the differences of their averages)
from data in which only some of the units have been exposed to the circumstances in which that
disposition is manifested. Rubin associates with each such dispositional quantity, Q, and each
value x of relevant treatment variable X, a random variable, Qx/=x, whose value for each unit in
the population is the value Q would have if that unit were to be given treatment x, or in other
words if the system were forced to have X value equal to x. If unit i is actually given treatment
lThis section is based on Spirtes, GIymour, Scheines, Meek, Fienberg and Slate, 1992.
xl and a value of Q is measured for that unit, the measured value of Q equals the value of
QXf=xl.
Experimentation may give a set of paired values <x, YXf=x>, where YXf=x is the value of the
random variable YXf=x, But for a unit i that is given treatment xl, we also want to know the
value of YXf=x2, YXf=x3, and so on for each possible value of X, representing respectively the
values for Y unit i is disposed to exhibit if unit i were exposed to treatment x2 or x3, that is, if
the X value for these units were forced to be x2 or x3 rather than xl. These unobserved values
depend on the causal structure of the system. For example, the value of Y that unit i is disposed
to exhibit on treatment x2 might depend on the treatments given to other units. We will suppose
that there is no dependence of this kind, but we will investigate in detail other sorts of
connections between causal structure and Rubin's counterfactual random variables.
A typical inference problem in Rubin's framework is to estimate the distribution of Yxf=x for
some value x of X, over all units in the population, from a sample in which only some members
have received the treatment x. A number of variations arise. Rather than forcing a unique value
on X, we may contemplate forcing some specified distribution of values on X, or we may
contemplate forcing different specified distributions on X depending on the (unforced) values of
some other variables Z; our "experiment" may be purely observational so that an observed value
q of variable Q for unit i when X is observed to have value x is not necessarily the same as
QXf=x, Answers to various problems such as these can be found in the papers cited. For
example, in our paraphrasing, Pratt and Schlaifer claim the following:
When all units are systems in which Y is an effect ofX and possibly of other variables, and no
causes ofY other than X are measured, in order for the conditional distribution ofY on X =x
to equal YXf=xfor all values x of X, it is sufficient and "almost necessary" that X and each of
the random variables YXf=x (where x ranges over all possible values of X) be statistically
independent.
In our terminology, when the conditional distribution of Y on X = x equals YXf=x for all values
x of X we say that the conditional distribution of Y on X is "invariant"; in their terminology it is
"observable." Pratt and Schlaifer's claim may be clarified with several examples, which will
also serve to illustrate some tacit assumptions in the application of the framework. Suppose X
and U, which is unobserved, are the only causes of Y, and they have no causal connection of
any kind with one another, a circumstance that we will represent by the graph in figure I.
Prediction 205
Figure 1
For simplicity let's suppose the dependencies are all linear, and that for all possible values of X,
Yand U, and all units, Y =X + U. LetXfrepresent values of X that could possibly beforced
on all units in the population. X is an observed variable; Xf is not. X is a random variable; Xf is
not. Consider values in Table 1.
Table 1
X Y U Xf UXf=1 YXf=1
1 1 0 1 0 1
I 2 1 1 1 2
1 3 2 1 2 3
2 2 0 1 0 1
2 3 1 1 1 2
2 4 2 1 2 3
Suppose for simplicity that each row (ignoring Xf, which is not a random variable) is equally
probable. Here the X and Y columns give possible values of the measured variables. The U
column gives possible values of the unmeasured variable U. Xfis a variable whose column
indicates values of X that might be forced on a unit; we have not continued the table beyond Xf
= 1. The UXj=l column represents the range of values of U when X is forced to have the value
1; the YXj=l gives the range of values of Y when X is forced to have the value 1. Notice that in
the table YXj=1 is uniquely determined by the value of Xfand the value of UXj=1 and is
independent of the value of X.
The table illustrates Pratt and Schlaifer's claim: YXj=1 is independent of X and the distribution
of Y conditional on X = I equals the distribution of YXj=1.
= =
We constructed the table by letting U UXf=1. and YXj=1 1 + UX/=1. In other words, we
obtained the table by assuming that save for the distribution of X, the causal structure and
probabilistic structure are completely unaltered if a value of X is forced on all units. By
applying the same procedure with YX/=2 = 2 + UX/=2, the table can be extended to obtain values
when X/ = 2 that satisfy Pratt and Schlaifer's claim.
Consider a different example in which, according to Pratt and Schlaifer's rule, the conditional
probability of Y on X is not invariant under direct manipulation. In this case X causes Yand U
causes Y, and there is no causal connection of any kind between X and U, as before, but in
addition an unmeasured variable V is a common cause of both X and Y, a situation represented
in figure 2.
1+---0
Figure 2
Consider the following distribution, with the same conventions as in Table I:
Table 2
X V U Y Xi VXf=l UXf=l YXf~l
0 0 0 0 I 0 0 I
0 0 I I I 0 I 2
0 0 2 2 I 0 2 3
0 0 3 3 I 0 3 4
I I 0 2 I I 0 2
I I I 3 I I I 3
I I 2 4 I I 2 4
I I 3 5 I I 3 5
Again, assume all rows are equally probable, ignoring the value of X/which is not a random
variable. Notice that YX/=l is now dependent on the value of X. And, just as Pratt and Schlaifer
require, the conditional distribution of Y on X = I is not equal to the distribution of YX/=l.
Prediction 207
The table was constructed so that when X = 1 is forced, and henceXf= I, the distributions of
UX/=1. and VX/=1 are independent of Xf. In other words, while the system of equations
y=X+V+U
X=V
was used to obtain the values of X, Y, and U, the assumptions UX!=1 = U, VX!=1 = V and the
equation
YX!=1 =Xf+ VX!=1 + UX!=1
were used to determine the values of UX/=1, VX/=1 and YX!=1. The forced system was treated
as if it were described by the diagram depicted in figure 3.
Figure 3
For another example, suppose Y =X + U, but there is also a variable V that is dependent on
both Y and X, so that the system can be depicted as in figure 4.
1..--0
Figure 4
Here is a table of values, obtained by assuming Y = X + U and V = Y + X, and these relations

are unaltered by a direct manipulation of X:
Table 3
x Y V U Xf VXf.=l UX'=l YXf=l
0 0 0 0 I 2 0 I
0 I I 1 1 3 1 2
0 2 2 2 1 4 2 3
1 1 2 0 1 2 0 1
I 2 3 1 1 3 1 2
1 3 4 2 1 4 2 3
Again assume all rows are equally probable. Note that YX/=l is independent of X, and YX/=l
has the same distribution as Y conditional on X = 1. So Pratt and Schlaifer's principle is again
satisfied, and in addition the conditional probability of Y on X is invariant. The table was
constructed by supposing the manipulated system satisfies the very same system of equations as
the unmanipulated system, and in effect that the graph of dependencies in figure 4 is unaltered
by forcing values on X.
Pratt and Schlaifer's rules, as we have reconstructed them, are consequences of the Markov
Condition. So are other examples described by Rubin. To make the connection explicit we
require some results. We will assume the technical defmitions introduced in Chapter 3, and we
will need some further definitions.
if G is a directed acyclic graph over a set of variables V u W, W is exogenous with respect to

V in G, Y and Z are disjoint subsets of V, P(V u W) is a distribution that satisfies the Markov
condition for G, and Manipulated(W) = X, then P(YIZ) is invariant under direct
manipulation of X in G by changing W from WI to W2 if and only if P(YIZ,W = WI) =
P(yIZ,W = wz) wherever they are both defined. Note that a sufficient condition for P(YIZ) to
be invariant under direct manipulation of X in G by changing W is that W be d-separated from
Y given Z in G. In a directed acyclic graph G containing Y and Z, ND(Y) is the set of all
vertices that do not have a descendant in Y. If Y n Z =0, then V is in IV(Y,Z) (informative
variables for Y given Z) if and only if V is d-connected to Y given Z, and V is not in ND(YZ).
(Note that this entails that V is not in Y u Z.) if Y n Z = 0, W is in IP(Y,Z) (W has a parent
who is an informative variable for Y given Z) if and only if W is a member of Z, and W has a
parent in IV(Y,Z) u Y. We will use the following result.
Prediction 209
Theorem 7.1: If GComb is a directed acyclic graph over V u W. W is exogenous

with respect to V in GComb. Y and Z are disjoint subsets of v. P(V u W) is a
distribution that satisfies the Markov condition for GComb. no member of X n Z is a
member of IP(Y.Z) in GUnman. and no member of X\Z is a member of IV(Y.Z) in
GUnman. then P(YIZ) is invariant under a direct manipulation of X in GComb by
changing W from WI to wz.
The importance of Theorem 7.1 is that whether P(YIZ) is invariant under a direct manipulation
of X in GComb by changing W from WI to Wz is determined by properties of GUnman alone.
Therefore. we will sometimes speak of the invariance of P(YIZ) under a direct manipulation of
X in GUnman without specifying W or GComb.
Each of the preceding examples. and Pratt and Schlaifer's general rule. are consequences of a
corollary to Theorem 7.1:
Corollary 7.1: If GComb is a directed acyclic graph over V u W. W is exogenous

with respect to V in GComb. X and Y are in V. and P(V u W) is a distribution that
satisfies the Markov condition for GComb. then P(YIX) is invariant under direct
manipulation of X in GComb by changing W from WI to Wz if in GUnman no
undirected path into X d-connects X and Y given the empty set of vertices.
Equivalently. if (1) Y is not a (direct or indirect) cause of X. and (2) there is no common
cause of X and Y in GUnman.
In graphical terms. Pratt and Schlaifer's claim amounts to requiring that for "observability"
(invariance) G and G'--the graph of a manipulated system obtained by removing from G all
edges into X--and their associated probabilities must give the same conditional distribution of Y
onX. Corollary 7.1 characterizes the sufficiency side of this claim. Pratt and Schlaifer say their
condition is "almost necessary." What they mean. we take it. is that there are cases in which the
antecedent of their condition fails to hold and the consequent does hold. and. furthermore. that
when the antecedent fails to hold the consequent will not hold unless a special constraint is
satisfied by the conditional probabilities. Parallel remarks apply to the graphical condition we
have given. There exist cases in which there are d-connecting paths between X and Y given the
empty set that are into X and the probability of Y when X is directly manipulated is equal to the
original conditional probability of Y on X. Again the antecedent will fail and the consequent will
hold only if a constraint is satisfied by the conditional probabilities. so the condition is "almost
necessary."
It may happen that the distribution of Y when a value is forced on X cannot be predicted from
the unforced conditional distribution of Y on X but, nonetheless, the conditional distribution of
Yon Z when a value is forced on X can be predicted from the unforced conditional distribution
of Y on X and Z. Pratt and Schlaifer consider the general case in which, besides X and Y,
some further variables Z are measured. Pratt and Schlaifer say that the law relating Y to X is
"observable with concomitant Z" when the unforced conditional distribution of Y on X and Z
equals the conditional distribution of Y on Z in the population in which X is forced to have a
particular value.
Pratt and Schlaifer claim sufficient and "almost necessary" conditions for observability with
concomitants, namely that for any value x of X the distribution of X be independent of the
conditional distribution ofYXf=x on the value of Z ofZXf=x when X is forced to have the value
x. This rule, too, is a special case of Theorem 7.1.
Consider an example due to Rubin. (Rubin'S X is Pratt and Schlaifer's Z; Rubin's T is Pratt
and Schlaifer's X). In an educational experiment in which reading program assignments Tare
assigned on the basis of a randomly sampled value of some pre-test variable X which shares
one or more unmeasured common causes, V, with Y, the score on a post-test, we wish to
predict the average difference TinY values if all students in the population were given treatment
T = I as against if all students were given treatment T = 2. The situation in the experiment is
represented in figure 5.
Figure 5
Provided the experimental sample is sufficiently representative, Rubin says that an unbiased
estimate of T can be obtained as follows: Let k range over values of X, from I to K, let Ylk be
the average value of Y conditional on T = I and X = k, and analogously for Y2k. Let nlk be
the number of units in the sample with T = I and X = k, and analogously for n2k. The numbers
nl and n2 represent the total number of units in the sample with T = I and T = 2 respectively.
Prediction 211
r
Let Tf =1 = expected value of Y if treatment 1 is forced on all units. According to Rubin,
r
estimate Tf =1 by:
K
L nlk + n2k Ylk
k=1 nl+n2
and estimate -r by:
K
L nlk + n2k (Ylk - Y2k)
k=1 nI+n2
The basis for this choice may not be apparent. If we look at the hypothetical population in
which every unit is forced to have T = 1, then it is clear from Rubin's tacit independence
assumptions that he treats the manipulated population as if it had the causal structure shown in
figure 6, as the following derivation shows.
-+
y Tf=l =LY x P(YTf=1) =
Y
-+ K
= = =
LYx LP(Yll =1 IXll=1 k,TTf=1 I)P(XTf=1 k1TTf=1 I)P(TTf=1 = =1) =
Y k=1
-+ K
LYX LP(YTf=1 IXTf=1 =k,TTf=1 =I)P(XTf =1 =k)
Y k=1
The second equality in the above equations hold because P(TTf=1 = 1) = 1, and XTf=1 and TTf=1
are independent according to the causal graph shown in figure 6. By Theorem 7.1, both
P(YTf=1IXTf=t.TTf=1) and P(XT/=1) are invariant under direct manipulation of T in the graph of
figure 5. This entails the following equation.
-+ K
rTf=1 =LY x LP(YTf=1 IXll=1 =k,TTf=1 =I)P(XTf=1 =k) =
Y k=1
fp(x =k)x iy xP(YIX =k,T = 1) = nlk+n2k xYlk

k=1 Y nI+n2
Y~0
Figure 6
Note that X and T, unlike X TJ=! and TTJ=! are not independent. Rubin's tacit assumption that
XTJ=! and TTJ=! are independent indicates that he is implicitly assuming that the causal graph of
the manipulated population is the graph of figure 6, not the graph of figure 5, which is the
causal structure of the unmanipulated population. YTj=2 can be derived in an analogous
fashion.
The reconstruction we have given to Rubin's theory assumes that all units in the population
have the same causal structure for the relevant variables, but not, of course, that the units are
otherwise homogenous. It is conceivable that someone might know the counterfactuals required
for prediction according to the Pratt and Schlaifer rules even though the relevant causal structure
in the population (and in the sample from which inferences are to be made) differs from unit to
unit. For example, it might somehow be known that A and B have no unmeasured common
cause and that B does not cause A, and the population might in fact be a mixture of systems in
which A causes B and systems in which A and B are independent. In that case the distribution
of B if A is forced to have the value A = a can be predicted from the conditional probability of B
given A = a, indeed the probabilities are the same. For this, and analogously for other cases of
prediction for populations with a mixture of causal structures, the predictions obtained by
applying Pratt and Schlaifer's rule can be derived from the Markov Condition by considering
whether the relevant conditional probabilities are invariant in each of the causally homogenous
subpopulations. Thus if A and B have no causal connection, P(BIA = a) equals the probability
of B when A is forced to have value a, and if A causes B, P(BIA =a) equals the probability of
B when A is forced to have value a, and so the probability is also the same in any mixture of
systems with these two causal structures.
Prediction 213
7.4 Prediction with Causal Sufficiency
The Rubin framework is specialized in two dimensions. It assumes known various

counterfactual (or causal) properties, and it addresses invariance of conditional probability. But
we very often don't know the causal structure or the counterfactuals before considering the
data, and we are interested not in invariance per se but only as an instrument in prediction. We
need to be clearer about the goal. We suppose that the investigator knows (or estimates) a
distribution PUnman(O) which is the marginal over 0 of a distribution faithful to an unknown
causal graph GUnman, with unknown vertex set V that includes O. She also knows the variable,
X, that is the member of 0 that will be directly manipulated, and the variables
Parents(GMan,x) that will be direct causes of X in GMan. She knows that X is the only
variable directly manipulated. Finally she knows what the manipulation will do to X, that is,
she knows P Man(XIParents( GMan,x». The distribution of Y conditional on Z is predictable
if in these circumstances PMan(YIZ) is uniquely determined no matter what the unknown causal
graph, no matter what the manipulated and unmanipulated distributions over unobserved
variables, and no matter how the manipulation is brought about consistent with the assumptions
just specified. The goal is to discover when the distribution of Y conditional on Z is
predictable, and how to obtain a prediction.
The assumption that PUnman(O) is the marginal over 0 of a distribution faithful to the
unmanipulated graph GUnman may fail for several reasons. First, it may fail because of the
particular parameters values of the distribution. If W is a set of policy variables, it also may fail
because the W2 (manipulated) subpopulation contains dependencies that are not in the WI
(unmanipulated) subpopulation. For example, suppose that a battery is connected to a light bulb
by a circuit that contains a switch. Let W be the state of the switch, WI be the unmanipulated
subpopulation where the switch is off and W2 be the manipulated subpopulation where the
switch is on. In the WI subpopulation the state of the light bulb (on or off) is independent of the
state of the battery (charged or not) because the bulb is always off. On the other hand in the W2
subpopulation the state of the light bulb is dependent on the state of the battery. Hence in
GComb there is an edge from the state of the battery to the state of the light bulb; it follows that
there is also an edge from the state of the battery to the state of the light bulb in GUnman (which
is the subgraph of GComb that leaves out W.) This implies that the joint distribution over the
state of the battery and the state of the light bulb in the WI subpopulation is not faithful to
GUnman.The results of the Prediction Algorithm are reliable only in circumstances where a
manipulation does not introduce additional dependencies (which mayor may not be part of
one's background knowledge.)
Suppose we wish to make a prediction of the effect of an intervention or policy from

observations of variables correctly believed to be causally sufficient for systems with a common
but unknown causal structure. In that case the sample and the PC (or some other) algorithm
determine a pattern representing a class of directed graphs, and properties of that class
determine whether the distribution of Y following a direct manipulation of X can be predicted.
Suppose for example that the pattern is X - Y - Z which represents the set of graphs in figure 7.
0 • II) .~
(i)
0· II) .~
(ii)
0· II). ~
(iii)
Figure 7
For each of these causal graphs, the distribution of Y after a direct manipulation of X can be
calculated, but the result is different for the first graph than for the two others. P Man(Y) for each
of the graphs can be calculated from the Manipulation Theorem and taking the appropriate
marginal; the results for each graph are given below:
-t
(i) PMan(Y) = L,PUnman(YIX)PMan(X)"# PUnman(Y)
X
(ii)
-t
(iii) PMan(Y) = L,PUnman (YIZ)PUnman (Z) = PUnman(Y)
Z
If every unit in the population is forced to have the same value of X, then for (i) the manipulated
distribution of Y does not equal the unmanipulated distribution of Y. For (ii) and (iii) the
manipulated distribution of Y equals the unmanipulated distribution. Since the pattern does not
Prediction 215
tell us which of these structures is correct, the distribution of Y on a manipulation of X cannot

be predicted.
If a different pattern had been obtained a prediction would have been possible; for example the
pattern U - X -> Y <- Z can represent either of the graphs in figure 8.
@] ·0 .[].
(i)
@]. 0 .[].
(ii)
Figure 8
P Man(Y) for each of the graphs can be calculated from the Manipulation Theorem and taking the
appropriate marginal; the results for each graph are given below:
--+
(i) PMan(Y) = L PUnman(YIX)PMan(X)
x
--+
(ii) PMan(Y) = L PUnman(YIX)PMan(X)
x
(Note however, that while PMan(Y) is the same for (i) and (ii), PMan(U,x,Y,'Z) is not the same
for (i) and (ii), so PMan(U,x,Y,'Z) is not predictable.)
When it is known that the structure is causally sufficient, we can decide the predictability of the
distribution of a variable (or conditional distribution of one set of variables on another set) by
fmding the pattern and applying the Manipulation Theorem and taking the appropriate marginal
for every graph represented by the pattern. If all graphs give the same result, that is the
prediction. Various computational shortcuts are possible, some of which are described in the
Prediction Algorithm stated in the next section.
7.S Prediction without Causal Sufficiency
We come finally to the most serious case, in which for all we know the causal structure of the
manipulated systems will be different from the causal structure of the observed systems, the
causal structure of the observed systems is unknown, and for all we know the observed
statistical dependencies may be due to unobserved common causes. This is the situation that
Mosteller and Tukey seem to think typical in non-experimental studies, and we agree. The
question is whether, nonetheless, prediction is sometimes possible, and if so when and how.
Consider the following trivial example. If we have measured only smoking and lung cancer, we
will find that they are correlated. The correlation could be produced by any of the three causal
graphs depicted in figure 9.
~IS~m!£ok~in!]gulr----t.~1 Lung cancer

(ii)
~mokin8 I. ILung cancer

(iii)
Figure 9
All three graphs yield the same maximally informative partially oriented inducing path graph,
shown in figure 10.
Prediction 217
I Smoking I 0'0-----0 fung Cancer
Figure 10
If Smoking is directly manipulated in graphs (i) or (iii), then P(Lung cancer) will not change;
but if Smoking is directly manipulated in graph (ii) then P(Lung cancer) will change. So it is not
possible to predict the effects of the direct manipulation of Smoking from the marginal
distribution of the measured variables.
In the causally sufficient case each complete orientation of the pattern yields a directed acyclic
graph G. According to the Manipulation Theorem, for each directed acyclic graph GUnman
when we factor the distribution into a product of terms of the form
PUnman(W)(VIParents(GUnman, V) we can calculate the effect of manipulating a variable X
simply by replacing PUnman(W)(XIParents(GUnman,x» with PMan(W)(XIParents(GMan.K)
(where GMan is the manipulated graph). This simple substitution works because each of the
terms in the factorization other than PUnman(W)(XIParents(GUnman,x» is guaranteed to be
invariant under any direct manipulation of X in GUnman, and hence can be estimated from
frequencies in the unmanipulated population.
Let us now try and generalize this strategy to the causally non-sufficient case, where P(O) is the
marginal of a distribution P(V) that is faithful to a directed acyclic graph GUnman, and 1ris the
partially oriented inducing path graph of GUnman. We could search for a factorization of the
distribution of P(O) that is a product of terms of the form PUnman(VIM(V) (where membership
in the set M(V) is a function of V) in which each of the terms except PUnman(XIM(X» is
invariant under all direct manipulations of X in all directed acyclic graphs for which 1r is a
partially oriented inducing path graph over O. If we fmd such a factorization, then we can
predict the effect of the manipulation by substituting the term PMan(XIParents(GMan,x)) for
PUnman(VIM(X)) (where GMan is the manipulated graph), just as we did in the causally
sufficient case. We will not know which of the many directed acyclic graphs for which 1r is a
partially oriented inducing path graph over 0 actually generated the distribution; however, it
will not matter, because P Man(YIZ) will be the same for each of them. This is essentially the
strategy that we adopt. However, the task of fmding such a factorization is considerably more
difficult in the causally non-sufficient case: unlike the causally sufficient case where we can
simply construct a factorization in which each term except P(XIParents(GUnman,x» is
invariant under direct manipulation of X in GUnman, in the causally non-sufficient case we have
to search among different factorizations in order to find a factorization in which each term
except PUnman(XIM(X) is invariant under all direct manipulations of X for all directed acyclic
graphs G that have partially oriented inducing path graph over 0 equal to 1r:. Fortunately, as we
will see, we do not have to search though every possible factorization of P(O).
We will flesh out the details of this strategy and provide examples. We will use the FC!
algorithm to construct a partially oriented inducing path graph 1r: over 0 of GUnman. Note that in
view of Verma and Pearl's example described in Chapter 6, it may be that some graphs for
which 1r: is a partially oriented inducing path graph over 0 may not represent any distribution
with marginal PUnman(O) because of non-independence constraints. From the theory developed
in this book, we cannot hope to provide a computational procedure that decides predictability
and obtains predictions whenever they are possible in principle, because we have no
understanding of all constraints that graphs may entail for marginal distributions. But by
considering only conditional independence constraints we can provide a sufficient condition for
predictability.
Here is an example that provides a more detailed illustration of the strategy: Suppose we
measure Genotype (G), Smoking (S), Income (I), Parents' smoking habits (PSH) and Lung
cancer (L). Suppose the unmanipulated distribution is faithful to the unmanipulated graph that
has the partially oriented inducing path graph shown in figure 11.
I ~L~~L.t-------~L.._ _ _-'
...-----./
IIncome
Parents'
smoking
habits
Figure 11
The partially oriented inducing path graph does not tell us whether Income and Smoking have a
common unmeasured cause, or Parents' smoking habits and Smoking have a common
Prediction 219
unmeasured cause, and so on. The measured distribution might be produced by any of several
structures, including, for example those in figure 12, where Tl and T2 are unmeasured.
If we directly manipulate Smoking so that Income and Parents' smoking habits are not parents
of Smoking in the manipulated graph, then no matter which graph produced the marginal
distribution, the partially oriented inducing path graph and the Manipulation Theorem tell us that
if Smoking is directly manipulated then in the manipulated population the resulting causal graph
will look like the graph shown in figure 13.
Income
Parents'
smoking
hobits
[Tn;:;:;;;;el--.r~~~~rit....Ij---------"ILUng cancer
Figure 12
jIncome S~m~o~k:!in~g:..J------------t.~ILung cancer

1..::1
Parents'
smoking
habits
Figure 13
In this case, we can determine the distribution of Lung cancer given a direct manipulation of
Smoking. Three steps are involved. Here, we simply give the results of carrying out each step.
How each step is carried out is explained in more detail in the next section.
First, from the partially oriented inducing path graph we find a way to factor the joint
distribution in the manipulated graph. Let PUnman be the distribution on the measured variables
and let PMan be the distribution that results from a direct manipulation of Smoking. It can be
determined from the partially oriented inducing path graph that
PMan(/, PSH, S, G, L) = PMan(l) X PMan(PSH) X PMan(S) X PMan(G) X PMan(L IG, S)
where I = Income, PSH = Parents' smoking habits, S = Smoking, G = Genotype, and L =

Lung cancer. This is the factorization of PMan corresponding to the immediately preceding
graph that represents the result of a direct manipulation of Smoking.
Second, we can determine from the partially oriented inducing path graph which factors in the
expression just given for the joint distribution are needed to calculate PMan(L). In this case
PMan(l) and PMan(PSH) prove irrelevant and we have:
Prediction 221
--+
PMan(L) = LPMan(S) X PMan(G) X PMan(LIG,S)
G,S
Third, we can determine from the partially oriented inducing path graph that PMan(G) and
PMan(LIG,S) are equal respectively to the corresponding unmanipulated probabilities,
PUnman(G) and PUnman(LIG,S). Furthermore, PMan(S) is assumed to be known, since it is the
quantity being manipulated. Hence, all three factors in the expression for PMan(L) are known,
and PMan(L) can be calculated.
Note that PMan(L) can be predicted even though P(L) is most definitely not invariant under a
direct manipulation of S. The example should be enough to show that while Mosteller and
Tukey's pessimism about prediction from observation may have been justified when they
wrote, it was not well-founded.
The algorithm sketched in the example is described more formally below, where we have
labeled each step by a letter for easy reference. Suppose PUnman(V) is the distribution before
the manipulation, P Man(V) the manipulation after the distribution, and a single variable X in X
is manipulated to have distribution PMan(XIParents(GMan,K), where GMan is the manipulated
graph. We assume that PUnman(V) is faithful to the unmanipulated graph GUnman, that
Parents(GMan,x) is known, that PMan(XIParents(GMan,x» is known, and that we are
interested in predicting PMan(YIZ). The Prediction Algorithm is simplified by the fact that if
PUnman(O) satisfies the Markov Condition for a graph GUnman, then so does PMan(O), and
hence any factorized expression for PUnman(YIZ) is also an expression for PMan(yIZ). Recall
that a total order Ord of variables in a graph G' is acceptable for G' if and only if whenever A
* B and there is a directed path from A to B in G', A precedes B in Ord. If tr is the FCI partially
oriented inducing path graph of Gover 0, then X is in Definite-Non-Descendants(Y) if
and only if there is no semi-directed path from any Y in Y to X in tr. Recall that a directed
acyclic graph G is a minimal I-map of distribution P if and only if P satisfies the Markov and
Minimality Conditions for G.
Prediction Algorithm
A). PMan(YIZ) = unknown.
B). Generate partially oriented inducing path graph trfrom PUnman(O).
C). For each ordering of variables acceptable for tr in which the predecessors of X in Ord
equals Parents(GMan,x) u Definite-Non-Descendants(X)
C1). Form the minimal I-map F of PUnman(O) for that ordering;
C2). Extract an expression for PUnman(yIZ) from F; call it E;

C3. If for each V *X, the termPUnman(VIParents(F,V» in E is invariant in GMan when X
is directly manipulated then
C3a). return PM a n (Y IZ) = E " where E' is equal to E except that
PUnman(XIParents(F,K) is replaced by PMan(XIParents(GMan,K»
C3b). exit
(The algorithm can also be applied to the case where a set X of variables is manipulated, as long
as it is possible to find an ordering of variables such that for each X in X all of the predecessors
of X are in Definite-Non-Descendants(X) or Parents(GMan,x), there are no causal
connections among the variables in X, and if some X in X is a parent of some variable V not in
X, then every member of X is a predecessor of V.) The description leaves out important
details. How can we find the partially oriented inducing path graph (step B), the graph for
which PUnman(V) satisfies the Minimality and Markov conditions for a given ordering of
variables (step CI), the expression E for PMan(YIZ) (step C2); how do we determine if a given
conditional probability term that appears in the expression for PUnman(YIZ) is invariant under a
direct manipulation of X in GUnman when we do not know what GUnman is (step C3)? The
details are described below.
Step B: We carry out step B) with the FCI Algorithm.
Step C: Say steps CI) and C2) are successful if they produce an expression for PUnman(YIZ)
in which for every V in O\{X}, PUnman(VIParents(F,V» is invariant under direct
manipulation of X in GUnman. We conjecture that if there is an ordering of variables for which
some directed acyclic graph makes CI) and C2) successful, then there is such an ordering that
is acceptable for 1r. (Notice that the correctness of the algorithm does not depend upon the
correctness of this conjecture, although if it is wrong the algorithm will be less informative than
some other algorithm that searches a larger set of variable orderings.)
Step CI: For a given ordering Ord, let Predecessors(Ord, V) be the predecessors of V in
Ord. For each V in F over 0, let Parents(V) be the smallest subset of Predecessors(V) such
that V is independent of Predecessors(Ord,V)\Parents(V) given Parents(V). Then F is a
minimal I-map of P(O). See Pearl (1988). Under the assumption that P(O) is the marginal of a
faithful distribution P (V) we can test whether V is independent of
Predecessors(Ord,V)\Parents(V) given Parents(V) by testing whether each member of
Predecessors(Ord, V)\Parents(V) is independent of V given Parents(V). This clearly
suggests testing whether small sets of variables are equal to Parents(V) first.
Prediction 223
For inducing path graph G' and acceptable total ordering Ord, W is in SP(Ord,G"V)
(separating predecessors of V in G' for ordering Ord) if and only if W precedes V in Ord and
there is an undirected path U between Wand V such that each vertex on U except for the
endpoints precedes V in Ord and is a collider on U. H G is a directed acyclic graph over V, GIP
is the inducing path graph of Gover 0, Ord is an ordering acceptable for GIP, and P(V) is
faithful to G, then the directed acyclic graph GMin in which for each X in ° Parents(X) =
SP(Ord,x) is a minimal I-map of P(O). Of course we are not generally given GIP. However,
we can construct a partially oriented inducing path graph and identify sets of variables that
narrow down the search for SP(Ord,x). For a partially oriented inducing path graph 1r and
ordering Ord acceptable for 1r, let V be in Possible-SP(Ord,x) if and only if V *- X and there
is an undirected path U in 1r between V and X such that every vertex on U except for X is a
predecessor of X in Ord; and no vertex on U except for the endpoints is a defInite-non-collider
°
on U. For a partially oriented inducing path graph 1r over and ordering Ord acceptable for 1r,
V is in Definite-SP(Ord,x) if and only if V *- X and there is an undirected path U in 1r
between V and X such that every vertex on U except for X is a predecessor of X in Ord, and
every vertex on U except for the endpoints is a collider on U.
Theorem 7.2: H P(O) is the marginal of a distribution faithful to G over V, 1r is a

partially oriented inducing path graph of Gover 0, and Ord is an ordering of variables in
° °
acceptable for some inducing path graph over with partially oriented inducing path
graph 1r, then there is a minimal I-map GMin of P(O) in which Definite-SP(Ord,x) in 1r
is included in Parents(GMin,x) which is included in Possible-SP(Ord,x) in 1r.
We can use Theorem 7.2 as a heuristic for searching for a minimal I-map of P(O). The
procedure is only a heuristic for the following reason. While from 1r we can identify orderings
°
that are not acceptable for any inducing path graph over with partially oriented inducing path
graph 1r, we cannot always definitely tell that some ordering acceptable for 1r is acceptable for
some inducing path graph over ° with partially oriented inducing path graph 1r. For orderings
not acceptable for any such inducing path graph over 0, it is possible that making SP(Ord,x)
the parents of X in GMin does not make GMin a minimal I-map, in which. case it may be that no
set M including Definite-SP(Ord,x) and included in Possible-SP(Ord,x) makes
Predecessors(Ord,V)\M independent of X given M. H that is the case, we must conduct a
wider search.
Step C2: If P satisfies the Markov condition for directed acyclic graph G, the following lemma
shows how to determine an expression E for P(YIZ). (For a related result see Geiger, Verma,
and Pearl 1990)
Lemma 3.3.5: If P satisfies the Markov condition for directed acyclic graph Gover
V, then
L
~
II P(WIParents(W»
P(YIZ) = IV(Y,Z) WEIV(Y,Z)uIP(Y,Z)uY
L
~
II
P(WIParents(W»
IV(Y,Z)uY WEIV(Y,Z)ulP(Y,Z)uY
for all values of V for which the conditional distributions in the factorization are
defined, and for which P(z) * o.
Step C3: We use Theorems 7.3 and 7.4 below to determine from 1C whether a given
conditional distribution is invariant under a direct manipulation of X in GUnman. If 1C is a
partially oriented inducing path graph over 0, then a vertex B on an undirected path U in a
partially oriented inducing path graph 1C over 0 is a definite non-collider on U if and only if
B is an endpoint of U or there are edges A *-* B *-* C, A *-* B -> C, or A <- B *-* Con U.
If A * B, and A and B are not in Z, then an undirected path U between A and B in a partially
oriented inducing path graph 1C over 0 is a possibly d-connecting path between A and B
given Z if and only if every collider on U is the source of a semi-directed path to a member of
Z, and every defmite non-collider is not in Z. If Y and Z are disjoint, then X is in Possibly-
IP(Y,Z) if and only if X is in Z, and there is a possibly d-connecting path between X and some
Yin Y given z\{X} that is not out of X. If Y and Z are disjoint, X is in Possibly-IV(Y,Z) if
and only if X is not in Z, there is a possibly d-connecting path between X and some Y in Y
given Z, and there is a semi-directed path from X to a member of Y u Z. Note that Theorems
7.3 and 7.4 also entail that if there is a directed acyclic graph G for which an ordering of
variables is acceptable that makes steps CI and C2 successful, then so does the minimal I-map
for which that ordering is acceptable.
Theorem 7.3: If G is a directed acyclic graph over V u W, W is exogenous with

respect to V in G, 0 is included in V, GUnman is the subgraph of G over V, 1C is the FCI
partially oriented inducing path graph over 0 of GUnman, Y and Z are included in 0, X is
included in Z, Y and Z are disjoint, and no X in X is in Possibly-IP(Y,Z) in 1C, then
Prediction 225
P(YIZ) is invariant under direct manipulation of X in G by changing the value of W from

WI to wz·
Theorem 7.4: If G is a directed acyclic graph over V u W, W is exogenous with respect

to V in G, 0 is included in V, GUnman is the subgraph of G over V, n is the FC! partially
oriented inducing path graph over 0 of GUnman, X, Y and Z are included in 0, X, Y and
Z are pairwise disjoint, and no X in X is in Possibly-IV(Y,Z) in n, then P(YIZ) is
invariant under direct manipulation of X in G by changing the value of W from WI to wz.
The Prediction Algorithm is based upon the construction of a partially oriented inducing path
graph from PUnman(W)(V). Consider the model in figure 14, where the relationships among X,
Z, and T are linear in graph GI, and W is a policy variable.
GraphG l GraphG I '
a= -b e
GraphG Z Graph G z'
Figure 14
Although the distribution over X, Z, and T is not faithful to GI when W =WI if a = -be, the
distribution over X and Z is faithful to GI'. In effect, although the distribution over X and Z
when W = wI is faithful to a directed acyclic graph, it is not faithful to the graph of the causal
process that generated the distribution. Graph G2 depicts the model when X is directly
manipulated by changing the value of W from WI to W2; this makes the coefficient of T in the
equation for X equal to 0, and imposes some new distribution upon X. The manipulated
distribution over X and Z does not satisfy the Markov condition for Gt'; rather it satisfies the
Markov condition for graph G2', which contains an edge between X and Z that Gl' does not
contain. If we were to construct a partially oriented inducing path graph from the unmanipulated
distribution over X and Z it would contain no edges, and make the prediction that the
distribution of Z would be the same in the manipulated and un manipulated distributions, we
would be wrong. Hence the Prediction Algorithm is only guaranteed to be correct when the
unmanipulated distribution is faithful to the unmanipulated graph (which includes the X -> Z
edge because the combined graph contains the X -> Z edge.)
This assumption is not as restrictive as it might first appear. Suppose that we perform an
experiment of the effects of Smoking upon Cancer. We decide to assign each subject a number
of cigarettes smoked per day in the following way. For each subject in the experiment, we roll a
die: if the die comes up I, they are assigned to smoke no cigarettes, if the die comes up 2, they
are assigned to smoke 10 cigarettes per day, etc. Let W = {Experiment} and V = {Die,
Smoking, Drinking, Cancer}. Figure 15 shows the causal graph for the combined population of
experimental and non-experimental subjects, and GUnman. The policy variable is Experiment: it
has the same value (0) for everyone in the non-experimental population, and the same value (1)
for everyone in the experimental population. Die is not a policy variable because it takes on
different values for members of the experimental population.
D~
~ISmoking ~"'.!!:C;'a-nc-e-r---'
GComb GUnman
Figure 15
In this case, the assumption that PUnman(V) is faithful to GUnman is clearly false because the
outcome of the roll of a die and the number of cigarettes smoked by a subject are independent in
the non-experimental population, but there is an edge between them in GUnman. Suppose,
however that we consider the subset of variables V' = {Smoking, Drinking, Cancer}. The
causal graphs that result from marginalizing over V' are shown in figure 16. In this case,
PUnman(V ' ) is faithful to GUnman. Since variables that are causes of Smoking in the
manipulated population but not in the unmanipulated population complicate the analysis, we will
in general simply not consider them. There is no problem in leaving them out of the causal
Prediction 227
graphs, as long as relative to the set of measured variables they are direct causes only of the
manipulated variable. This guarantees that the set of variables that remain after they are removed
is causally sufficient
D~
~moking ~r~c:;-a-n-ce-r---'
GComb GUnman
Figure 16
Theorem 7.5: If G is a directed acyclic graph over V u W, W is exogenous with respect

to V in G, GUnman is the subgraph of G over V, PUnman(W)(V) = P(VIW = Wi) is faithful
to GUnman, and changing the value of W from Wi to W2 is a direct manipulation of X in G,
then the Prediction Algorithm is correct.
The Prediction Algorithm is not complete; it may say that PMan(YIZ) is unknown when it is
calculable in principle.
7.6 Examples
First we consider our hypothetical example from the previous chapter, with the directed acyclic
graph depicted in figure 17, and the partially oriented inducing path graph 11: over 0 = {Income,
Parents' smoking habits, Smoking, Cilia damage, Heart disease, Lung capacity, Measured
breathing dysfunction} depicted in figure 18. We assume that PUnman is faithful to GUnman,
and that in the manipulated graph that Income and Parents' smoking habits are not parents of
Smoking. We will use the Prediction Algorithm to draw our conclusions.
Figure 17
Income Parents'smoking
habits
Figure 18
Prediction 229
We will show in some detail the process of determining that the entire joint distribution of
{Income, Parents' smoking habits, Heart disease, Lung capacity and Measured breathing
dysfunction} is predictable given a direct manipulation of Smoking. Let us abbreviate the names
of the variables in the following way:
Income I
Parents'Smoking Habits PSH
Smoking S
Cilia damage C
Heart disease H
Measured breathing dysfunction M
Lung capacity L
We begin by choosing an ordering for the variables. There are two constraints we impose upon
the orderings. First, the only variables that precede S are those variables that are in Definite-
Non-Descendant(S), and second, the ordering is acceptable for the partially oriented
inducing path graph. That means that I, PSH, and H precede S. Second, in order to be
acceptable for the partially oriented inducing path graph, S, C, L, and M have to occur in that
order. We arbitrarily choose one ordering Ord compatible with these constraints: I, PSH, H, S,
C, L M. (Note that the ordering among the variables that are predecessors of the directly
manipulated variable never matters because each term containing only variables that are
predecessors of the directly manipulated variable is always invariant.)
We generate a directed graph for which PUnman(l,PSH,S,C,H,M,LC) satisfies the Minimality

and Markov conditions. In this case we can determine that any ordering acceptable for the
partially oriented inducing path graph in figure 18 is also an ordering acceptable for the inducing
path graph. Hence, we can apply Theorem 7.2. The resulting factorization is PUnman(l) x
PUnman(PSH) x PUnman(H) x PUnman(SII,PSH) x PUnman(CIS,H) x PUnman(LIC,H,S) x
PUnman(MIC,H ,L).
We now determine which terms in the factorized distribution are needed in order to predict the
conditional distribution under consideration. Because we are predicting the entire joint
distribution, it is trivial that we need every term in the factorized distribution.
Finally, we use the partially oriented inducing path graph to test whether each of the terms
except PUnman(SII,PSH) in the factorized distribution is invariant under direct manipulation of S
in GUnman. PUnman(l), PUnman(PSH), and PUnman(H) are invariant by Theorem 7.4 because
there are no semi-directed paths from S to /, H, or PSH. PUnman(CIS,H) is invariant by

Theorem 7.3 because every path possibly d-connecting path S to C given H is out of S.
PUnman(LIC,S,H), is invariant by Theorem 7.3 because every path possibly d-connecting path
between Sand L given C and H is out of S. Finally PUnman(MIC,H,L) is invariant by Theorem
7.4 because there is no possibly d-connecting path between Sand M given C, H, and L.
Hence, PMan(/,PSH,H,S,C,L,M) = PUnman(/) x PUnman(PSH) x PUnman(H) x PMan(S) X

PUnman(CIS,H) x PUnman(LIC,H,S) x PUnman(MIC,H,L).
In this case, the search was simple because for the given ordering of variables, every term in the
expression for PUnman(/'pSH,H,S,C,L,M) except for PMan(S) is invariant under direct
manipulation of Smoking in GUnman. If the expression had failed this test we would have
repeated the process by generating different orderings of variables, until we had found a
factorized expression of P(/,PSH,H,S,C,L,M) in which each term except PMan(S) was
invariant or we ran out of orderings.
For the next example, consider three alternative models of the relationship between Smoking
and Lung cancer depicted in figure 19. In G}, Smoking causes Lung cancer, and there is a
common cause of Smoking and Lung cancer; in G2, Smoking does not cause Lung cancer, but
there is a common cause of Lung cancer and Smoking; and in G3, Smoking causes Lung
cancer, but there is no common cause of Smoking and Lung cancer.
Prediction 231
Genotype
Income
Parents'
~• Smoking
Tar
• Lung cancer
smoking • deposits
habits
Cilia damage
..........Graph G 1
Genotype
Income
Parents'
smoking
~
• ..........
Smoking - .
Tar
. deposits Lung cancer
habits
Cilia damage Graph G 2
Genotype
Income
Parents'
smoking
habits
~hng -
Tar
• deposits
Cilia damage
GraphG 3
Figure 19
The maximally informative partially oriented inducing path graph of Gl, G2, and G3 over 0 =
{Smoking, Lung cancer} is shown in figure 20.
~moking 1o1----~01 Lung cancer
Figure 20
From this partially oriented inducing path graph it is impossible to determine whether Smoking
causes Lung cancer (as in G3) or Smoking does not cause Lung cancer but there is a common
cause of Smoking and Lung cancer (as in Oz.), or Smoking causes Lung cancer and there is also
a common cause (as in Gl). In addition, we cannot predict the distribution of Lung cancer when
Smoking is directly manipulated. If we try the ordering of variables <.Smoking ,Lung cancer>
then in order to apply the Prediction Algorithm, we need to show that P(Lung cancerlSmoking)
is invariant under direct manipulation of Smoking in GUnlMn. But we cannot use Theorem 7.3
to show that P(Lung cancerlSmoking) is invariant because the Smoking 0-0 Lung cancer edge
guarantees that there is a possibly d-connecting path between Smoking and Lung cancer given
the empty set that is not out of Smoking. This is a quite general feature of the method; it cannot
be used to predict a conditional distribution of Y whenever there is an edge between the variable
X being directly manipulated and Y that has a "0" at the X end. Of course, this feature does not
of itself show that P(Lung cancer) is not predictable by some other method (although in this
example it clearly is not)
Suppose, however, that 0 = {Smoking, Lung cancer,lncome}. If the true graph is G2, the
partially oriented inducing path graph is shown in figure 21.
IIncome r---+ISmoking t'l•1..._--------.. . ogung cancer

Figure 21
By the results of the previous chapter, we can conclude that Smoking does not cause Lung
cancer, because there is no semi-directed path from Smoking to Lung cancer. In this case
P(Lung cancer) is invariant under direct manipulation of Smoking in GUnlMn, so PMan(Lung
cancer) is predictable.
The partially oriented inducing path graphs for Gl and G3 over 0 = {Lung cancer, Smoking,
Income} (shown in figure 22) do not contain enough information in order to determine whether
Prediction 233
Smoking causes Lung cancer. Because in each case there is a Smoking 0-0 Lung cancer edge it
follows that we cannot use the Prediction Algorithm to predictPMan(Lung cancer).
'Income 'o~Smoking 10>----------0 '-_ _ _.....
Partially Oriented Inducing Path Graph of Gl

Over 0 ={Lung Cancer, Smoking, Income}
Vncome '0-----4Smoking lo--------~ogung cancer I
Partially Oriented Inducing Path Graph of G3

Over 0 ={Lung Cancer, Smoking, Income}
Figure 22
If the true graph is G3 it is possible to determine that Smoking causes Lung cancer by also
measuring two causes of Smoking that are not directly connected in the partially oriented
inducing path graph, as in figure 23. Because there is a directed path from Smoking to Lung
cancer in the partially oriented inducing path graph, by the results of the preceding chapter there
is a directed path from Smoking to Lung cancer in the causal graph of the process that generated
the data, and Smoking causes Lung cancer. The output of the Prediction Algorithm is:
-+
PMan(Lung Cancer) = L PMan (Smoking)PUnman (Lung CancerlSmoking)
Smoking
Note that it is not necessary that Parents' Smoking Habits and Income be uncorrelated, or direct
parents of Smoking. The Smoking to Lung cancer edge is oriented by any pair of variables that
have edges that collide at a third variable V, that are not adjacent in the partially oriented
inducing path graph, and such that there is a directed path U from V to Smoking and for every
subpath <X,y,z> of U, X, Y, and Z do not form a triangle.
IIncome ./Lung cancer /
Figure 23
Unfortunately, it is more difficult to determine whether Smoking is a cause of Lung cancer if

Gl is the true causal graph. If 0 = {Smoking, Lung cancer, Income, Parents' Smoking Habits}
and Gl is the true causal graph, without further background knowledge we cannot determine
whether Smoking causes Lung cancer. Figure 24 shows that in the partially oriented inducing
path graph the Smoking to Lung cancer edge is in triangles with Income and Parents' smoking
habits and hence is oriented with an '0' at each end. It follows from the existence of the
Smoking 0-0 Lung cancer edge that we cannot use the Prediction Algorithm to predict P(Lung
cancer) when Smoking is directly manipulated.
Figure 24
It is plausible that Income does not cause Lung cancer directly. If we know from background
knowledge that if there is a causal connection between Income and Lung cancer it contains a
causal path from Smoking to Lung cancer, then we can conclude from the partially oriented
inducing path graph that Smoking does cause Lung cancer.
Prediction 235
Alternatively, if Gl is the correct model, we might try to determine that Smoking is a cause of
cancer by measuring a variable such as Tar deposits, that is causally between Smoking and
Lung cancer. While there is still an induced edge between Income and Lung Cancer in the
partially oriented inducing path graph, Income, Smoking, and Tar deposits are not in a triangle,
and the edge from Smoking to Tar deposits can be oriented. Unfortunately, as figure 25
illustrates, this now leaves one end of the edge between Tar deposits and Lung cancer oriented
with a "0" at one end, so the partially oriented inducing path graph still does not entail that
Smoking causes Lung cancer. And because there is a Smoking 0-0 Lung cancer edge,
PMan(Lung cancer) is not predictable using the Prediction Algorithm.
Figure 25
However, if Gl is the correct model, and we measure a variable between Smoking and Lung
cancer, such as Tar deposits, and another cause of Tar deposits, such as Cilia dmnage, we can
determine that Smoking causes Lung cancer. See figure 26. However, we cannot predict
PMan(Lung cancer) using the Prediction Algorithm because of the Smoking 0-> Lung cancer
edge.
ICilia damage I
Figure 26
We can also determine that Smoking is a cause of Lung cancer by breaking the Income-
Smoking-Lung cancer triangle by measuring all of the common causes of Smoking and Lung
cancer (in this case, Genotype). By measuring all of the common causes of Smoking and Lung
cancer, the edge between Income and Lung cancer is removed from the partially oriented
inducing path graph. This breaks triangles involving Income, Smoking, and Lung cancer, so
that the Smoking to Lung cancer edge can be oriented by the edge between Income and
Smoking, as in figure 27. In addition, PMan(Lung cancer) is predictable. The output of the
Prediction Algorithm is:
PMan(Lung Cancer) =
~
LPMan(Smoking)PUnman(Genotype)PUnman (Lung CancerISmoking,Genotype)

Smoking ,Genotype
Of course, measuring all of the common causes of Smoking and Lung cancer may be difficult
both because of the number of such common causes, and because of measurement difficulties
(as in the case of Genotype). So long as even one common cause remains unmeasured, the
inducing path graph has an Income - Smoking - Lung cancer triangle, and the edge between
Smoking and Lung cancer cannot be oriented.
Although we cannot determine from the partially oriented inducing path graph in figure 27
whether Genotype is a common cause of Smoking and Lung cancer, we can determine that
there is some common cause of Smoking and Lung cancer.
Prediction 237
IGenotype
Parents'
Smoking
Hsbits
Figure 27
7.7 Conclusion
The results developed here show that there exist possible cases in which predictions of the
effects of manipulations can be obtained from observations of unmanipulated systems, and
predictions of experimental outcomes can be made from uncontrolled observations. Some
examples from real data analysis problems will be considered in the next chapter. We do not
know whether our sufficient conditions for prediction are close to maximally informative, and a
good deal of theoretical work remains to be done on the question.
7.S Background Notes
While there are of course many applications that conform to it, we have not been able to find
anticipations of the theory developed in this chapter except in the tradition of work inaugurated
by Rubin. The special case of the Manipulation Theorem that applies when an intervention
makes a single directly manipulated variable X independent of its parents was independently
conjectured by Fienberg in a seminar in 1991.
Chapter 8
Regression, Causation and Prediction
Regression is a special case, not a special subject. The problems of causal inference in
regression studies are instances of the problems we have considered in the previous chapters,
and the solutions are to be found there as well. What is singular about regression is only that a
technique so ill suited to causal inference should have found such wide employment to that
purpose.
8.1 When Regression Fails to Measure Influence
Regression models are commonly used to estimate the "influence" that regressors X have on an
outcome variable, Y.1 If the relations among the variables are linear then for each Xi the
expected change in Y that would be produced by a unit change in Xi if all other X variables are
forced to be constant can be represented by a parameter, say ai. It is obvious and widely noted
(see, for example, Fox, 1984) that the regression estimate of ai will be incorrect if Xi and Y
have one or more unmeasured common causes, or in more conventional statistical terminology,
the estimate will be biased and inconsistent if the error variable for Y is correlated with Xi. To
avoid such errors, it is often recommended (Pratt and Schlaifer, 1988) that investigators enlarge
the set of potential regressors and determine if the regression coefficients for the original
1In linear regression, we understand the "direct influence" of Xi on Y to mean (i) the change in value of a variable
Y that would be produced in each member of a population by a unit change in Xi. with all other X variables
forced to be unchanged. Other meanings might be given, for example: (ii) the population average change in Y for
unit change in Xi. with all other X variables forced to be unchanged; (iii) the change in Y in each member of the
population for unit change in Xi; (iv) the population average change in Y for unit change in Xi; etc. Under
interpretations (iii) and (iv) the regression coefficient is an unreliable estimate whenever Xi also influences other
regressors that influence Y. Interpretation (ii) is equivalent to (i) if the units are homogeneous and the stochastic
properties are due to sampling; otherwise. regression will be unreliable under interpretation (i) except in special
cases. e.g .• when the linear coefficients, as random variables. are independently distributed (in which case the
analysis given here still applies (Glymour. Spirtes and Scheines. 1991a».
Regression, Causation, and Prediction 239
regressors remain stable, in the hope that confounding common causes, if any, will thereby be
measured and revealed. Regression estimates are known often to be unstable when the number
of regressors is enlarged, because, for example, additional regressors may be common causes
of previous regressors and the outcome variable (Mosteller and Tukey, 1977). The stability of a
regression coefficient for X when other regressors are added is taken to be evidence that X and
the outcome variable have no common cause.
It does not seem to be recognized, however, that when regressors are statistically dependent,
the existence of an unmeasured common cause of regressor Xj and outcome variable Y may
bias estimates of the influence of other regressors, Xk; variables having no influence on Y
whatsoever, nor even a common cause with Y, may thereby be given significant regression
coefficients. The error may be quite large. The strategy of regressing on a larger set of variables
and checking stability may compound rather than remedy this problem. A similar difficulty may
arise if one of the measured candidate regressors is an effect, rather than a cause, of Y, a
circumstance that we think may sometimes occur in uncontrolled studies.
To illustrate the problem, consider the linear structures in figure 1, where for concreteness we
specify that exogenous and error variables are all uncorrelated and jointly normally distributed,
the error variables have zero means, and linear coefficients are not zero. Only the X variables
and Y are assumed to be measured. Each set of linear equations is accompanied by a directed
graph illustrating the assumed causal and functional dependencies among the non-error
variables. In large samples, for data from each of these structures linear multiple regression will
give all variables in the set {X1, X2, X3, Xs} non-zero regression coefficients, even though X2
has no direct influence on Y in any of these structures, and X3 has no influence direct or
indirect on Y in structures (i), and (ii), and the effect of X3 in structures (iii) and (iv) is
confounded by an unmeasured common cause. The regression estimates of the influences of X2
and X3 will in all four cases be incorrect. If a specification search for regressors had selected
X 1 alone or X 1 and Xs in (i) or (ii), or Xs alone in (i), (ii), (iii), or (iv), a regression on these
variables would give consistent, unbiased estimates of their influence on Y. But the textbook
procedures in commercial statistical packages will in all of these cases fail to identify {X 1} or
{Xs} or {X 1. Xs} as the appropriate subset of regressors.
Y=alXl +a2X5+ey Y=al Xl + a2X5 + a3 T + ey

Xl = a3X2 + ll4X4 + £1 Xl =a4X2 + a5 X4 + £1
X3 = a5 X2 + a6 Y + E3 X3 = a6 X2 +a7 T + E3
(i) (ii)
Y =al Xl + a2XS + a3 T2 + a4 X3 + ey Y= alXl + a2XS + a3 T2 +a4 X3 + ey

Xl = a5 X2 + a6 X4 + £1 Xl = a5 X2 + a6 X4 + £1
X2=a7 T l +£2 X2=a7 T l +£2
X3 = ag Tl + a9 T2 + £3 X3 = ag Tl + a9 T2 + E3
Xs = alO X6 + au X7 + £S
(iii) (iv)
Figure 1
It is easy to produce examples of the difficulty by simulation. Using structure (i), twenty sets of
values for the linear coefficients were generated, half positive and half negative, each with
absolute value greater than .5. For each system of coefficient values a random sample of 5,000
units was generated by substituting those values for the coefficients of structure (i) and using
uncorrelated standard normal distributions for the exogenous variables. Each sample was given
to MINITAB, and in all cases MINITAB found that {Xl,X2,X3,XS} is the set ofregressors
with significant regression coefficients. The STEPWISE procedure in MlNITAB selected the
same set in approximately half the cases and in the others added X4 to boot; selection by the
lowest value of Mallow's Cp or adjusted R2 gave results similar to the STEPWISE procedure.
The difficulty can be remedied if one measures all common causes of the outcome variable and
the candidate regressors, but unfortunately nothing in regression methods informs one as to
when that condition has been reached. And the addition of extra candidate regressors may create
the problem rather than remedy it; in structures (i) and (ii), if X3 were not measured the
regression estimate of X2 would be consistent and unbiased.
The problem we have illustrated is quite general; it will lead to error in the estimate of the
influence of any regressor Xk that directly causes or has a common direct unmeasured common
cause with any regressor Xi such that Xi and Yhave an unmeasured common cause (or Xi is an
effect of Y). Depending on the true structure and coefficient values the error may be quite large.
It is easy to construct cases in which a variable with no influence on the outcome variable has a
standardized regression coefficient larger than any other single regressor. Completely parallel
problems arise for categorical data.. Recall Theorem 3.4:
Theorem 3.4: If P is faithful to some graph, then P is faithful to G if and only if

(i) for all vertices, X, Y of G, X and Y are adjacent if and only if X and Y are
dependent conditional on every set of vertices of G that does not include X or Y;
and
(ii) for all vertices X, Y, Z such that X is adjacent to Y and Y is adjacent to Z and
X and Z are not adjacent, X -> Y <- Z is a subgraph of'G if and only if X, Z are
Consideration of the first part of this theorem explains why in structure (i) in figure 1 regression
procedures incorrectly select X2 as a variable directly influencing Y: The structure and
distribution satisfy the Markov and Faithfulness conditions, but linear regression takes a
variable Xi to influence Y provided the partial correlation of Xi and Y controlling for all of the
other X variables does not vanish. Part (i) of Theorem 3.4 shows that the regression criterion is
insufficient. It follows immediately from Theorem 3.4 that, assuming the Markov and
Faithfulness Conditions, regression of Y on a set X of variables will only yield an unbiased or
consistent estimate of the influences of the X variables provided in the true structure no X
variable is the effect of Y or has a common unmeasured cause with Y.
Since typical empirical data sets to which multiple regression methods are applied have some
correlated regressors, and in uncontrolled studies it is rare to know that unmeasured common
causes are not acting on both the outcome variable and the regressors, the problem is endemic.
One of the most common uses of statistical methods thus appears to be little more than elaborate
guessing.
8.2 A Solution and Its Application
Assuming the right variables have been measured, there is a straightforward solution to these
problems: apply the PC, FCI, or other reliable algorithm, and appropriate theorems from the
preceding chapters, to determine which X variables influence the outcome Y, which do not, and
for which the question cannot be answered from the measurements; then estimate the
dependencies by whatever methods seem appropriate and apply the results of the previous
chapter to obtain predictions of the effect of manipulating the X variables. No extra theory is
required. We will give a number of illustrations, both empirical and simulated.
We begin by noting that for the twenty samples from structure (i), in every case our
implementation of the PC algorithm--which of course assumes there are no latent variables--
selects {X!. Xs} as the variables that directly influence Y. Our implementation of the FCI
algorithm, which makes no such assumption, in every case says that Xl directly influences Y,
thatXs may, and that the other variables do not.
In each of the other three structures in figure I with sufficiently large samples multiple
regression methods will make comparable errors, always including X2 and X3 among the
"significant" or "best" or "important" variables. In contrast the FCI algorithm together with
Theorems 6.5 through 6.8 give the following results:
Structure Direct Influence No Direct Influence Undetermined

(ii) X2,X3,X4
(iii) X4
(iv) X4,X6,X7
In all of these cases the FCI procedure either determines definitely that X2 and X3 have no
direct influence on Y, or determines that it cannot be decided whether they have any
unconfounded direct influence.
8.2.1 Components of tbe Armed Forces Qualification Test
The AFQT is a test battery used by the United States armed forces. It bas a number of
component tests, including those listed below:
Arithmetical Reasoning (AR)

Numerical Operations (NO)
Word Knowledge (WK)
In addition a number of other tests, including those listed below, are not part of the AFQT but
are correlated with it and with its components:
Mathematical Knowledge (MK)

Electronics Information (El)
General Science (GS)
Mechanical Comprebension (Me)
Given scores for these 8 measures on 6224 armed forces personnel, a linear multiple regression
of AFQT on the other seven variables gives significant regression coefficients to all seven and
thus fails to distinguish the tests that are in fact linear components of AFQT. The covariance
matrix is shown below.
n=6224
AFQT NO WK AR MK EI Me GS
253.9850
29.6490 51.7649
60.3604 6.2931 41.967
57.6566 14.5143 16.0226 40.9329
29.3763 18.2701 13.2055 20.6052 40.7386
36.2318 2.10733 22.6958 16.3664 12.1773 63.1039
35.8244 4.45539 17.4155 20.3952 16.459 35.1981 62.9647
38.2510 5.61516 27.1492 14.7402 14.8442 29.9095 26.6842 48.9300
Given tbe prior information that AFQT is not a cause of any of the other variables, the PC
algorithm in TETRAD II correctly picks out {AR, NO, WK} as the only variables adjacent to
AFQT, and hence the only variables that can be components of AFQT. (Spirtes, Glymour,
Scheines and Sorensen, 1990),2
8.2.2 The Causes of Spartina Biomass
A recent textbook on regression (Rawlings 1988) skillfully illustrates regression principles and
techniques for a biological study in which it is reasonable to think there is a causal process at
work relating the variables. The question at issue is plainly causal: among a set of 14 variables,
which have the most influence on an outcome variable, the weight of Spartina grass? Since the
example is the principal application given for an entire textbook on regression, the reader who
reaches the 13th chapter may be surprised to find that the methods yield almost no useful
information about that question.
According to Rawlings, Linthurst (1979) obtained five samples of Spartina grass and soil from
each of nine sites on the Cape Fear Estuary of North Carolina. Besides the mass of Spartina
(BID), fourteen variables were measured for each sample:
Free Sulfide (H2S)

Salinity (SAL)
Redox potentials at pH 7 (EH7)
Soil pH in water (PH)
Buffer acidity at pH 6.6 (BUF)
Phosphorus concentration (P)
Potassium concentration (K)
Calcium concentration (CA)
Magnesium concentration (MG)
Sodium concentration (NA)
Manganeseconcentration(MN)
Zinc concentration (ZN)
Copper concentration (CU)
Ammonium concentration (NH4)
21n fact, we were inadvertently misinfonned that all seven tests are components of AFQT and we frrst discovered
otherwise with the SGS algorithm.
The correlation matrix is as follows 3:
BJO H7.S SAL EH7 PH BUF P K CA MG NA MN ZN CU NH4

1.0
.33 1.0
-.10 .10 1.0
.05 .40 .31 1.0
.77 .27 -.05 .09 1.0
-.73 -.37 -.01 -.15 -.95 1.0
-.35 -.12 -.19 -.31 -.40 .38 1.0
-.20 .07 -.02 .42 .02 -.07 -.23 1.0
.64 .09 .09 -.04 .88 -.79 -.31 -.26 1.0
-.38 -.11 -.01 .30 -.18 .13 -.06 .86 -.42 1.0
-.27 .00 .16 .34 -.04 -.06 -.16 .79 -.25 .90 1.0
-.35 .14 -.25 -.11 -.48 .42 .50 -.35 -.31 -.22 -.31 1.00
-.62 -.27 -.42 -.23 -.72 .71 .56 .07 -.70 .35 .12 .60 1.0
.09 .01 -.27 .09 .18 -.14 -.05 .69 -.11 .71 .56 -.23 .21 1.0
-.63 -.43 -.16 -.24 -.75 .85 .49 -.12 -.58 .11 -.11 .53 .72 .01 1.0
The aim of the data analysis was to detennine for a later experimental study which of these
variables most influenced the biomass of Spartina in the wild. Greenhouse experiments would
then try to estimate causal dependencies out in the wild. In the best case one might hope that the
statistical analyses of the observational study would correctly select variables that influence the
growth of Spartina in the greenhouse. In the worst case, one supposes, the observational study
would find the wrong causal structure, or would find variables that influence growth in the wild
(e.g., by inhibiting or promoting growth of a competing species) but have no influence in the
greenhouse.
Using the SAS statistical package, Rawlings analyzed the variable set with a multiple regression
and then with two stepwise regression procedures. A search through all possible subsets of
regressors was not carried out, presumably because the candidate set of regressors is too large.
The results were as follows:
3Tbe correlation matrix given in Rawlings (1988) incorrectly gives tbe correlation between CU and NH4 as
0.93.
(i) a multiple regression of BIO on all other variables gives only K and CU significant
regression coefficients;
(ii) two stepwise regression procedures4 both yield a model with PH, MG, CA and CU as
the only regressors, and multiple regression on these variables alone gives them all
significant coefficients;
(iii) simple regressions one variable at a time give significant coefficients to PH, BUP, CA,
ZN andNH4.
What is one to think? Rawling's reports that "None of the results was satisfying to the biologist;
the inconsistencies of the results were confusing and variables expected to be biologically
important were not showing significant effects." (p. 361). This analysis is supplemented by a
ridge regression, which increases the stability of the estimates of coefficients, but the results for
the point at issue--identifying the important variables--are much the same as with least squares.
Rawlings also provides a principal components factor analysis and various geometrical plots of
the components. These calculations provide no information about which of the measured
variables influence Spartina growth.
Noting that PH, for example, is highly correlated with BUP, and using BUP instead of PH
along with MG, CA and CU would also result in significant coefficients, Rawlings effectively
gives up on this use of the procedures his book is about:
Ordinary least squares regression tends either to indicate that none of the variables in a
correlated complex is important when all variables are in the model, or to arbitrarily
choose one of the variables to represent the complex when an automated variable
selection technique is used. A truly important variable may appear unimportant because
its contribution is being usurped by variables with which it is correlated. Conversely,
unimportant variables may appear important because of their associations with the real
causal factors. It is particularly dangerous in the presence of collinearity to use the
regression results to impart a "relative importance," whether in a causal sense or not, to
the independent variables. (p. 362)
Rawling's conclusion is correct about multiple regression and about conventional methods for
choosing regressors, but it is not true of more reliable inference procedures. If we apply the PC
4The "maximum R-square" and "stepwise" options in PROC REG in the SAS program.
algorithm to the Linthurst data then there is one robust conclusion: the only variable that may
directly influence biomass in this population5 is PH; PH is distinguished from all other
variables by the fact that the correlation of every other variable (except MG) with BIO vanishes
or vanishes when PH is controlled for. 6 The relation is not symmetric; the correlation of PH
and BIO, for example, does not vanish when BUF is controlled. The algorithm finds PH to be
the only variable adjacent to BIO no matter whether we use a significance level of .05 to test for
vanishing partial correlations, or a level of 0.1, or a level of 0.2. In all of these cases, the PC
algorithm or the FCI algorithm yield the result that PH and only PH can be directly connected
with BIO. If the system is linear normal and the Causal Markov Condition obtains, then in this
population any influence of the other regressors on BIO would be blocked if PH were held
constant. Of course, over a larger range of values of the variables there is little reason to think
that BIO depends linearly on the regressors, or that factors that have no influence in producing
variation within this sample would continue to have no influence. Nor can the analysis
determine whether the relationship between PH and BIO is confounded by one or more
unmeasured common causes, but the principles of the theory in this case suggest otherwise. If
PH and BIO have a common unmeasured cause T, say, and any other variable, Z, among the
13 others either causes PH or has a common unmeasured cause with PH, then Z and BIO
should be correlated conditional on PH, which appears not to be the case.
The program and theory lead us to expect that if PH is forced to have values like those in the
sample--which are almost all either below PH 5 or above PH 7-- then manipulations of other
variables within the ranges evidenced in the sample will have no effect on the growth of
Spartina. The inference is a little risky, since growing plants in a greenhouse under controlled
conditions may not be a direct manipulation of the variables relevant to growth in the wild. If
for example, in the wild variations in PH affect Spartina growth chiefly through their influence
on the growth of competing species not present in the greenhouse, a greenhouse experiment
will not be a direct manipulation of PH for the system.
The fourth chapter of Linthurst's thesis partly confirms the PC algorithm's analysis. In the
experiment Linthurst describes, samples of Spartina were collected from a salt marsh creekbank
(presumably at a different site than those used in the observational study). Using a 3 x 4 x 2
(PH x SAL x AERATION) randomized complete block design with four blocks, after
transplantation to a greenhouse the plants were given a common nutrient solution with varying
5Although the defmition of the population in this case is unclear, and must in any case be drawn quite narrowly.
6More exactly, at .05, with the exception of MG the partial correlation of every regressor with RIO vanishes
when SOrTU! set containing PH is controlled for; the correlation of MG with RIO vanishes when CA is controlled
for.
values PH and SAL and AERATION. The AERATION variable turned out not to matter in this
experiment. Acidity values were PH 4, 6 and 8. SAL for the nutrient solutions was adjusted to
15, 25, 35 and 45 %0.
Linthurst found that growth varied with SAL at PH 6 but not at the other PH values, 4 and 8,
while growth varied with PH at all values of SAL (p. 104). Each variable was correlated with
plant mineral levels. Linthurst considered a variety of mechanisms by which extreme PH values
might control plant growth:
At pH 4 and 8, salinity had little effect on the performance of the species. The pH appeared
to be more dominant in determining the growth response. However, there appears to be no
evidence for any causal effects of high or low tissue concentrations on plant performance
unless the effects of pH and salinity are also accounted for. (p.108)
The overall effect of pH at the two extremes is suggestive of damage to the root directly,
thereby modifying its membrane permeability and subsequently its capacity for selective
uptake. (p. 109).
A comparison of the observational and experimental data suggests that the PC Algorithm result
was essentially correct and can be extrapolated through the variation in the populations sampled
in the two procedures, but cannot be extrapolated through PH values that approach neutrality.
The result of the PC search was that in the non-experimental sample, observed variations in
aerial biomass were perhaps caused by variations in PH, but were not caused by variations in
other variables. In the observational data Rawlings reports (p. 358) almost all SAL
measurements are around 30--the extremes are 24 and 38. Compared to the experimental study
rather restricted variation was observed in the wild sample. The observed values of PH in the
wild, however, are clustered at the two extremes; only four observations are within half a PH
unit of 6, and no observations at all occurred at PH values between 5.6 and 7.1. For the
observed values of PH and SAL, the experimental results appear to be in very good agreement
with our results from the observational study: small variations in SAL have no effect on
Spartina growth if the PH value is extreme.
8.2.3 The Effects of Foreign Investment on Political Repression
Timberlake and Williams (1984) used regression to claim foreign investment in third-world
countries promotes dictatorship. They measured political exclusion (PO) (i.e., dictatorship),
foreign investment penetration in 1973 (F/), energy development in 1975 (EN), and civil
liberties (CV). Civil liberties was measured on an ordered scale from 1 to 7, with lower values
indicating greater civil liberties. Their correlations for 72 "non-core" countries are:
PO PI EN CV
1.0
-.175 1.0
-.480 .330 1.0
.868 -.391 -.430 1.0
Their inference is unwarranted. Their model and the model obtained from the SGS algorithm
using a .12 significance level to test for vanishing partial correlations) are shown in figure 2. 7
Regression Model SGS Algorithm

Model
Figure 2
The SGS Algorithm will not orient the FJ-EN and EN-PO, edges, or determine whether they
are due to at least one unmeasured common cause. Maximum likelihood estimates of any of the
SGS Algorithm models require that the influence of FJ on PO (if any) be negative, and the
models easily pass a likelihood ratio test with the EQS program. If one of the SGS Algorithm
models is correct, Timberlake and William's regression model appears to be a case in which an
effect of the outcome variable is taken as a regressor, as in structure (i) of figure 1.
This analysis of the data assumes their are no unmeasured common causes. If we run the
correlations through the FCI algorithm using the same significance level, we obtain the
following partially oriented inducing path graph:
7Searches at lower significance levels remove the adjacency between Fl and EN.
~ €r-+"I' Jl
-----------------
Figure 3
The graph together with the required signs of the dependencies, says that foreign investment
and energy consumption have a common cause, as do foreign investment and civil liberties, that
energy development has no influence on political exclusion, but political exclusion may have a
negative effect on energy development, and that foreign investment has no influence, direct or
indirect, on political exclusion.
8.2.4 More Simulation Studies
In the following simulation study we used data generated from the graph of figure 4, which
illustrates some of the confusions that seem to be present in the regression produced by
Timberlake and Williams.
Figure 4
For both the linear and the discrete cases with binary variables, one hundred trials were run at
each of sample sizes 2,000 and 10,000 using the SGS algorithm. A similar set was run using
the PC algorithm for linear and ternary variables. (Each of these algorithms assumes causal
sufficiency.) Results were scored separately for errors concerning the existence and the
directions of edges, and for correct choice of regressors. Let us call the pattern of the graph in
figure 4 the true pattern. Recall that an edge existence error of commission (Co) occurs when
any pair of variables are adjacent in the output pattern but not in the true pattern. An edge
direction error of commission occurs when in an edge occurring in both the true pattern and the
output pattern there is an arrowhead in the output pattern but not the true pattern. Errors of
omission (Om) are defined analogously in each case. The results are tabulated as the average
over the trial distributions of the ratio of the number of actual errors to the number of possible
errors of each kind. The proportion of trials in which both (Both) actual causes of Y were
correctly identified (with no incorrect causes), and in which one (One) but not both causes of Y
were correctly identified (again with no incorrect causes) were recorded for each sample size:
Variable #trials n %Edge Existence %Edge Direction %Both %One

Type Correct Correct
Co Om Co Om
SGS
Linear 100 2000 1.4 3.6 3.0 5.4 85.7 3.6
Linear 100 10000 1.6 1.0 2.7 2.2 90.0 7.0
Binary 100 2000 0.6 16.6 29.5 21.8 38.0 34.0
Binary 100 10000 1.2 7.4 30.0 9.1 60.0 25.0
PC
Linear 100 2000 6.0 2.0 1.0 6.2 80.0 15.0
Linear 100 10000 0.0 1.0 2.5 2.9 95.0 0.0
Ternary 100 2000 3.0 1.0 29.1 8.3 65.0 35.0
Ternary 100 10000 3.0 2.0 10.8 1.2 85.0 15.0
The differences in the results with the SGS and PC algorithms for discrete data are due to the
choice of binary variables in the former case and ternary variables in the latter case. The tests for
statistical independence with discrete variables appear to have more power when variables can
have more than two values.
For purposes of prediction and policy, the numbers in the last two columns suggest that the
procedure quite reliably finds real causes of the outcome variable when the statistical
assumptions of the simulations are met, the sample is large and a causal structure like that in
figure 4 obtains. Regression will in these cases find that all of the regressors influence the
outcome variable.
8.3 Error Probabilities for Specification Searches
We have shown that various algorithms for specifying causal structure from the data are correct
if the requisite statistical decisions are correctly made, but we have given no results about the
probability of various sorts of errors in small and medium size samples. The Neyman-Pearson
account of testing has made popular two measures of error: the probability of rejecting the null
hypothesis when it is true (type I), and the probability of not rejecting the null hypothesis when
an alternative is true (type ll). Correspondingly, when a search procedure yields a model M
from a sample, we can ask for the probability that, were the model M true, the procedure would
not fmd it on samples of that size, and given an alternative M', we can ask for the probability
that were M' true the search procedure would find M on samples of that size. We shall also
refer to the error probabilities for the outcomes of search procedures as probabilities of type I
and type n errors respectively. Especially in small samples, the significance levels and powers
of the tests used in deciding conditional independence may not be reliable indicators of the
probabilities of corresponding errors in the search procedure.
Error probabilities for search procedures are nearly impossible to obtain analytically, and we
have recommended that Monte Carlo methods be used instead. When a procedure yields M
from a sample of size n, estimate M and use the estimated model to generate a number of
samples of size n, run the search procedure on each and count the frequency with which
something other than M is found. For plausible or interesting-alternative models M', estimate
M', use the estimated model to generate a number of samples of size n, run the search
procedure on each and count the frequency with which M is found. We will illustrate the
determination of error probabilities for specification searches with a case in which probability of
type IT error is quite high.
Weisberg, (1985) illustrates a procedure for detecting outliers with an experimental study for
which regression produces anomalous results:
An experiment was conducted to investigate the amount of a particular drug present in

the liver of a rat. Nineteen rats were randomly selected, weighed, placed under light
ether anesthesia and given an oral dose of the drug. Because it was felt that large livers
would absorb more of a given dose than smaller liver, the actual dose an animal received
was approximately determined as 40 mg of the drug per kilogram of body weight. ...
The experimental hypothesis was that, for the method of determining the dose, there is
no relationship between the percentage of the dose in the liver (Y) and the body weight
(XI), liver weight (X2), and relative dose (X3). (p. 121-124)
Regressing Yon (X 10 X2, X3) gives a result not in agreement with the hypothesis; the
coefficients of Y on body weight (X I) and dose (X3) are both significant, even though one is
determined by the other. We find the following regression values for Weisberg's data (standard
errors are parenthesized and below the coefficients, and t statistics are shown just below the
standard errors):
Y = -3.902*XI + .197*X2 + 3.995*X3 + e

(1.345) (.218) (1.336)
-2.901 .903 2.989
A multiple regression not including X2 also yields significant regressors for X I and X3 at the
.05 level. Yet, Weisberg observes that no individual regression of Yon anyone of the X
variables is significant at that level. The results of the several statistical decisions are therefore
= = =
inconsistent; we have, for example, that PXt Y 0, PX2Y 0, PX3Y 0 but PXtY.X3 *" o. One
might take any of several views about such inconsistencies. On.e is that it is largely an artifact of
the particular significance level used. If the .01 level were used to reject hypotheses of
vanishing correlations and partial correlations, the correlations with Y would vanish and so
would the partial correlations controlling for one other variable. But the partial correlation of X I
with Y controlling for both of the other regressors could not be rejected, and an inconsistency
would remain. Another view is that inconsistencies in the outcomes of sequential statistical
decisions are to be expected, especially in small samples, and where possible, inferences should
be based on the statistical decisions that are most reliable. In the case at hand the power of any
of the statistical tests is low because of the sample size, but the lower the order of the partial
correlation the greater the power. The rule of thumb is that to control for an extra variable is to
throwaway a data point. Thus in this case the PC algorithm never considers the partial
correlations and concludes solely from the vanishing correlations that none of the X variables
cause the Y variable. Weisberg instead recommends excluding one of the 19 data points, after
which a multiple regression using the remaining data gives no (.05) significant regression
coefficients.
From the experimental setup we can assume that body weight and liver weight are causally prior
to dose, which itself is prior to the outcome, i.e., the amount of the drug found in the rat's
liver. Applying the PC algorithm to the original data set with this background knowledge, and
using the .05 significance level in the program's tests, we get the pattern in figure 5.
The PC algorithm gives the supposed correct result in this case because no correlation of an X
variable with Y is significant, and that is all the program needs to decide absence of influence.
The regression of Y against each individual X variable alone is an essentially equivalent test To
estimate the type I error of the PC search, we obtained a maximum likelihood estimate
(assuming normal distributions) of the model shown in figure 5 and used it to generate 100
simulated data sets each of size 19. The PC algorithm was then applied to each data set. In four
of the 100 samples the procedure erroneously introduced an edge between Yand one or another
of the X variables.
X I = Body Weight
X 2 = Liver Weight
X3 = Dose
Y = Drug Found in Liver
Model Output by the PC

Algorithm
Figure 5
To investigate the power of the procedure against alternatives, we consider three models in
which Y is connected to at least one X variable. The first is simply the regression model with
correlations among the regressors estimated from the sample. With a correlation of about .99
between Xl and X3, the regression model with correlated errors is nearly unfaithful, and we
should expect the search to be liable not to fmd the structure. We generated 100 samples at each
of the sizes 19,50, 100 and 1000. We then ran the PC algorithm on each sample, counting the
output as a type 2 error if it included no edge between Y and some X variable. Figure 6 gives
the results for the first three sample sizes.
10
90-
80-
70-
60-
50-
40-
30-
20-
10-
I I I
20 40 60 80 100
Sample size
Figure 6: Percentage of Samples from the Regression Model for which PC

Omits at Least One Edge
In 100 trials at sample size 1,000 PC never makes a type 2 error against this alternative.
The second alternative is an elaboration of the original PC output. We add an edge from body
weight to the outcome, giving the graph in figure 7.
X 1 = Body Weight
X 2 = Liver Weight I--~~
X3=Dose
Y = Drug Found in Liver
Figure 7
We estimated this model with the EQS program, which found a value of .228 for the linear
coefficient associated with the Xl -> Y edge, and then used the estimated model to again
generate 100 samples at each of the four sample sizes. The results for the first three sample
sizes are shown in figure 8.
100-
90 -
80 -
70 -
60 -
50 -
40 -
30 -
20 -
10 -
I I I
20 40 60 80 100
Sample size
Figure 8: Percentage of samples from the model of figure 7 for which PC omits
Xl -> Y
Even at sample size 1,000 the search makes an error of type 2 against this alternative in 55% of
the cases. "Small" influences of body weight on Y cannot be detected. We would expect the
same to be true of dose.
In the third case we increased the linear coefficient connecting Xl and Y in the model in figure 7
to 1.0.
100
90
80
70
60
50
40
30
20
10
20 40 60 80 100
Sample size
Figure 9: Proportion of samples from the model of figure 7 for which PC omits
Xl .> Y
At sample size 1,000 the search makes an error against this alternative in 2% of the cases.
For this problem at small sample sizes the search has little power against some alternatives, and
little power even at large sample sizes against alternatives that may not be implausible.
8.4 Conclusion
In the absence of very strong prior causal knowledge, multiple regression should not be used to
select the variables that influence an outcome or criterion variable in data from uncontrolled
studies. So far as we can tell, the popular automatic regression search procedures should not be
used at all in contexts where causal inferences are at stake. Such contexts require improved
versions of algorithms like those described here to select those variables whose influence on an
outcome can be reliably estimated by regression. In applications, the power of the specification
searches against reasonable alternative explanations of the data is easy to determine by
simulation and ought to be investigated.
It should be noted that the present state of the algorithm is scarcely the last word on selecting
direct causes. There are cases in which a partially oriented inducing path graph of a directed
acyclic graph Gover 0 contains a directed edge from X to Y even though X is not a direct
cause of Y relative to 0 (although of course there is a directed path from X to Y in G.)
However, Theorem 6.8 states a sufficient condition for a directed edge in a partially oriented
inducing path graph to entail that X is a direct cause of Y. In some cases tests based on
constraints such as Verma and Pearl's, noted in section 6.9, would help with the problem, but
they have not been developed or implemented.
Chapter 9
The Design of Empirical Studies
Simple extensions of the results of the preceding chapters are relevant to the design of empirical
studies. In this chapter we consider only a few fundamental issues. They include a comparison
of the powers of observational and experimental designs, some implications for sampling and
variable selection, and some considerations regarding ethical experimental design. We conclude
with a reconsideration from the present perspective of the famous dispute over the causal
conclusions that could legitimately be drawn from epidemiological studies of smoking and
health.
9.1 Observational or Experimental Study?
There are any number of practical issues about both experimental and non-experimental studies
that will not concern us here. Questions of the practical difficulty of obtaining an adequate
random sample aside, when can alternative possible causal structures be distinguished without
experiment and when only by experiment?
Suppose that one is interested in whether a treatment T causes an outcome O. According to

Fisher, R.Fisher (1959) one important advantage of a randomized experiment is that it
eliminates from consideration several alternatives to the causal hypothesis to be tested. If the
value ofT is assigned randomly, then the hypothesis that 0 causes T or that there is an
unmeasured common cause of 0 and T can be eliminated. Fisher argues that the elimination of
this alternative hypothesis greatly simplifies causal inference; the question of whether T causes
o is reduced to the question of whether T is statistically dependent on O. (This assumes, of
course, instances of the Markov and Faithfulness Conditions.)
Critics of randomized experiments, e.g. Howson and Urbach (1989), have correctly questioned
whether randomization in all cases does eliminate this alternative hypothesis. The treatments
given to people are typically very complex and change the values of many random variables.
For example, suppose one is interested in the question of whether inhaling tobacco smoke from
cigarettes causes lung cancer. Imagine a randomized experiment in which one group of people
is randomly assigned to a control group (not allowed to smoke) and another group is randomly
assigned to a treatment group (forced to smoke 20 cigarettes a day.) Further imagine that the
experimenter does not know that an unrecorded feature of the cigarettes, such as the presence of
a chemical in some of the paper wrappings of the cigarettes, is the actual cause of lung cancer,
and inhaling tobacco smoke does not cause lung cancer. In that case lung cancer and inhaling
tobacco smoke from cigarettes are statistically dependent even though inhaling tobacco smoke
from cigarettes does not cause lung cancer. They are dependent because assignment to the
treatment group is a common cause of inhaling tobacco smoke from cigarettes and of lung
cancer.
Fisher (1951, p. 20) suggests that "the random choice of the objects to be treated in different
ways would be a complete guarantee of the validity of the test of significance, if these
treatments were the last in time of the stages in the physical history of the objects which might
affect their experimental reaction." But this does not explain how an experimenter who does not
even suspect that cigarette paper might be treated with some cancer causing chemical could
know that he had not eliminated all common causes of lung cancer and inhaling tobacco smoke
from cigarettes, even though he had randomized assignment to the treatment group. This is an
important and difficult question about randomization, made more difficult by the fact that
randomization often produces deterministic relationships between such variables as drug dosage
and treatment group, producing violations of the Faithfulness Condition.
In this section we will put aside this question, and simply assume that an experimenter has
some method that correctly eliminates the possibility that 0 causes T or that there are common
causes of 0 and T. In general, causal inferences from experiments are based on the principles
described in chapters 6 and 7. The theory applies uniformly to inferences from experimental and
from non-experimental data. Inferences to causal structure are often more informative when
experimental data is available, not because causation is somehow logically tied to experimental
manipulations, but because the experimental setup provides relevant background causal
knowledge that is not available about non-experimental data. (See Pearl and Verma 1991 for a
similar point.)
The Design of Empirical Studies 261
There are, of course, besides the argument that randomization eliminates some alternative causal
hypotheses from consideration, a variety of other arguments that have been given for
randomization. It has been argued that it reduces observer bias; that it warrants the experimenter
assigning a probability distribution to the outcomes conditional on the null (causal) hypothesis
being true, thereby allowing him to perform a statistical test and calculate the probability of type
I error; that for discrete random variables it can increase the power of a test by simulating
continuity; and that by bypassing 'nuisance factors' it provides a basis for precise confidence
levels. We will not address these arguments for randomization here; for a discussion of these
arguments from a Bayesian perspective see e.g. Kadane and Seidenfeld (1990).
Consider three alternative causal structures, and let us suppose for the moment that they exhaust
the possibilities and are mutually exclusive: (i) A causes C, (ii) some third variable B causes
both A and C, or (iii) C causes A. If by experimental manipulation we can produce a known
distribution on A not caused by B or C, and if we can produce a known distribution on C not
caused by A or B, we can distinguish these causal structures. In the experiment, all of the edges
into A in the causal graph of the non-experimental population are broken, and replaced by an
edge from U to A; furthermore there is no non-empty undirected path between U and any other
variable in the graph that does not contain the edge from U to A. Any procedure in which A is
caused only by a variable U with these properties we will call a controlled experiment. In a
controlled experiment we know three useful facts about U: U causes A, there is no common
cause of U and C, and if U causes C it does so by a mechanism that is blocked if A is held
constant (Le. in the causal graph if there is a directed path from U to C it contains A.). As we
noted in Chapter 7, U is not a policy variable and is not included in the combined, manipulated
or unmanipulated causal graphs.
The controlled experimental setups for the three alternative causal structures are shown in figure
1, where an A-experiment represents a manipulation of A breaking the edges into A, and a C-
experiment represents a manipulation of C breaking edges into C. If we do an A-experiment and
find partially oriented inducing path graph (ia*) over {A,C} then we know that A causes C
because we know that we have broken all edges into A. (We do not include U (or V) in the
partially oriented inducing path graphs in figure I because including them does not strengthen
the conclusions that can be drawn in this case, but does complicate the analysis because of the
possible deterministic relationships between U and A.) Similarly, if we perform a C-experiment
and find partially oriented inducing path graph (iiic*) then we know that C causes A. If we
perform an A-experiment and get (iia*) and a C-experiment and get (iic*) then we know that
there is a latent common cause of A and C (assuming that A and C are dependent in the non-
experimental population.)
Partially Partially
Model A-Experiment Oriented C-Experiment Oriented
Inducing Inducing
@J4-0 ~
Path Graph Path Graph
~~ ~-+@] ~o-o[£] ~ [£]

(i) (ia) (ia*) (ie) (ie*)
d\ (ii)
~\ (iia)
~ @]
(iia*)
1@J4-0 ~ (iie)
@]
(iie*)
~+0 ~ @] ~ @] ~~ ~o-o@]
(iii) (iiia) (iiia*) (iiie) (iiie*)
Figure 1
Now suppose that in the non-experimental population there are variables U and V known to
bear the same relations to A and C respectively as in the experimental setup. (We assume in the
non-experimental population that A is not a deterministic function of U, and C is not a
deterministic function of V.) That is, U causes A, there is no common cause of U and C, and if
there is any directed path from U to C it contains A; also, V causes C, there is no common
cause of V and A, and if there is any directed path from V to A it contains C. Can we still
distinguish (i), (ii), and (iii) from each other without an experiment? The answer is yes. In
figure 2, (io*), (iio*) and (iiio*) are the partially oriented inducing path graphs corresponding to
(io), (iio), and (iiio) respectively. Suppose the FCI algorithm constructs (io*). If it is known
that U causes A, then from the fact that the edge between U and A and the edge between A and
C do not collide, we can conclude that the edge between A and C is oriented as A -> C in the
inducing path graph. It follows that A causes C. Similarly, if the FCI algorithm constructs
(iiio*) ideally we can conclude that C causes A. The partially oriented inducing path graph in
(iio*) indicates by Theorem 6.9 that there is a latent common cause of A and C, and by
Theorem 6.6 that A does not cause C, and C does not cause A.
(iio) (iio*)
(iiio) (iiio*)
Figure 2
Note that if we had measured variables such as W, U, V, and X in figure 3 then the
corresponding partially oriented inducing path graphs would enable us to distinguish (i), (ii),
and (iii) without experimentation and without the use of any prior knowledge about the causal
relations among the variables.
ffiJ,," ./fKl
~~ (io)
@]~~[]
(io**)
ffiJ~ ./fKl
@]~~~[]
(iio**)
(iiio) (iiio**)
Figure 3
Consider now the more complex cases in which the possibilities are (i) A causes C and there is
a latent common cause B of A and C, (ii) there is a latent common cause B of A and C, and (iii)
C causes A and there is a latent common cause B of A and C. Each of the structures (i), (ii) and
(iii) can be distinguished from the others by experimental manipulations in which for a sample
of systems we break the edges into A and impose a distribution on A and for another sample we
break the edges into C and impose a distribution on C. The corresponding graphs are presented
in figure 4, and the analysis of the experiment is essentially the same as in the previous case.
Partially Partially
Model A-Experiment Oriented C-Experiment Oriented
Inducing Inducing
A ~
[]~ lTIo-o@] []
! 1£14-0 lTI @]
(i) (ia) (ia*) (ic) (ic*)
~ ~~ E1 11£14-0 E1
(ii) (iia)
@]
(iia*) (iic)
@]
(iic*)
A ~~ !
(iii) (iiia)
~ @] ~~ @]o-o@]
(iiia*) (iii c) (iiic*)
Figure 4
The analysis of the corresponding non-experimental case is more complicated. Assume that
there is a variable U and it is known that U causes A, there is no common cause of U and A,
and if there is any directed path from U to C it contains A, and that there is a variable V and it is
known that V causes C, there is no common cause of V and A, and if there is any directed path
from V to A it contains C. The directed acyclic graphs and their corresponding partially oriented
inducing path graphs are shown in figure 5. Now suppose that the directed acyclic graphs are
true of an observed non-experimental population. Can we still distinguish (i), (ii), and (iii) from
each other?
~ (io)
~~~ (io*)
(iio) (iio*)
(iiio) (iiio*)
Figure 5
Once again the answer is yes. For example, suppose that an application of the FeI algorithm
produces (io*). The existence of the U 0-> C edge entails that either there is a common cause of
U and C or a directed path from U to C. By assumption, there is no common cause of U and C,
so there is a directed path from U to C. Also by assumption, all directed paths from U to C
contain A, so there is a directed path from A to C. Given that there is an edge between U and C
in the partially oriented inducing path graph, and the same background knowledge, it also
follows that there is a latent common cause of A and C. (The proof is somewhat complex and
we have placed it in an Appendix to this chapter.) Similarly, if we obtain (ilio*) then we know
that C causes A and there is a latent common cause of A and C. If we obtain (iio*) then we
know that A and C have a latent common cause but that A does not cause C and C does not
cause A. It is also possible to distinguish (i), (ii), and (iii) from each other without any prior
knowledge of particular causal relations, but it requires a more complex pattern of measured
variables, as shown in figure 6. If we obtain (io**) then we know without using any such prior
knowledge about the causal relationships between the variables that A causes C and that there is
a latent common cause of A and C, and similarly for (iio**) and (ilio**).
There is an important advantage to experimentation over passive observation in one of these

cases. By performing an experiment we can make a quantitative prediction about the
consequences of manipulating A in (i), (ii), and (ili). But if (i) is the correct causal model, we
cannot use the Prediction Algorithm to make a quantitative prediction of the effects of
manipulating A. (In the linear case, a prediction could be made because U serves as an
"instrumental variable.")
(io) (io*)
(iio) (iio*)
(iiio) (iiio*)
Figure 6
Suppose fmally that we want to know whether there are two causal pathways that lead from A
to C. More specifically, suppose we want to distinguish which of (i), (ii) and (iii) in figure 7
obtains, remembering again that B is unmeasured.
,
Partially Partially
Model A.Experiment C·Experiment Oriented
Oriented
Inducing Inducing
A ~
~ II]<H>[£] II] ~ II] [£]
(i) (ia) (ia*) (ie) (ie*)
~ ~\ E1 ,~ E1
(ii) (iia)
@]
(iia*) (iie)
[£]
(iie*)
A ~ ,~ (iii) (iiia)
~<H>@]
(iiia*) (iiie)
~ ~
(iiie*)
Figure 7
The question is fairly close to Blyth's version of Simpson's paradox. By experimental

manipulation that breaks the edges directing into A and imposes a distribution on A, we can
distinguish structures (i) and (iii) from structure (ii) but not from one another. Note that in
figure 7 the partially oriented inducing path graph (ia*) is identical to (ilia*) and (ic*) is identical
to (ilic*).
Assume once again that in a non-experimental population it is known that U causes A, there is
no common cause of U and C, and if there is any path from U to C it contains A, and V causes
C, there is no common cause of V and A, and if there is any path from V to A it contains C. The
directed acyclic graphs and their corresponding partially oriented inducing path graphs are
shown in figure 8.
~ (io) (io*)
~ (iio) (iio*)
(iiio) (iiio*)
Figure 8
Unlike the controlled experimental case, where (i) and (iii) cannot be distinguished, in the non-
experimental case they can be distinguished. Suppose we obtain (iiio*). We know from the
background knowledge that U causes A, and from (iiio*) that the edge between U and A does
not collide with the edge between A and C. Hence in the corresponding inducing path graph
there is an edge from A to C and in the corresponding directed acyclic graph there is a path from
A to C. (Of course we cannot tell how many paths from A to C there are; (iiio*) is compatible
with a graph like (iiio) but in which the <A,B,C> path does not exist.) We also know that there
is no latent common cause of A and C because (iiio*) together with our background knowledge
entails that there is no path in the inducing path graph between A and C that is into A. Suppose
on the other hand that we obtain (io*). Recall that the background knowledge together with the
partially oriented inducing path graph entail that A is a cause of C and that there is a latent
common cause of A and C. (We have placed the proof in an Appendix to this chapter.)
Once again if more variables are measured, it is also possible to distinguish these three cases
without any background knowledge about the causal relationships among the variables, as
shown in figure 9.
(io)
(iio) (iio*)
(iiio) (iiio*)
Figure 9
Thus all three structures can be distinguished without experimental manipulation or prior
knowledge.
It may seem extraordinary to claim that structure (i) in figure 7 cannot be distinguished from
structure (iii) by a controlled experiment, but can be distinguished without experimental control
if the structure is appropriately embedded in a larger structure whose variables are measured. It
runs against common sense to claim that when A causes C, a controlled experiment cannot
distinguish A and C also having an unmeasured common cause from A also having a second
mechanism through which it effects B, but that observation without experiment sometimes can
distinguish these situations. But controlled experimental manipulation that forces a distribution
on A breaks the dependency (in the experimental sample) of A on B in structure (i), and thus
information that is essential to distinguish the two structures is lost.
While a controlled experiment alone cannot distinguish (i) from (iii) in figure 7 the combination
of a simple observational study and controlled experimentation can distinguish (i) from (iii). We
can detennine from an A-experiment that there is a path from A to C, and hence no path from C
to A. We know if P(CIA) is not invariant under manipulation of A then there is a trek between C
and A that is into A. Hence if P(CIA) is different in the non-experimental population and the A-
experimental population we can conclude that there is a common cause of A and C.1f P(CIA) is
invariant under manipulation of A then we know that either there is no common cause of A and
C or the particular parameter values of the model "coincidentally" produce the invariance. By
combining information from an observational study and an experimental study it is sometimes
possible to infer causal relations that cannot be inferred from either alone. This is often done in
an informal way. For example, suppose that in both an A-experiment and a C-experiment A and
C are independent. This indicates that there is no directed path from A to C or C to A. But it
does not distinguish between the case where there is no common cause of A and C (i.e. there is
no trek at all between A and C) and the case where there is a common cause of A and C. Of
course in practice these two models are distinguished by determining whether A and C are
independent in the non-experimental population; assuming faithfulness, there is a trek between
A and C if and only if A is not independent of C.
In view of these facts the advantages of experimental procedures in identifying (as distinct from
measuring) causal relations need to be recast. There are, of course, well known practical
difficulties in obtaining adequate non-experimental random samples without missing values but
we are interested in issues of principle. One disadvantage of non-experimental studies is that in
order to make the distinctions in structure just illustrated either one has to know something in
advance about some of the causal relations of some of the measured variables, or else one must
be lucky in actually measuring variables that stand in the right causal relations. The chief
advantage of experimentation is that we sometimes know how to create the appropriate causal
relations. A further advantage to experimental studies is in identifying causal structures in mixed
samples. In the experimental population the causal relation between a manipulating variable and
a manipulated variable is known to be common to every system so treated. Mixing different
causal structures acts like the introduction of a latent variable, which makes inferences about
other causal relations from a partially oriented inducing path graph more difficult. Similar
conclusions apply to cases in which experimental and statistical controls are combined.
In the "controlled" experiments we have discussed thus far, we have assumed that the
experimental manipulation breaks all of the edges into A in the causal graph of the non-
experimental population, and that the variable U used to manipulate the value of A has no
common cause with C. However, it is possible to do informative experiments that satisfy
neither of these assumptions. Suppose, for example, the causal graph of figure 10 describes a
non-experimental population.
Figure 10
Suppose that in an experiment in which we manipulate A, we force a distribution upon P(AIU).

In this case the causal graph of the experimental population is the same as the causal graph of
the non-experimental population, although of course the parametrization of the graph is
different in the two populations. This kind of experiment does not break the edges into A. More
generally, we assume that there is a set of variables U used to influence the value of A such that
any direct cause V of A that is not in U is connected to some outcome variable C only by
undirected paths that contain some member of U as a non-collider. (This may occur for
example, if U is a proper subset of the variables used to fix the value of A, and the other
variables used to fix the value of A are directly connected only to A.) These are just the
conditions that we need in order to guarantee the invariance of the distribution of a variable C
given U and A, and hence allows the use of the Prediction Algorithm. (A more extensive
discussion of this kind of experiment is given in section 9.4.) With an experiment of this kind,
it is possible to distinguish model (i) from model (iii) in figure 7. Of course, with the same
background knowledge assumptions it is also possible to distinguish (i) from (iii) in a non-
experimental study in which the distribution of P(AIU) is not changed. Indeed with this kind of
experiment, the only difference between the analysis of the experimental population and a non-
experimental population lies in the background knowledge employed.
9.2 Selecting Variables
The selection of variables is the part of inference that at present depends almost entirely on
human judgment. We have seen that poor variable selection will usually not of itself lead to
incorrect causal conclusions, but can very well result in a loss of information. Discretizing
continuous variables and using continuous approximations for discrete variables both risk
altering the results of statistical decisions about conditional independence.
One fundamental new consideration in the selection of variables is that in the absence of prior
knowledge of the causal structure, empirical studies that aim to measure the influence, if any, of
A on C or of C on A, should try to measure at least two variables correlated with A that are
thought to influence C, if at all, only through A, and likewise for C. As the previous section
illustrates, variables with these properties are especially informative about whether A causes C,
or C causes A, or neither causes the other but there is a common cause of A and C.
The strategy of measuring every variable that might be a common cause of A and B and
conditioning on all such variables is hazardous. If one of the additional variables is in fact an
effect of B, or shares a common cause with A and a common cause with B, conditioning on that
variable will produce a spurious dependency between A and B. That is not to say that extra
variables should not be measured if it is thought that they may be common causes; but if they
are measured, they should be analyzed by the methods of chapters 5 and 6 rather than by
multiple regression.
Finally, if methods like those described in chapters 5, 6 and 7 are to be employed, we offer the
obvious but unusual suggestion that variables be selected for which good conditional
independence tests are available. At present our simulations suggest that is a reason to avoid
binary variables where possible--for example in psychometric and sociometric test design.
9.3 Sampling
We can view many sampling designs as procedures that specify a property S, which may have
two values or several, and from subpopulations with particular S values draw a sample in
which the distribution of values of the jth unit drawn is distributed independently of and
identically to the distribution of all other sample places from that subpopulation. In the simplest
case S can be viewed as a binary variable with the value I indicating that a unit has the sample
property, which of course does not mean that the unit occurs in any particular actual sample.
We distinguish the sample property S from any treatments that might be applied to members of
the sample. In sampling according to a property S we obtain information directly not about the
general population but about the segments of the population that have various values of S. Our
general questions therefore concern when conditioning on any value of S in the population
leaves unaltered the conditional probabilities or conditional independence relations for variables
in the causal graph G describing the causal structure of each unit in the population. That is,
suppose there is a population in which the causal structure of all units is described by a directed
graph G, and let the values of the variables be distributed as P, where P is faithful to G. What
are the causal and statistical constraints a sampling property S must satisfy in order that a sub-
population consisting of all units with a given value of S will accurately reflect the conditional
independence relations in P--and thus the causal structure G--and under what conditions will the
conditional probabilities for such sub-populations be as in P? The answers to these questions
bear on a number of familiar questions about sampling, including the appropriateness of
retrospective versus prospective sampling and of random sampling as against other sampling
arrangements. We will not consider questions about the sampling distributions obtained by
imposing various constraints on the distribution of values of S in a sample. Our discussion
assumes that S (which may be identical to one of the variables in G) is not determined by any
subset of the other variables in G.
We assume in our discussion that S is defined in such a way that if the sampling procedure
necessarily excludes any part of the population from occurring in a sample, then the excluded
units have the same S value. For example, if a sample is to be drawn from the sub-population
of people over six feet tall, then we will assume that S = 0 corresponds to people six feet tall or
under and S = 1 corresponds to people over 6 feet tall.
The causal graph G relating the variables of interest can be expanded to a graph G(S) that
includes S and whatever causal relations S and the other variables realize. We assume a
distribution P(S) faithful to G(S) whose marginal distribution summing over S values will of
course be P. We suppose that the sampling distribution is determined by the conditional
distribution P( I S). Our questions are then, more precisely, when this conditional distribution
has the same conditional probabilities and conditional independence relations as P. We require,
moreover, that the answer be given in terms of the properties of the graph G(S). The following
theorem is obvious and will not be proved.
Theorem 9.1 If P(S) is faithful to G(S), and X and Y are sets of variables in G(S) not
containing S, then P(YIX) = P(YIX,S) if and only if X d-separates Y and S in G(S).
Our sampling property should not be the direct or indirect cause or effect of Y save through a
mechanism blocked by X, and X should not be the effect, direct or indirect of both Y and the
sampling property. (The second clause in effect guarantees that Simpson's paradox is avoided
in a faithful distribution). The theorem is essentially the observation that P(YIX u Z) = P(YIX
u Zu (S)) if and only if in P Y and S are independent conditional on X u Z. It entails, for
example, that if we wish to estimate the conditional probability of Y on X from a sample of
units with an S property (say, S = 1), we should try to ensure that there is
(i) no direct edge between any Yin Y and S,
(ii) no trek between any Yin Y and S that does not contain some X in X, and
(iii) no pair of directed paths from any Yin Y to an X in X and from S to X.
Figure 11 illustrates some of the ways estimation from the sampling property can bias estimates
of the conditional probability of Y given X and Z.
(i) (ii)
(iii)
Figure 11
Cases (i) and (iii) are typical of retrospective designs. In case (ii) the sampling property biases
estimates of P(YIX,Z) because Y and the sample property S are dependent conditional on
{X,Z}. Theorem 9.1 amounts to a (very) partial justification of the notion that: "prospective"
sampling is more reliable than "retrospective" sampling, if by the former is meant a procedure
that selects by a property that causes or is caused by Y, the effect, if at all only through X, the
cause, and by the latter is meant a procedure that selects by a property that causes or is caused
by X only through Y. In a prospective sampling design in which X is the only direct cause of
S, and S does not cause any variable, the estimate of P(YIX,z) is not biased. But case (ii)
shows that under some conditions prospective samples can bias estimates as well.
Similar conclusions should be drawn about random sampling. Suppose as before that the goal
is to estimate the conditional probability P(YIX) in distribution P. In drawing a random sample
of units from P we attempt to sample according to a property S that is entirely disconnected with
the variables of interest in the system. If we succeed in doing that then we ensure that S has no
causal connections that can bias the estimate of the conditional probability. Of course even a
random sample may fail if the very property of being selected for a study (a property, it should
be noted, different from having some particular value of S) affects the outcome, which is part of
the reason for blinding treatments. Further, any property that has the same causal disconnection
will do as a basis for sampling; there is nothing special in this respect about randomization,
except that a random S is believed to be causally disconnected from other variables.
When the aim is only to determine the causal structure, and not to estimate the distribution P or
the conditional probabilities in P, the asymmetry between prospective and retrospective
sampling vanishes.
In models (i) and (iii) of figure 11, which are examples of retrospective design, for any three
disjoint sets of variables A, B, and C not containing S, A is d~separated from B given C if and
only if A is d-separated from B given C u S. So these cases in which conditional probability in
P cannot be determined from S samples are nonetheless cases in which conditional
independence in P, and hence causal structure, can be determined from S samples.
Theorem 9.2 states conditions under which the set of conditional independence relations true in
the population as a whole is different from the set of conditional independence relations true in a
subpopulation with a constant value of S. In Theorem 9.2 let Z be any set of variables in G not
including X and Y.
Theorem 9.2: For a joint distribution P, faithful to graph G, exactly one of <Y 11 XIZ;
Y 11 XIZ u {S}> is true in P if and only if the corresponding member and only that
member of <Z d-separates X;Y; Z u {S} d-separates X, Y> is true in G.
Although Theorem 9.2 is no more than a restatement of Theorem 3.3, its consequences are
rather intricate. Suppose that X and Y are independent conditional on Z in distribution P. When
will sample property S make it appear that X and Yare instead dependent conditional on Z?
The answer is exactly when X, Y are dependent conditional on Z u S in P(S). This
circumstance--conditional independence in P and conditional dependence in P(S)--can occur for

faithful distributions when and only when there exists an undirected path U from X in X to Y in
Y such that
i. no noncollider on U is in Z u IS};
ii. every collider on U has a descendant in Z u IS};
iii. some collider on U does not have a descendant in Z.
The converse error involves conditional dependence in P and conditional independence in P(S).
That can happen in a faithful distribution when and only when there exists an undirected path U
from X to Y such that
i. every collider on U has a descendant in Z;
ii. no noncollider in U is in Z;
and S is a noncollider on every such path. Again, asymptotically both of these errors can be
avoided by sampling randomly, or by any property S that is unconnected with the variables of
interest.
In experimental designs the aim is sometimes to sample from an ambient population, apply a
spectrum of treatments to the sampled units, and then infer from the outcome the effect a policy
of treatment would have if applied to the general population. In the next section we consider
some relations between experimental design, policy prediction, and causal reasoning.
9.4 Ethical Issues in Experimental Design
Clinical trials of alternative therapies have at least two ethical problems. (1) In the course of the
trials (or even beforehand) suspicion may grow to near certainty that some treatments are better
than others; is it ethical to assign people to treatments, or to continue treatments, found to be
less efficacious? (2) In clinical trials, whether randomized or other, patients are generally
assigned to treatment categories; if the patients were not part of an experimental design,
presumably they would be free to choose their treatment (free, that is, if they could pay for it, or
persuade their insurer to); is it ethical to ask or induce patients to forego choosing? Suppose the
answer one gives to each of these questions is negative. Are there experimental designs for
clinical trials that avoid or mitigate the ethical problems but still permit reasonable predictions of
the effects of treatment throughout the population from which the experimental subjects are
obtained?
Kadane and Sedransk (1980) describes a design (jointly proposed by Kadane. Sedransk. and
Seidenfeld) to meet the fIrst problem. Their design has been used in trials of drugs for post-
operative heart patients and in other applications. The inferences Kadane and Seidenfeld (1990)
make in explaining the design are in accord with the Markov Condition. and indeed follow from
it. and the case nicely illustrates the role of causal reasoning in experimental design.
Furthermore. combining the Markov and Faithfulness Conditions. and using the Manipulation
Theorem and Theorem 7.1 leads to two novel conclusions:
1. The effIcacy of treatments in an experiment can be reliably assessed in an experimental

procedure that takes patient preference into account in allocating treatment, but except in special
cases the knowledge so acquired could not be used to predict the effects of a general policy of
treatment.
2. Perhaps of more practical interest. given comparatively weak causal knowledge on the part of
the experts. another design in which treatment allocation depends on patient preference can be
used to determine whether or not patient self-assignment in the experiment will confound
prediction of the outcome of a treatment policy in the general population. When all influences
are linear. the effects of treatment policy can be predicted even if confounding occurs.
9.4.1 The Kadane/SedransklSeidenfeld Design
In the Kadane/Sedransk/Seidenfeld experimental design (described in Kadane and Seidenfeld

(1990». for each member of a panel of experts. degrees of belief are elicited about the outcome
a of each treatment T conditional on each profIle of values of XI .. X n. The elicited judgments
are used to specify some prior distribution over parameters in a model of the treatment process.
For each experimental subject. the panel of experts receives information on the variables Xl •
...• Xn. Nothing else about the patient is known to the experts. Based on the values of Xl •...•
Xn each expert i recommends a preferred treatment Pi(X) to the patient, and the patient is
assigned to treatment by some rule T = h(X. PI •... ,pk) that is a function of the X values and the
experts' treatment preferences (pt. ...Pk) for patients described by X. and perhaps some random
factor. The rule guarantees that no patient is given a treatment unless at least one expert
recommends it for patients with that profIle. The model determines the likelihood. for each
vector of parameter values. of outcomes conditional on X and T values. As data are collected on
patients. the prior distribution over the parameters is updated by conditioning. If the evidence
reaches a stage at which all the experts agree that some treatment T for patients with profIle X is
not the best treatment for the patient, then treatment T is suspended for such patients. As
evidence accrues, the experts' degrees of belief about the parameter values of the likelihood
model should converge.
Let Xj be a vector of observed characteristics of the}h patient, "including those that are used as
a basis for deciding what treatment each patient is to receive, and possibly other characteristics
as well." (We do not place Xj in boldface in order to match Kadane and Seidenfeld's notation.)
Let Tj be the treatment assigned to patientj. Let OJ be the outcome for patientj. Let Pj = (OJ, Tj,
Xj, OJ-I, Tj-l, Xj_}, ... , XI) be the past evidence up to and including what is known about
e
patient j. Let be a vector of the parameters of interest, those that determine the probabilities of
outcomes OJ for a patient j given characteristics Xj and treatment Tj. For example, the degrees
of belief of an expert might be represented by a mixture of linear models parametrized by
exogenous variances, means and linear coefficients. A unique value of these parameters then
"determines" the probability of an outcome given X values. For reasons that will become clear,
e
it is essential to the definition of that alternative values for the parameter not give alternative
specifications of the distribution of X variables.
The expression t e(.P]) represents the expert's conditional degree of belief, given e, that the total
evidence is p]. Kadane and Seidenfeld add that "It is part of the definition of eas the parameter
that
What this means is that econtains all the information contained in Pj-l that might be useful for
predicting OJ from Tj and Xj." The factorization of degree of belief,
t B(P]) = [TItB(OjITj,Xj,Pj-l)][(ItB(TjIXj,Pj-l)][TItB(XjIPj-l)]
J=1 J=1 J=1
2 3
follows by the definition of conditional probability. The terms are marked 1, 2, and 3. Kadane
and Seidenfeld claim that term 3 does not depend on e if one believes that the features,
treatments and outcomes for earlier subjects in the experimental trial have no influence on "the
kinds of people" who subsequently become subjects. (Recall that parameters relevant to the
distribution of Xl .. Xn are not included in e.). Kadane and Seidenfeld also say that term 2 does
not depend on ebecause there is a fixed rule for treatment assignment as a function of X values
and the history of the experimental outcomes.
It follows from (1) that
J J
IIf9(OjIT j ,Xj ,Pj - 1 ) = IIf9(OjIT j ,Xj )
j=l j=l
Kadane and Seidenfeld say the proportionality given by this term:
J
f9(PJ )oc IIf9(OjIT j ,Xj )
j=l
"is the form that we use to evaluate the results of a clinical trial of the kind considered here."
That is for each value 9; of 9, multiplying f 9, (PJ) by the prior density of 9; gives a quantity
proportional to the posterior density of 9;. The ratios of the posterior densities of two values of
9; can therefore be found.
Now for a new case, each value of 9 determines a probability of treatment outcome given an X
profile and a treatment T, and so the posterior distribution ,of 9 yields, for anyone expert,
degrees of belief in the outcomes of various treatment regimes to various classes of patients.
Although Kadane and Seidenfeld say nothing explicit about predicting the effects of policies of
treatment, these degrees of belief may be transformed into expected values if outcome is
somehow quantified. In any case, an expert who began believing that a rule of treatment given
by T = k(X) would most often result in a successful outcome, may come instead to predict that
a different rule of treatment, say T =g(X) will more often be successful.
Why can't the experiment let the subjects simply choose their own treatments, and seek any
advice that they want? Kadane and Seidenfeld give two reasons. One is that if patients were to
determine their own treatment, the argument that term 2 in the factorization does not depend on
9 would no longer hold. The other is that "It would now be necessary to explain statistically the
behavior of patients in choosing their treatments, and there might well be contamination
between these choices and the effect of the treatment itself." We will consider the force of these
considerations in the next subsection.
9.4.2 Causal Reasoning in the Experimental Design
What is it that the experts believe that warrants this analysis of the experiment, or the derivation
of any predictions? The expert surely entertains the possibility that some unknown common
causes U may influence both the X features of a patient and the outcome of the patient's
treatment. And yet the analysis assumes that in the expert's degrees of belief, treatment is
independent of any such U conditional on X values. That is implicit in the claim that term 2 in
the factorization is independent of 9. Why should treatment T and unknown causes U be
conditionally independent given X? The reason, clearly, is that in the experiment the only
factors that influence the treatment a patient receives are the X values for that patient and Pj-l;
any such U, should it exist, has no influence on T except through X. A causal fact is the basis
for independence of probabilities.
Aspects of the expert's understanding of the experimental set-up are pictured in figure 12
(where we have made Xj a single variable.)
Figure 12
The expert may not be at all sure that the edges with "?" correspond to real influences, but she is
sure there is no influence of the kind in figure 13 in boldface.
Figure 13
The experimental design, which makes treatment assignment a function of the X variables and
Pj-l only, is contrived to exclude such influences. The expert's thought seems to be that if U
influences T only through X, then U and T are independent conditional on X. That thought is
an instance of the Markov Condition. The probabilities in the Markov Condition can be
understood either objectively or subjectively. But in the Kadane/SedranskiSeidenfeld design the
probability in the Markov Condition cannot be the expert's unconditional degrees of belief,
because those probabilities are mixtures over 0 of distributions conditional on 0, and mixtures
of distributions satisfying the Markov Condition do not always satisfy the Markov Condition.
We will assume the distributions conditional on Odo so.
Consider another feature of the idealized expert belief. An idealized expert in Kadane and
Seidenfeld's experiment changes his probability distribution for a parameter whose values
specify a model of the experimental process. At the end of the experiment the expert has a view
not only about the outcome to be expected for a new patient with profile X if that patient were
assigned treatment according to the rule T = h(X'pj-l) used in the experiment, but also about the
outcome to be expected for a new patient with profile X if that patient were assigned treatment
according to the rule T =g(X) that the expert now, in light of the evidence, prefers. In
principle, the expert's probabilities for outcomes if the new patient were treated by the
experimental rule T = h(X'pj-l) is easy to compute because that probability is determinate for
each value of 0, and we know the expert's posterior distribution of O. But what determines the
expert's probability for outcomes if the patient with profile X is now treated according the
preferred rule, T = g(X)? Why doesn't changing the rule change the dependence of 0 on X and
T? The sensible answer implicit in Kadane and Seidenfeld's analysis, is that the outcome for
any patient depends on the X profile of the patient and the treatment given to the patient, but not
on the "rule" by which treatments are assigned. Changing the assignment rule changes the
probability of treatment T given profile X, but has no effect on other relevant conditional
probabilities; the probability 0 given T and X is unaltered. We can derive this more formally in
the following way.
If for a fixed value of e the distribution !£I0j,Tj)(j'pj_l) satisfies the Markov condition for
graphs of the type in figure 12, then Theorem 7.1 entails that!£I0jITj)(j) is invariant under a
manipulation of Tj. According to Theorem 7.1, in a distribution that satisfies the Markov
condition for graphs of the type in figure 12'/£1OjITj)(j) is invariant under a manipulation of Tj
if there is no path that d-connects OJ and Xj given Tj that is into Tj. Every undirected path
between Tj and OJ that contains some Xj variable satisfies this condition because some member
of Xj is a non-collider on such a path. There are no undirected paths between Tj and OJ that
contain Pj-l. Hence!£I0jITj)(j) is invariant under manipulation ofTj-
But does the Markov condition reasonably apply in an experiment designed according to the
Kadane/SedranskiSeidenfeld specifications in which !()(OJ,Tj)(j,Pj-l) does not represent
frequencies but an expert's opinions? We are concerned with the circumstance in which the
experiment is concluded, and the expert's degrees of belief, we will suppose, have converged
so far as they will. The expert is uncertain as to whether there are common causes of X and
outcome, or how many there are, but all of the causal structures she entertains are like figure 12
e
and none are like figure 13. Conditional on and any particular causal hypothesis we suppose
the Markov condition is satisfied, but the expert's actual degrees of belief are some mixture over
different causal structures. Should !£I0jITj)(j) be invariant under manipulation of Tj in the
e
opinion of the expert when her distribution for a given value of is a mixture of several
different causal hypotheses? The answer is yes, as the following argument shows.
Let us call the experimental (unmanipulated in the sense of Chapter 7) population Exp, and the
hypothetical population subjected to some policy based on the results of the experiment Pol. Let
!(),Exp(Oj,Tj)(j,Pj-l) represent the expert's degrees of belief about OJ, Tj, Xj, and Pj-l
conditional on ein the experimental population, and!fJ,po/(Oj,Tj)(j'pj_l) represent the expert's
degrees of belief about OJ, Tj, Xj, and Pj-l conditional on e in the hypothetical population
subjected to some policy. Let CS be a random variable that denotes a causal structure. We have
already noted that it follows from Theorem 7.1 that
Because edetermines the density of OJ conditional on Tj and Xj,

/9,Exp(OjIT j ,Xj ,CS) = / 9,Exp(OjITj ,X j)

/9,Po/(OjIT j ,Xj ,CS) = /9,Po/(OjIT j ,Xj )
Hence,
Consider next the question of "bias" raised by Kadane and Seidenfeld. The very notion requires
us to consider not just degrees of belief but also some facts and some potential facts. We
suppose there is really a correct (or nearly correct) value for the parameters in the likelihood
model, and the true values describe features of the process that go on in the experiment. We
suppose the expert converges to the truth, so that her posterior distribution is concentrated
around the true values. What the public that pays for these experiments cares about is whether
the expert's views about the best treatment are correct: Would a policy that puts in place the
expert's pre/erred rule o/treatment, say T = g(X), result in better outcomes than alternative
policies under consideration? One way to look at that question is to ask if the expert's expected
values for outcome conditional on X profile and treatment roughly equal the population mean
for outcome under these conditions. If degrees of belief accord with population distributions,
that is just to ask when the frequency of 0 conditional on T ,and X that would result if every
relevant person in the population were treated on the basis of the experimental assignment rule T
= h(X'pj-l) would be roughly the same as the frequency of 0 conditional on T and X for the
general population if a revised rule T =g(X) for assigning treatments were used. In other
words: Will the frequency 0/0 conditional on T and X be invariant under a direct manipulation
ofT!
As we have just seen for this case, the Markov Condition and Theorem 7.1 entail that for the
graph in figure 12, and all others like it (those in which every trek whose source is a common
cause of OJ and Tj contains an Xj variable, OJ does not cause Tj, and every common cause of an
Xj variable and Tj is an Xj variable) the probability of OJ on Tj and Xj is invariant under a direct
manipulation of Tj. No other assumptions are required. The example is a very special case of a
general sufficient condition for the invariance of conditional probabilities under a direct
manipulation.
Consider next why the experimental design forbids that treatment assignment depend directly on
"unrecorded" features of the patients, such as Y. Suppose such assignments were allowed;
Kadane and Seidenfeld say the outcome might be "contaminated" which we understand to mean
that some unrecorded causes of the patient preference, and hence of Tj, might also be causes of
OJ. So that we have the causal picture in figure 14.
Figure 14
The question marks indicate that, we (or the expert), are uncertain about whether the
corresponding causal influences exist. Suppose the directed edges from Uj to Y and from Uj to
OJ exist. Then there is an undirected path between OJ and Tj that contains Y. In this case the
Markov condition entails that the probability of OJ conditional on Tj and the Xj variables is not
invariant under a direct manipulation of Tj except for "coincidental" parameter values. So if
unrecorded Y values were allowed to influence assignments then for all we or the experts
know, the expert's prediction of the effects of his proposed rule T = g(X) would be wrong.
Let us now return to the question of why patient preference cannot be used to influence
treatment. The reason why patient preferences cannot be used to determine treatment assignment
in the experiment is not only because there may be, for all one knows, a causal interaction
between patient preference and treatment outcome. It is true that if it were known that no such
confounding occurs, then patient preference could be used in treatment assignment, but why
cannot such assignment be used even if there is confounding? In order to make treatment
assignments depend on patient preference (and presumably also on other features, such as the
Xj variables) the patients' preferences must be ascertained. If the preferences are known, why
not conditionalize outcome on Preference, T and the Xj variables, just as we have
conditionalized outcome on T and the Xj variables? The probability of OJ conditional on
Preference, the Xj variables, and Tj has no formally different role than the probability of OJ
conditional on the Xj variables and Tj. If in figure 14 we make Y =Preference, then according
to the Markov condition and Theorem 7.1, the probability of OJ conditional on Tj,Xj and
Preference is invariant under a manipulation of Tj. Of course some precautions would have to
be taken in the course of an experiment that allows patient preference to influence treatment
assignment. There is not much point to allowing a patient to choose his or her treatment unless
the choice is informed. In the experimental setting, it would be necessary to standardize the
information and advice that each patient received.
Could this design actually be used to predict the effects of a policy of treatment? At the end of
the study the expert has a density function of outcomes given T, X, and Preference; subsequent
patients' preferences for treatment would have to be recorded (but not necessarily used in
determining treatment). If the announced results of the study alter Preference, the probability of
outcome conditional on Preference, X and T depends on whether influences represented by the
edges adjacent to U in figure 12 exist. But if patients are informed of the experimental results
we must certainly expect in many cases that their preferences will be changed, i.e. announcing
the result of the experiment is a direct manipulation of Preference. Reliable predictions could
only be made if the experimental outcome were kept secret!
The reason that patient preferences cannot be used for assigning treatment is therefore, not just
because their preferences might have complicated interactions with the outcome of the treatment-
-so might the X variables. No analysis that does not also consider how a policy changes
variables that were relevant to outcome in an experimental study can give a complete account of
when predictions can be relied upon. As Kadane has pointed out l , it is more likely that the
causes of Preference in the experimental population are different from the causes of Preference
in the non-experimental population than it is that the causes of X in the experimental popUlation
are different from the causes of X in the non-experimental population. In the case we are
considering, announcements of experimental results (or of recommendations) about policies that
use patient preferences for assigning treatment can generally be expected to directly change
those very preferences--whereas policies that use the X variables for assigning treatment do not
generally change the values of the X variables for people in the popUlation. (Of course, there
may be instances where the results of a study that does not base treatment on patient preference
also directly manipulates the distribution of the X variables, in which case the prediction of
outcome conditional on treatment and the X variables would also be unreliable. Suppose,
fancifully, in experimental trials that assign treatment as a function of cholesterol levels, it were
found that a certain drug is very effective against cancer for subjects with low cholesterol.)
IPersonal communication.
The Kadane/SedranskiSeidenfeld design thus reveals an ethical conundrum that conventional

methodological prejudices against non-randomized trials has hidden. There is an obligation to
find the most effective and cost effective treatments, and an obligation to take into account in
treatment the preferences of people who participate as subjects in clinical trials. Both can be
satisfied. But there is also an obligation fully to inform patients about the relevant scientific
results that bear on decisions about their treatment. This obligation is incompatible with the
others.
9.4.3 Towards Ethical Trials
Finally, we can use the causal analysis to obtain some more optimistic results about patient
selection of treatment in experimental trials. Suppose in an experiment treatment assignment is a
function T = h(Xj, Preference,Pj_j), and every undirected path between Preference and OJ
contains some member of Xj as a non-collider. (If this is the case then we will say that
Preference is not confounded with 0.) Then it can be shown strictly as a consequence of the
Markov and Faithfulness Conditions that the probability of OJ conditional on Tj and Xj is
invariant under a direct manipulation of Tj; we mayor may not conditionalize on Preference, or
take Preference into account in the treatment rule used in policy, and in that case whether or not
the announced experimental results changes the distribution of preferences is irrelevant to the
accuracy of predictions. Now in some cases it may very well be that patient preference is not
confounded with treatment outcome. If investigators could in fact discover that Preference is not
confounded with 0, then they could let such preferences be a factor both in treatment
assignments in the experimental protocol and in the recommended policy. Kadane and
Seidenfeld say that such dependencies, if they exist, are undetectable. If the experts are
completely ignorant about what factors do not influence the outcome of treatment, Kadane and
Seidenfeld are right; but if the experts know something, anything, that varies with patients and
that has no effect on outcome except through treatment and has no common cause with
outcome, we disagree. The something could be the phase of the moon on the patient's last
birthday, the angular separation of the sun and Jupiter on the celestial sphere on the day of the
patient's mother's day of birth, or simply the output of some randomizing device assigned to
each patient. How is that?
In any distribution faithful to the graph of figure 15, E and C are dependent conditional on B.
The relation is necessary in linear models, without assuming faithfulness.
rv Figure 15
Now, returning to our problem, let Z be any feature whatsoever that varies from patient to
patient, and that the experts agree in regarding as independent of patients' preferences for
treatments and as affecting outcome only through treatment. Adopt a rule in the experiment that
makes treatment a function of Preference, the patient's X profile, Pj-l, and the patient's Z value.
Then the expert view of the causal process in the experiment looks something like figure 16:
Figure 16
If OJ and Zj are independent conditional on Tj and Xj then there is (assuming faithfulness) no

path d-connecting Preferencej and OJ given Xj that is into Preferencej. A confounding relation
between Preferencej and OJ, if it exists, can be discovered from the experimental data.
(Similarly, if Tj and OJ are dependent given Xj because the experimental population consists of
a mixture of causal structures, then OJ and Zj are dependent conditional on Tj and Xj unless
some particular set of parameter values produces a "coincidental" independence.) Indeed, on the
rather brave assumption that all dependencies are linear, Zj is an instrumental variable (Bowden
and Turkington, 1984) and the linear coefficient representing the influence of T conditional on
o and X can be calculated from the correlations and partial correlations.
This suggests that it is possible to do a pilot study to determine whether Preference is
confounded with 0 in the experimental population. In the pilot study, Preference can be a factor
influencing, but not completely determining, T. If the results of the pilot study indicate that
Preference is not confounded with 0 in the experimental popUlation, a larger study in which
Preference completely determines T can be done; otherwise, the KadaneiSedransklSeidenfeld
design can be employed.
The goal of a medical experiment might be to predict outcomes in a population where a policy of
assigning treatments without consulting patient preference is adopted. For example, the
question might be "What would the death rate be if only halothane were used as a general
anesthetic?" In this case, patient preference has little or nothing to do with the assigned
treatment If patient preference is not used to assign treatment in the policy population there is
no reason to think that predictions of P(OIX,T) in the policy population will be inaccurate when
based upon experiments in which patients choose (or at least influence) their treatment, and
Preference and 0 are not confounded.
It might be, however, that the goal of the experiment is to predict P(OIX,T) in the policy
population, and to let the patients choose or at least influence the choice of the treatment they
receive. For example, in choosing between lumpectomy and mastectomy, patient preference
may be the deciding factor. In this case there are a number of reasons to question the accuracy
of a prediction of P(OIX,T) in the policy population based upon the design we propose. But in
this case every design meets the same difficulties, whether or not patients have assigned
themselves in experimental treatments. These are equally good reasons for questioning the
accuracy of a prediction of P(OIX,T) (interpreted as frequencies or propensities and not as
degrees of belief) in the policy population based upon the Kadane/Sedransk/Seidenfeld or a
classical randomized design. The fundamental problem is that there are any number of plausible
ways in which the causal relationships among preference and other variables in the experimental
population may be different from the causal relationships in the policy population.2 For
example, in the experimental population the assignment of treatment will not depend on the
patient's income. However, in the actual population, the choice of treatment may very well
depend upon income. There could easily be a common causal pathway connecting income and
outcome that does not contain any variable in the patient's X profIle. Again, in the experimental
population the information and advice patients receive can be standardized. We can also try and
ensure that the advice given is a function only of the patient's X profIle. In the policy
population, however, the advice and information that patients receive cannot be controlled in
this way. If this is the case, the determination of preference may be a mixture of different causal
structures in the policy population. Finally, the determination of patient preference in the policy
2We thank Jay Kadane for pointing out that the causaJ relationship between Preference and other variables might
be different in the experimental and non-experimental populations, even if Preference is not directly manipulated.
population could easily be unstable. There are fads and fashions among patients, and also fads
and fashions among doctors. New information could be released, or an intensive advertising
campaign introduced. Any of these might create a trek between Preference and 0, and hence
between T and 0, that does not contain any member of the X profile as a non-collider.
So even if in the experimental population Preference and 0 are not confounded, they very well
might be in the policy population. If they are confounded in the policy population then
P(OIX.,T) will not be the same in the experimental and policy populations (unless the parameters
of the different causal structures coincidentally have values that make them equal.) Note that the
same is true of predictions of P(OIX.,T) based on the Kadane/SedransklSeidenfeld design or of
a prediction of P( 01T) based upon a randomized experiment. This does not mean that no useful
predictions can be made in situations where patient preference will be used to influence
treatment in the policy population. It is still possible to inform the patient what P( OIX.,T) would
be if a particular treatment were given without patient choice The patient can use this
information to help make an informed decision. And with the design we have propo&ed, this
(counterfactual) prediction is accurate as long as Preference and 0 are not confounded in the
experimental population, regardless of how Preference is causally connected to 0 in the policy
population.
Suppose then that we are merely trying to predict P(OIX.,T) in a population in which everyone is
assigned a treatment. Is the design we have suggested practical? One potential problem is the
obligation to give patients who are experimental subjects advice and information about their
treatments; if this were not done the experiment would be unethical. If patients have access to
advice from physicians, the advice is likely to be based upon their X profile. Even if the only
information subjects receive is that all of the experts agree that they should not choose treatment
Tl, their X profile is a cause of their preference. Will giving this advice and information make it
unlikely that P( OIX.,T) is invariant under manipulation of T? It is true that the variables in the X
profiles were chosen to be variables thought to be causes of 0 or to have common causes with
O. Hence advice of this kind is very likely to create a common cause of Preference and 0 in the
experimental population. Hence, it is likely that in the experimental population there will be a
trek between T and 0 that is into T and contains Preference. However, such a trek would not
d-connect T and 0 given X because it would also contain some member of X as a non-collider.
(See figure 17.) Hence such a trek does not invalidate the invariance of P(OIX.,T) under
manipulation of T. Moreover, there is no problem with changing the advice as the experiment
progresses, under the assumption that Pj-l is causally connected to OJ only through Preferencej.
j -1
Figure 17
Can we let patients choose their own treatment, or merely influence the choice of treatment? As
long as we are merely trying to predict P(OIX,1) in a population in which everyone is assigned
a treatment we can let patients choose their own treatment, as long as this doesn't result in all
patients with a particular X profile always failing to choose some treatment Tl. In that case,
P(OIX,T = TI) is undefined in the experimental population and cannot be used to predict
P(OIX,T = Tl) in a population where that quantity is defined.
In summary, so long as the goal is to predict P(OIX,1) in a population where everyone is

assigned a treatment, and there is enough variation of choice of treatments among the patients,
and the experimental population is not a mixture of causal structures, and Preference and 0 are
not confounded in the experimental population (an issue that must be decided empirically) it is
possible to make accurate predictions from an experimental population in which informed
patients choose their own treatment. If it is really important to let patient preferences influence
their treatment in experiments, then it is worth risking some cost to realize that condition if it is
possible to do so consistent with reliable prediction. How much it is worth, either in money or
in degradation in confidence about the reliability of predictions, is not for us to say. But a
simple modification of the Kadane/Sedransk/Seidenfeld design which has initial trials base
treatment assignment onXj, Pj-I, Zj' and Preferencej, and then allows patient self-assignment if
Prejerencej and OJ are discovered to be unconfounded, would in some cases permit
investigators to conduct clinical trials that conform to ethical requirements of autonomy and
informed consent.
9.5 An Example: Smoking and Lung Cancer
The fascinating history of the debates over smoking and lung cancer illustrates the difficulties of
causal inference and prediction from policy studies, and also illustrates some common mistakes.
Perhaps no other hypothetical cause and effect relationship has been so thoroughly studied by
non-experimental methods or has so neatly divided the professions of medicine and statistics
into opposing camps. The theoretical results of this and the preceding chapters provide some
insight into the logic and fallacies of the dispute.
The thumbnail sketch is as follows: In the 1950s a retrospective study by Doll and Hill (1952)
found a strong correlation between cigarette smoking and lung cancer. That initial research
prompted a number of other studies, both retrospective and prospective, in the United States,
the United Kingdom, and soon after in other nations, all of which found strong correlations
between cigarette smoking and lung cancer, and more generally between cigarette "Smoking and
cancer and between cigarette smoking and mortality. The correlations prompted health activists
and some of the medical press to conclude that cigarette smoking causes death, cancer, and
most particularly, lung cancer. Sir Ronald Fisher took very strong exception to the inference,
preferring a theory in which smoking behavior and lung ,cancer are causally connected only
through genetics. Fisher wrote letters, essays, and eventually a book against the inference from
the statistical dependencies to the causal conclusion. Neyman ventured a criticism of the
evidence from retrospective studies. The heavyweights of the statistical profession were thus
allied against the methods of the medical community. A review of the evidence containing a
response to Fisher and Neyman was published in 1959 by Cornfield, Haenszel, Hammond,
Lilienfeld, Shirnkin, and Wynder. The Cornfield paper became part of the blueprint for the
Report of the Surgeon General on Smoking and Health in 1964, which effectively established
that as a political fact smoking would be treated as an unconfounded cause of lung cancer, and
set in motion a public health campaign that is with us still. Brownlee (1965) reviewed the 1964
report in the Journal of the American Statistical Association and rejected its arguments as
statistically unsound for many of the reasons one can imagine Fisher would have given. In
1979, the Surgeon General published a second report on smoking and health, repeating the
arguments of the first report but with more extensive data, but offering no serious response to
Brownlee's criticisms. The report made strong claims from the evidence, in particular that
cigarette smoking was the largest preventable cause of death in the United States. The foreword
to the report, by Joseph Califano, was downright vicious, and claimed that any criticism of the
conclusions of the report was an attack on science itself. That did not stop P. Burch (1983), a
physicist turned theoretical biologist turned statistician, from publishing a lengthy criticism of
the second report, again on grounds that were detailed extensions of Fisher's criticisms, but
buttressed as well by the rust reports of randomized clinical trials of the effects of smoking
intervention, all of which were either null or actually suggested that intervention programs
increased mortality. Burch's remarks brought a reply by A. Lilienfeld (1983), which began and
ended with an ad hominem attack on Burch.
Fisher's criticisms were directed against the claim that uncontrolled observations of a correlation
between smoking and cancer, no matter whether retrospective or prospective, provided
evidence that smoking causes lung cancer, as against the alternative hypothesis that there are
one or more common causes of smoking and lung cancer. His strong views can be understood
in the light of features of his career. Fisher had been largely responsible for the introduction of
randomized experimental designs, one of the very points of which was to obtain statistical
dependencies between a hypothetical cause and effect that could not be explained by the action
of unmeasured common causes. Another point of randomization was to insure a well-defined
distribution for tests of hypotheses, something Fisher may have doubted was available in
observational studies. Throughout his adult life Fisher's research interests had been in heredity,
and he had been a strong advocate of the eugenics movement. He was therefore disposed to
believe in genetic causes of very detailed features of human behavior and disease. Fisher
thought a likely explanation of the correlation of lung cancer and smoking was that a substantial
fraction of the population had a genetic predisposition both to smoke and to get lung cancer.
One of Fisher's (1959, p. 8) fundamental criticisms of these epidemiological arguments was

that correlation underdeterrnines causation: besides smoking causing cancer, wrote Fisher "there
are two classes of alternative theories which any statistical association, observed without the
precautions of a definite experiment, always allows--namely, (1) that the supposed effect is
really the cause, or in this case that incipient cancer, or a pre-cancerous condition with chronic
inflammation, is a factor in inducing the smoking of cigarettes, or (2) that cigarette smoking and
lung cancer, though not mutually causative, are both influenced by a common cause, in this
case the individual genotype." Not even Fisher took (1) seriously. To these must be added
others Fisher did not mention, for example that smoking and lung cancer have several distinct
unmeasured common causes, or that while smoking causes cancer, something unmeasured also
causes both smoking and cancer.
If we interpret "statistical association" as statistical dependence, Fisher is correct that given

observation only of a statistical dependence between smoking and lung cancer in an
uncontrolled study, the possibility that smoking does not cause lung cancer cannot be ruled out.
However, he does not mention the possibility that this hypothesis, if true, could have been
established without experimentation by finding a factor associated with smoking but

independent, or conditionally independent (on variables other than smoking) of cancer. By the
1960s a number of personal and social factors associated with smoking had been identified, and
several causes of lung cancer (principally associated with occupational hazards and radiation)
potentially independent of smoking had been identified, but their potential bearing on questions
of common causes of smoking and lung cancer seems to have gone unnoticed. The more
difficult cases to distinguish are the hypotheses that smoking is an unconfounded cause of lung
cancer versus the joint hypotheses that smoking causes cancer and that there is also an
unmeasured common cause--or causes--of smoking and cancer.
Fisher's hypothesis that genotype causes both smoking behavior and cancer was speculative,
but it wasn't a will-o-the-wisp. Fisher obtained evidence that the smoking behavior of
monozygotic twins was more alike than the smoking behavior of dizygotic twins. As his critics
pointed out, the fact could be explained on the supposition that monozygotic twins are more
encouraged by everyone about them to do things alike than are dizygotic twins, but-Fisher was
surely correct that it could also be explained by a genetic disposition to smoke. On the other
side, Fisher could refer to evidence that some forms of cancer have genetic causes.
The paper by Cornfield, et al. (including Lilienfeld) argued that while lung cancer may well
have other causes besides, cigarette smoking causes lung cancer. This view had already been
announced by official study groups in the United States and Great Britain. Cornfield's paper is
of more scientific interest than the Surgeon General's report five years later, in part because the
former is not primarily a political document. Cornfield, et al. claimed the existing data showed
several things:
1. Carcinomas of the lung found at autopsy had systematically increased since 1900, although
different studies gave different rates of increase. Lung cancers are found to increase
monotonically with the amount of cigarette smoking and to be higher in current than in former
cigarette smokers. In large prospective studies diagnoses of lung cancer may have an unknown
error rate, but the total death rate also increases monotonically with cigarette smoking.
2. Lung cancer mortality rates are higher in urban than in rural populations, and rural people
smoke less than city people, but in both populations smokers have higher death rates from lung
cancer than do non-smokers.
3. Men have much higher death rates from lung cancer than women, especially among persons
over 55, but women smoked much less and as a class had taken up the habit much later than
men.
4. There are a host of causes of lung cancer, including a variety of industrial pollutants and
unknown circumstances associated with socioeconomic class, with the poorer and less well off
more likely than the better off to contract the disease, but no more likely to smoke. Cornfield, et
al. emphasize that "The population exposed to established industrial carcinogens is small, and
these agents cannot account for the increasing lung-cancer risk in the remainder of the
population. Also, the effects associated with socioeconomic class and related characteristics are
smaller than those noted for smoking history, and the smoking class differences cannot be
accounted for in terms of these other effects." (p.179). This passage states that the difference in
cancer rates for smokers and non-smokers could not be explained by socioeconomic
differences. While this claim was very likely true, no analysis was given in support of it, and
the central question of whether smoking and lung cancer were independent or nearly
independent conditional on all subsets of the known risk factors that are not effects of smoking
and cancer--area of residence, exposure to known carcinogens, socioeconomic class, and so
on, was not considered. Instead, Cornfield et al. note that different studies measured different
variables and "The important fact is that in all studies when other variables are held constant,
cigarette smoking retains its high association with lung cancer."
5. Cigarette smoking is not associated with increased cancer of the upper respiratory tract, the
mouth tissues or the fingers. Carcinoma of the trachea, for example, is a rarity. But, Cornfield,
et al. point out, "There is no a priori reason why a carcinogen that produces bronchogenic
cancer in man should also produce neoplastic changes in the anspharynx or in other sites." (p.
186).
6. Experimental evidence shows that cigarette smoke inhibits the action of the cilia in cows, rats
and rabbits. Inhibition of the cilia interferes with the removal of foreign material from the
surface of the bronchia. Damage to ciliated cells is more frequent in smokers than in
nonsmokers.
7. Application of cigarette tar directly to the bronchia of dogs produced changes in the cells, and
in some but not other experiments applications of tobacco tar to the skin of mice produced
cancers. Exposure of mice to cigarette smoke for up to 200 days produced cell changes but no
cancers.
8. A number of aromatic polycyclic compounds have been isolated in tobacco smoke, and one
of them, the a form of benzopyrene, was known to be a carcinogen.
Perhaps the most original technical part of the argument was an a kind of sensitivity analysis of
the hypothesis that smoking causes lung cancer. Cornfield, et al. considered a single
hypothetical binary latent variable causing lung cancer and statistically dependent on smoking
behavior. They argued such a latent cause would have to be almost perfectly associated with
lung cancer and strongly associated with smoking to account for the observed association. The
argument neglected, however, the reasonable possibility of multiple common causes of
smoking and lung cancer, and had no clear bearing on the hypothesis that the observed
association of smoking and lung cancer is due both to a direct influence and to common causes.
In sum, Cornfield, et al. thought they could show a mechanism for smoking to cause cancer,
and claimed evidence from animal studies, although their position in that regard tended to trip
over itself (compare items 5 and 7). They didn't put the statistical case entirely clear!y, but their
position seems to have been that lung cancer is also caused by a number of measurable factors
that are not plausibly regarded as effects of smoking but which may cause smoking, and that
smoking and cancer remain statistically dependent conditional on these factors. Against Fisher
they argued as follows:
The difficulties with the constitutional hypothesis include the following considerations:
(a) changes in lung-cancer mortality over the last half century; (b) the carcinogenicity of
tobacco tars for experimental animals; (c) the existence of a large effect from pipe and
cigar tobacco on cancer of the buccal cavity and larynx but not on cancer of the lung; (d)
the reduced lung-cancer mortality among discontinued cigarette smokers. No one of
these considerations is perhaps sufficient by itself to counter the constitutional
hypothesis, ad hoc modification of which can accommodate each additional piece of
evidence. A point is reached, however, when a continuously modified hypothesis
becomes difficult to entertain seriously. (p. 191)
Logically, Cornfield et al. visited every part of the map. The evidence was supposed to be
inconsistent with a common cause of smoking and lung cancer, but also consistent with it.
Objections that a study involved self-selection--as Fisher and company would object to (d)--was
counted as an "ad hoc modification" of the common cause hypothesis. The same response was
in effect given to the unstated but genuine objections that the time series argument ignored the
combined effects of dramatic improvements in diagnosis of lung cancer, a tendency of
physicians to bias diagnoses of lung cancer for heavy smokers and to overlook such a diagnosis
for light smokers, and the systematic increase in the same period of other factors implicated in
lung cancer, such as urbanization. The rhetoric of Cornfield, et al. converted reasonable
demands for sound study designs into ad hoc hypotheses. In fact none of the evidence adduced
was inconsistent with the "constitutional hypothesis."
A reading of the Cornfield paper suggests that their real objection to a genetic explanation was
that it would require a very close correlation between genotypic differences and differences in
smoking behavior and liability to various forms of cancer. Pipe and cigar smokers would have
to differ genotypically from cigarette smokers; light cigarette smokers would have to differ
genotypically from heavy cigarette smokers; those who quit cigarette smoking would have to
differ genotypically from those who did not. Later the Surgeon General would add that
Mormons would have to differ genotypically from non-Mormons and Seventh Day Adventists
from non-Seventh Day Adventists. The physicians simply didn't believe it. Their skepticism
was in keeping with the spirit of a time in which genetic explanations of behavioral differences
were increasingly regarded as politically and morally incorrect, and the moribund eugenics
movement was coming to be viewed in retrospect as an embarrassing bit of racism.
In 1964 the Surgeon General's report reviewed many of the same studies and arguments as had
Cornfield, but it added a set of "Epidemiological Criteria for Causality," said to be sufficient for
establishing a causal connection and claimed that smoking and cancer met the criteria. The
criteria were indefensible, and they did not promote any good scientific assessment of the case.
The criteria were the "consistency" of the association, the "strength" of the association, the
"specificity" of the association, the temporal relationship of the association and the "coherence"
of the association.
All of these criteria were left quite vague, but no way of making them precise would suffice for
reliably discriminating causal from common causal structures. Consistency meant that separate
studies should give the "same" results, but in what respects results should be the same was not
specified. Different studies of the relative risk of cigarette smoking gave very different
multipliers depending on the gender, age and nationality of the subjects. The results of most
studies were the same in that they were all positive; they were plainly not nearly the same in the
seriousness of the risk. Why stronger associations should be more likely to indicate causes than
weaker associations was not made clear by the report. Specificity meant the putative cause,
smoking, should be associated almost uniquely with the putative effect, lung cancer. Cornfield,
et al. had rejected this requirement on causes for good reason, and it was palpably violated in
the smoking data presented by the Surgeon General's report. "Coherence" in the jargon of the
report meant that no other explanation of the data was possible, a criterion the observational data
did not meet in this case. The temporal issue concerned the correlation between increase in
cigarette smoking and increase in lung cancer, with a lag of many years. Critics pointed out that
the time series were confounded with urbanization, diagnostic changes and other factors, and
that the very criterion Cornfield, et al. had used to avoid the issue of the unreliability of
diagnoses, namely total mortality, was, when age-adjusted, uncorrelated with cigarette
consumption over the century.
Brownlee (1965) made many of these points in his review of the report in the Journal o/the
American Statistical Association. His contempt for the level of argument in the report was plain,
and his conclusion was that Fisher's alternative hypothesis had not been eliminated or even very
seriously addressed. In Brownlee's view, the Surgeon General's report had only two
arguments against a genetic common cause: (a) the genetic hypothesis would allegedly have to
be very complicated to explain the dose/response data, and (b) the rapid historical rise in lung
cancer following by about 20 years a rapid historical rise in cigarette smoking. Brownlee did
not address (a), but he argued strongly that (b) is poor evidence because of changes in
diagnostics, changes in other factors of known and unknown relevance, and because of
changes in the survival rate of weak neonates whom, as adults, might be more prone to lung
cancer.
One of the more interesting aspects of the review was Brownlee's "very simplified" proposal
for a statistical analysis of "E2 causes E1" which was that Eland E2 be dependent conditional
on every possible vector of values for all other variables of the system. Brownlee realized, of
course, that his condition did not separate "E2 causes El" from E 1 causes E2," but that was not
a problem with smoking and cancer. But even ignoring the direction of causation, Brownlee's
condition--perhaps suggested to him by the fact that the same principle is erroneously used in
regression--is quite wrong. It would be satisfied, for example, if, El and E2 had no causal
connection whatsoever provided some measured variable Ej were a direct effect of both Eland
E2·
Brownlee thought his way of considering the matter was important for prediction and
intervention:
If the inequality holds only for, say, one particular subset Ej, ... , Ek, and for all other
subsets equality holds, and if the subset Ej, ... ,Ek occurs in the population with low
probability, then Pr{EIIE2}, while not strictly equal to Pr{EIIE2C }, will be numerically
close to it, and then E2 as a cause of El may be of small practical importance. These
considerations are related to the Committee's responsibility for assessment of the
magnitude of the health hazard (page 8). Further complexities arise when we distinguish
between cases in which one of the required secondary conditions Ej, ... ,Ek is, on the one
hand, presumably controllable by the individual, e.g., the eating of parsnips, or
uncontrollable, e.g., the presence of some genetic property. In the latter case, it further
makes a difference whether the genetic property is identifiable or non-identifiable: for
example it could be brown eyes which is the significant subsidiary condition Ej, and we
could tell everybody with not-brown eyes it was safe for them to smoke. (p. 725)
No one seems to have given any better thought than this to the question of how to predict the
effects of public policy intervention against smoking. Brownlee regretted that the Surgeon
General's report made no explicit attempt to estimate the expected increase in life expectancy
from not smoking or from quitting after various histories.
Fifteen years later, in 1979, the second Surgeon General's Report on Smoking and Health was
able to report studies that showed a monotonic increase in mortality rates with virtually every
feature of smoking practice that increased smoke in the lungs: number of cigarettes smoked per
day, number of years of smoking, inhaling versus not inhaling, low tar and nicotine versus
high tar and nicotine, length of cigarette habitually left unsmoked. The monotonic increase in
mortality rates with cigarette smoking had been shown in England, the continental United
States, Hawaii, Japan, Scandinavia and elsewhere, for whites and blacks, for men and women.
The report dismissed Fisher's hypothesis in a single paragraph by citing a Scandinavian study
(Cederlof, Friberg and Lundman, 1977) that included monozygotic and dizygotic twins:
When smokers and nonsmokers among the dizygotic pairs were compared, a mortality
ratio of 1,45 for males and 1.21 for females was observed. Corresponding mortality
ratios for the monozygotic pairs were 1.5 for males and 1.222 for females. Commenting
on the constitutional hypothesis and lung cancer, the authors observed that "the
constitutional hypothesis as advanced by Fisher and still supported by a few, has here
been tested in twin studies. The results from the Swedish monozygotic twin series
speak strongly against the constitutional hypothesis."
The second Surgeon General's report claimed that tobacco smoking is responsible for 30% of
all cancer deaths; cigarette smoking is responsible for 85% of all lung cancer deaths.
A year before the report appeared, in a paper for the British Statistical Association P. Burch
(1978) had used the example of smoking and lung cancer to illustrate the problems of
distinguishing causes from common causes without experiment. In 1982 he published a full
fledged assault on the second Surgeon General's report. The criticisms of the argument of the
report were similar to Brownlee's criticisms of the 1964 report, but Burch was less restrained
and his objections more pointed. His first criticism was that while all of the studies showed a
increase in risk of mortality with cigarette smoking, the degree of increase varied widely from
study to study. In some studies the age adjusted multiple regression of mortality on cigarettes,
beer, wine and liquor consumption gave a smaller partial correlation with cigarettes than with
beer drinking. Burch gave no explanation of why the regression model should be an even
approximately correct account of the causal relations. Burch thought the fact that the apparent
dose/response curve for various culturally, geographically and ethnically distinct groups were
very different indicated that the effect of cigarettes was significantly confounded with
environmental or genetic causes. He wanted the Surgeon General to produce a unified theory of
the causes of lung cancer, with confidence intervals for any relevant parameter estimates:
Where, he asked, did the 85% figure come from?
Burch pointed out, correctly, that the cohort of 1487 dizygotic and 572 monozygotic twins in
the Scandinavian study born between 1901 and 1925 gave no support at all to the claim that the
constitutional explanation of the connection between smoking and lung cancer had been refuted,
despite the announcements of the authors of that study. The study showed that of the dizygotes
exactly 2 nonsmokers or infrequent smokers had died of lung cancer and 10 heavy smokers had
died of lung cancer; of the monozygotes, 2 low non smokers and 2 heavy smokers had died of
the disease. The numbers were useless, but if they suggested anything, it was that if genetic
variation was controlled there is no difference in lung cancer rates between smokers and
nonsmokers. The Surgeon General's report of the conclusion of the Scandinavian study was
accurate, but not the less misleading for that.
Burch also gave a novel discussion of the time series data, arguing that it virtually refuted the
causal hypothesis. The Surgeon General and others had used the time series in a direct way. In
the U.K. for example, male cigarette consumption per capita had increased roughly a
hundredfold between 1890 and 1960, with a slight decrease thereafter. The age-standardized
male death rate from lung cancer began to increase steeply about 1920, suggesting a thirty year
lag, consistent with the fact that people often begin smoking in their twenties and typically
present lung cancer in their fifties. According to Burch's data, the onset of cigarette smoking for
women lagged behind males by some years, and did not begin until the 1920s. The Surgeon
General's report noted that the death rate from lung cancer for women had also increased
dramatically between 1920 and 1980. Burch pointed out that the autocorrelations for the male
series and female series didn't mesh: there was no lag in death rates for the women. Using
U.K. data, Burch plotted the percentage change in the age-standardized death rate from lung
cancer for both men and women from 1900 to 1980. The curves matched perfectly until 1960.
Burch's conclusion is that whatever caused the increase in death rates from lung cancer affected
both men and women at the same time, from the beginning of the century on, although
whatever it is had a smaller absolute effect on women than on men. But then the whatever-it-
was could not have been cigarette smoking, since increases in women's consumption of
cigarettes lagged twenty to thirty years behind male increases.
Burch was relentless. The Surgeon General's report had cited the low occurrence of lung cancer
among Mormons. Burch pointed out that Mormon's in Utah not only have lower age-adjusted
incidences of cancer than the general population, but also have higher incidences than non-
Mormon nonsmokers in Utah. Evidently their lower lung cancer rates could not be simply
attributed to their smoking habits.
Abraham Lilienfeld, who only shortly before had written a textbook on epidemiology and who
had been involved with the smoking and cancer issue for more than twenty years, published a
reply to Burch that is of some interest. Lilienfeld gives the impression of being at once
defensive and disdainful. His defense of the Surgeon General's report began with an ad
hominem attack, suggesting that Burch was so out of fashion as to be a crank, and ended with
another ad hominem, demanding that if Burch wanted to criticize others' inferences from their
data he go get his own. The most substantive reply Lilienfeld offered is that the detailed
correlation of lung cancer with smoking habits in one subpopulation after another makes it seem
very implausible that the association is due to a common cause. Lilienfeld said, citing himself,
that the conclusion that 85% of lung cancer deaths are due to cigarettes is based on the relative
risk for cigarette smokers and the frequency of cigarette smoking in the population, predicting,
in effect, that if cigarette smoking ceased the death rate from lung cancer would decline by that
percentage. (The prediction would only be correct, Burch pointed out in response, provided
cigarette smoking is a completely unconfounded cause oflung cancer.) Lilienfeld challenged the
source of Burch's data on female cigarette consumption early in the century, which Burch
subsequently admitted were estimates.
Both Burch and Lilienfeld discussed a then recent report by Rose et al. (1982) on a ten year
randomized smoking intervention study. The Rose study, and another that appeared at nearly
the same time with virtually the same results, illustrates the hazards of prediction. Middle-aged
male smokers were assigned randomly to a treatment or non-treatment group. The treatment
group was encouraged to quit smoking and given counseling and support to that end. By self-
report, a large proportion of the treatment group either quit or reduced cigarette smoking. The
difference in self-reported smoking levels between the treatment and non-treatment groups was
thus considerable, although the difference declined toward the end of the ten year study. To
most everyone's dismay, Rose found that there was no statistically significant difference in lung
cancer between the two groups after ten years (or after five), but there was a difference in
overall mortality--the group that had been encouraged to quit smoking, and had in part done so,
suffered higher mortality.
Fully ignoring their own evidence, the authors of the Rose study concluded nonetheless that
smokers should be encouraged to give up smoking, which makes one wonder why they
bothered with a randomized trial. Burch found the Rose report unsurprising; Lilienfeld claimed
the numbers of lung cancer deaths in the sample are too small to be reliable, although he did not
fault the Surgeon General's report for using the Scandinavian data, where the numbers are even
smaller, and he simply quoted the conclusion of the report. which seems almost disingenuous.
To Burch's evident delight, as Lilienfeld's defense of the Surgeon General appeared so did yet
further experimental evidence that intervening in smoker's behavior has no benign effect on
lung cancer rates. The Multiple Risk Factor Intervention Trial Research Group (1982) reported
the results after six years of a much larger randomized experimental intervention study
producing roughly three times the number of lung cancer deaths as in the Rose study. But the
intervention group showed more lung cancer deaths than the usual care group! The absolute
numbers were small in both studies but there could be no doubt that nothing like the results
expected by the epidemiological community had materialized.
The results of the controlled intervention trials illustrate how naive it is to think that
experimentation always produces unambiguous results, or frees one from requirements of prior
knowledge. One possible explanation for the null effects of intervention on lung cancer, for
example, is that the reduced smoking produced by intervention was concentrated among those
whose lungs were already in poor health and who were most likely to get lung cancer in any
case. (Rose, et al. gave insufficient information for an analysis of the correlation of smoking
behavior and lung cancer within the intervention group.) This possibility could have been tested
by experiments using blocks more finely selected by health of the subjects.
In retrospect the general lines of the dispute were fairly simple. The statistical community
focused on the want of a good scientific argument against a hypothesis given prestige by one of
their own; the medical community acted like Bayesians who gave the "constitutional"
hypothesis necessary to account for the dose/response data so Iowa prior that it did not merit
serious consideration. Neither side understood what uncontrolled studies could and could not
determine about causal relations and the effects of interventions. The statisticians pretended to
an understanding of causality and correlation they did not have; the epidemiologists resorted to
informal and often irrelevant criteria, appeals to plausibility, and in the worst case to ad
hominem.
Fisher's prestige as well as his arguments set the line for statisticians, and the line was that
uncontrolled observations cannot distinguish among three cases: smoking causes cancer,
something causes smoking and cancer, or something causes smoking and cancer and smoking
causes cancer. The most likely candidate for the "something" was genotype. Fisher was wrong
about the logic of the matter, but the issue never was satisfactorily clarified, even though some
statisticians, notably Brownlee and Burch, tried unsuccessfully to characterize more precisely
the connection between probability and causality. While the statisticians didn't get the
connection between causality and probability right, the Surgeon General's "epidemiological
criteria for causality" were an intellectual disgrace, and the level of argument in defense of the
conclusions of the Surgeon General's Report was sometimes more worthy of literary critics
than scientists. The real view of the medical community seems to have been that it was just too
implausible to suppose that genotype strongly influenced how much one smoked, whether one
smoked at all, whether one smoked cigarettes as against a cigar or pipe, whether one was a
Mormon or a Seventh day Adventist, and whether one quit smoking or not. After Cornfield's
survey the medical and public health communities gave the common cause hypothesis more
invective than serious consideration. And, finally, in contrast to Burch, who was an outsider
and maverick, leading epidemiologists, such as Lilienfeld, seem simply not to have understood
that if the relation between smoking and cancer is confounded by one or more common causes,
the effects of abolishing smoking cannot be predicted from the "risk ratios," i.e., from sample
conditional probabilities. The subsequent controlled smoking intervention studies gave evidence
of how very bad were the expectations based on uncontrolled observation of the relative risks of
lung cancer in those who quit smoking compared to those who did not.
9.6 Appendix
(io*)
Figure 18
We will prove that the partially oriented inducing path graph (io*) in figure 18, together with the
assumptions that U causes A, that there is no common cause of U and C, and that every
directed path from U to C contains A, entail that A causes C and that there is a latent common
cause of A and C. We assume that A is not a deterministic function of U.
Let 0 = {A,C,U,V}, and G be the directed acyclic graph that generated (io*). The U 0-> C
edge in (io*) entails that in the inducing path graph of G either U -> C or U <-> C. If there is a
U <-> C edge, then there is a latent common cause of U and C, contrary to our assumption.
Hence the inducing path graph contains a U -> C edge. It follows that in G there is a directed
path from U to C. Because every directed path from U to C contains A, there is a directed path
from A to C in G. Hence A causes C.
The U -> C edge in the inducing path graph of G entails that there is an inducing path Z relative
to 0 that is out of U and into C in G. If Z does not contain a collider then Z is a directed path
from U to C and hence it contains A. But then A is a non-collider on Z, and Z is not an inducing
path relative to 0, (because Z contains a member of 0, namely A, that is a non-collider on Z)
contrary to our assumption. Hence Z contains a collider.
We will now show that no collider on Z is an ancestor of U. Suppose, on the contrary that there
is a collider on Z that is an ancestor of U; let M be the closest such collider on Z to C. No
directed path from M to U contains C, because there is a directed path from U to C, and hence
no directed path from C to U. There are two cases.
Suppose fust that there is no collider between M and C. Then there is a variable Q on Z, such
that Z(Q,C) is a directed path from Q to C and Z(Q,M) is a directed path from Q to M. (As in
the proofs in Chapter 13, we adopt the convention that on an acyclic path Z containing Q and C,
*
Z(Q,C) represents the subpath of Z between Q and C.) U M because M is a collider on Z and
U is not. U does not lie on Z(Q,C) or Z(Q,M) because Z is acyclic. The concatenation of
Z(Q,M) and a directed path from M to U contains a directed path from Q to U that does not
contain C. Z(Q,C) is a directed path from Q to C that does not contain U. Q is a non-collider on
Z, and because Z is an inducing path relative to 0, Q is not in O. Hence Q is a latent common
cause of U and C, contrary to our assumption.
Suppose next that there is a collider between M and C, and N is the collider on Z closest to M
and between M and C. Then there is a variable Q on Z, such that Z(Q,N) is a directed path from
*
Q to N and Z(Q,M) is a directed path from Q to M. U M because M is a collider on Z and U
is not. U does not lie on Z(Q,N) or Z(Q,M) because Z is acyclic. The concatenation of Z(Q,M)
and a directed path from M to U contains a directed path from Q to U that does not contain C.
There is a directed path from N to C, and by hypothesis no such directed path contains U. The
concatenation of Z(Q,N) and a directed path from N to C contains a directed path from Q to C
that does not contain U. Q is a non-collider on Z, and because Z is an inducing path relative to
0, Q is not in O. Hence Q is a latent common cause of U and C, contrary to our assumption.
It follows that no collider on Z is an ancestor of U.
Let X be the collider on Z closest to U. There is a directed path from X to C. Z(U,x) is a

directed path from U to X. The concatenation of Z(U,x) and a directed path from X to C
contains a directed path from U to C. By assumption, such a path contains A. A does not lie
between U and X on Z, because every vertex between U and X on Z is a non-collider, and if A
occurs on Z it is a collider on Z. Hence A lies on every directed path from X to C. Hence there
exists a collider on Z that is the source of a directed path to C that contains A. Let R be the
collider on Z closest to C such that there is a directed path D from R to C that contains A. There
are again two cases.
If there is no collider between R and C on Z, then there is a vertex Q on Z such that Z(Q,C) is a
directed path from Q to C and Z(Q,R) is a directed path from Q to R. A does not lie on Z(Q,C)
because no vertex on Z(Q,C) is a collider on Z. C does not lie on D(R,A) because the directed
graph is acyclic. C"* Q because Z has an edge into C but not Q. C "* R because R is a collider
on Z and C is not. Hence, C does not lie on Z(Q,R) because Z is acyclic. The concatenation of
Z(Q,R) and D(R,A) contains a directed path from Q to A that does not contain C. Q is not a
collider on Z, so it not in O. Hence Q is a latent common cause of A and C.
Suppose next that there is a collider between R and C on Z, and N is the closest such collider to
R on Z. Then there is a vertex Q on Z such that Z(Q,N) is a directed path from Q to N and
Z(Q,R) is a directed path from Q to R. Q "* N because by hypothesis there is a path from N to C
that does not contain A. A does not lie on Z(Q,N) because no vertex on Z(Q,N) except N is a
collider on Z. There is a directed path from N to C, but it does not contain A by hypothesis.
Hence the concatenation of Z(Q,N) and a directed path from N to C contains a directed path that
does not contain A. C does not lie on D(R,A) because the directed graph is acyclic. C"* Q
because Z has an edge into C but not Q. C "* R because R is a collider on Z and C is not.
Hence, C does not lie on Z(Q,R) because Z is acyclic. The concatenation of Z(Q,R) and D(R,A)
contains a directed path from Q to A that does not contain C. Q is not a collider on Z, so it not in
O. Hence Q is a latent common cause of A and C.
Hence in either case, A and C have a latent common cause in G.

Chapter 10
The Structure of the Unobserved
10.1 Introduction
Many theories suppose there are variables that have not been measured but that influence
measured variables. In studies in econometrics, psychometrics, sociology and elsewhere the
principal aim may be to uncover the causal relations among such "latent" variables. In such
cases it is usually assumed that one knows that the measured variables (e.g., responses to
questionnaire items) are not themselves causes of unmeasured variables of interest (e.g.,
attitude), and the measuring instruments are often designed with fairly definite ideas as to which
measured items are caused by which unmeasured variables. Survey questionnaires may involve
hundreds of items, and the very number of variables is ordinarily an impediment to drawing
useful conclusions about structure. Although there are a number of procedures commonly used
for such problems, their reliability is doubtful. A common practice, for example, is to form
aggregated scales by averaging measures of variables that are held to be proxies for the same
unmeasured variable, and then to study the correlations of the scales. The correlations thus
obtained have no simple systematic connection with causal relations among the unmeasured
variables.
What can a mixture of substantive knowledge about the measured indicators and statistical
observations of those indicators reveal about the causal structure of the unobserved variables?
And under what assumptions about distributions, linearity, etc.? This chapter begins to address
these questions. The procedures for forming scales, or "pure measurement models," that we
will describe in this chapter have found empirical application in the study of large psychometric
data sets (Callahan and Sorensen 1992).
The Structure of the Unobserved 307
10.2 An Outline of the Algorithm
Consider the problem of determining the causal structure among a set of unmeasured variables
of interest in linear pseudo-indeterministic models, commonly called "structural equation
models with latent variables." Assume the distributions are linearly faithful. Structural equation
models with latent variables are sometimes presented in two parts: the "measurement model",
and the "structural model" (see figure 1). The structural model involves only the causal
connections among the latent variables; the remainder is the measurement model. From a
mathematical point of view, the distinction marks only a difference in the investigator's interests
and access and not any distinction in formal properties. The same principles connecting graphs,
probabilities and causes apply to the measurement model as to the structural model. In figure 1
we give an example of a latent variable model in which the measured variables (QI-Q12) might
be answers to survey questions.
Let T be a set of latent variables and Va set of measured variables. We will assume that Tis
causally sufficient, although that is clearly not the general case. We let C denote the set of
"nuisance" latent common causes, that is, unobserved common causes, not in T, of two or
more variables in T u V. Call a subgraph of G that contains all of the edges in G except for
edges between members of T a measurement model of G.
In actual research the set V is often chosen so that for each Ti in T, a subset of V is intended to
measure Ti. In Kohn's (1969) study of class and attitude in America, for example, various
questionnaire items where chosen with the intent of measUring the same attitude; factor analysis
of the data largely agreed with the clustering one might expect on intuitive grounds.
Accordingly, we suppose the investigator can partition V into V(Tj), such that for each i the
variables in V(Tj) are direct effects of Tj. We then seek to eliminate those members of V(Tj)
that are impure measures of Tj, either because they are also the effects of some other
unmeasured variable in T, because they are also causes or effects of some other measured
variable, or because they share an unmeasured common cause in C with another measured
variable.
Full Latent
Variable Model
Measurement
Model
Structural Model
Figure 1
In the class of models we are considering, a measured variable can be an impure measure for
four reasons, which are exhaustive:
(i) If there is a directed edge from some Tj in T to some V in V (Tj) but also a trek
between V and Tj that does not contain Tj or any member of V except V then V is
latent-measured impure.
(ii) If there is a trek between a pair of measured variables Vi, Vz from the same cluster
V(Tj) that does not contain any member of T then Vi and Vz are intra-construct
impure.
(iii) IT there is a trek between a pair of measured variables Vb V2 from distinct clusters
V(Ti) and V(Tj) that does not contain any member of T then we say VI and V2 are
cross-construct impure.
(iv) IT there is a variable in C that is the source of a trek between Ti and some member V
of V (Ti) we say V is common cause impure.
In figure 2, for example, if V(Tl) = {X 1,x2,x3} and V(T2) = {X4,xS,x6} then X4 is latent-
measured impure, Xl and X2 are intra-construct impure, X2 andXS are cross-construct impure,
and X6 is common cause impure. Only X3 is a pure measure of Tt.
Figure 2
We say that a measurement model is almost pure if the only kind of impurities among the
measured variables are common cause impurities. An almost pure latent variable graph is
a directed acyclic graph with an almost pure measurement model. In an almost pure latent
variable graph we continue to refer throughout the rest of this chapter to the set of measured
variables as V, a subset of the latent variables as T, and the "nuisance" latent variables that are
common causes of members of T and V as C.
The strategy that we employ has three steps:

(i) Eliminate measured variables until the variables that remain form the largest almost
pure measurement model with at least two indicators for each latent variable.
(ii) Use vanishing tetrad differences among variables in the measurement model from (i)
to determine the zero and first order independence relations among the variables in T.
(iii) Use the PC algorithm to construct a pattern from the zero and first order
independence relations among the variables in T.
The next section describes a procedure for identifying the appropriate measured variables. The
details are rather intricate; the reader should bear in mind that the procedures have all been
automated, that they work very well in simulation tests, and they all derive from fundamental
structural principles. Given the population correlations the inference techniques would be
reliable (in large samples) for any conditions under which the Tetrad Representation Theorem
holds. The statistical decisions involve a substantial number of joint tests, and no doubt could
be improved. We occasionally resort to heuristics for cases in which each latent variable has a
large number of measured indicators.
10.3 Finding Almost Pure Measurement Models
If G is the true model over V v T v C with measurement model GM, then our task in this
section is to find a subset P of V (the larger the better) such that the sub-model of GM on vertex
set P v T v C is an almost pure measurement model, if one exists with at least two indicators
per latent variable. Our strategy is to use different types of foursomes of variables to
sequentially eliminate impure measures.
Intra-Construct Cross-Construct
RX
~~ ~ IX41
Modell Model 2
Figure 3
As in figure 3, we call four measured variables an intra-construct foursome if all four are
in V(Tj) for some Tj in T; otherwise call it a cross-construct foursome:
10.3.1 Intra-Construct Foursomes
In this section we discuss what can be learned about the measurement model for Tj from V(Tj)
alone. We take advantage of the following principle, which is a consequence of the Tetrad
Representation Theorem.
(P-l) If a directed acyclic graph linearly implies all tetrad differences among the
variables in V(Tj) vanish, then no pair of variables in V(Tj) is intra-construct impure.
So given a set, V(Tj), of variables that measure Tj, we seek the largest subset, P(Tj), of V(Tj)
such that all tetrad differences are judged to vanish among P(Tj). The number of subsets of
V (Tj) is 2IV(T.~, so it is not generally feasible to examine each of them. Further, in realistic
samples we won't find a sizable subset in which all tetrad differences are judged to vanish. A
more feasible strategy is to prune the set iteratively, removing at each stage the variable that
improves the performance of the remaining set P(Tj) on easily computable heuristic criteria
derived from principle P-I. In practice, if the set V (Tj) is large, some small subset of V (Tj)
may by chance do well on these two criteria. For example, if V(Tj) has 12 variables, then there
are 495 subsets of size 4, each of which has only 3 possible vanishing tetrad differences. There
are 792 subsets of size 5, but there are 15 possible tetrad differences that must all be judged to
vanish among each set instead of 3. Because the larger the size of P(Tj) the more unlikely it is
that all tetrad differences among P(Tj) will be judged to vanish by chance, and because we
might eliminate variables from P(Tj) later in the process, we want P(Tj) to be as large as
possible. On the other hand, no matter how well a set P(Tj) does on the first criterion above,
some subset of it will do at least as well or better. Thus, in order to avoid always choosing the
smallest possible subsets we have to penalize smaller sets.
We use the following simple algorithm. We initialize P(Tj) to V(Tj). If the set of tetrad
differences among variables in P(Tj) passes a statistical test, we exit. (We count a set of n tetrad
differences as passing a statistical test at a given significance level Sig if each individual tetrad
difference passes a statistical test at significance level Sig/n. The details of the statistical tests
that we employ on individual tetrad differences are described in Chapter 11.) If the set does not
pass a statistical test, we look for a variable to eliminate from P(Tj). We score each measured
variable X in the following way. For each tetrad difference t among variables in P(Tj) in which
X appears we give X credit if t passes a statistical test, and discredit if t fails a statistical test.
We then eliminate the variable with the lowest score from P(Tj). We repeat this process until we
arrive at a set P(Tj) that passes the statistical test, or we run out of variables.
10.3.2 Cross-Construct Foursomes
Having found, for each latent variable Tj, a subset P(Tj) of V(Tj) in which no variables are
intra-construct impure, we form a subset P of V such that
P= UP(Ti ).
T;ET
We next eliminate members of P that are cross-construct impure.
2x2 foursomes involve two measured variables from P(Tj) and two from P(Tj), where i and
j are distinct. A 2x2 foursome in a pure latent variable model linearly implies exactly one tetrad
equation, regardless of the nature of the causal connection between Tj and Tj in the structural
model. For example, the graph in figure 4 linearly implies the vanishing tetrad difference
PXYPWZ - PXWPYZ = O. Graphs in which Tj causes Tj and graphs in which Tj and Tj are not
causally connected (i.e. there is no trek between them) also linearly imply PXYPWZ - PxwPYZ
= O.
Figure 4
If one variable in V(Tj) is latent-measured impure because of a trek containing Tj, and one
variable in V(Tj) is latent-measured impure because of a trek containing Tj, then the tetrad
differences among the foursome are not linearly implied to vanish by the graph. If Tj and Tj are
connected by some trek and a pair of variables in V(Tj) and V(Tj) respectively are cross-
construct impure then again the tetrad difference is not linearly implied to vanish by the graph.
(The case where Tj and Tj are not connected by some trek is considered below.) In figure 5, for
example, model (i) implies the tetrad equation PXYPWZ = PXWPYZ but models (ii) and (iii) do
not.
(i) (ii)
(iii)
Figure 5
So if we test a 2x2 foursome FI and the hypothesis that the appropriate tetrad difference
vanishes can be rejected, then we know that in at least one of the four pairs in which there is a
measured variable from each construct, both members of the pair are impure. We don't yet
know which pair. We can fmd out by testing other 2x2 foursomes that share variables with Fl.
Suppose the largest subgraph of the true model containing P(TI) and P(T2) is the graph in
figure 6.
Figure 6
Only 2x2 foursomes involving the pair <W,Y> will be recognizably impure. When we test
vanishing tetrad differences in the foursome FI = <X,W,Y02>, we won't know which of the
pairs <W,Z>, <X,Y>, <X,Z>, <W,Y> is impure. When we test the foursome F2 =
<X,W,Z,V>, however, we find that no pair among <X,Z>, <X,V>, <W,z>, or <W,V> is
impure. We know therefore that the pairs <x,z> and <W,z> are not impure in Fl. By testing
the foursome F3 = <U,x,y,z> we fmd that <X,Y> is not impure, entailing <W,Y> is impure
in Fl. If there are at least two pure indicators within each construct, then we can detect exactly
which of the other indicators are impure in this way.
By testing all the 2x2 foursomes in P, we can in principle eliminate all variables that are cross-
construct impure. We cannot yet eliminate all the variables that are latent-measured impure,
because if there is only one such variable it is undetectable from 2x2 foursomes.
Foursomes that involve three measured variables from P(Tj) and one from P(Tj), where i andj
are distinct, are called 3d foursomes. All 3xl foursomes in a pure measurement model
linearly imply all three possible vanishing tetrad differences (see model (i) in figure 7 for
example), IW matter what the causal connection between Tj and Tj. If the variable from P(Tj) in
a 3xl foursome is impure because it measures both latents (model (ii) in figure 7), then Tj is
still a choke point and all three equations are linearly implied. If a variable Z from P(Tj) is
impure because it measures both latents (model (iii) in figure 7), however, then the latent
variable model does not linearly imply that the tetrad differences containing the pair <Z, W>
vanish. This entails that a non-vanishing tetrad differences among the variables in a 3xl
foursome can identify a unique measured variable as latent-measured impure.
Also if Tj and Tj are not trek-connected and a pair of variables VI and V2 in P(Tj) and P(Tj)
respectively are cross-construct impure, then the correlation between VI and V2 does not
vanish, and a tetrad difference among a 3xl foursome that contains VI and V2 is not linearly
implied to vanish; hence the impure member of P(Tj) will be recognized.
If there are least three variables in P(Tj) for each i, then when we fmish examining all 3xl
foursomes we will have a subset P of V such that the sub-model of the true measurement model
over P (which we call Gp) is an almost pure measurement model.
(i)
(ii) (iii)
Figure 7
10.4 Facts about the Unobserved Determined by the Observed
In an almost pure latent variable model constraints on the correlation matrix among the
measured variables determine
(i) for each pair A, B, of latent variables, whether A, Bare uncorrelated,
(ii) for each triple A, B, C of latent variables, whether A and Bare d-separated given {C}.
Part (i) is obvious: two measured variables are uncorrelated in an almost pure latent variable
model if and only if they are effects of distinct unmeasured variables that are not trek connected
(i.e. there is no trek between them) and hence are d-separated given the empty set of variables.
Part (ii) is less obvious, but in fact certain d-separation facts are determined by vanishing tetrad
differences among the measured variables.
Theorem 10.1 is a consequence of the Tetrad Representation Theorem:

Theorem 10.1: If G is an almost pure latent variable graph over V u T u C, T is

causally sufficient, and each latent variable in T has at least two measured indicators,
then latent variables Tl and T3, whose measured indicators include J and L respectively,
are d-separated given latent variable Tz, whose measured indicators include I and K, if
and only if G linearly implies PJlPLK =PJLPKI =PJKP1L.
For example, in the model in figure 8, the fact that Tl and T3 are d-separated given TZ is
entailed by the fact that for all m, n, 0, and p between 1 and 3, where 0 and p are distinct:
Figure 8
By testing for such vanishing tetrad differences we can test for first order d-separability
relations among the unmeasured variables in an almost pure latent variable model. (If A and B
are d-separated given D, we call the number of variables in D the order of the d-separability
relation.) These zero and first order d-separation relations can then be used as input to the PC
algorithm, or to some other procedure, to obtain information about the causal structure among
the latent variables. In the ideal case, the pattern among the latents that is output will always
contain the pattern that would result from applying the PC algorithm directly to d-separation
facts among the latents, but it may contain extra edges and fewer orientations.
10.5 Unifying the Pieces
Suppose the true but unknown graph is shown in figure 9.

Figure 9: True Causal Structure
We assume that a researcher can accurately cluster the variables in the specified measurement
model, e.g., figure lO.
Figure 10: Specified Measurement Model
The actual measurement model is then the graph in figure 11:

Figure 11: Actual Measurement Model
Figure 12 shows a subset of the variables in G (one that leaves out X .. X6 and X 14) that do
fonn an almost pure measurement model.
Figure 12: Almost Pure Measurement Model
Assuming the sequence of vanishing tetrad difference tests finds such an almost pure
measurement model, a sequence of tests of lx2xl vanishing tetrad difference tests then decides
some d-separability facts for the PC or other algorithm through Theorem 10.1. Since in figure
12 there are many lx2xl tetrad tests with measured variables drawn respectively from the
clusters for Tl, T2 and T3, the results of the tests must somehow be aggregated. For each
lx2xl tetrad difference among variables in V(Tl), V(T2), and V(T3) we give credit if the tetrad
difference passes a significance test and discredit if it fails a significance test; if the final score is
greater than 0, we judge that Tl and T3 are d-separated by h
With two slight modifications, the PC algorithm can be applied to the zero and first order d-
separation relations determined by the vanishing tetrad differences. The first modification is of
course that the algorithm never tries to test any d-separation relation of order greater than I (i.e.
in the loop in step B) of the PC Algorithm the maximum value of n is 1.) The second is that in
step D) of the PC algorithm we do not orient edges to avoid cycles.
Without all of the d-separability facts available, the PC algorithm may not find the correct
pattern of the graph. It may include extra edges and fail to orient some edges. However, it is
possible to recognize from the pattern that some edges are definitely in the graph that generated
the pattern, while others mayor may not be. We add the following step to the PC algorithm to
label with a "?" edges that mayor may not be in the graph. Y is a definite non-collider on
an undirected path U in pattern n if and only if either X *-* Y -> Z, or X <- Y *-* Z are
subpaths of U, or X and Z are not adjacent and notX -> Y <- Z on U.
E.) Let P be the set of all undirected paths in n between X and Y of length ~ 2. If X
and Y are adjacent in n, then mark the edge between X and Y with a "?" unless either
(i) no paths are in P, or
(ii) every path in P contains a collider, or
(iii) there exists a vertex Z such that Z is a defmite non-collider on every path in P, or
(iv) every path in P contains the same subpath <A,B,C>.
We refer to the combined procedure as the Multiple Indicator Model Building (MIMBuild)
Algorithm.
Theorem 10.2: If G is an almost pure latent variable graph over V u T u C, Tis

causally sufficient, each variable in T has at least two measured indicators, the input to
MIMBuild is a list of all vanishing zero and first order correlations among the latent
variables linearly implied by G, and n is the output of the MIMBuild Algorithm then:
A-I) If X and Yare not adjacent in n, then they are not adjacent in G.
A-2) If X and Yare adjacent in nand the edge is not labeled with a "?", then X and
Y are adjacent in G.
0-1) If X -> Y is in n, then every trek in G between X and Y is into Y.
0-2) If X -> Y is in n and the edge between X and Y is not labeled with a "?", then
X -> Y is in G.
The algorithm's complexity is bounded by the number of tetrad differences it must test, which
in turn is bounded by the number of foursomes of measured variables. If there are n measured
variables the number of foursomes is O(n4). We do not test each possible foursome, however,
and the actual complexity depends on the number of latent variables and how many variables
measure each latent. If there are m latent variables and s measured variables for each, then the
number of foursomes is O(m3 xs4). Since m x s = n, this is O(n3 xs).
10.6 Simulation Tests
The procedure we have sketched has been fully automated in the TE1RAD II program, with
sensible but rather arbitrary weighting principles where required. To test the behavior of the
procedure we generated data from the causal graph in figure 13, which has 11 impure
indicators.
The distribution for the exogenous variables is standard normal. For each sample, the linear
coefficients were chosen randomly between .5 and 1.5.
We conducted 20 trials each at sample sizes of 100, 500, and 2000. We counted errors of
commission and errors of omission for detecting uncorrelated latents (O-order d-separation) and
for detecting 1st-order d-separation. In each case we counted how many errors the procedure
could have made and how many it actually made. We also give the number of samples in which
the algorithm identified the d-separations perfectly. The results are shown in Table 1, where the
proportions in each case indicate the number of errors of a given kind over all samples divided
by the number of possible errors of that kind over all samples.
Extensive simulation tests with a variety of latent topologies for as many as six latent variables,
and 60 normally distributed measured variables, show that for a given sample size the reliability
of the procedure is determined by the number of indicators of each latent and the proportion of
indicators that are confounded. Increased numbers of almost pure indicators make decisions
about d-separability more reliable, but increased proportions of confounded variables makes
identifying the almost pure indicators more difficult. For large samples with ten indicators per
latent the procedure gives good results until more than half of the indicators are confounded.
Figure 13 Impure Indicators = {X .. X 2, X 4, X ~O, X 12, X 14, X 18, X 23, X 25,

X27,X30}
Table 1
O-order O-order 1st-Order 1st-Order
Sample Size Corrunission Perfect
Omission Corrunission Omission
100 2.50% 0.00% 3.20% 5.00% 65.00%
500 1.25% 0.00% 0.90% 0.00% 95.00%
2000 0.00% 0.00% 0.00% 0.00% 100.00%

10.7 Conclusion
Alternative strategies are available. One could, for example, purify the measurement sets, and
specify a "theoretical model" in which each pair of latent variables is directly correlated. A
maximum likelihood estimate of this structure will then give an estimate of the correlation matrix
for the latents. The correlation matrix could then be used as input to the PC or FCI algorithms.
The strategy has two apparent disadvantages. One is that these estimates typically depend on an
assumption of normality. The other is that in preliminary simulation studies with normal
variates and using LISREL to estimate the latent correlations, we have found the strategy less
reliable than the procedure described in this chapter. Decisions about d-separation facts among
latent variables seem to be more reliable if they are founded on a weighted average of a number
of decisions about vanishing tetrad differences based on measured correlations than if they are
founded on decisions about vanishing partial correlations based on estimated correlations.
The MIMBuild algorithm assumes T is causally sufficient; an interesting open question is

whether there are reliable algorithms that do not make this assumption. In addition, although the
algorithm is correct, it is incomplete in a number of distinct ways. There is further orientation
information linearly implied by the zero and first order vanishing partial correlations. Further,
we do not know whether there is further information about which edges defInitely exist (Le.
should not be marked with a "?") that is linearly implied by the vanishing zero and first order
partial correlations. Moreover, it is sometimes the case that for each edge labeled with a "?"in
the MIMBuild output there exists a pattern compatible with the vanishing zero and fust order
partial correlations that does not contain that edge, but no pattern compatible with the vanishing
zero and first order partial correlations that does not contain two or more of the edges so
labeled.
Finally, and most importantly, the strategy we have described is not very informative about
latent structures that have multiple causal pathways among variables. An extension of the
strategy might be more informative and merits investigation. In addition to tetrad differences,
one could test for higher-order constraints on measured correlations (e.g. algebraic
combinations of five or more correlations) and use the resulting decisions to determine higher-
order d-separation relations among the latent variables. The necessary theory has not been
developed.
Chapter 11
Elaborating Linear Theories with Unmeasured Variables 1
11.1 Introduction
In many cases investigators have a causal theory in which they place some confidence, but they
are unsure whether the model contains all important causal connections, or they believe it to be
incomplete but don't know which dependencies are missing. How can further unknown causal
connections be discovered? The same sort of question arises for the output of the PC or FC!
algorithms when, for example, two correlated variables are disconnected in the p4ttern; in that
case we may think that some mechanism not represented in the pattern accounts for the
dependency, and the pattern needs to be elaborated. In this chapter we consider a special case of
the "elaboration problem," confined to linear theories with unmeasured common causes each
having one or more measured indicators. The general str~tegy we develop for addressing the
elaboration problem can be adapted to models without latent variables, and also to models for
discrete variables. Other strategies than those we consider here are also promising; the Bayesian
methods of Cooper and Herskovits, in particular, could be adapted to the elaboration problem.
The problem of elaborating incomplete "structural equation models" has been addressed in at
least two commercial computer packages, the LISREL program (Joreskog and Sorbom 1984)
and the EQS program (Bentler 1985). We will describe detailed tests of the reliabilities of the
automated search procedures in these packages. Generally speaking, we find them to be very
unreliable, but not quite useless, and the analysis of why they fail when other methods succeed
suggests an important general lesson about computerized search in statistics: in specification
search computation matters, and in large search spaces it matters far more than does using tests
that would, were computation free, be optimal.
lThis chapter is an abbreviated version of Spines. Scheines and Glymour (1990). and is reprinted with the
permission of Sage Publications.
We will compare the EQS and LISREL searches with a search procedure based on tests of
vanishing tetrad differences. In principle, the collection of tetrad tests is less infonnative than
maximum likelihood tests of an entire model used by the LISREL and EQS searches. In
practice, this disadvantage is overwhelmed by the computational advantages of the tetrad
procedure. Under some general assumptions, the procedure we describe gives correct (but not
necessarily fully infonnative) answers if correct decisions are made about vanishing tetrad
differences in the population. We demonstrate that for many problems the procedure obtains
very reliable conclusions from samples of realistic sizes.
11.2 The Procedure
The procedure we will describe is implemented in the TETRAD IT program. It takes as input:
(i) a sample size,
(ii) a correlation or covariance matrix, and
(iii) the directed acyclic graph of an initial linear structural equation model.
A number of specifications of internal parameters can also be input. The graph is given to the
program simply by specifying a list of paired causes and effects. The algorithm can be divided
into two parts, a scoring procedure and a search procedure.
11.2.1 Scoring
The procedure uses the following methodological principles.
Falsification Principle: Other things being equal, prefer models that do not linearly imply
constraints that are judged not to hold in the population.
Explanatory Principle: Other things being equal, prefer models that linearly imply constraints
that are judged to hold in the population.
Simplicity Principle: Other things being equal, prefer simpler models.

Elaborating Theories with Unmeasured Variables 325
The intuition behind the Explanatory Principle is that an explanation of a constraint based on the
causal structure of a model is superior to an explanation that depends upon special values of the
free parameters of a model. This intuition has been widely shared in the natural sciences; it was
used to argue for the Copernican theory of the solar system, the General Theory of Relativity,
and the atomic hypothesis. A more complete discussion of the Explanatory Principle can be
found in Glymour et. al. (1987), Scheines (1988), and Glymour (1983). As with vanishing
partial correlations, the set of values of linear coefficients associated with Ute edges of a graph
that generate a vanishing tetrad difference not linearly implied by the graph has Lebesgue
measure zero.
Unfortunately, the principles can conflict. Suppose, for example, that model M' is a
modification of model M, formed by adding an extra edge to M. Suppose further that M'
linearly implies fewer constraints that are judged to hold in the population, but also linearly
implies fewer constraints that are judged not to hold in the population. Then M' is superior to M
with respect to the Falsification Principle, but inferior to M with respect to the Simplicity and
Explanatory Principles. The procedure we use introduces a heuristic scoring function that
balances these dimensions. 2
In order to calculate the Tetrad-score we first calculate the associated probability P(t) of a
vanishing tetrad difference, which is the probability of obtairting a tetrad difference as large or
larger than the one actually observed in the sample, under the assumption that the tetrad
difference vanishes in the population. Assuming normal variates, Wishart (1928) showed that
the variance of the sampling distribution of the vanishing tetrad difference PIJPKL - PILPjK is
equal to
D1P34(N+I) D
(N-I)(N-2)
where D is the determinant of the population correlation matrix of the four variables J, J, K, and
L, D12 is the determinant of the two-dimensional upper left-comer submatrix, D34 is the
determinant of the lower right-comer submatrix and J, J, K, and L, have a joint normal
distribution. In calculating P(t) we substitute the sample covariances for the corresponding
population covariances in the formula. P(t) is determined by lookup in a chart for the standard
normal distribution. An asymptotically distribution free test has been described by Bollen
(1990).
2 The original TETRAD program (Glymour, Scheines, Spirtes and Kelly, 1987) had no such scoring function.
It was left to the user to balance the Explanatory and Falsification principles.
Among any four distinct measured variables I, J, K and L we compute three tetrad differences:
tl = PIJPKL - PILPJK
t2 = PILPJK -PIKPJL
t3 = PIKPJL -PIJPKL
and their associated probabilities P(ti) on the hypothesis that the tetrad difference vanishes. If
P(ti) is larger than the given significance level, the procedure takes the tetrad difference to
vanish in the population. If P(ti) is smaller than the significance level, but the other two tetrad
differences have associated probabilities higher than the significance level, then ti is ignored.
Otherwise, if P(ti) is smaller than the significance level, the tetrad difference is judged not to
vanish in the population.
Let ImpliedH be the set of vanishing tetrads linearly implied by a model M that are judged to
hold in the population and ImpJied.H be the set of vanishing tetrads linearly implied by M that
are judged not to hold in the population. Let Tetrad-score be the score of model M for a given
significance level assigned by the algorithm, and let weight be a parameter (whose significance
is explained below). Then we define
T= LP(t) L weight*(l- P(t»)

telmpIiedH telmpIied_H
The first term implements the explanatory principle while the second term implements the
falsification principle. The simplicity principle is implemented by preferring, among models
with identical Tetrad-scores, those with fewer free parameters--which amounts to preferring
graphs with fewer edges. The weight determines how conflicts between the explanatory and
falsification principles are resolved by determining the relative importance of explanation
relative to residual reduction.
The scoring function is controlled by two parameters. The significance level is used to judge
when a given tetrad difference is zero in the population. The weight is used to determine the
relative importance of the Explanatory and Falsification Principles. The scoring function has
several desirable asymptotic properties, but we do not know whether the particular value for
weight we use is optimal.
11.2.2 Search
The TETRAD II procedure searches a tree of elaborations of an initial model. The search is
comparatively fast because there is an easy algorithm for determining the vanishing tetrad
differences linearly implied by a graph (using the Tetrad Representation Theorem), because
most of the computational work required to evaluate a model can be stored and reused to
evaluate elaborations, and because the scoring function is such that if a model can be
conclusively eliminated from consideration because of a poor score, so can any elaboration of
it.
The search generates each possible one-edge elaboration of the initial model, orders them by the
tetrad score, and eliminates any that score poorly. It then repeats this process recursively on
each model generated, until no improvements can be made to a model.
The search is guided by a quantity called T -maxscore, which for a given model M represents
the maximum Tetrad-score that could possibly be obtained by any elaboration of M. T-
maxscore is equal to:
T - maxscore = L,P(t)
tElmpliedH
The use we make of this quantity is justified by the following theorem.
Theorem 11.1. If G is a sub graph of directed acyclic graph G', than the set of tetrad
equations among variables of G that are linearly implied by G' is a subset of those linearly
implied by G.
In order to keep the following example small, suppose that there are just 4 edges, el, e2, e3, or
e4 which could be added to the initial model. The example illustrates the search procedure in a
case where each possible elaboration of the initial model is considered. Node 1 in figure 1
represents the initial model. Each node in the graph represents the model generated by adding
the edge next to the node to its parent. For example, node 2 represents the initial model + et;
node 7 represents node 2 + e4, which is the initial model + el + e4. We will say that a program
visits a node when it creates the model M corresponding to the node and then determines
whether any elaboration of M has a higher Tetrad-score than M. (Note that the algorithm can
generate a model M without visiting M if it generates M but does not determine whether any
elaboration of M has a higher Tetrad-score than M.) The numbers inside each node indicate the
order in which the models are visited. Thus for example, when the algorithm visits node 2, it
first generates all possible one edge additions of the initial model + el, and orders them
according to their T-maxscore. It then first visits the one with the highest T-maxscore (in this
case, node 3 that represents the initial model + el + ez). Note that the program does not visit the
initial model + ez (node 10) until after it has visited all elaborations of the initial model + el.
Figure 1
In practice, this kind of complete search could not possibly be carried out in a reasonable
amount of time. Fortunately we are able to eliminate many models from consideration without
actually visiting them. Addition of edges to a graph may defeat the linear implication of tetrad
equations, but in view of Theorem 11.1 will never cause more tetrad equations to be linearly
implied by the resulting graph. If the T-maxscore of a model M is less than the Tetrad-score of
some model M' already visited, then we know that neither M nor any elaboration of M has a
Tetrad-score as high as that of M'. Hence we need never visit M or any of its elaborations. This
is illustrated in figure 2. If we fInd that T-maxscore of the initial model + e4 is lower than the
Tetrad-score of the initial model + el, we can eliminate from the search all models that contain
the edge e4.
Figure 2
In some cases in the simulation study described later, the procedure described here !s too slow
to be practical. In those cases the time spent on a search is limited by restricting the depth of
search. (We made sure that in every case the depth restriction was large enough that the
program had a chance to err by overfitting.) The program adjusts the search to a depth that can
be searched in a reasonable amount of time; in many of the Monte Carlo simulation cases no
restriction on depth was necessary.3
11.3 The LISREL and EQS Procedures
LISREL VI and EQS are computer programs that perform a variety of functions, such as
providing maximum likelihood estimates of the free parameters in a structural equation model.
The feature we will consider automatically suggests modifications to underspecified models.
11.3.1 Input and Output
Both programs take as input:

(i) a sample size,
3 We have also implemented heuristic search procedures that are theoretically less reliable than that described here
but are much faster and in practice about equally reliable.
(ii) a sample covariance matrix,

(iii) initial estimates of the variances of independent variables,
(iv) initial estimates of the linear coefficients,
(v) an initial causal model (specified by fixing at zero the linear coefficient of A in the
equation for B if and only if A is not a direct cause of B), in the form of equations
(EQS) or a system of matrices (LISREL VI)
(vi) a list of parameters not to be freed during the course of the search,
(vii) a significance level, and
(viii) a bound on the number of iterations in the estimation of parameters.
The output of both programs includes a single estimated model that is an elaboration of the
initial causal model, various diagnostic information as well as a X2 value for the suggested
revision, and the associated probability of the X2 measure.
11.3.2 Scoring
LISREL VI and EQS provide maximum likelihood estimates of the free parameters in a
structural equation model. More precisely, the estimates are chosen to minimize the fitting
function
F =logl~ + tr(SL- 1) -logl51- t
where S is the sample covariance matrix, 1: is the predicted covariance matrix, t is the total
number of indicators, and if A is a square matrix then IAI is the determinant of A and tr (A) is
the trace of A. In the limit, the parameters that minimize the fitting function F also maximize the
likelihood of the covariance matrix for the given causal structure.
After estimating the parameters in a given model, LISREL VI and EQS test the null hypothesis
that 1: is of the form implied by the model against the hypothesis that 1: is unconstrained. If the
associated probability is greater than the chosen significance level, the null hypothesis is
accepted, and the discrepancy is attributed to sample error; if the probability is less than the
significance level, the null hypothesis is rejected, and the discrepancy is attributed to the falsity
of M. For a "nested" series of models MI. ... ,Mk in which for all models M; in the sequence the
free parameters of M; are a subset of the free parameters of M;+I. asymptotically, the difference
between the X2 values of two nested models also has a X2 distribution, with degrees of freedom
equal to the difference between the degrees of freedom of the two nested models.
11.3.3 The LISREL VI Search4
The LlSREL VI search is guided by the "modification indices" of the fixed parameters. Each
modification index is a function of the derivatives of the fitting function with respect to a given
fixed parameter. More precisely, the modification index of a given fixed parameter is defined to
be N/2 times the ratio between the squared first-order derivative and the second-order derivative
(where N is the sample size). Each modification index provides a lower bound on the decrease
in the %2 obtained if that parameter is freed and all previously estimated parameters are kept at
their previously estimated values. s (Note that if the coefficient for variable A in the linear
equation for B is fixed at zero, then freeing that coefficient amounts to adding an edge from A to
B to the graph of the model.) LlSREL VI first makes the starting model the current best model
in its search. It then calculates the modification indices for all of the fixed parameters6 in the
starting model. If LlSREL VI estimates that the difference between the %2 statistics of M, the
current best model, and M', the model obtained from M by freeing the parameter with the
largest modification index, is not significant, then the search ends, and LlSREL VI suggests
model M. Otherwise, it makes M' the current best model and repeats the process.
11.3.4 The EQS Search
EQS computes a Lagrange Multiplier statistic, which is asymptotically distributed as %2.7 EQS
performs univariate Lagrange Multiplier tests to determine the approximate separate effects on
the X2 statistic of freeing each fixed parameter in a set specified by the user. It frees the
parameter that it estimates will result in the largest decrease in the %2 value. The program repeats
th~s procedure until it estimates that there are no parameters that will significantly decrease the
X2. Unlike LlSREL VI, when EQS frees a parameter it does not reestimate the model. 8
4LISREL vn retains the same architecture but with an altered modification index.
5 LISREL VI outputs a number of other measures that could be used to suggest modifications to a starting
model, but these are not used in the automatic search. See Costner and Herting (1985).
6 As long as they are not in the list of parameters not to be freed.
7 Since the Lagrange Multiplier statistic, like the modification indices of LISREL VI, estimates the effect on the
'1.2 of freeing a parameter, in subsequent sections we will use the tenn "modification index" to refer to either of
these statistics.
8 EQS allows the user to specify several different types of searches. We have only described the one used in our
Monte Carlo simulation tests.
It should be noted that both LISREL VI and EQS are by now quite complicated programs. An
understanding of their flexibility and limitations can only be obtained through experimentation
with the programs.
11.4 The Primary Study
Eighty data sets, forty with a sample size of 200 and forty with a sample size of 2,000, were
generated by Monte Carlo methods from each of nine different structural equation models with
latent variables. The models were chosen because they involve the kinds of causal structures
that are often thought to arise in social and psychological scientific work. In each case part of
the model used to generate the data was omitted and the remainder, together in tum with each of
the data sets for that model, was given to the LISREL VI, EQS, and TETRAD II programs. A
variety of specification errors are represented in the nine cases. Linear coefficient values used in
the true models were generated at random to avoid biasing the tests in favor of one or another of
the procedures. In addition, a number of ancillary studies were suggested by the primary
studies and bear on the reliability of the three programs.
11.4.1 The Design of Comparative Simulation Studies
To study the reliability of automatic respecification procedure under conditions in which the
general structural equation modeling assumptions are met, the following factors should be
varied independently.
(i) the causal structure of the true model;
(ii) the magnitudes and signs of the parameters of the true model;
(iii) how the starting model is misspecified;
(iv) the sample size.
In addition, an ideal study should:

(i) Compare fully algorithmic procedures, rather than procedures that require judgment on
the part of the user. Procedures that require judgment can only adequately be tested by
carefully blinding the user to the true model; further, results obtained by one user may
not transfer to other users. With fully algorithmic procedures, neither of these problems
arises.
(ii) Examine causal structures that are of a kind postulated in empirical research, or that
there are substantive reasons to think occur in real domains.
(iii) Generate coefficients in the models randomly. Costner and Herting showed that the size
of the parameters affects LlSREL's performance. Further, the reliability of TETRAD n
depends on whether vanishing tetrad differences hold in a sample because of the
particular numerical values of the coefficients rather than because of the causal structure,
and it is important not to bias the study either for or against this possibility.
(iv) Ensure insofar as possible that all programs compared must search the same space of
alternative models.
11.4.2 Study Design
11.4.2.1 Selection of Causal Structures
The nine causal structures studied are illustrated in figures 3, 4 and 5. For simplicity of
depiction we have omitted uncorrelated error terms in the figures, but such terms were included
in the linear models. The heavier directed or undirected lines in each figure represent
relationships that were included in the model used to generate'simulated data, but were omitted
from the models given to the three programs; i.e., they represent the dependencies that the
programs were to attempt to recover. The starting models are shown in figure 6. The models
studied include a one factor model with five measured variables, seven multiple indicator
models each with eight measured variables and two latent variables, and one multiple indicator
model with three latent variables and eight measured variables.
One factor models commonly arise in psychometric and personality studies (see Kohn 1969);
two latent factor models are common in longitudinal studies in which the same measures are
taken at different times (see McPherson et. al. 1977), and also arise in psychometric studies; the
triangular arrangement oflatent variables is a typical geometry (see Wheaton et al. 1977).
The set of alternative structures determines the search space. Each program was forced to search
the same space of alternative elaborations of the initial model, and the set of alternatives was
chosen to be as large as possible consistent with that requirement
Modell
Model 2
-.6
Tl~--------~--------~
~ ~L'~
'---- 1.2
e _ 1.92 .E
5'---"'" 6
Model 3
-.87
~ ~
" -_______
- ~1~.2____- -
Figure 3
Model 4
U,X~· ., ,
71. ~
~~[;]~
63
\
~ 6
t
£2
t
£3
t
£5
'---- ">.,. .818 =- ~
4.77
£7
--"
ModelS
-.87
r---------------------~T2
,~
[;]-+6 ~ ~
" -_______
- ~2~.2~1__---~
Model 6
Figure 4
Model 7
Model 8
Model 9
Figure 5
Start for Model 1
Start for Models 2 - 8
Start for Model 9
Figure 6
11.4.2.2 Selection of Connections to be Recovered
The connections to be recovered include:

(i) Directed edges from latent variables to latent variables; relations of this kind are often the
principal point of empirical research. See Maruyama and McGarvey (1980) for an
example.
(ii) Edges from latent variables to measured variables; connections of this kind may arise
when measures are impure, and in other contexts. See Costner and Schoenberg (1973)
for an example.
(iii) Correlated errors between measured variables; relationships of this kind are perhaps the
most frequent form of respecification.
(iv) Directed edges from measured variables to measured variables. Such relations cannot
obtain, for example, between social indices, but they may very well obtain between
responses to surveyor psychometric instruments (see Campbell et. aI. 1966), and of
course between measured variables such as interest rates and housing sales.
We have not included cases that we know beforehand cannot be recovered by one or another of
the programs. Details are given in a later section.
11.4.2.3 Selection of Starting Models
Only three starting models were used in the nine cases. The starting models are, in causal
modeling terms, pure factor models or pure multiple indicator models. In graph theoretic terms
they are trees.
11.4.2.4 Selection of Parameters
In the figures showing the true models the numbers next to directed edges represent the values
given to the associated linear coefficients. The numbers next to undirected lines represent the
values of specified covariances. In all cases, save for models 1 and 5, the coefficients were
chosen by random selection from a uniform distribution between .5 and 2.5. The value obtained
was then randomly given a sign, positive or negative.
In modell, all linear coefficients were made positive. The values of the causal connections
between indicators were specified non-randomly. The case was constructed to simulate a
psychometric or other study in which the loadings on the latent factor are known to be positive,
and in which the direct interactions between measured variables are comparatively small.
Model 5 was chosen to provide a comparison with model 3 in which the coefficients of the
measured-measured edges were deliberately chosen to be large relative to those in model 3.
11.4.2.5 Generation of Data
For each of the nine cases, twenty data sets with sample size 200 and twenty data sets with
sample 2,000 were generated by Monte Carlo simulation methods.
Pseudo-random samples were generated by the method described in Chapter 5. In order to

optimize the performance of each of the programs, we assumed that all of the exogenous
variables had a standard normal distribution. This assumption made it possible to fix a value for
each exogenous variable for each unit in the population by pseudo random sampling from a
standard normal distribution. Correlated errors were obtained in the simulation by introducing a
new exogenous common cause of the variables associated with the error terms.
11.4.2.6 Data Conditioning
The entire study we discuss here was performed twice. In the original study, we gave LISREL
VI and EQS positive starting values for all parameters. If either program had difficulty
estimating the starting model, we reran the case with the initial values set to the correct sign.
LISREL and EQS employ iterative procedures to estimate the free parameters of a model. These
procedures are sensitive to "poorly conditioned" variables and will not perform optimally unless
the data are transformed. For example, it is a rule of thumb with these procedures that no two
variances should vary by more than an order of magnitude in the measured variables. After
generating data in the way we describe above, a small but significant percentage of our
covariance matrices were ill conditioned in this way.
To check the possibility that the low reliability we obtained in the first study for the LISREL VI
and EQS procedures was due to "ill-conditioned" data, the entire study was repeated. Sample
covariances were transformed into sample correlations by dividing each cell [J,J] in the
covariance matrix by SIS], where SI is the sample standard deviation of l. To avoid sample
variances of widely varying magnitudes, we transformed each cell [J,J] in the sample
covariance matrix by dividing it by UJu] where UI is the population standard deviation of J9 We
call the result of this transformation the pseudnco"elation matrix. The transformation makes all
of the variances of the measured variables almost equal, without using a data-dependent
transformation. Of course in empirical research, this transformation could not be performed,
since the population parameters would not be known.
In practice, we found that conditioning the data and giving the population parameters as starting
values did little to change the performance of LISREL VI or EQS. The performance of the
TETRAD II procedure was essentially the same in both cases. Conditioning the data improved
LISREL VI's reliability very slightly for small samples, and degraded it slightly for large
samples.
11.4.2.7 Starting Values for tbe Parameters
We selected the linear coefficients for our models randomly, allowing some to be negative and
some to be positive. Models with negative parameters actually represent a harder case for the
TETRAD procedures. If a model implies a vanishing tetrad difference then the signs of its
parameters make no difference. If a model does not imply that a tetrad difference vanishes,
however, but instead implies that the tetrad difference is equal to the sum of two or more terms,
then it is possible, if not all of the model's parameters are positive, that these terms sum to zero.
Thus, in data generated by a model with negative parameters, we are more likely to observe
vanishing tetrad differences that are not linearly implied by the model.
The iterative estimation procedures for LISREL and EQS begin with a vector of parameters (J.
They update this vector until the likelihood function converges to a local maximum. Inevitably,
the iterative procedures are sensitive to starting values. Given the same model and data, but two
different starting vectors (Ji and (Jj, the procedures might converge for one but not for the
other. This is especially true when the parameters are of mixed signs. To give LISREL and
EQS the best chance possible in the second study, we set the starting values of each parameter
to its actual value whenever possible. For the linear coefficients that correspond to edges in the
9 We are indebted to Peter Bentler for suggesting Ibis transfonnation.

generating model left out of the starting model, we assigned a starting value of O. For all other
parameters, however, we started LISREL and EQS with the exact value in th£ population. 10
1l.4.2.8 Significance Levels
EQS and LISREL VI continue to free parameters as long as the associated probability of the
decrease in X2 exceeds the user-specified significance level. For both LISREL and EQS, we set
the significance level to .01. (This is the default value for LISREL; the default value for EQS is
.05.) The lower the significance level, the fewer the parameters that each program tends to set
free. Since both LISREL and EQS both tend to overfit even at .01, we did not attempt to set the
significance level any higher. (It may appear in our results that LISREL VI and EQS both
underfit more than they overfit, but almost all of the "underfitting" was due to aborted searches
that did not employ the normal stopping criterion.)
1l.4.2.9 Number of Iterations
The default number of maximum iterations for estimating parameters for LISREL VI on a
personal computer is 250. We set the number of maximum iterations to 250 for both our
LISREL VI and EQS tests.
1l.4.2.10 Specifying Starting Models in LISREL VI
LISREL VI, like previous editions of the program, requires the user to put variables into
distinct matrices according to whether they are exogenous, endogenous but unmeasured,
measured but dependent on exogenous latent, measured but dependent on endogenous latent,
and so forth. Variables in certain of these categories cannot have effects on variables in other
categories. When formulated as recommended in the LISREL manual, LISREL VI would be in
principle unable to detect many of the effects considered in this study. However, these
restrictions can in most cases be overcome by a system of substitutions of phantom variables in
lOWe did not provide LISREL or EQS with the values of the parameters in the original models that generated
our covariance matrices because the input to LISREL and EQS was a pseudocorrelation matrix. not the original
covariance matrix. We therefore provided the programs with the population parameters of ttansformed models that
would generate the pseudo correlation matrices. The detailed transformations are given in Spines (1990).
which measured variables are actually represented as endogenous latent variables.ll In the
current study, we were not able to get LISREL VI to accept changing ~ variables, which are
exogenous and latent, to T/ variables, which are endogenous and latent. This had the unfortunate
effect that LISREL would not consider adding any edges into Tl (represented by the ~
variable). To ensure a comparable search problem, we restricted TETRAD IT and EQS in the
same way.
11.4.2.11 Implementation
The LISREL VI runs were performed with the personal computer version of the program, run
on a Compaq 386 computer with a math coprocessor. EQS runs were performed on an mM XT
clone with a math coprocessor. All TETRAD II runs were performed on Sun 3/50
workstations. For TETRAD II, which also runs on mM clones, the processing time for the
Compaq 386 and the Sun 3/50 are roughly the same.
11.4.2.12 Specification of TETRAD II Parameters
TETRAD II requires that the user set a value of the weight parameter, a value for the
significance level used in the test for vanishing tetrad differences, and a value for a percentage
parameter that bounds the search. In all cases we set the significance level at 0.05. At sample
size 2000, we set the weight to.l and the percentage to 0.95.
At smaller sample size the estimates of the population covariances are less reliable, and more
tetrad differences are incorrectly judged to vanish in the population. This makes judgments
about the Explanatory Principle less reliable. For this reason, at sample size 200, we set the
weight to 1, in order to place greater importance upon the Falsification Principle. Less reliable
judgments about the Explanatory Principle also make lowering the percentage for small sample
sizes helpful. At sample size 200, we set the percentage to 0.90. We do not know if these
parameter settings are optimal.
11 For LISREL IV. the details of this procedure are described in Discovering Causal Structure. The same
procedure works for LISREL VI with the exception of the Beta matrix. See Joreskog and Sorbom (1984).
11.5 Results
For each data set and initial model, TETRAD II produces a set of best alternative elaborations.
In some cases that set consists of a single model; typically it consists of two or three
alternatives. EQS and LISREL VI, when run in their automatic search mode, produce as output
a single model elaborating the initial model. The information provided by each program is
scored "correct" when the output contains the true model. But it is important to see how the
various programs err when their output is not correct, and we have provided a more detailed
classification of various kinds of error. We have classified the output of TETRAD II as follows
(where a model is in TETRAD's top group if and only if it is tied for the highest Tetrad score,
and no model with the same Tetrad-score has fewer edges):
Correct--the true model is in TETRAD's top group.

Width--the average number of alternatives in TETRAD's top group.
Errors:
Overfit--TETRAD's top group does not contain the true .model but contains a model that is
an elaboration of the true model.
Underfit--TETRAD's top group does not contain the true model but does contain a model of
which the true model is an elaboration.
Other--none of the previous categories apply to the output.
We have scored the output of the LISREL VI and EQS programs as follows:
Correct--the true model is recommended by the program.
Errors:
In TETRAD's Top Group--the recommended model is not correct, but is among the best
alternatives suggested by the TETRAD II program for the same data.
Overfit--the recommended model is an elaboration of the true model.
Underfit--the true model is an elaboration of the recommended model.
Right Variable Pairs--the recommended model is not in any of the previous categories, but it
does connect the same pairs of variables as were connected in the omitted parts of the
true model.
Other--none of the previous categories apply to the output.
In most cases no estimation problems occurred for either LISREL VI or EQS. In a number of
data sets for cases 3 and 5, LISREL VI and EQS either issued warnings about estimation
problems or aborted the search due to computational problems. Since our input files were built
to minimize convergence problems, we ignored such warnings in our tabulation of the results.
If either program recommended freeing a parameter, we counted that parameter as freed
regardless of what warnings or estimation problems occurred before or after freeing it. If either
program failed to recommend freeing any parameters because of estimation problems in the
starting model, we counted it as an underfit. The results are shown in the next table and figure.
TABLE 1: Case by Case Width of Set Suggestions

Width, n=2000
Case 1 2 3 4 5 6 7 8 9
LISREL VI 1 1 I 1 I 1 1 1 1
EQS 1 1 1 1 I I I I 1
TETRAD 4 2.1 2 1 1.1 3 7.1 11.3 2.9
Width, n=200
Case 1 2 3 4 5 6 7 8 9
LISREL VI 1 1 1 I I 1 I I 1
EQS 1 I 1 1 I 1 I 1 I
TETRAD 1.9 3.5 1.5 1 1 3.2 5.9 8.4 3
U\,
90
TETRAD =
LlSREL = Ii2
EQS = ES1
•
Rr
70 70
% 60 60
Correct 50 50
40 40
30 30
20 20
0
O-
n =
tt2000
~
N;j
n = 200
10
o
n = 2000
tOO 1ETRAD
90
L1SREL ~
EQS ISS]
80
70 ------------------------------------------- 70
%
60
50 -------------------------------------------- 50
40
30
20
10
0
In TClJ'ad's Right Variable Other
~ Underfit
Correct Top Group Pairs ~
v
Incorrect
n = 200
lOO TETRAD
90 L1SREL
0 EQS
70 r------------------------------------------------- 70
%
60 r------------------------------------------ 60
50 --------------------------------------------- 50
40 40
30 30
20 20
10 10
0 ........J...lo..L-O
Overfit Underfit In TClJ'ad's Right Variable Other
Correct '----- Top Group Pairs ~
-------~v
Incorrect
Figure 8
For a sample size of 2000, TETRAD II's set included the correct respecification in 95% of the
cases. LISREL VI found the right model 18.8% of the time and EQS 13.3%. For a sample size
of 200, TETRAD II's set included the correct respecification 52.2% of the time, while LISREL
VI corrected the misspecification 15.0% of the time, and EQS corrected the misspecification
10.0 % of the time. A more detailed characterization of the errors is given in figure 8.
11.6 Reliability and Informativeness
There are two criteria by which the suggestions of each of these programs can be judged. The
first is reliability. Let the reliability of a program be defined as the probability that its set of
suggested models includes the correct one. In these cases, the TETRAD search procedures are
clearly more reliable than either LISREL VI and EQS. One can achieve higher reliability simply
by increasing the number of guesses. A program that outputs the top million models might be
quite reliable, but its suggestions would be uninformative. Thus we call the second criterion
boldness. Let the boldness of a program's suggestions be the reciprocal of the number of
models suggested. On this measure, our procedure does worse than LISREL VI or EQS in
seven of the nine cases considered.
Since neither our procedure nor the modification index procedures dominate on both of these
criteria, it is natural to ask whether the greater reliability of the former is due simply to reduced
boldness. This question can be interpreted in at least two ways:
(i) If TETRAD II were to increase its boldness to match LISREL VI and EQS, Le., if it
were to output a single model, would it be more or less reliable than LISREL VI or
EQS?
(ii) If LISREL VI or EQS were to decrease their boldness to match TETRAD II, i.e., were
they to output a set of models as large as does TETRAD II, would they be more or less
reliable than TETRAD II?
If we have no reason to believe that anyone model in the TETRAD II output is more likely than
any other to be correct, we could simply choose a model at random. We can calculate the
expected single model reliability of our procedure in the following way. We assume that
when TETRAD II outputs a list of n models for a given covariance matrix, the probability of
selecting any particular one of the models as the best guess is lin. So instead of counting a list
of length n that contains the correct model as a single correct answer, we would count it as lin
correct answers.12 Then simply divide the expected number of correct answers by the number
of trial runs.
Were TETRAD II to be as bold as LISREL VI or EQS, its single model reliability at sample size
2000 would drop from 95% to about 42.3%. On our data, LISREL VI has a reliability of
18.8% and EQS has a reliability of 13.3%. At sample size 200 the TETRAD II single model
reliability is 30.2% LISREL has a reliability of 15.0% for sample size 200 and EQS 10.0%. In
a more realistic setting one might have substantive reasons to prefer one model over another. If
substantive knowledge is worth anything, and we use it to select a single model M, then Mis
more likely to be true than a model selected at random from TETRAD Irs set of suggested
models. Thus, in a sense the numbers given in the paragraph above are worst case.
An alternative strategy is to cut down the size of the set before one picks a model. We can often
eliminate some of the TETRAD II suggestions by running them through EQS or LISREL VI
and discarding those that were not tied for the highest associated probability. There is little
effect. We raise the (worst case) single model reliability of TETRAD II at sample size 2000
from 42.3% to about 46%, and at sample size 200 from 30.2 to approximately 32%.
There are a number of good reasons to want a list of equally good suggestions rather than a
single guess. All have to do with the reliability and informativeness of the output.
First, it is important for the user of a program to have a good idea of how reliable the output of
a program is. At sample size 2000, in the range of cases that we considered, the reliability of the
TETRAD II output was very stable, ranging from a low of 90% to a high of 100%. For reasons
explained below, the single model output by LISREL VI andEQS is at best in effect a random
selection from a list of models that contains all of the models whose associated probabilities are
equal to that of the true model (and possibly others of lower associated probabilities as well).
Unfortunately, the size of the list from which the suggested model is randomly selected varies a
great deal depending on the structure of the model, and is not known to the user. Thus, even
ignoring the cases where LISREL VI had substantial computational difficulties, the reliability of
LISREL VI's output at sample size 2000 ranged from 0 out of 20 to 11 out of 20. So it is rather
difficult for a user of LISREL VI or EQS to know how much confidence to have in the
suggested models.
12To simplify the calculations, we assumed that the length of the lists output by 1ElRAD II for all of the
covariance matrices generated by a single model was in each case equal to the average length of the lists. This is
a fairly good approximation in most cases.
Second, more than one model in a suggested set might lead to the same conclusion. For
example, many of the models suggested by TETRAD II might overlap, i.e., they might agree
on a substantial number of the causal connections. If one's research concerns are located within
those parts of the models that agree, then choosing a single model is not necessary. In this case
one need not sacrifice reliability by increasing boldness, because all competitors agree.
Finally, having a well-defmed list of plausible alternatives is more useful than a single less
reliable suggestion for guiding further research. In designing experiments and in gathering more
data it is useful to know exactly what competing models have to be eliminated in order to
establish a conclusive result. For example, consider case 3. The correct model contains edges
from X I to Xs and Xs to X6. TETRAD II suggests the correct model, as well as a model
containing edges from Xl to Xs and Xl to X6. An experiment which varied XI, and examined
the effect on X6 would not distinguish between these two alternatives (since both predict that
varying Xl would cause X6 to change), but an experiment which varied Xs and examined the
effect on X6 would distinguish between these alternatives. Only by knowing the plausible
alternatives can we decide which of these experiments is more useful.
If LISREL VI or EQS were to output a set of models as large as does our procedure, would
they be as reliable? The answer depends upon how the rest of the models in the set were
chosen. In many cases LISREL VI and EQS fmd several parameters tied, or almost tied, for the
highest modification index. Currently both programs select one, and only one, of these
parameters to free, on the basis of an arbitrary ordering of parameters. For example, if after
evaluating the initial model it found that X3 -> Xs and X3 C XS13 were tied for the highest
modification indices, LISREL VI or EQS would choose one of them, (say X3 -> Xs) and
continue until the search found no more parameters to free. Then they would suggest the single
model that had the highest associated probability. If LISREL VI or EQS searched all branches
corresponding to tied modification indices, instead of arbitrarily choosing one, their reliability
would undoubtedly increase substantially. For example, after freeing X3 -> Xs and then freeing
parameters until no more should be freed, LISREL VI or EQS could return to the initial model,
free X3 C Xs, and again continue freeing parameters until no more should be freed. They could
then suggest all of the models tied for the highest associated probability. This is essentially the
search strategy followed by the TETRAD II program.
If the LISREL VI search were expanded in this way on case 1 at sample size 2000, it would
increase the number of correct outputs from 3 to 16 out of 20. In other cases, this strategy
13 The expression "X C Y" means that the error terms for X and Y are correlated, or, equivalently, that there is
an additional, common cause of X and Y
would not improve the performance of LISREL VI or EQS much at all. For example, in case 5
at sample size 2000, LISREL VI was incorrect on every sample in part because of a variety of
convergence and computational problems, while TETRAD IT was correct in every case. In case
4 at sample size 2000, LISREL VI missed the correct answer on nine samples (while TETRAD
n missed the correct answer on only two samples) for reasons having nothing to do with the
method of breaking ties.
LISREL VI and EQS would pay a substantial price for expanding their searches; their
processing time would increase dramatically. A branching procedure that retained three
alternatives at each stage and which stopped on all branches after freeing two parameters in the
initial model, would increase the time required by about a factor of 7. In general, the time
required for a branching search increases exponentially as the number of alternatives considered
at each stage. Could such a search be run in a reasonable amount of time? Without a math
coprocessor, a typical LISREL VI run on a Compaq 386 took roughly 20 minutes; with a math
coprocessor it took about 4 minutes. EQS runs were done on a LEADING EDGE (anmM XT
clone that is considerably slower than the COMPAQ 386) with a math coprocessor and the
average EQS run was about 5 minutes. This suggests that a branching strategy is possible for
LISREL VI even for medium-sized models only on relatively fast machines; a branching search
is practical on slower machines for the faster, but less reliable EQS search.
11.7 Using LISREL and EQS as Adjuncts to Search
There are two ways in which the sort of search TETRAD n illustrates can profitably be used in
conjunction with LISREL VI or EQS. A procedure such as ours can be used to generate a list of
alternative revisions of an initial model, which can then be estimated by LISREL or EQS,
discarding those alternatives that have very low, or comparatively low associated
probabilities. 14 We found that in only three cases could the associated probabilities distinguish
among models suggested by TE1RAD n. In case 6, one of the three models 'suggested by
TETRAD n had a lower associated probability that the other two. In case 7, one of the six
models suggested by TETRAD n had a lower associated probability that the other five. The
largest reduction in TE1RAD II's suggestions came in case 8, where 8 of the 12 models
suggested by TE1RAD IT had associated probabilities lower than the top four. These results
141ETRAD IT will, on request, automatically generate EQS input files for all models that it suggests.
were obtained when LISREL VI was given the correct starting values for all of the edges in the
true model, and a starting value of zero for edges not in the true model; in previous tests when
LISREL VI was not given the true parameters as initial values, it often suffered convergence
problems.
It is also instructive to run the both the automatic searches of TETRAD II and LISREL VI or
EQS together. When LISREL VI and TETRAD II agree (that is when the model suggested by
LISREL VI is in TETRAD II's top group) both programs are correct a higher percentage of
times than their respective averages; conversely when they disagree, both programs are wrong a
higher percentage of times than their average. The same holds true of EQS when used in
conjunction with TETRAD II. Indeed, at sample size 2000, neither EQS nor LISREL VI was
ever correct when it disagreed with TETRAD II. In contrast, at sample size 2000 LISREL VI
was correct 61.8% of the time when it agreed with TETRAD II, and EQS was correct 53.3% of
the time when it agreed with TETRAD II. Again, at sample size 2000, TETRAD II was always
correct when it agreed with either LISREL VI or EQS. At sample size 200, while TETRAD II
was correct on average 52.2% of the time, when it agreed with LISREL VI it was correct
75.7% of the time, and when it agreed with EQS it was correct 75.0% of the time. These results
are summarized below:
Sample size 2000:
P(TETRAD correct) 95.0

P(LISREL VI correct) 18.8
P(EQS correct) 13.3
P(TETRAD correct I LISREL VI agree) 100.0
P(TETRAD correct I LISREL VI disagree) 92.1
P(TETRAD correct I EQS agree) 1000
P(TETRAD correct I EQS disagree) 92.6
P(LISREL VI correct I TETRAD II agree) 61.8
P(LISREL VI correct I TETRAD II disagree) 0.0
P(EQS correct I TETRAD II agree) 53.3
P(EQS correct I TETRAD II disagree) 0.0
Sample size 200:
P(TETRAD correct) 52.2

P(LISREL VI correct) 15.0
P(EQS correct) 10.0
P(TETRAD correct I LISREL VI agree) 75.7
P(TETRAD correct I LISREL VI disagree) 46.9
P(TETRAD correct I EQS agree) 75.0
P(TE1RAD correct I EQS disagree) 47.2
P(LISREL VI correct I TETRAD IT agree) 39.4
P(LISREL VI correct I TETRAD IT disagree) 9.5
P(EQS correct I TETRAD IT agree) 43.7
P(EQS correct I TETRAD IT disagree) 2.7
11.8 Limitations of the TETRAD II Elaboration Search
The TETRAD II procedure cannot find the correct model if there are a large number of
vanishing TETRAD differences that are not linearly implied by the true model, but hold because
of coincidental values of the free parameters. Our study indicates that this occurrence is
unusual, at least given the uniform distribution that we placed on the linear coefficients in the
models that generated our data, but it certainly does occur. The same results can be expected for
any other "natural" distribution on the parameters. Further, the search does not guarantee that it
will fmd all of the models that have the highest Tetrad-score. But in many cases, depending
upon the size of the model, the amount of background knowledge, the structure of the model,
and the sample size, the search space is so large that a search that guarantees fmding the models
with the highest Tetrad-score is not practical. One way the procedure limits search is through
the application of the simplicity principle. This is a substantive assumption that may be false.
The simplicity assumption is not needed for some smaIl models, but in many problems with
more variables there may be a large number of models that have maximal scores but contain
many redundant edges that do not contribute to the score. Without the use of the simplicity
principle, it is often difficult to search this space of models and if it is searched, there may be so
many models tied for the highest score that the output is uninformative. If a model with
"redundant" edges is correct, then our procedure will not find it. Typically these structures are
underidentified, and so they could not be found by either LISREL VI or EQS.
The search procedure we have described here is practical for no more than several dozen
variables. However, for larger numbers of variables, the MIMBuild algorithm described in
Chapter 10 may be applicable.
Finally, there exist many latent variable models that cannot be distinguished by the vanishing
tetrad differences they imply, but are nonetheless in principle statistically distinguishable. More
reliable versions of the LISREL or EQS procedures might succeed in discovering such
structures when the TETRAD procedures fail.
11.9 Some Morals for Statistical Search
There were three reasons why the TETRAD II procedure proved more reliable over the
problems considered here than either of the other search procedures.
(i) TETRAD II, unlike LISREL VI or EQS, does not need to estimate any parameters in
order to conduct its search. Because the parameter estimation must be performed on an
initial model that is wrong, LISREL VI and EQS often failed to converge, or calculated
highly inaccurate parameter estimates. This in turn, led to problems in their respective
searches.
(ii) In the TETRAD II search, when the scores of several different models are tied, the
program considers elaborations of each model. In contrast, LISREL VI and EQS
arbitrarily chose a single model to elaborate.
(iii) Both LISREL VI and EQS are less reliable than TETRAD II in deciding when to stop
adding edges.
The morals for statistical search are evident: avoid iterative numerical procedures wherever
possible; structure search so that it is feasible to branch when alternative steps seem equally
good; find structural properties that permit reliable pruning of the search tree; for computational
efficiency use local properties whenever possible; don't rely on statistical tests as stopping
criteria without good evidence that they are reliable in that role.
Statistical searches cannot be adequately evaluated without clarity about the goals of search. We
think in the social, medical and psychological uses of statistics the goals are often to find and
estimate causal influence. The final moral for search is simple: once the goals are clearly and
candidly given, if theoretical justifications of reliability are unavailable for the short run or even
the long run, the computer offers the opportunity to subject the procedures to experimental tests
of reliability under controlled conditions.
Chapter 12
Open Problems
A number of questions have been raised and not answered in the course of this book. Foremost
among these are issues concerning extensions of the reliabilities and informativeness of the
search algorithms. We record here a number of other questions that seem important. Some of
the problems and questions may be quite easy, or may follow from results already available but
unknown to us. Others have been worked at for some time by ourselves or others and appear to
be quite difficult. Some are not particularly difficult but require work we have not done. All of
the issues seem to us important to help fill out our understanding of the relations between causal
structure and probability, and of the possibilities of causal inference and prediction.
12.1 Feedback, Reciprocal Causation, and Cyclic Graphs
"Reciprocal causation" may arise through treating A and B as variables in a population in which
some A events cause B events and some B events cause A events. Consider, for example, a
population of samples of ideal gasses, some of which have obtained their state by manipulation
of temperature holding volume constant, and some by manipulation of volume holding
temperature constant. In such cases the population under study is a mixture in the sense used
throughout this book. If all causal connections are "reciprocal" in this way, then applying the
PC Algorithm to such a population will generally yield a complete graph, and the FCI procedure
will generally yield a complete graph with an "0" at each end of each edge. An interesting
question is whether graphs with cycles may usefully represent causal systems in which some,
but not all variables reciprocally cause each other.
In many applications statistical models are produced with a set of simultaneous equations,
usually linear, in which some variable X, is specified as a function of Y and other variables,
and reciprocally, Y is specified as a function of X and other variables. Such models are
Open Problems 355
standardly called "non-recursive." They are a clear indication that the relevant mathematical
structure is not given by the algebra and probabilities alone, but involves directed graphical
relationships, and in these cases graphs with cycles. The cyclic graphs are sometimes given
explicitly. In contemporary cognitive science certain models of how humans compute have a
related structure. In these theories computation is carried out through cyclic directed graphs
whose nodes are random variables, in some cases variables taking only discrete values.
These mathematical structures are intended somehow to represent both the causal relationships
among systems with feedback and the "equilibrium" probability relations among the variables of
such systems. Just what that means is unclear, and part of the aim of this section is to try to
clarify the matter a little. The analyses of representation, indistinguishability and inference
developed earlier for acyclic representations of causal processes and probability distributions
need to be extended to the cyclic case.
12.1.1 Mason's Theorem
For linear simultaneous equation systems in which each equation is accompanied by an "error
term," assumed to be independent of other error terms, the algebra entails that each variable X is
a function of various error terms only. The error terms and combinations of linear coefficients
that occur in such an equation for a variable X depend on the cycles in which X occurs, the
cycles in which variables in cycles with X occur, the cycles in which the variables in these
neighboring cycles occur and so on. Nearly forty years ago, Samuel Mason gave a general
analysis, entirely in terms of graphical properties, of how each variable in a cyclic graph
depends on error variables and linear coefficients. For any linear system associated with a cyclic
graph, taking appropriate expectations using Mason's formulas for the variables then gives the
correlations entailed by the simultaneous equation system. So for the linear case there is a
developed theory that can give us some information about when a cyclic graph linearly implies
vanishing correlations or partial correlations. For a useful review and references see Heise
(1975).
Mason's results enable us to use the linear case to examine how cyclic graphs may be
interpreted as representations of the "equilibrium" or limiting results of time series, and whether
the Markov Condition, factorization, Faithfulness, and d-separability are sensible or informative
conditions for cyclic graphs.
12.1.2 Time Series and Cyclic Graphs
Since cyclic graphical models are meant to represent equilibria that result from feedback
processes, each such model must correspond to the limiting distribution of some class of time
series models. What is the correspondence? Consider a simple case:
Figure 1
We have the equations
Y=aX+Ey
X = bY +Ex
which by Mason's rules yield
a~x +b~y
_
YXY - (l-abl
Consider the time series
X t = bYt-k + EXt _ m
Yt = aXt_j+ EYt_n
where for all t, EXt _ m are i.i.d. and for all t, EYt _ n are Li.d. For simplicity take k = m = j = n
= 1. A finite segment of the corresponding directed graph is
Open Problems 357
Figure 2
As t increases without bound this graph gives Mason's formula for the correlation of X and Y
obtained from the simultaneous equations. If the ex and Eyvariables are constant, it also gives
Mason's formula for the distribution of X and of Y. The same limit will of course result if at
some finite time the values (or distributions) of the ex and Eyvariables change but remain the
same thereafter.
We can also have a "shock" model in which ex and ey represent changes to a system in
equilibrium. According to this model, for each individual Xo and YO have some arbitrary initial
value, and ex and Ey represent changes to the system introduced (independent of XO, Yo, and
each other) at time tl.
Xl =XO+ex
YI = YO + ey
X t - Xt-I = b(Yt-I - Yt-2)
Yt - Yt-I = a(Xt-I - X t-2)
The latter two equations we can write in a cyclical form:
dY=aM
!lX =MY
Mason's formula entails that the new equilibrium values of X and Y, Xe and Y e respectively
are:
X -X ex + bey
e- 0+ I -ab
Y _Y ey+aex
e- 0+ I -ab
If ex. and ey have zero covariance it follows that the change diTI' in the covariance of X and
Y, resulting from the "shocks" has the following form:
aCT~X +b~y
drxy - --""'----,.......
- (1-ab)2
A finite segment of the corresponding directed graph is depicted in figure 3.
Figure 3
Other representations are also possible.

Open Problems 359
12.1.3 The Markov Condition, Factorizability and Faithfulness
The Markov Condition makes sense for a cyclic graph, but it may not be informative. Consider
the graph:
Y
/\ W
\z/ Figure 4
The Markov condition entails no independence or conditional independence relations among the
four variables, because each variable is a descendent of each of the others. But the associated
undirected independence graph represents two conditional independence claims: X 11 z
I{Y,W} and Y 11 W I{X,z}. Moreover, if we introduce an errdr variable adjacent to each of X,
Y, Z and W, treat the errors as independent, and write each variable as a linear function of its
error term and its parent, the partial correlation of X, Z controlling for Y, W vanishes, and so
does the partial correlation of Y, W controlling for X, Z.
In the acyclic case the Markov condition is essentially equivalent to the requirement that d-
separation implies conditional independence. The example just considered shows the
equivalence does not obtain in cyclic graphs, where the d-separability condition is more
informative. The vanishing partial correlations obtained from figure 4 are exactly those that
result if we assume that d-separation implies conditional independence, or, in the linear case,
vanishing partial correlation. Further
Theorem 12.1: If A and Bare d-separated by C in a directed graph (cyclic or acyclic)

G, then G linearly implies that the partial correlation of A, B controlling for C vanishes.
These results suggest the following conjecture.

Conjecture 1: if A and Bare d-separated by set C in a directed graph (cyclic or acyclic)

G, then G linearly implies that the partial correlation of A, B controlling for C vanishes.
The conjecture gains support from the reflection that the translations of cyclic graphs into time
series models in the previous section preserves a form of d-separation. If in the finite cyclic
graph A and Bare d-separated by C, then in the infinite non-cyclic time series graph every At
and Bt+k are d-separated by some occurrences of variables in C, and one would expect the
corresponding partial correlation to vanish in the limit. If the conjecture is true, it would provide
a strong case for adopting the convention for graphs of discrete variables that d-separation
entails conditional independence.
Suppose that we adopt the convention that a cyclic directed graph G represents a distribution P
if and only if whenever X and Y are d-separated given Z in G, X and Y are independent given
Z in P. Unlike the acyclic case, it does not follow from this that for a positive distribution that a
distribution represented by G can be written as the product of the conditional distribution of
each variable given its parents in G. For example, suppose in G, X is a parent of Y, and Y is a
parent of X. G represents any distribution over X and Y. However if the density! satisfies
!(XY) = ftXIY)!(YIX)
then it follows, simply by writing out the definitions of conditional density on the right hand
side, that X and Yare independent. (Note that in linear models, if we adopt the convention that
a cycle containing X and Y means X is a linear function of Y, and Y is a linear function of X,
this does not entail that X and Yare uncorrelated.)
12.1.4 Discovery Procedures
Open Problem 1: Find a computationally feasible algorithm for inferring graphical

structure when the measured variable set may be causally insufficient and there may be
feedback.
Open Problems 361
12.2 Indistinguisbability Relations
Assume for the moment the convention that d-separation implies conditional independence. For
acyclic graphs the Faithfulness Condition is equivalent to the converse: if A and B are
conditionally independent on C (or have vanishing partial correlation in the linear case) then C
d-separates A, B. Likewise, we understand Faithfulness as a principle connecting conditional
independence facts in a distribution with d-separation facts in a graph.
These assumptions and conventions make the Faithfulness Condition intelligible for cyclic as
well as for acyclic graphs, and the definition of faithful indistinguishability (f.i.) from Chapter 4
therefore extends to cyclic graphs. Let us say that a cyclic path in a directed graph is chordless
if and only if for each variable V in the cyclic path there is at most one directed edge out of V
and into another variable in the cycle.
Conjecture 2: No directed cyclic graph with a chordless directed cycle of length 4 or more
is faithfully indistinguishable from any directed acyclic graph on the same vertex set
Given a system of linear equations with a cyclic graph, Masc;m's theorem permits one to write
an "equivalent" system of equations in which each variable is a function of exogenous
variables, and the corresponding graph is acyclic. But the exogenous variables will include the
error terms, and so the acyclic model will have unmeasured common causes. In Chapter 4 we
defmed indistinguishability relations only for causally sufficient sets of variables. The first
problem is therefore to say precisely what "equivalent" means.
Definition: Let G and H be directed graphs, and let 0 be a subset of the vertex sets of
both graphs. G and H are faithfully indistinguishable over 0 if and only if for every
distribution faithful to G there is a distribution faithful to H with the same 0 marginal, and
vice versa.
A parallel defmition of linearly faithful indistinguishable over 0 is obvious. Now the

acyclic model obtained by applying Mason's theorem to a cyclic graph will not in general be
linearly faithfully indistinguishable from the cyclic model. Thus in figure 4, if we give each of
the variables a disturbance term, applying Mason's rules we obtain a model in which each
measured variable is a function of each unmeasured disturbance, with a corresponding graph:
Figure 5
Notice that the graph in figure 5 does not entail the conditional independencies required by the
graph in figure 4, and the two graphs are therefore not faithfully indistinguishable over 0 =
{X,Y,Z,W}. The graph of figure 5 is indistinguishable over 0 from the graph of figure 4 only
in the weak sense that for any choice of linear coefficients and distributions on the error
variables for figure 4, there exists a choice of linear coefficients and distributions on the e
variables of figure 5 producing the same distribution on O. But while the conditional
independence relations among X, Y, Z and W will be stable for the model of figure 4 under a
small variation of linear coefficients, they will not be stable for the model of figure 5.
There are cases in which acyclic graphs are faithfully indistinguishable from cyclic graphs. For
example graphs 1 and 2 in figure 6 are faithfully indistinguishable.
(1) W 4-X •
(2) W 4-X
Figure 6
On the other hand, the graph in figure 7 is not faithfully indistinguishable from any acyclic
graph on the same vertex set. It is, however, faithfully indistinguishable from an acyclic graph
with unmeasured common causes.
Open Problems 363
Figure 7
Conjecture 3: Every directed graph without directed cycles of length greater than 2 is
faithfully indistinguishable from an acyclic directed graph, possibly with extra vertices.
Open Problem 2: Give necessary and sufficient conditions for two directed acyclic
graphs sharing a subset 0 of vertices to be faithfully indistinguishable over O.
Open Problem 3: Give necessary and sufficient conditions for two directed graphs,
cyclic or acyclic, sharing a subset 0 of vertices to be faithfully indistinguishable over O.
Open Problem 4: Give necessary and sufficient conditions for two directed acyclic
graphs sharing a subset 0 of vertices to be linearly faithfully indistinguishable over O.
Open Problem 5: Give necessary and sufficient conditions for two directed graphs,
cyclic or acyclic, sharing a subset 0 of vertices to be linearly faithfully indistinguishable
overO.
12.3 Time series and Granger Causality
Econometricians often deal with data in which the sample consists of the same variables
measured at a discrete sequence of times. A variety of techniques have been developed to reduce
inferences from such data to linear regression problems. Each series may be differenced to
produce an approximation of stationarity, so that the values obtained at different times can be
regarded as samples from the same population. (A stochastic process is said to be strictly
stationary if the joint distribution of Zt is the same as the joint distribution of Zt-k for all time
points t and all time lags k.) Variables may be transformed or data "fIltered" in the attempt to
approximate linearity and reduce autocorrelation.
The properties relevant to unit t in such a sample are not only the values Xt. Yt. Zt of variables X,
Y and Z at t, but also the values these variables had at times t -k, for each k, that is, Xt-k. Yt-k
and Zt-k. An outcome variable of interest, say Y, is then regressed on Yt-k. Xt-k and Zt-k.
although in principle the choices of lags may be different for each series. The causal picture in
one such regression model is illustrated in figure 8, where error terms are omitted:
Figure 8
The regression equation in this case specifies Yt as a function of the lags Yt-l. Y t-2. X t-I,X t-2.
Zt-l and an error term. The influence of X on Y, for example, is estimated by testing for
vanishing partial correlations of Y t with the X lags, controlling for the Y lags and the Z lags, or
by testing simultaneously for vanishing linear coefficients associated with the X lags (Geweke,
Meese and Dent, 1983). When there are multiple series, stepwise regression has been
recommended and used for selecting the influences among series variables (Hsiao, 1981).
Granger's (1969) account of the notion of causal influence between time series is sometimes
used in econometrics to justify the view that regression procedures result in reasonable
estimates of causal influence. Granger assumes the series are stationary, denotes by Pt<AIB)
"the optimum, unbiased, least-squares predictor of A using the set of values Bt" and he denotes
the prediction error by etfAIB) = At - Pt(AIB), with variance (J'2(AIB). "Let V t be all the
information in the universe accumulated since time t-l and let Vt-Y denote all this information
apart from the specified series Yt."(p. 428) His definition is then that Y causes X if (J'2(XIU) <
a 2(XI!l:D, where Il;X, is the V information prior to t absent the information in the Y series
prior to t. Granger also proposes that instead of the imaginary set V, the set of all "relevant"
information could be used.
The use of regression procedures to estimate causal influence in econometric time series usually
realizes Granger's proposal by taking linear least squares estimators to be optimal and by taking
Open Problems 365
the "relevant" variables to be known a priori or to be determined by stepwise regression. One

could quibble with Granger's definition, but the fundamental problem about common methods
of causal inference from time series data appears to rest in these further assumptions. The
problems with regression as a means for determining causal influence seem essentially the same
in time series studies as elsewhere. First, tests of regression parameters waste degrees of
freedom at the cost in small samples of power against alternatives. Since in many cases the
number of observations is of the order of the number of parameters, whatever can be done to
increase reliability should be. Second, it appears that while the time series setting removes
ambiguities about the direction of dependencies, or edges, it does not remedy problems about
unmeasured common causes of the outcome and regressors, and thus even asymptotically
regression may yield significant coefficients for variables that are neither direct nor indirect
causes of the outcome. These remarks are, however, entirely informal, and hence the next open
question:
Open Problem 6: Do the methods of this book extend to causal relations among
stationary time series?
12.4 Model Specification and Parameter Estimation from the Same

Data Base
One of the routine objections to specification searches is that the error probabilities for
parameter estimates in such models are no longer given by the conventional formulas. By
simulation, Freedman, Navidi and Peters (1986) examine the estimates of mean squared error,
mean square prediction error, and R2 for bootstrap, jackknife and cross validation estimators in
models specified as a function of sample data. On the whole, they find cross-validation works
best but not especially well. Their case is extreme: they use 75 variables and 100 data points in
each trial. A subset of the variables is selected by choosing the siguificant regressors in a
multiple regression at the .25 level of significance.
It would be difficult to form any warranted conclusion about methodology for most applications
from this single, nearly worst-case example, but the simulation does raise an interesting
question. Freedman, et aI, vary the success criteria and the estimation methods in their
simulations, but they do not vary the method of model specification. (In fact they use a method
that is not correct for causal inference unless prior assumptions about the causal structure are
made which are often not justified in applications.) The error probabilities of tests of parameter
hypotheses are functions of the information in the data, represented by the number of data
points and the number of free parameters in the model; intuitively, the model specification uses
up some of that information. But not all specification searches must use the data in the same
way or to the same extent. For example, if the true graph is sparse many of the methods we
have described will use only low order tests of conditional independence, or tests of vanishing
partial correlations and vanishing tetrad differences.
Open Problem 7: How do error probabilities of estimates vary for parameters in models
specified by the PC or other algorithms, as a function of sample size and sparseness of the
true graph?
A similar question applies to the search procedures for multiple indicator models using tests of
vanishing tetrad differences.
12.5 Conditional Independence Tests
The examples in this book have used either approximately continuous variables assumed to be
linearly related, or discrete variables, but never both, and never non-linear functional forms.
The reasons for these restrictions are entirely statistical: the availability of relevant tests of
conditional independence. Many empirical studies measure both discrete and continuous
variables, and the continuous variables may sometimes be regarded as causes, and sometimes
as effects, of the discrete variables. For the special case in which continuous variables never
cause discrete variables, directed graphical treatments have been given in terms of joint
distributions that are "conditionally Gaussian," that is the continuous variables are normally
distributed conditional on each vector of values of their discrete parents. But in many instances
where logistic regression is applied, and in many psychometric models, the cause is continuous
and the effect discrete. Expanding the portfolio of tests of conditional independence to permit
reliable decisions about conditional independence would likewise expand the range of
applications for which causal inferences can reliably be made.
Chapter 13
Proofs of Theorems
We will adopt the following notational conventions. "w.l.g." abbreviates "without loss of
generality", "r.h.s." abbreviates "right hand side", and "l.h.s." abbreviates "left hand side".
Any sum over the empty set is equal to 0 and any product over the empty set is 1. R(/,J)
represents a directed path from / to J. If U is an undirected path from A to B, and X and Y
occur on U, then we will denote the subpath of U between X to Yas U(X,Y). T(/,J) represents
a trek in T(/,J). The defmitions of all technical terms in this chapter that have not been defined
in Chapters 2 or 3 have been placed in a glossary following the chapter.
13.1 Theorem 2.1
Theorem 2.1: If P(V) is a positive distribution, then for any ordering of the variables in V, P
satisfies the Markov and Minimality conditions for the directed independence graph of P(V) for
that ordering.
Proof. See Pearl (1988).
13.2 Theorem 3.1
Theorem 3.1: If S is an LCT, and S' is a random coefficient LCT with the same directed
acyclic graph, the same set of non-coefficient random variables, the same variances for each
non-coefficient exogenous variable, and for each random coefficient a'1J in S', E(a'IJ) =alJ in
S, then a partial correlation is equal to zero in S if and only if it is equal to 0 in S.
Let a linear causal theory be (LCT) be <<R,M,E>, (aloP), EQ,L,Err> where

i. (aloP) is a probability space, where a is the sample space,fis a sigma-field over a,
and P is a probability distribution overf.
ii. <R,M,E> is a directed acyclic graph. R is a set of random variables over (aloP).
iii. The variables in R have a joint distribution. Every variable in R has a non-zero variance.
E is a set of directed edges between variables in R. (M is the set of marks that occur in a
directed graph, i.e. {EM, >}.
iv. EQ is a consistent set of independent homogeneous linear equations in random
variables in R. For each Xi in R of positive indegree there is an equation in EQ of the form
Xi = I,aij Xj
XjEParents(Xj)
where each aij is a non-zero real number and each Xi is in R. This implies that each vertex
Xi in R of positive indegree can be expressed as a linear function of all and only its parents.
There are no other equations in EQ. A non-zero value of aij is the equation coefficient
of Xj in the equation for Xi.
v. If vertices (random variables) Xi and Xj are exogenous, then Xi and Xj are pairwise
statistically independent
vi. L is a function with domain E such that for each e in E, L(e) =aij iff head(e) =Xi and
tail(e) = Xj. L(e) will be called the label of e. By extension, the product of labels of edges
in any acyclic undirected path U will be denoted by L(U), and L(U) will be called the label
of U. The label of an empty path is fixed at 1.
vii. There is a subset of S of R called the error variables, each of indegree 0 and
outdegree 1. For every Xi in R of indegree '# 0 there is exactly one error variable with an
edge into Xi. We assume that the partial correlations of all orders involving only non-error
variables are defined.
Note that the variance of any endogenous variable I conditional on any set of variables that does
not contain the error variable of I is not equal to zero.
The definition of a random coefficient linear causal theory is the same as that of a linear
causal theory except that each linear coefficient is a random variable independent of the set of all
other random variables in the model.
Proofs 369
A linear causal form is an unestimated LCT in which the linear coefficients and the
variances of the exogenous variables are real variables instead of constants. This entails that an
edge label in an LCF is a real variable instead of a constant (except that the label of an edge
from an error variable is fixed at one.) More formally, let a linear causal form (LCF) be
<<R,M,E>, C, V, EQ,L,Err> where
i. <R,M,E> is a directed acyclic graph. Err is a subset of R called the error variables.
Each error variable is of indegree 0 and outdegree 1. For every Xi in R of indegree ,;: 0
there is exactly one error variable with an edge into Xi.
ii. Cij is a unique real variable associated with an edge from Xj to Xj , and C is the set of
Cji V is the set of variables ci{, where Xi is an exogenous variable in <R,M,E> and ci{ is
a variable that ranges over the positive real numbers.

iii. L is a function with domain E such that for each e in E, L(e) = cij iff head(e) = Xi and
tail(e) =Xj. L(e) will be called the label of e. By extension, the product of labels of edges
iv. EQ is a consistent set of independent homogeneous linear equationals in variables in R.
For each Xi in R of positive indegree there is an equation in EQ of the form
Xi = !,cij Xj
XjEParents(Xj) .
where each Cjj is a real variable in C and each Xi is in R. There are no other equations in
EQ. Cji is the equation coefficient of Xj in the equation for Xi.
An LCT S is an instance of an LCF F if and only if the directed acyclic graph of S is

isomorphic to the directed acyclic graph of F. In an LCF, a quantity (e.g. a covariance) X is
equivalent to a polynomial in the coefficients and variances of exogenous
variables if and only if for each LCF F = <<R,M,E>, C, V, EQ,L,Err> and in every LCT
S =«R',M',E'>, (fJf,P), EQ',L',Err'> that is an instance of F, there is a polynomial in
the variables in C and V such that X is equal to the result of substituting the linear coefficients
of S in as values for the corresponding variables in C, and the variances of the exogenous
variables in S as values for the corresponding variables in V.
In an LCT or LCF S , a variable Xi is independent iff Xj has zero indegree (i.e. there are no
edges directed into it); otherwise it is dependent. Note that the property of independence is
completely distinct from the relation of statistical independence. The context will make clear in
which of these senses the term is used. For a directed acyclic graph G,Ind is the set of
independent variables in G. Given a directed acyclic graph G, D(Xi,xj) is the set of all directed
paths from Xj to Xj. In an LCF «R,M,E>, C, V, EQ,L,S> an equation is an independent
equational for a dependent variable Xj if and only if it is implied by EQ and the
variables in R which appear on the r.h.s. are independent and occur at most once. IndaIJ is the
coefficient of] in the independent equational for I.
Lemma 3.1.1: In an LCF S, if] is an independent variable, then
lndau = LL(U)
UeD(J,l)
Proof. This is a special case of Mason's rule for calculating the "total effect" of a variable] on
a variable I. See Glymour et. al. (1987). :.
The following two lemmas show how to calculate the variance of random variables and
covariances between random variables in terms of the covariances between other random
variables. The proofs of these lemmas can be found in Freund and Walpole (1980). We denote
the covariance of I and] by "/If, the variance of I by aT, the correlation of I and] by Pu, the
partial correlation of I and] given the set H by PU.H, and the partial covariance of I and] given
H by "/If.H. The correlation of two subscripted variables such as Xi and Xj we will write as Pij
for legibility, and similarly for partial correlations, etc.
Lemma 3.1.2: If Q is a set of random variables with a joint probability distribution and
and
Z= LaZl]
feQ
then
ryZ = L LaYlaZlrU
leQfeQ
Lemmas 3.1.3, 3.1.5, and 3.1.7 are not used in the proof of Theorem 3.1, but they are used in
later theorems, and we include them here because they follow easily from the other lemmas in
this section.
Proofs 371
Lemma 3.1.3: If Q is a set of random variables with a joint probability distribution and
then
In an LCF S, Ux is the set of all independent variables that are the source of a directed path to
X. (Note that if X is independent then X E Ux since there is an empty path from every vertex
to itself.) In an LCF S, UXy is Ux n Uy.
Lemma 3.1.4: If S is an LCF,
and
then
ryz = 2. Inday/ndaZJaT
leUyz
Proof. Ind is a set of independent variables. It follows that rlJ is equal to 0 if I "* i, and rlJ is
equal to a: if 1= i. Substituting these values for 'YIl into the r.h.s. of the equation for iYz in
lemma 3.1.2 shows that
(1) ryZ = 2. Inday/ndaZJaT

leInd
If I is in Ind, but I is not in U yz then there is no pair of directed acyclic paths from I to Y and
z. By lemma 3.1.1, if there is no pair of directed acyclic paths from I to Y and Z, then the
coefficient of I in the independent equation for either Y or Z is zero. So, the only non-zero
terms in equation 1 are for lEU yz. :.
~/nd
Y =,LJ aylI
leInd
then
_2 ~ Ind 2"
CIy =,LJ aYl C1J
leUy
Proof. Ind is a set of independent random variables. It follows that 'YIJ is equal to 0 if I*- J,
and 'YtJ is equal to a:
if 1= J. Substituting these values for 'YtJ into the r.h.s. of the equation for
a;in lemma 3.1.3 proves that
_2 _ ~ Ind 2?·
(2) CIy -,LJ aYI C1J
leInd
If I is in Ind, but I is not in Uy, then there is no directed path from I to Y. It follows from
lemma 3.1.1 that aYI is zero. Hence the only non-zero terms in equation 2 come from lEU y.
'YIJ = L L LL(R)L(R')ak
KeU/J ReD(K,I)R'eD(K,J)
Proof. This follows immediately from lemmas 3.1.2 and 3.1.4. :.
a7 = L
KeUI
[( LL(R) 2 ak
ReD(K,L)
J 1
Theorem 3.1: If S is an LCf, and S' is a random coefficient LCf with the same directed
acyclic graph, the same set of non-coefficient random variables, the same variances for each
exogenous variable, and for each random coefficient a'lJ in S', E(a'lJ) = au in S, then a partial
correlation is equal to zero in S if and only if it is equal to 0 in S'.
Proofs 373
Proof. Because S is an instance of an LCF, by lemma 3.1.6
rIJ = I, I, I,L(R)L(R')oi
KeU/J ReD(K,I)R'eD(K,J)
The label of a path is equal to the product of the labels of the edges and because the random
coefficients are independent of each other and all the random variables that are not coefficients,
it follows that
E[ n
edgeeU
L(edge») = n
edgeeU
E(L(edge»
Transform all of the variables so that they have mean 0; this does not affect the value of any of
the covariances. In T, rIJ = E(/J) and
E(lJ) = E[ I, I, I, I,L(U)L(V)HF] =
HeU I UeD(H,X) FeUJ VeD(F,Y)
I, I, I,E(L(U)L(V)H 2 ) =
HeU IJ UeD(H,X) VeD(H,Y)
I, I,
HeU /J UeD(H,X)VeD(H,Y)
I, E( n L(edge)
edgeeU
n L(edge)H2» =
edgeeV
I, I, I,
HeU/J UeD(H,X)VeD(H,Y)edgeeU
n E(L(edge» n E(L(edge»E(H 2 )
edgeeV
because for exogenous variables E(HF) = 0 unless Ii =F.
By hypothesis, E(L(edge» in S' = L(edge) in S. Hence the expression rIJ is the same for both
random and constant coefficients. The partial correlations are a function of the covariance
matrix so the partial correlations are the same in S and S'. :.
13.3 Theorem 3.2
Theorem 3.2: Let M be an LCF with n free linear coefficients at, ... , an and k positive
variances Vt, •.• , Vk. Let M«ut, ••. ,un. un+t"",Un+k» be the distributions consistent with
specifying values <Ut •... 'un. un+t, ... ,Un+k> for at, ... , an and Vt, ... Vk. Let n be the set of
probability measures P on the space 9{ n+k of values of the parameters of M such that for every
subset V of 9{n+k having Lebesgue measure zero, P(V) =o. Let Q be the set of vectors of
coefficient and variance values such that for all q in Q every probability distribution consistent
with M(q) has a vanishing partial correlation that is not linearly implied by M. Then for all P in
II, P(Q) = O.
Lemma 3.2.1: In an LCF S, Pij.X = 0 is equivalent to a polynomial equation in the linear

coefficients and variances of the independent variables.
Proof. We will prove more generally that a polynomial equation in partial covariances is
equivalent to a polynomial equation in the linear coefficients and variances of the independent
variables. If X contains n distinct variables, then say Pij.X is a partial correlation of order n. Let
the pc-order (partial covariance order) of a polynomial in partial covariances be the highest
order of any partial covariance appearing in the polynomial. The proof is by induction on the
pc-order of the polynomials.
Base Case. If polynomial Q is of pc-order 0, then by lemma 3.1.2, Q is equivalent to a

polynomial equation in the linear coefficients and variances of the independent variables.
Induction Case. Suppose that the lemma is true for polynomials of pc-order n-l, and let Q be a
polynomial of pc-order n. The recursion formula for partial covariances is
Yir.yrjr.Y
rij.Yur = Yij.Y
rrr.Y
Form Q' by using this recursion formula to replace each covariance of pc-order n appearing in
Q by an algebraic combination of covariances of pc-order n-l. Form Q" by multiplying Q' by
the lowest common denominator of all of the terms in Q', producing a polynomial of pc-order
n-1. By the induction hypothesis, Q" is equivalent to a polynomiaI equation in the linear
coefficients and variances of the independent variables. Hence, a polynomial equation in partial
covarlances is equivalent to a polynomial equation in the linear coefficients and variances of the
independent variables.
Proofs 375
By defmition,
Pij.X =- ; = =rij.X
:f-;==
.../ru.x ~rjj.x
so Pij.X = 0 iff 'Yij.x = O. Since the latter is a polynomial equation in partial covariances, it is
equivalent to a polynomial equation in the linear coefficients and variances of the independent
variables. It follows that the former is also equivalent to a polynomial equation in the linear
coefficients and variances of the independent variables. :.
Theorem 3.2: Let M be an LCF with n free linear coefficients aI, ... , an and k positive
variances VI, ... , vk. Let M«UI, ... ,Un. Un+I, ... ,Un+k» be the distributions consistent with
specifying values <UI, ... ,u n • un+I, ... 'Un+k> for aI, ... , an and VI, ... Vk. Let II be the set of
probability measures P on the space 9t n+k of values of the parameters of M such that for every
subset V of 9t n+k having Lebesgue measure zero, P(V) = O. Let Q be the set of vectors of
coefficient and variance values such that for all q in Q every probability distribution consistent
with M(q) has a vanishing partial correlation that is not linearly implied by M. Then for all P in
II, P(Q) = O.
Proof. For any LCF, each partial correlation is equivalent to a polynomial in the linear
coefficients and the variances of the exogenous variables: the rest of the features of the
distribution have no bearing on the partial correlation. Hence for a vanishing partial correlation
to be linearly implied by the directed acyclic graph of the theory, it is necessary and sufficient
that the corresponding polynomial in the linear coefficient and variance parameters vanish
identically. Thus any vanishing partial correlation not linearly implied by an LCF represents a
polynomial P in variables consisting of the linear coefficients and variances of that theory, and
the polynomial does not vanish identically. So the set of linear coefficient and variance values
satisfying P is an algebraic variety in 9t n +k. Any connected component of such a variety has
Lebesgue measure zero. But an algebraic variety has at most a finite number of connected
components (Whitney, 1957). :.
13.4 Theorem 3.3
Theorem 3.3 P(V) is faithful to directed acyclic graph G with vertex set V if and only if for
all disjoint sets of vertices X, Y, and Z, X and Y are independent conditional on Z if and only
if X and Y are d-separated given Z.
The "if" portion of the theorem was first proved in Verma (1986) and the "only if" portion of
the theorem was first proved in Geiger and Pearl (1 989a). The proof produced here is
considerably different, but since the bulk of it is a series of lemmas that we also need to prove
other theorems, we state it here.
G' is an inducing path graph over 0 for directed acyclic graph G if and only if 0 is
a subset of the vertices in G, there is an edge between variables A and B with an arrowhead at
A if and only if A and B are in 0, and there is an inducing path in G between A and B relative
to 0 that is into A. (Using the notation of Chapter 2, the set of marks in an inducing path graph
is I>, EM}.) We will refer to the variables in 0 as observed variables. Unlike a directed
acyclic graph, an inducing path graph can contain double-headed arrows. However, it does not
contain any edges with no arrowheads. If there is an inducing path between A and B in G that
is into A, then the edge between A and BinG' is into A. However, if there is an inducing path
between A and B in G that is out of A, it does not follow that the edge in G' between A and B
is out of A. Only if no inducing path between A and BinG is into A is the edge between A and
B in G' out of A. The definitions of directed path, d-separability, inducing path, collider,
ancestor, and descendant are the same as those for directed graphs, i.e. a directed path in an
inducing path graph, as in an acyclic directed graph, contains only directed edges (e.g. A -> B).
However, an undirected path in an inducing path graph can contain either directed edges, or bi-
directed edges (e.g. C <-> D.) Also, if A <-> B in an inducing path graph, A is not a parent of
B. Note that if G is a directed acyclic graph, and G' the inducing path graph for Gover 0, then
there are no directed cycles in G'.
Lemma 3.3.1 states a method for constructing a path between X and Y that d-connects X and Y
given Z out of a sequence of paths.
Lemma 3.3.1: In a directed acyclic graph G (or an inducing path graph G) over V, if X and
Y are not in Z, there is a sequence S of distinct vertices in V from X to Y, and there is a set T
of undirected paths such that
Proofs 377
(i). for each pair of adjacent vertices V and Win S there is a unique undirected path in T that
d-connects V and W given z\{V,W}, and
(ii). if a vertex Q in S is in Z, then the paths in T that contain Q as an endpoint collide at Q,
and
(iii). if for three vertices V, W, Q occurring in that order in S the d-connecting paths in T
between V and W, and Wand Q collide at W then W has a descendant in Z,
then there is a path U in G that d-connects X and Y given Z. In addition, if all of the edges in
all of the paths in T that contain X are into (out of) X then U is into (out of) X, and similarly for
Y.
Proof. Let U' be the concatenation of all of the paths in T in the order of the sequence S. U'
may not be an acyclic undirected path, because it may contain some vertices more than once.
Let U be the result of removing all of the cycles from U. If each edge in U that contains X is
into (out of) X, then U is into (out of) X, because each edge in U is an edge in U'. Similarly, if
each edge in U' that contains Y is into (out of) Y, then U is into (out of) Y, because each edge
in U is an edge in U'. We will prove that U d-connects X and Y given Z.
We will call an edge in U containing a given vertex Van endpoint edge if V is in the sequence
S, and the edge containing V occurs on the path in T between V and its predecessor or
successor in S; otherwise the edge is an internal edge.
First we prove that every member R of Z that is on U is a collider on U. If there is an endpoint

edge containing R on U then it is into R because by assumption the paths in T containing R
collide at R. If an edge on U is an internal edge with endpoint R then it is into R because it is an
edge on a path that d-connects two variables A and B not equal to R given Z\{ A,B}, and R is in
Z. All of the edges on paths in T are into R, and hence the subset of those edges that occur on
U are intoR.
Next we show that every collider R on U has a descendant in Z. R is not equal to either of the
endpoints X or Y, because the endpoints of a path are not colliders along the path. If R is a
collider on any of the paths in T then R has a descendant in Z because it is an edge on a path
that d-connects two variables A and B not equal to R given Z\{A,B}. If R is a collider on two
endpoint edges then it has a descendant in Z by hypothesis. Suppose then that R is not a
collider on the path in T between A and B, and not a collider on the path in T between C and D,
but after cycles have been removed from U', R is a collider on U. In that case U' contains an
undirected cycle containing R. Because G is acyclic, the undirected cycle contains a collider.
Hence R has a descendant that is a collider on U'. Each collider on U' has a descendant in Z.
Hence R has a descendant in Z. :.
Lemma 3.3.2: If G is a directed acyclic graph (or an inducing path graph), R is d-connected
to Y given Z by undirected path U, and Wand X are distinct vertices on U not in Z, then
U(W,K) d-connects Wand X given Z = Z\{W,x}.
Proof. Suppose G is a directed acyclic graph, R is d-connected to Y given Z by undirected
path U, and Wand X are distinct vertices on U not in Z. Each non-collider on U(W,K) except
for the endpoints is a non-collider on U, and hence not in Z. Every collider on U(W,K) has a
descendant in Z because each collider on U(W,K) is a collider on U, which d-connects R and Y
given Z. It follows that U(W,K) d-connects Wand X given Z = Z\{W,x}. :.
Lemma 3.3.3: If G is a directed acyclic graph (or an inducing path graph), R is d-connected
to Y given Z by undirected path U, there is a directed path D from R to X that does not contain
any member of Z, and X is not on U, then X is d-connected to Y given Z by a path U' that is
into X. If D does not contain Y, then U' is into Y if and only if U is.
Proof. Let D be a directed path from R to X that does not contain any member of Z, and U an
undirected path that d-connects R and Y given Z and does not contain X. Let Q be the point of
intersection of D and U that is closest to Yon U. Q is not in Z because it is on D.
If D does contain Y, then Y = Q, and D(Y,K) is a path into X that d-connects X and Y given Z
because it contains no colliders and no members of Z.
If D does not contain Y then Q * Y. X * Q because X is not on U and Q is. By lemma 3.3.2
U(Q,Y) d-connects Q and Y given Z\{Q,Y} = Z. Also, D(Q,x) d-connects Q and X given
Z\{Q,x} = Z. D(Q,x) is out of Q, and Q is not in Z. By lemma 3.3.1, there is a path U' that
d-connects X and Y given Z that is into X. If Y is not on D, then all of the edges containing Y
in U' are in U(Q,Y), and hence by lemma 3.3.1 U' is into Y if and only if U is. :.
In a directed acyclic graph G, ND(Y) is the set of all vertices that do not have a descendant in
Y
Lemma 3.3.4: If P(V) satisfies the Markov condition for directed acyclic graph G over V, S
is a subset of V, and ND(Y) is included in S, then
f( n
S VeV
p(VIParents(V») = f ( n
S\ND(Y) VeV \ND(Y)
p(VIParents(V»)
Proofs 379
Proof. S can be partitioned into S\ND(Y) and S n ND(Y) = ND(Y). If V is in V\ND(Y)

then no variable occurring in the term P(VIParents(V» occurs in ND(Y); hence for each V in
V\ND(Y), P(VIParents(V» can be removed from the scope of the summation over the values
of variables in ND(Y).
f( II
S
p(VIParents(V») =
VEV
(1)
f (
S\ND(Y) VEV\ND(Y)
II P(VIParents(V» xl f [ II p(vIParents(v»)ll
ND(Y) VEND(Y)
We will now show that
f [
ND(Y) VEND(Y)
II p(VIParents(V») = 1
unless for some value of S\ND(Y) the set of values of ND(Y) such that P(V1Parents(V» is
defIned for each V in ND(Y) is empty, in which case on the l.h.s of (1) no term containing that
value of S\ND(Y) appears in the sum, and on the r.h.s.of (1) every term in the scope of the
summation over S\ND(Y) that contains that value of S\ND(Y) is zero.
Let P(WIParents(W»be a term in the factorization such that W does not occur in any other
term, i.e. W is not the parent of any other variable. If ND(Y) is not empty W is in ND(Y).
i ( II p(VIParents(V»)
NDCY) VeNDCY)
=
i ( II p(VIParents(V»)x (iP(W'Parents(W»)
NDCY)\{W) VeNDCY)\{W) W
The latter expression can now be written as
i (
NDCY)\{W)
II p(VIParents(V»)
VeNDCY)\{W)
because iP(WIParents(W» is equal to one. Now some element in ND(Y)\W is not a parent
w
of any other member of ND(Y)\{ W}, and the process can be repeated until each element is
removed from ND(Y). :.
In a directed acyclic graph G, if Y (") Z = 0, then Vis in IV(Y,Z) (informative variables for Y
given Z) if and only if V is d-connected to Y given Z, and V is not in ND(YZ). (This entails
that V is not in Y u Z by defmition of d-connection.) In a directed acyclic graph G, if Y (") Z =
0, W is in JP(Y,Z) (W has a parent that is an informative variable for Y given Z) if and only if
W is a member of Z, and W has a parent in IV(Y,Z) u Y.
Lemma 3.3.5: If P satisfies the Markov condition for directed acyclic graph G over V, then
-?
L II
P(WIParents(W»
P(YIZ) = IV(Y,Z) WEIV(Y,Z)uIP(Y,Z)uY
-?
L II
P(WIParents(W»
IV(Y,Z)uY WEIV(Y,Z)uIP(Y,Z)uY
for all values of V for which the conditional distributions in the factorization are defined, and
for which P(z) '# O.
Proof. Let V' = V\ND(YZ), i.e. the subset of V with descendants in YZ. It follows from the
defmition of conditional probability that
-?
L II P(WIParents(W»
P(YIZ) = P(YZ) = ...;.V...;.;\Y:..:Z'-W:.;...E=..cV'--_ _ __
P(Z) -?
L II P(WIParents(W»
V\Z WEV
By lemma 3.3.4,
-? -?
L II P(WIParents(W» L II P(WIParents(W»
V\YZ WEV V'\YZ WEV'
-? -?
L II P(WIParents(W» L II P(WIParents(W»
V\Z WEV V'\Z WEV
Proofs 381
First we will show that we can factor the numerator and the denominator into a product of two
sums. The second term in both the numerator and the denominator is the same, so it cancels. In
the case of the denominator, we show that
L IIP(WIParents(W» =
~
V'\Z WeV'
L L
~ ~
IIP(WIParents(W» x II
P(WIParents(W»
IV(Y,Z)uY WeIV(Y,Z)uIP(Y,Z)uY V'\(IV(Y,Z)uYZ) WeV'\(IV(Y,Z)uIP(Y,Z)uY)
by demonstrating that if W is in IV (Y ,Z) u IP(Y,Z) u Y, then neither W nor any parent of W

occurs in the scope of the summation over V'\(IV(Y,Z) u YZ), and also that if W is in
V'\(IV(Y,Z) u IP(Y,Z) u Y) then neither W nor any parent of W is in the scope of the
summation over IV(y,Z) u Y.
First we demonstrate that if W is in IV(Y,Z) u IP(Y,Z) u Y then W is not in V,\(IV(Y,Z) u

YZ). If W is in IV(Y,Z) u Y then trivially it is not in V'\(IV(Y,Z) u YZ). If W is in
IP(Y,Z) then W is in Z, so W is not in V'\(IV(Y,Z) u YZ).
Now we will demonstrate if W is in IV(Y,Z) u IP(Y,Z) u Y then no parent of W is in

V'\(IV (Y ,Z) u YZ). Suppose first that W is in IV (Y ,Z) and T is a parent of W. If T is in
IV(Y,Z) u IP(Y,Z) u Y this reduces to the previous case. Assume then that T is not in
IV(Y,Z) u IP(Y,Z) u Y. We will show that T is in YZ. T is not d-connected to Y given Z.
However, W, a child of T, is d-connected to Y given Z by some path U. If T is on U then Tis
d-connected to Y given Z, contrary to our assumption, unless T is in YZ. If T is not on U, and
U is not into W, then the concatenation of the edge between T and W with U d-connects T and
Y given Z, contrary to our assumption, unless T is in YZ. If T is not on U, but U is into W,
then because W is in IV (Y ,Z) it has a descendant in YZ. If W has a descendant in Z, then W is
a collider on the concatentation of the edge 'between T and W with U, and has a descendant in
Z; hence T is d-connected to Y given Z, contrary to our assumption, unless T is in YZ. If W
does not have a descendant in Z, then there is a directed path D from W to Y that does not
contain any member of Z. The concatenation of the edge from T to W and D d-connects T and
Y given Z, contrary to our assumption, unless T is in YZ. In any case, T is in YZ, and not in
V'\(IV(Y,Z) u YZ).
Suppose next that W is in IP(Y,Z) and T is a parent of W. It follows that some parent R of W
is in IV(Y,Z) or in Y, and Wis in Z. If Tis in IV(Y,Z) u IP(Y,Z) u Y this reduces to the
previous case. Assume then that T is not in IV (Y ,Z) u IP(y ,Z) u Y. If R is in Y, then T is d-
connected to Y given Z by the concatenation of the edge from R to W and the edge from W to
T, contrary to our assumption, unless T is in YZ. Hence T is in YZ, and not in V,\(IV(Y,Z)
u YZ). Assume then that R is in IV (Y ,Z). R is d-connected to Y given Z by some path U. If
T is on U then T is d-connected to Y given Z unless T is in YZ. If W is on U, but T is not,
then W is a collider on U, because W is in Z. W is also a collider on the concatenation of the
edge from T to W with the subpath of U from W to Y; hence this path d-connects T and Y
given Z unless T is in YZ. If neither T nor W is on U, then the concatentation of the edge
between T and W, the edge between Wand R, and U, is a path on which W is a collider and R
is not a collider (because R is a parent of W); hence this path d-connects T and Y given Z,
unless W is in YZ. By hypothesis, T is not d-connected to Y given Z because T is not in
IV(Y 7;); it follows that T is in YZ. Hence T is not in V'\(IV(Y,Z) u YZ).
Sr.ppose finally that W is in Y and T is a parent of W. It follows that T is d-connected to Y

;;,iven Z unless T is in YZ. By hypothesis, T is not d-connected to Y given Z because T is not
in IV(Y,Z) so T is in YZ. Hence T is not in V,\(IV(Y,Z) u YZ).
Now we will demonstrate by contraposition that if W is in V'\(IV(Y,Z) u IP(Y,Z) u Y) then

neither W nor any parent of W is in the scope of the summation over IV(Y,Z) u Y. Suppose
Wor some parent T of W is in IV(Y,Z) u Y. If W is in IV(Y,Z) u Y it follows trivially that
W is not in V'\(IV(Y,Z) u IP(Y,Z) u V). Suppose T is in IV(Y,Z) u Y but W is not. We
will show that W is in YZ. If T is in Y, then W is d-connected to Y given Z, contrary to our
assumption, unless T is in YZ. If T is in IV (Y ,Z) it follows that there is a path U d-connecting
T and Y given Z. If W is on U, then W is d-connected to Y given Z, contrary to our
hypothesis, unless W is in YZ. If W is not on U, then the concatenation of the edge between W
and T with U d-connects Wand Y given Z (because T is not a collider and not in Z), contrary
to our hypothesis, unless W is in YZ. It follows that W is in YZ. If W is in Z, then W is in
IP(Y,Z), and hence not in V,\(IV(Y,Z) u IP(Y,Z) u V). If W is in Y, then W is not in
V,\(IV(Y,Z) u IP(Y,Z) u V). Hence by contraposition, if W is in V'\(IV(Y,Z) u IP(Y,Z)
u Y) then neither W nor any parent of W is in the scope of the summation over IV(Y,Z) u Y.
The prooffor the numerator is essentially the same. Hence,

Proofs 383
IT
~
I, P(WIParents(W))
V'\YZ WEV
IT
~
I, P(WIParents(W))
V'\Z WEV'
IT P(WIParents(W))
~
I,
-=.IV.:..;("'Y-'=,Z"")--'W.:...E;::I:..:.V-'-(Y"-',=Zt..;;)u:.,:I:::..P.>..:(Y:...:,Z=)'-'u--'Y'----_ _ _ X
IT
~
I, P(WIParents(W))
IT
~
I, P(WIParents(W))
V'\(IV (Y ,Z)u YZ) WE V' \(IV (Y ,Z)uIP(Y ,Z)U Y)
IT
~
I, P(WIParents(W))
V'\(IV(Y,Z)uYZ) WEV\(IV(Y,Z)uIP(Y,Z)uY)
IT P(WIParents(W))
~
I,
IV(Y,Z) WEIV(Y,Z)uIP(Y,Z)uY
IT
~
I, P(WIParents(W))
Lemma 3.3.6: In a directed acyclic graph G, if V is d-connected to Y given Z, and X is d-

separated from Y given Z, then V is d-connected to Y given XZ.
Proof. Suppose X is d-separated from Y given Z. If V is d-separated from Y given XZ, but
d-connected to Y given Z, then there is a path U that d-connects V and some Yin Y given Z,
but not given XZ. It follows that some non-collider X on U is in X. Bence U(X,y) d-connects
X and Y given Z. :.
Lemma 3.3.7: In a directed acyclic graph G, if V is d-connected to Y given XZ, and X is d-

separated from Y given Z, then V is d-connected to Y given Z.
Proof. Suppose X is d-separated from Y given Z. If V is d-separated from Y given Z, but d-
connected to Y given XZ, then there is a path U that d-connects V and Y given XZ, but not
given Z. Some vertex on U is a collider with a descendant in X, but not in Z. Let C be the
vertex on U closest to Y that is the source of a directed path to some X in X that contains no
member of Z. C is d-connected to Y given Z. If X is on U then U(X,Y) d-connects X and Y
given Z. If X is not on V, then there is a directed path from C to X that does not contain any
member of Z, and hence X is d-connected to Y given Z, contrary to our assumption. :.
Lemma 3.3.8: In a directed acyclic graph G, if X is d-separated from Y given Z, and P

satisfies the Markov condition for G, then X is independent of Y given Z.
Proof. We will show if X is d-separated from Y given Z that P(YIXZ) = P(YIZ) by showing
that IV(Y,XZ) = IV(Y,Z) and IP(Y,XZ) = IP(Y,Z) and applying lemma 3.3.5.
Suppose that V is in IV(Y,Z). V is d-connected to Y given Z and has a descendant in YZ.

Hence V has a descendant in XYZ. It follows by lemma 3.3.6 that V is d-connected to Y given
XZ. Hence V is in IV(Y,XZ).
Suppose then that V is in IV(Y,XZ); we will show that V is also in IV(Y,Z). Because V is in
IV(Y,XZ), V is not in XYZ, V has a descendant in XYZ and is d-connected to Y given XZ.
Because Vis not in XYZ it is not in XZ. By lemma 3.3.7 V is d-connected to Y given Z. If V
has a member X of X as a descendant, but no member of YZ as a descendant then there is a
directed path from V to X that contains no member of Y or Z. It follows by lemma 3.3.3 that X
is d-connected to Y given Z, contrary to our hypothesis. Hence V has a member of YZ as a
descendant, and is in IV(Y,Z).
Suppose that V is in IP(Y,Z). If V has a parent in Y, then V is in IP(Y,XZ). IfV has a parent
Tin IV(Y,Z) then T is in IV(Y,XZ) because IV(Y,Z) = IV(Y,XZ). Hence V is in
IP(Y,XZ).
Suppose that V is in IP(Y,XZ). Because V is in IP(Y,XZ) V is in XZ and has a parent in

IV(Y,XZ) u Y. We have already shown that IV(Y,XZ) u Y = IV(Y,Z) u Y. We will now
show that V is not in X. If V is in X and has a member of Y as a parent, then X is d-connected
to Y given Z, contrary to our hypothesis. If V is in X and has some Win IV(Y,XZ) as a
parent, then W is in IV (Y ,Z). It follows that X is d-connected to Y given Z, contrary to our
hypothesis. Hence V is not in X, and IP(Y,XZ) = IP(Y,Z).
By lemma 3.3.5, P(YIXZ) = P(YIZ), and hence X is independent of Y given Z. :.
Lemma 3.3.9: In a directed acyclic graph G, if X is not a descendant of Y, and X and Y are
not adjacent, then X and Yare d-separated by Parents(y).
Proof. (A slight variant of this is stated in Pearl (1989». Suppose on the contrary that some
undirected path V d-connects X and Y given Parents(X). If V is into Y then it contains some
Proofs 385
member of Parents(y) not equal to X as a non-collider. Hence it does not d-connectX and Y
given Parents(y), contrary to our assumption. If U is out of Y, then because X is not a
descendant of Y, U contains a collider. Let C be the collider on U closest to Y. If U d-connects
X and Y given Parents(y) then C has a descendant in Parents(Y). But then C is an ancestor
of Y, and Y is an ancestor of C, so G is cyclic, contrary to our assumption. Hence no
undirected path between X and Y d-connects X and Y given Parents(y). :.
Theorem 3.3: P(V) is faithful to directed acyclic graph G with vertex set V if and only if for
all disjoint sets of vertices X, Y, and Z, X and Y are independent conditional on Z if and only
if X and Y are d-separated given Z.
Proof. :::>Suppose that P is faithful to G. It follows that P satisfies the Markov condition for
G. By lemma 3.3.8 if X and Y are d-separated given Z then X and Y are independent
conditional on Z. By lemma 3.5.8 (proved below) there is a distribution P' that satisfies the
Markov condition for G such that if X and Y are not d-separated given Z then X and Y are not
independent conditional on Z. It follows that if X and Y are not d-separated given Z then the
Markov condition does not entail that X and Y independent conditional on Z.
<=Suppose that X and Y are independent conditional on Z in P if and only if X and Y are d-
separated given Z. It follows from lemma 3.3.9 that that P satisfies the Markov condition for G
because Parents(V) d-separates V from V\(Descendants(V) u Parents(V». Hence all of
the conditional independence relations entailed by the Markov condition are true of P. If the
independence of X and Y conditional on Z is not entailed by the Markov condition for G then
by lemma 3.5.8 X and Y are not d-separated in G, and X and Yare not independent
conditional on Z. It follows that P is faithful to G. :.
13.5 Theorem 3.4
Theorem 3.4: If P(V) is faithful to some directed acyclic graph, then P(V) is faithful to
directed acyclic graph G with vertex set V if and only if
(i) for all vertices X, Y of G, X and Yare adjacent if and only if X and Yare dependent
conditional on every set of vertices of G that does not include X or Y; and
(ii) for all vertices X, Y, Z such that X is adjacent to Yand Y is adjacent to Z and X and
Z are not adjacent, X -> Y <- Z is a subgraph of G if and only if X, Z are dependent
conditional on every set containing Y but not X or Z.
Proof. The theorem follows from a theorem first proved in Verma and Pearl (1990b). :.
13.6 Theorem 3.5
Theorem 3.5: Let S be an LCT with directed acyclic graph G over the set of non-error
variables V. Then for any two non-error vertices A, B in V and any subset H of V\{A,B}, G
linearly implies that PAB.H = 0 if and only if A, Bare d-separated given H .
The distributed form of an expression or equation E is the result of carrying out every
multiplication, but no additions, subtractions, or divisions in E. If there are no divisions in an
equation then its distributed form is a sum of terms. For example, the distributed form of the
equation u = (a + b)(e + d)v is u = aev + adv + bev + bdv. In an LCF or LCT T, if an
expression is equal to ee, where c is a non-zero constant, and e is a product of equation
coefficients raised to positive integral powers, then e is the equation coefficient
factor(e.c2.f.) of ee, and c is the constant factor (c.f.) of ceo
An acyclic directed graph G over V is an I-map of probability distribution P(V) iff for every
X, Y, and Z that are disjoint sets of random variables in V, if X is d-separated from Y given Z
in G then X is independent of Y given Z in P(V). An acyclic graph G is a minimal I-map of
probability distribution P iff G is an I-map of P, and no proper subgraph of G is an I-map of P.
An acyclic graph G over V is a D-map of probability distribution P(V) iff for every X, Y,
and Z that are disjoint sets of random variables in V, if X is not d-separated from Y given Z in
G then X is not independent of Y given Z in P(V). However, when minimal I-map, I-map, or
D-map is applied to the graph in an LCT or LCF, the quantifiers in the definitions apply ouly to
sets of non-error variables.
A trek T(/,.I) between two distinct vertices / and J is an unordered pair of acyclic directed paths
from some vertex K to / and J respectively that intersect only at K. The source of the paths in
the trek is called the source of the trek. / and J are called the termini of the trek. Given a
trek T(/,.I) between / and J, J(T(/,J) will denote the path in T(/,.I) from the source of T(/,.I) to
/ and J(T(/,.I) will denote the path in T(/,.I) from the source ofT(/,.I) to J. One of the paths in a
trek may be an empty path. However, since the termini of a trek are distinct, only one path in a
trek can be empty. T(/,.I) is the set of all treks between / and J. T(/,.I) will represent a trek in
T(/,.I). S(T(/,.I) represents the source of the trek T(/,.I).
Proofs 387
The proofs of the following two lemmas are trivial.
Lemma 3.5.1: In a directed acyclic graph G, every undirected path V = <VI,v2, ... V n-I,Vn>
without colliders contains a vertex Vk such that <Vk, ... ,VI> and <Vk, ... ,Vn> are directed
subpaths of V that intersect only at Vk.
Hence, corresponding to each undirected path V = <VI,V2, ... Vn-I, V n> without colliders is a
trek T = «Vk, ... ,VI>,<Vk, ... ,vn». When V is a directed path, one of the paths is empty; for
example, Vk = VI.
Lemma 3.5.2: In a directed acyclic graph G, for every trek «VI, ... ,Vn>,<Vt, ... ,Vm», the
concatenation of <Vn, ... ,VI> with <Vt, ... ,Vm> is an undirected path from Vn to Vm without
colliders.
We will say that a directed acyclic graph has error variables if every vertex of indegree pot equal
to 0 has an edge into it from a vertex of indegree 0 and outdegree 1. If each independent
random variable in an LCT S is normally distributed, then the joint distribution of the set of all
random variables in the LCT is multi-variate normal. We will say the random variables in such
an LCT have a linear multi-variate normal distribution. The next series of lemmas demonstrate
that every directed acyclic graph with error variables is faithful to some LCT S in which the
joint distribution Q of the random variables in S is linear multi-variate normal.
Lemma 3.5.3: If S is an acyclic multi-variate normal LCT with directed acyclic graph G' and
distribution P, V is the set of non-error terms in S, G is the subgraph of G' over V, and the
exogenous variables are jointly independent, then G is a minimal I-map of P(V).
Proof. Let V be the set of non-error terms in S, and G be the subgraph of G' over V. First we
will show that if A and B are distinct variables in V, and B is not a descendant of A or a parent
of A in G, then A is independent of B given Parents(G,A). eA is normally distributed and
uncorrelated with any of the parents of A or B. B is not a linear function of Parents(G,A)
because the distribution is positive. Hence, if we write A as a linear function of Parents(G,A),
B, and eA, this is a regression model of A. The coefficient of B in such an equation is zero. The
coefficient of B in such a linear equation for A is zero if and only if A and B are independent
conditional on Parents(G,A). (See Whittaker 1990.). Hence B is independent of A given
Parents(G ,A). Because the joint distribution is normal, it follows that A is independent of the
set of its non-parental non-descendants given its parents. Hence G is an I-map of P(V).
We will now show that P(V) satisfies the Minirnality Condition for G. Suppose, on the
contrary, that G is not a minimal I-map of P(V). It follows that some some subgraph of G is an
I-map of P(V). Let GSub be a subgraph of G that is an I-map of P(V), and in which the only
difference between G and GSub is that X is a parent of Y in G, but not in GSub. Because
Parents(GSub,Y) u {X} = Parents(G,x), when Y is written as a linear function of
Parents(GSub,Y), X, and Ey, the coefficient of X is not zero. But because X is not a parent of
Yin GSub, and not a descendant of Y in GSub, it follows that X and Y are d-separated given
Parents(GSub,Y). Because GSub is an I-map of P(V), X and Y are independent given
Parents(GSub,Y). But this entails that the coefficient of X in the linear equation for Y in terms
of Parents(G,Y) and Ey is zero, which is a contradiction. :.
Lemma 3.5.4: If a polynomial equation Q in real variables <Xt. ... ,xn> is not an identity,
then for every solution a of Q, and for every E > 0 there is a non-solution b of Q such that Ib - al
<E.
Proof. The proof is by induction on the number n of variables in Q.
Base case: If n = 1, then there are only a finite number of solutions of Q. It follows that for
every solution a of Q, and for every E> 0 there is a non-solution b of Q such that Ib - al < E.
Induction case: Suppose that Q is a polynomial equation in <X ..... ,xn>' Q is not an identity,
and the lemma is true for n-l. Take an arbitrary solution <a ..... ,an> of Q. Transform Q into a
polynomial equation Q' inXn by fixing the variables <Xl"",xn-l> at the value <ab ... ,an.l>.
There are two cases.
In the fust case, Q' is not an identity. Hence, by the induction hypothesis, there is a non-
solution of Q' whose distance from an is < E. Let a'n be this non-solution of Q'. Then a' =
<al, ...an-l,a'n> is a non-solution of Q, and la - a'i < E.
In the second case, Q' is an identity. Rewrite Q so that it is of the form
where each Qm is a polynomial in at mostXl,... ,xn-l.
For each m, the equation Qm =0 is a polynomial equation in less than n variables. If Q' is an
identity, then when terms of the same power of Xn are added together, the coefficient of each
Proofs 389
power of Xn is zero. This implies that <al, ... ,an-I> is a solution to Qm = 0 for each m. If, for
each m, Qm = 0 is an identity, then so is Q; hence for some m, Qm = 0 is not an identity. For
this value of m, by the induction hypothesis, there is a non-solution <a'I, ... ,a'n-l> to Qm = 0
that is less than distance e from <a1, ... ,an_I>. If <a'l, ... ,a'n-I> is substituted for <Xl, ... ,xn-
I> in Q, the resulting polynomial equation in Xn is not an identity. This reduces to the first
case. :.
Lemma 3.5.5: If G' is a subgraph of G, and there is some LeT S' with directed acyclic
*
graph G and distribution P' such that PIJ.Z 0 in P', then there is some LeT S containing G
and distribution P such that PIJ.z * 0 in P.
Proof. By lemma 3.2.1 in S' PIJ.z= 0 is equivalent to a polynomial equation in the linear
coefficients and variances of independent variables in S'. Since there is some LeT S'
*
containing G' such that PIJ.z 0 in S', tthe polynomial equation is not an identity.
Let S be an LCT with directed acyclic graph G such that for all variables J, /, if the coefficient
c' of J in the equation for / in S' is not equal to zero, then the coefficient of J in the equation for
/ in S is equal to c'. In S, PIJ.Z = 0 is equivalent to a polynomial equation E in the linear
coefficients and variances of independent variables in S. When labels of the edges in G but not
in G' are set to zero, the polynomial in E equals the polynomial in E'. No label of an edge in G
but not in G' occurs in E'. Hence when the labels of the edges in G but not in G' are set to non-
zero values, the polynomial in E contains all of the terms that are in E' and possibly some extra
terms. Let us say that two terms in a polynomial equation are like terms if they contain the
same variables raised to the same powers. Each of the terms that are in E but not E' contain
some linear coefficient that does not appear in any term in E'; hence each of the additional terms
in E is not like any term in E'.
If E were an identity, then the sum of the coefficients of like terms in E would be equal to zero.
Since E' is not an identity, there are like terms in E' such that the sum of their coefficients is not
zero. These same like terms appear in E. Furthermore, since the only additional terms in E that
are not in E' are not like any term in E', it follows that if the sum of the coefficients of like
terms in E' is not zero, then the sum of the coefficients of the same like terms in E is not
identically zero. Hence E is not identically zero, and there is some LCT S containing G such
that PIJ.Z * 0 in S. :.
The next lemma states that given a set Z of partial correlations and a directed acyclic graph G, if
it is possible to construct a set S of LCTs with directed acyclic graph G such that each Z in Z
fails to vanish for some one of the LCTs in S, then it is possible to construct a single LCT with
directed acyclic graph G such that all of the Z in Z fail to vanish.
Lemma 3.5.6: Given a set of partial correlations Z and a directed acyclic graph G, if for all Z
*
in Z there exists an LCT S' with directed acyclic graph G and distribution P' such that Z 0 in
P', then there exists a single LCT S with directed acyclic graph G and distribution P such that
for allZ in Z, Z *0 in P.
Proof. The proof is by induction on the cardinality of Z.
Base Case: If the only member of Z is Z, then by assumption there is an LCT S containing G
such that Z = O.
Induction Case: Suppose that the lemma is true for each set of cardinality n-l, Z is of
cardinality n, and for each Zi in Z, there is an LCT S' with directed acyclic graph G and
*
distribution P' such that Zi 0 in P'. By the induction hypothesis, there is an LCT S with
directed acyclic graph G and distribution P such that Zi * 0, i SIS n-l. Let V be a set of
values for the linear coefficients and variances of independent variables such that Zi * 0, i SIS
n-l. The valuation V either makes Zn equal to zero or it doesn't If it doesn't, then the proof is
done. If it does, we will show how to perturb V by a small amount to make Zn * 0, while
keeping each Zi * 0, i SIS n-l.
By lemma 3.2.1, each of the partial correlations in Zi in Z is equivalent to a polynomial Qi in
the linear coefficients and the variances of independent variables in G. Suppose that the smallest
non-zero value for any of the Qi under the valuation V is 8. By lemma 3.5.4, for arbitrarily
small e there is a non-solution V' to Zn = 0 within distance e of V. Choose an e small enough
so that the largest possible change in any of the Qi is less than 8. For the valuation V' then Zi *
0, i S 1 Sn. :.
Recall that if a graph with error variables is a D-map of some distribution P, then we consider
only dependencies among the non-error variables.
Lemma 3.5.7: For every directed acyclic graph G with error variables, there'is an LCT S
with directed acyclic graph G and joint linear multi-variate normal distribution Q, such that G is
a D-rnap of Q.
Proof. In order to show that G is a D-rnap of Q, we must show that for all disjoint sets of
variables X, Y, and Z, if X and Y are not d-separated inG, then X is not independent of Y
given Z in Q. In a linear multi-variate normal distribution, if X, Y, and Z are disjoint sets of
Proofs 391
variables, then X 11 YIZ iff X 11 YIZ for each X in X and Yin Y; similarly if X, Y, and Z are
disjoint sets of variables then X and Y are d-separated given Z iff for all X in X and Yin Y, X
and Y are d-separated given Z. Hence, we need consider only dependency statements of the
form X and Y are not independent given Z, where X and Y are individual variables. Also in a
linear multi-variate normal distribution, PXy.z = 0 iff X 11 YIZ. So it suffices to prove that
there is an LCf S with directed acyclic graph G and distribution P such that for each X, Y, and
Z in G such that X and Y are not d-separated given Z in G, PXy.z ¢ 0 in P. The proof is by
induction. We assume that in all of the LCfs constructed, the independent random variables are
normally distributed.
Base Case. If Z is empty, then by lemma 3.5.1, X and Yare not d-separated given Z iff there
is a trek connecting them. Form a subgraph G' and a sub-LCf S' with directed acyclic graph
G' and distribution P', such that there is exactly one trek between X and Y. It was proved in
Glymour et al. (1987) that in this case the covariance between X and Y is equal to the product
of the labels of the edges in the trek (the linear coefficients) times the variance of the source of
the trek. If each of these quantities is non-zero, so is the covariance, and also the correlation in
P'. By lemma 3.5.5 if PXy is not identically zero in S' it is also not identically zero in some
LCT S with directed acyclic graph G. By lemma 3.5.6 there exists a LCT containing G in
which for all X and Y, if X and Y are not d-separated by the empty set then the correlation
between X and Y is not zero.
Induction Case. Suppose that there is an LCT S with directed acyclic graph G and distribution
P such that for each X, Y, and for each A of cardinality less than n that does not contains X or
Y, such that X and Y are not d-separated given A in G, PXY.A ¢ 0 in P. Let Z be of cardinality
n. Suppose that X and Y are not d-separated by Z in G. It follows that there is an undirected
path U between X and Y such that every vertex without a collider is not in Z, and every vertex
Vi on U that is a collider is the source of a directed path Ui. from Vi to a variable in Z. Form a
subgraph G', such that G' contains only the undirected path U, one directed path Ui from each
collider Vi on U, the vertices in those paths, and the vertices in Z. Shorten each Ui so that it
contains only one variable in Z. Finally, if two variables Vn and Vm that are colliders on U are
the sources of directed paths Un and Um that intersect, let F be the first point of intersection of
Un and Um. Replace the subpath of U from Vn to Vm by the concatenation of the subpaths of
Un(Vn,F) and Um(F,vm), and replace Un and Um by Un(F,2), where Z is in Z. The new path
has one fewer collider than the old path. Repeat this process until none of the Ui intersect each
other or there are no colliders on U. There are two cases.
In the fIrst case, U contains no vertices with a collider, and hence no vertices in Z. By lemma
3.5.1 there is a trek between X and Y that contains no vertices in Z. Let R be an arbitrary vertex
in Z, and W = Z\{R}. There is a trek between X and Y that contains no vertices in W. It
follows that W does not d-separate X and Y, so by the induction hypothesis, there is an LCf
with directed acyclic graph G' and distribution P' such that PxY.w ¢ O. It follows from lemma
= =
3.5.3 that in P' that PxR.w 0 and PYR.W 0 because by construction there are no undirected
paths from X to R or Y to R. By the recursion formula for partial correlation, PXY.w = 0 iff
PXY.W = PXR.W X PYR.W' But PXY.W is non-zero in P', and PXR.W x PYR.W is zero in P'.
Hence PXY.z ¢ 0 in P'. By lemma 3.5.5, there is some LeT S" with directed acyclic graph G
and distribution P" such that PXY.z ¢ 0 in P".
In the second case, U contains vertices with colliders, but every vertex that is not a collider is
not in Z. See fIgure 1.
X • A B c • Y
T D
Z= {D, E }
T E
Figure I
Let E be the vertex that is the sink of the directed path from the collider closest to Yon U, and
W = Z\{E}. Since by construction there is a trek between Yand E that does not contain any
variables in W, Yand E are not d-separated by W. There is also an undirected path from X to E
such that every vertex that is not a collider is not in W, and every vertex that does contain a
collider has a descendant in W. Hence X and E are not d-separated by W. By the induction
hypothesis, there is an LCf S' with directed acyclic graph G' and distribution P' such that
PXE.W ¢ 0, and PYE.W ¢ 0 in P'.
On the other hand, since path U was constructed so that each vertex that is a collider has only
one descendant in Z, and W does not contain E, X and Y are d-separated by W. Hence by
lemma 3.5.3 PXY.W = 0 in P'.
Proofs 393
PXY.W = 0 iff PXY.w = PXE.W XPYE.W· Since PXY.w = 0, while PXE.W x PYE.W ¢ 0,
PXY.z ¢ 0 in P'. By lemma 3.5.5, there is an LCT S" with directed acyclic graph G and
distribution P" such that PXY.Z ¢ 0 in P".
Since for each triple X, Y, Z such that X and Yare not d-separated given ZinG there is an
LCT S' with directed acyclic graph G and distribution P' such that PXY.Z ¢ 0 in P', by lemma
3.5.6 there is an LCT S" with directed acyclic graph G and distribution P" such that for each
triple X, Y, Z for which X and Yare not d-separated given Z in G, PXY.Z ¢ 0 in P". Because
the LCTs constructed in lemmas 3.5.5 and 3.5.6 don't change the normality of the independent
variables, the joint distribution of the random variables in S is linear multi-variate normal.
Hence there is an LCT S such that Q is a linear multi-variate normal distribution and G is a D-
map ofQ. :.
Lemma 3.5.8: For every directed acyclic graph G with error variables, there is an LCT S
containing G with a linear multivariate normal distribution Q such that G is faithful ttl Q.
The next theorem states that the d-separability relations between sets of non-error variables can
be determined from a subgraph that does not include error terms.
Lemma 3.5.9: In an acyclic LCT S with directed acyclic graph G, let G' be the subgraph of
G over the non-error variables. Given three disjoint sets X, Y, and Z of non-error variables, X
is d-separated from Y given Z in G iff X is d-separated from Y given ZinG'.
Proof. If an error variable occurs on an undirected path, then that error variable is either the
source or the sink of the undirected path. Hence, error variables do not occur on any undirected
path between non-error variables. It follows that the undirected paths in G and G' between non-
error variables are exactly the same. The lemma then follows from the definition of d-
separability. :.
A directed acyclic graph G linearly implies PAB." = 0 if and only if PAB." = 0 in all
distributions linearly represented by G. (We assume all partial correlations exist for the
distribution.) Kiiveri and Speed (1982) explicitly notes the connection between the Markov
Condition and zero partial correlations.
Lemma 3.5.10: In an LCT S with directed acyclic graph G over the set of non-error
variables V and the distribution P(V), if Y d-separates X and Z, then S linearly implies that
pxz.Y =0.
Proof. Suppose Y d-separates X and Zin G. The values of the partial correlations in P(V) are
completely determined by the values of the linear coefficients and the variances of the
independent variables. Consider a multi-variate normal distribution P'(V) in the LCT with the
same linear coefficients and the same variances of independent variables as S, but in which the
independent variables are normally distributed and jointly independent. By lemma 3.5.3, Gis
an I-map of P'(V), and because Y d-separates X and Z, X 11ZIY in P'(V). Because P'(V) is
a multi-variate normal distribution, X 11ZIY if and only pxz.Y = O. It follows that pxz.Y = 0
in P'(V), and hence pxz.Y =0 in P(V). :.
Theorem 3.5: Let S be an LCT with directed acyclic graph G over the set of non-error
variables V. Then for any two non-error vertices A, B in V and any subset H of V\{ A,B}, G
linearly implies that PAB.H = 0 if and only if A, B are d-separated given H.
Proof. The if clause follows from Lemma 3.5.10.
The only if clause follows from lemma 3.5.7. By lemma 3.5.7 there is an LCT S such that Q,
the joint distribution of the random variables is linear multi-variate normal, and G is a D-rnap of
Q. In S, if A and B are not d-separated given H, then A and B are not independent given H,
and PAB.H ¢ O. Hence if A and B are not d-separated given H, G does not linearly imply that
PAB.H =O. :.
Corollary 3.5.1: In an LCT S = <G, (D/,P), EQ, L> in which the exogenous variables are
jointly independent, if X and Z are distinct non-error variables, and Y is a set of non-error
variables not including X and Z, if pxz.y is linearly implied to vanish then X,z 11 Y.
Corollary 3.5.2: In an LCT S = <G, (DJ,P), EQ, L>, if P is faithful to G,X and Z are
distinct non-error variables, and Y is a set of non-error variables not including X and Z , G
linearly implies that PxZ.Y = 0 if and only if X 11 ZIY.
Proofs 395
13.7 Theorem 3.6 (Manipulation Theorem)
Tbeorem 3.6 (ManipUlation Theorem): Given directed acyclic graph GComb over vertex set V
u W and distribution P(V u W) that satisfies the Markov condition for GComb, if changing the
value of W from WI to wz is a manipUlation of GComb with respect to V, GUnman is the
unmanipUlated graph, GMan is the manipulated graph, and
PUnman(W)(V) = II PUnman(W) (XIParents(Gunman ,X»

XeV
for all values of V for which the conditional distributions are defined, then
PMan(W) (V) =
IIPMan(W) (XIParents(GMan ,X» X
XeManipulated(W)
II PUnman(W) (X IParents(Gunman ,X»

x eV\Manipulated(W)
for all values of V for which each of the conditional distributiops is defined.
If G is a directed acyclic graph over a set of variables V u W, and V ("'\ W = 0, then W is

exogenous witb respect to V in G if and only if there is no directed edge from any
member of V to any member of W. If GComb is a directed acyclic graph over a set of variables
V u W, and P(V u W) satisfies the Markov condition for GComb, then changing the value of
W from WI to Wz is a manipulation of GComb with respect to V if and only if W is
exogenous with respect to V, and P(VIW = WI) ¢ P(VIW = wz).
We defme PUnman(W)(V) =P(VIW =WI), and PMan(W)(V) =P(VIW =wz), and similarly for
various marginal and conditional distributions formed from P(V).
We refer to GComb as the combined grapb, and the subgraph of GComb over V as the
unmanipulated grapb GUnman.
V is in Manipulated(W) (that is, V is a variable directly influenced by one of the manipUlation

variables) if and only if V is in Cbildren(W) ("'\ V; we will also say that the variables in

policy variables.
The manipulated graph, GMan is a subgraph of GUnman for which PMan(W)(V) satisfies the
Markov Condition and which differs from GUnman in at most the parents of members of
Manipulated(W).
Lemmas 3.6.1 and 3.6.2 show that distributions satisfying the antecedent of Theorem 3.6
exist.
In a directed acyclic graph G over V, X is in Non-Descendants(G ,Y) if and only if X is in
V and there is no directed path from any member of Y to X in G.
Lemma 3.6.1: Given directed acyclic graph GComb over vertex set V u Wand distribution
P(V u W) that satisfies the Markov condition for G, if changing the value of W from WI to
W2 is a manipulation of GComb with respect to V, and GUnman is the unmanipulated graph,
then PUnman(W)(V) satisfies the Markov Condition for GUnman.
Proof. PUnman(W)(V) satisfies the Markov Condition for GUnman if for each vertex V in V, V
is independent of Non-Descendants(GUnman,v)\Parents(GUnman,V) conditional on
Parents(GUnman, V) u W. Suppose that on the contrary that for some V in V, V is dependent
on Non-Descendants(GUnman, V)\Parents(GUnman, V) conditional on Parents(GUnman, V)
u W. It follows that there is some path U in GComb that d-connects V and some member X in
Non-Descendants(GUnman,V) given Parents(GUnman,V) u W. Every member of W that
occurs on U is a collider on U because U d-connects X and V given Parents(GUnman, V) u
W. Because W is exogenous to V, U contains no member ofW. It follows that no collider on
U has a descendant in W. Hence U d-connects V and X given Parents(GUnman,V) in GComb.
The path corresponding to U in GUnman also d-connects V and X given Parents(GUnman,V).
But this contradicts lemma 3.3.9. :.
Lemma 3.6.2: Given directed acyclic graph GComb over vertex set V u Wand distribution
P(V u W) that satisfies the Markov condition for GComb , if changing the value of W from WI
to W2 is a manipulation of GComb with respect to V, and GUnman is the unmanipulated graph,
then PMan(W)(V) satisfies the Markov Condition for some subgraph of GUnman.
Proof. The proof that PMan(W)(V) satisfies the Markov Condition for GUnman is essentially
the same as that of lemma 3.6.1. Because GUnman is an (improper) subgraph of itself,
PMan(W)(V) satisfies the Markov Condition for some subgraph of GUnman.
Proofs 397
Theorem 3.6 (Manipulation Theorem): Given directed acyclic graph GComb over vertex set V
u Wand distribution P(V u W) that satisfies the Markov condition for GComb, if changing the
value of W from WI to wz is a manipulation of GComb with respect to V, GUnman is the
unmanipulated graph, GMan is the manipulated graph, and
PUnman(W)(V) = II PUnman(W)(XIParents(Gunman'X))
XEV
for all values of V for which the conditional distributions are defmed, then
PMan(W)(V) =
IIPMan(W) (XIParents(GMan,X)) X
X EManipulated(W)
II
PUnman(W) (XIParents(Gunman'X))
XEV\Manipulated(W)
for all values of V for which each of the conditional distributions is defined.
Proof. By assumption, PMan(W)(V) satisfies the Markov Condition for GMan. Hence
PMan(W) = IIP(XI Parents(GMan , X)) =

XEV
IIP(XIParents(GMan , X)) x II
P(XI Parents(GMan , X))
XEManipulated(W) XEV\Manipulated(W)
for all values of V for which the conditional distributions exist. No member of W is a
descendant of any variable in V in GComb, so for each V in V\Manipulated(W), W is d-
separated from V given Parents(GComb,V) in GComb. For any member X of
V\Manipulated(W), Parents(GComb,x) = Parents(GUnman,x) = Parents(GMan,x). It
follows that P(VIParents(GMan,X),W = wz) = P(VIParents(GMan,X))
P(VIParents(GMan,x),W = WI) = P(VIParents(GUnman,x),W = WI). Hence
PMan(W) (V) =
II
PMan(W) (XIParents(GMan,X)) X II
PUnman(W) (XI Parents(Gunman' X))
X EManipulated(W) XEV\Manipulated(W)
for all values of V for which the conditional distributions are defined. :.
13.8 Theorem 3.7
Theorem 3.7: If G is a directed acyclic graph over V, X, Y, and Z are disjoint subsets of V,
and P(V) satisfies the Markov condition for G and the deterministic relations in
Deterministic(V) then if X and Y are D-separated given Z and Deterministic(V), X and Y
are independent given Z in P.
We will say that a set of variables Z determines the set of variables A, when every variable in
function of any proper subset of Z. Suppose G is a directed acyclic graph over V, and
Deterministic(V) is a set of ordered tuples of variables in V, where for each tuple D in
Deterministic(V), if Dis <VI, ... ,Vn> then Vn is a deterministic function of VI, ... , Vn-I and
is not a deterministic function of any subset of VI, ... , Vn-I; we also say (VI, ... ,Vn-tl
determines Vn.For a given Deterministic(V), if Z is included in V, then Det(Z) is the set
of variables determined by any subset of Z. Note that Z is included in Det(Z).
If G is a directed acyclic graph over V, and Z is included in V, then G' is in Mod(G) relative to
Deterministic(V) and Z if and only if for each V in V
(i) if there exists a set of vertices included in Z that are non-descendants of V in G and
that determine V, then Parents(G', V)= X, where X is some set of vertices included in
Z that are non-descendants of V in G and that determine V;
(ii) if there is no set X of vertices included in Z that are non-descendants of V in G and
that determine V, then Parents(G',V) = Parents(G,V).
If G is a directed acyclic graph with vertex set V, Z is a set of vertices not containing X or Y,
*'
and X Y, then X and Y are D-separated given Z and Deterministic(V) if and only if
there is no undirected path U in G between X and Y such that each collider on U has a
*'
descendant in Z, and no other vertex on U is in Det(Z); otherwise if X Y and X and Y are
not in Z, then X and Yare D-connected given Z and Deterministic(V). Similarly, if X, Y,
and Z are disjoint sets of variables, and X and Y are non-empty, then X and Y are D-separated
given Z and Deterministic(V) if and only if each pair <X,Y> in the Cartesian product of X
and Y are D-separated given Z and Deterministic(V); otherwise if X, Y, and Z are
disjoint, and X and Y are non-empty, then X and Y are D-connected given Z and
Deterministic(V).
Proofs 399
If G is a directed acyclic graph over V, Z is a subset of V that does not contain X or Y, and X
#' Y,then X and Y are det-separated given Z and Deterministic(V) if and only if either X
and Yare d-separated given Z v Det(Z) in some Mod(G) relative to Deterministic(V) and
Z, or X or Y is in Det(Z); otherwise if X#' Y and X and Yare not in Z, then X and Yare det-
connected given Z and Deterministic(V). If X, Y and Z are disjoint sets of variables in V,
and X and Y are non-empty, then X and Yare det-separated given Z if and only if every
member X of X and every member Y of Yare det-separated given Z; otherise if X, Y and Z
are disjoint sets of variables in V, and X and Y are non-empty, then X and Y are det-
connected given Z and Deterministic(V).
Lemma 3.7.1: Let G be a directed acyclic graph with vertex set V, Ord an ordering of
variables in V such that if A is before B in Ord then A is not a descendant of B in G,
Predecessors(Ord, V) the set of all vertices before V in Ofd, and P(V) a distribution over V.
P(V) satisfies the Minimality and Markov Conditions for G if and only if for each V in V, V is
independent of Predecessors(Ord,V)\Parents(G,V) given Parents(G,V) and for);lo proper
subset X(V) of Parents(G,V), V is independent of Predecessors(Ord,V)\X(V) given X(V).
Proof. See Pearl(1988). :.
Lemma 3.7.2: If G is a directed acyclic graph over V, and X, Y, and Z are disjoint subsets
of V, and P(V) satisfies the Markov condition for G and the deterministic relations in
Deterministic(V), then if X and Y are det-separated given Z and Deterministic(V), X and
Yare independent given Z in P.
Proof. First we will prove that P(V) satisfies the Markov condition for each directed acylic
graph G' in Mod(G). First form an acceptable ordering Ord of the variables in V for G. Let
Predecessors(Ord,V) be the variables that precede V in Ord. From lemma 3.7.1 it follows
that if G' is a directed acyclic graph in which for each V in V, V is independent of
Predecessors(V)\Parents(V) given Parents(V), then G' is an I-map of P(V). If X is a
subset of Parents(V) that determines V, it follows that V is independent of
Predecessors(V)\X given X. Hence if in G' Parents(V) = X, G' is still an I-map of P(V).
If either X or Y is included in Det(Z), it follows that X and Y are independent given Z v

Det(Z). Suppose then that neither X nor Y is included in Det(Z). By definition of det-
separability, X\Det(Z) and Y\Det(Z) are d-separated given Z v Det(Z). Hence
P«X v Y) \ Det(Z)IZ v Det(Z» = P(X \ Det(Z)IZ v Det(Z»P(Y \ Det(Z)IZ v Det(Z»
It now follows that X is independent of Y given Z because

P(X u YIZ) =P(X u YIZ u Det(Z» =P«X u Y) \ Det(Z)IZ u Det(Z» =

P(X \ Det(Z)IZ u Det(Z»P(Y \ Det(Z)IZ u Det(Z»=
P(XIZ u Det(Z»P(YIZ u Det(Z» = P(XIZ)P(YIZ)
Theorem 3.7: If G is a directed acyclic graph over V, X, Y, and Z are disjoint subsets of V,
and P(V) satisfies the Markov condition for G and the deterministic relations in
Deterministic(G) then if X and Y are D-separated given Z and Deterministic(V), X and Y
are independent given Z in P.
Proof. We will prove that if X and Y are det-connected given Z and Deterministic(V), then
X and Y are D-connected given Z and Deterministic(V). It follows then that if X and Y are
D-separated given Z and Deterministic(V), then X and Y are det-separated given Z and
Deterministic(V), and by lemma 3.7.1, X and Y are independent given Z in P.
Suppose some X in X is det-connected to some Y in Y given Z and Deterministic(V). It

follows by definition that X and Y are not in Z and not in Det(Z). Because X and Y are det-
connected given Z there is an undirected path U' that d-connects X and Y given Z in some
graph G' in Mod(G).
First, we will show that the path U corresponding to U' exists in G; then we will show that U
D-connects X and Y given Z and Deterministic(V) in G.
No member of Det(Z) is a non-collider on U' because U' d-connects X and Y given Z u

Det(Z). Hence for each non-collider A on U', Parents(G',A) equals Parents(G,A). It
follows that if there is an edge into A in G', there is a corresponding edge into A in G.
Suppose then that A is a collider on U'. If there is an edge into A in G' that does not exist in G,
then every parent of A is in Z. It follows that either the endpoints of U' are in Z, or some non-
collider on U' is in Z. But then U' does not d-connectX and Y given Z u Det(Z) in G'. Hence
if there is an edge into A on U', then the corresponding edge exists in G.
It follows that the path U in G corresponding to U' in G' exists.
The endpoints of U are not in Z u Det(Z), because they are equal to the endpoints of U',
which are not in Z u Det(Z).
Proofs 401
No non-collider on U is in Z v Det(Z), because each non-collider on U is a non-collider on

U', and no non-collier on U' is in Z v Det(Z).
Finally suppose that A is a collider on U'. It follows that A has a descendant in Z v Det(Z) in
G'. There are two cases.
If A has a descendant in Z in G', then it has a descendant in ZinG. Suppose that A has a
descendant X in Z in G, and let D(A,x) be a directed path from A to X in G'. Let Z be the
member of Z closest to A on D(A,x). Every edge that is in G' but not in G is out of a member
of Z. D(A,z) has no edges out of a member of Z. Hence every edge in D(A,2) exists in G, and
A has a descendant in Z in G.
Suppose A does not have a descendant in Z in G'. It follows that there is a directed path
D(A,x) from A to a member X of Det(Z)\Z in G'. If A itself is in Det(Z) then it has parents
not in Z, because U' d-connects X and Y given Z v Det(Z). Because G' is in Mod(G), it
follows from the fact that A has a parent not in Z that A has a descendant in ZinG. If A is not
in Det(Z) then D(A,x) is not an empty path, and it does not contain any member of Z. Hence
X has a parent that is not in Z. Because G' is in Mod(G), it follows from the fact that X has a
parent not in Z that X has a descendant in Z in G. D(A,x) exists in G because every edge in G'
but not in G is out of a member of Z, and D(A,x) contains no member of Z. Hence A has a
descendant in Z in G.
It follows that U D-connects X and Y given Z and Deterministic(V) in G.:.
13.9 Theorem 4.1
Theorem 4.1: Two directed acyclic graphs Gt, G2, are strongly statistically indistinguishable
if and only if (i) they have the same vertex set V, (ii) vertices VI and V2 are adjacent in GI if
and only if they are adjacent in G2, and (iii) for every triple VI, V2, V3 in V, the graph VI ->
V2 <- V3 is a subgraph of GI if and only if it is a subgraph of G2.
Proof. <= Suppose two directed acyclic graphs GI and G2 contain the same vertices, the same
adjacencies and the same colliders, and GI is a minimal I-map of P. By Theorem 3.4 the same
distributions are faithful to GI and Gz so they have the same d-separability relations, and hence
G2 is also an I-map of P.
G2 is also minimal. Every subgraph of Gl has the same d-separability relations as does the
corresponding subgraph of G2 because removing corresponding vertices and adjacencies from
both graphs leaves subgraphs that contain the same vertices, adjacencies and colliders. Hence,
if a subgraph of G2 is an I-map of P, then the corresponding subgraph of Gl is an I-map of P.
But by supposition, no proper subgraph of Gl is an I-map of P. Hence no proper subgraph of
G2 is an I-map of P. By defInition, G2 is a minimal I-map of P. It follows that Gl and G2 are
s.s.i.
=> Now consider the case where Gl and G2 differ either in their sets of vertices, their
adjacencies, or their colliders. We will show that there exists a distribution P such that Gl is a
minimal I-map of P, while G2 is not. By defInition, it follows that Gl and Gz are not s.s.i.
Case 1. Suppose fIrst that Gl and G2 differ in their sets of vertices. By defInition they are not
s.s.i.
Case 2. Suppose that Gl and G2 differ in their adjacencies. Suppose without loss of generality
that Gl contains an adjacency not in G2. Then there is a pair of vertices X and Y such that X
and Y are d-separated given a subset S in G2, while X and Y are not d-separated given S in
Gl. There is a distribution P faithful to Gl. Gl is also a minimal I-map of P. In Gl, X and Y
are dependent conditional on S. But because X and Y are d-separated given a subset S in G2,
G2 is not an I-map of P. Hence Gl and G2 are not s.s.i.
Case 3. Suppose that Gl and Gz differ in their unshielded colliders but not in any adjacencies.
Let Y be an unshielded collider on the path <X,y,z> in Gt. but not in G2. Let P be a
distribution faithful to Gl. It follows that Gl is a minimal I-map of P. In G2, X and Z are d-
separated given a set S containing Y, while in Gl X and Z are not d-separated given S. Since
Gl is faithful to P, X and Z are dependent conditional on S. Hence G2 is not a minimal I-map
of P, and Gl and G2 are not s.s.i.
Case 4. Finally, suppose that Gl and G2 differ in their shielded colliders but not in any
adjacencies or unshielded colliders. Let Y be a shielded collider on the path <X,Y,z> in Gl,
but not in G2. Suppose Gz' is the subgraph of G2 with the edge between X and Z removed.
Gz' is faithful to some distribution P. G2 is not a minimal I-map of P (because it contains a
subgraph which is an I-map of P). We will now show that Gl is a minimal I-map of P.
Proofs 403
First, Gl is an I-map of P. Gl is f.i. to G2. G2 is a proper supergraph of G2', and so the d-

separation relations true of G2 are included in the d-separation relations true of G2'; hence the
d-separation relations true of Gl are included in the d-separation relations true of G2'. It
follows that Gl is an I-map of P.
Gl is also minimal. If Gl' is a subgraph obtained by deleting from Gl any edge other than the
X - Z edge, by Case 2, the subgraph is not an I-map of P. If Gl' is a subgraph obtained by
deleting from Gl just the X - Z edge, then Gl' contains an unshielded collider at Y that does not
occur in G2'. By Case 3, Gl' is not an I-map of P.
Because Gl is a minimal I-map of P, and Gz is not, Gl and G2 are not s.s.i. :.
13.10 Theorem 4.2
Theorem 4.2: Two directed acyclic graphs G and H are faithfully indistinguishable if and
only if (i) they have the same vertex set, (ii) any two vertices are adjacent in G if and only if
they are adjacent in H, and (iii) any three vertices, X, Y, Z, such that X is adjacent to Yand Y
is adjacent to Z but X is not adjacent to Z in G or H, are oriented as X -> Y <- Z in G if and
only if they are so oriented in H.
Proof. This was proved in Venna and Pearl (1990b).
It also follows directly from Theorem 3.4. :.
13.11 Theorem 4.3
Theorem 4.3: Two directed acyclic graphs are faithfully indistinguishable if and only if some
distribution faithful to one is faithful to the other and conversely; i.e. they are f.i. if and only if
they are w.f.i.
Proof. Suppose Gl and G2 are f.i. By lemma 3.5.8 there is some distribution P faithful to
Gl. Hence P is faithful to G2, and Gl and G2 are w.f.i.
Suppose that Gl and G2 are w.f.i. Then there is some distribution P faithful to Gl and G2. It
follows that Gl and G2 have the same d-separation relations, so any distribution faithful to G1
is also faithful to G2 and vice-versa. :.
13.12 Tbeorem 4.4
Tbeorem 4.4: If probability distribution P satisfies the Markov Condition for directed acyclic
graphs G and H, and P is faithful to H, then for all vertices X, Y, if X, Yare adjacent in H they
are adjacent in G.
Proof. If P is faithful to H then X is adjacent to Yin H only if X, Yare dependent conditional
on every set of vertices not containing X or Y. Suppose then that P satisfies the Markov
condition for G but, contrary to the claim, X and Y are not adjacent in G. Then X is not a
parent of Y and Y is not a parent of X. Either X is not a descendant of Y or Y is not a
descendant of X; suppose without loss of generality that X is not a descendant of Y. Then by
the Markov Condition, X and Y are independent in P conditional on the set of all parents of Y,
which is a contradiction. :.
13.13 Tbeorem 4.5
Tbeorem 4.5: If probability distribution P satisfies the Markov and Minimality Conditions
for directed acyclic graphs G , and P is faithful to graph H, then (i) for all X, Y, Z such that X
-> Y <- Z is in H and X is not adjacent to Z in H, either X -> Y <- ZinG or X, Z are adjacent
in G and (ii) for every triple X, Y, Z of vertices such that X -> Y <- Z is in G and X is not
adjacent to ZinG, if X is adjacent to Yin Hand Y is adjacent to Z in H then X -> Y <- Z in H.
Proof. (i) Suppose that P satisfies the Markov and Minirnality Conditions for directed acyclic
graphs G , and P is faithful to graph H. Suppose X -> Y <- Z is in H and X is not adjacent to Z
in H. By Theorem 4.4, X is adjacent to Y and Y is adjacent to ZinG. Suppose Y is not a
collider on <X, Y, Z> in G and X and Z are not adjacent in G. Then by the Markov Condition
X and Z are independent conditional on some set containing Y; but since H is faithful, this is
impossible.
Proofs 405
(li) Suppose Y is an unshielded collider on the path <X ,Y,Z> in G. Then X and Z are d-
separated in G given some set of vertices, and hence d-separated given Parents(G,x) or
Parents(G,Z). It follows that X and Z are independent given Parents(G,x) or
Parents(G,'Z) in P. Y is not a parent of X or ZinG; hence in p,x and Z are independent given
some set not containing Y. But if X, Y and Y, Z are adjacent in Hand Y is not a collider on
<X, Y, Z>, then there is a trek between X and Z containing only X, Y, and Z; hence in H, X
and Z are not d-separated given any set of variables not containing Y. Because P is faithful to
H, X and Z are not independent given any set of variables containing Y. This is a
contradiction. :.
Corollary 4.1: If probability distribution P satisfies the Markov condition for directed acyclic
graph G and P is faithful to directed acyclic graph H and G and H agree on an ordering of the
variables (as, for example, by time) such that X - > Y only if X < Y in the order, then H is a
subgraph of G.
Proof. An immediate consequence of Theorem 4.4.
13.14 Theorem 4.6
Theorem 4.6: No two distinct s.s.i. directed acyclic graphs with the same vertex set are
rigidly statistically indistinguishable.
Proof. Suppose G1 and G2 are distinct s.s.i. directed acyclic graphs with vertex set V.
Because they are s.s.i they have the same adjacencies; hence if they are distinct graphs there is
some edge A -> B in G1 and B -> A in G2. Let U 1 and U2 be variables not in V. Embed G 1
and ~ in H1 and H2respectively by adding edges from U1 toA and U2 toB. ThenH1 and H2
are not s.s.i because they have different colliders. :.
13.15 Theorem 5.1
Theorem 5.1: If the input to the PC, SOS, PC-I, PC-2, PC* or 10 algorithms is data
faithful to directed acyclic graph G, the output is a pattern that represents G.
In a graph G, let V be in Undirected(X,y) if and only if V lies on some undirected path

between X and Y.
Lemma 5.1.1: In a directed acyclic graph G, if X is not a descendant of Y, and Y and X are
not adjacent in G, then X is d-separated from Y given Parents(y) n Undirected(X,Y).
Proof. Suppose on the contrary that some undirected path U d-connects X and Y given
Parents(X) n Undirected(X,y). If U is into Y then it contains some member of Parents(y)
n Undirected(X,y) not equal to X as a non-collider. Hence it does not d-connect X and Y
given Parents(Y) n Undirected(X,y), contrary to our assumption. If U is out of Y, then
because X is not a descendant of Y, U contains a collider in Undirected(X,Y). Let C be the
collider on U closest to Y. If U d-connects X and Y given Parents(y) n Undirected(X,y)
then C has a descendant in Parents(y) n Undirected(X,Y). But then C is an ancestor of Y,
and Y is an ancestor of C, so G is cyclic, contrary to our assumption. Hence no undirected path
between X and Y d-connects X and Y given Parents(y) n Undirected(X,Y). :.
Lemma 5.1.2: In a directed acyclic graph G, if X is adjacent to Y, and Y is adjacent to Z, and

X is not adjacent to Z, then the edges are oriented as X -> Y <- Z if and only for every subset S
of V,X is d-connected to Z given {Y} u S\{X,z}.
Proof. This follows from Theorem 3.4. :.
Lemma 5.l.3 was suggested in Pearl(1990a).
Lemma 5.1.3: In a directed acyclic graph G, if X is adjacent to Y, and Y is adjacent to Z, and

X is not adjacent to Z, then either Y is in every set of variables that d-separates X and Z, or it is
in no set of variables that d-separates X and Z.
Proof. Assume that in G, X, Z are not adjacent but X is adjacent to Y and Y is adjacent to Z.
Since X, Z are not adjacent, they are d-separated given some subset S\{X,z}. In G, the X - Y
and Y - Z edges collide at Y if and only if there is no set S containing Y and not X or Z such
that X, Z are d-separated given S. If the X - Yand Y - Z edges do not collide at Y, then there is
an undirected path U between X and Z that contains no colliders (including Y). Any set
S\{X,z} that does not contain Y will fail to d-separate X and Z because of this path. :.
Proofs 407
Theorem 5.1: If the input to the PC, SOS, PC-I, PC-2, PC* or 10 algorithms is data
faithful to directed acyclic graph G, the output is a pattern that represents G.
Proof. The correctness of the SOS algorithm is evident from Theorem 3.4 since the procedure
simply verifies the conditions for faithfulness given in that theorem.
Let G' be the output of one any of the algorithms except SOS. Suppose that X and Y are not
adjacent in G'. None of the algorithms removes an edge between X and Y unless X and Yare
d-separated given some subset of v\{X,Y}. If X and Yare d-separated given some subset of
v\{X,Y}, then they are not adjacent in G. Hence if X and Y are not adjacent in G', X and Y
are not adjacent in G.
Suppose X and Yare adjacent in the output G' of any of the algorithms except PC*. It follows
that in G X and Y are not d-separated given any subset of the adjacencies of X or any of the
adjacencies of Yin G' From what we have just proved, the adjacencies of X in G' are a
superset of Parents(G,x) and the adjacencies of Yin G' are a superset of P.arents(G,Y).
Hence X and Y are not d-separated given Parents(X,G) or Parents(Y,G) in G . It follows
from lemma 3.5.9 that X and Yare adjacent in G.
Suppose X and Y are adjacent in the output G' of PC*. Undirected(X,Y) in G' is a superset
of Undirected(X,y) in G. This, together with lemmas '3.5.9 and 5.1.1 entails that X and Y
are adjacent in G.
We will show by induction on the number of applications of orientation rules in the repeat loop
of the algorithm that the orientations are correct in the output G'.
Base Case: Suppose that X -> Y is oriented by the rule that if X is adjacent to Y, and Y is
adjacent to Z, and X is not adjacent to Z, then the edges are oriented as X -> Y <- Z if and only
Y is not in Sepset(X,2). This is a correct orientation by lemmas 5.1.2 and 5.1.3.
Induction Case: Suppose that the orientations of G' after n applications of orientation rules are
correct Suppose first that X -> Yis oriented because there is a directed path from X to Yin G'.
It follows from the induction hypothesis that there is a directed path from X to Y in G, and
hence X -> Yin G because G is acyclic. Suppose next that X -> Y is oriented because there is
an edge Z -> X and the edge between X and Yin G' has no arrowhead at X. It follows that Y is
in Sepset(X,Z), and hence Y is not a collider on the path <X,Y,Z> in G. Also by the
induction hypothesis Z -> X in G, and hence X -> Yin G. :.
13.16 Theorem 6.1
Theorem 6.1: (Verma and Pearl): If V is a set of vertices, 0 is a subset of V containing A

and B, and G is a directed acyclic graph over V (or an inducing path graph over 0) then A and
B are not d-separated by any subset of O\{A,B} if and only if there is an inducing path over the
subset 0 between A and B.
(Theorem 6.1 was first stated and proved in Venna and Pearl 1990 for directed acyclic graphs,
but that paper did not include the parts of the lemmas relating the existence of an inducing path
that is into (or out of) its endpoints to the existence of d-connecting paths that are into (or out
of) their endpoints.)
If G is a directed acyclic graph over a set of variables V, 0 is a subset of V containing A and

*
B, and A B, then an undirected path U between A and B is an inducing path relative to
o if and only if every member of 0 on U except for the endpoints is a collider on U, and every
collider on U is an ancestor of either A or B. We will sometimes refer to members of 0 as
observed variables. In a graph G, an edge between A and B is into A if and only if the mark
at the A end of edge is an ">". If an undirected path U between A and B contains an edge into A
we will say that U is into A. In a graph G, an edge between A and B is out of A if and only if
the mark at the A endpoint is the empty mark. If an undirected path U between A and B
contains an edge out of A we will say that U is out of A.
Lemma 6.1.1: If V is a set of vertices, 0 is a subset of V, G is a directed acyclic graph over

V (or an inducing path graph over 0) if there is an inducing path relative to 0 between A and B
that is out of A and into B, then for any subset Z of O\{ A,B} there is an undirected path C that
d-connects A and B given Z that is out of A and into B.
Proof. Let U be an inducing path over 0 between A and B that is out of A and into B. Every
observed vertex on U except for the endpoints is a collider, and every collider is an ancestor of
either A or B.
If every collider on U has a descendant in Z, then let C = U. Cd-connects A and B given Z

because every collider has a descendant in Z, and no non-collider is in Z. C is out of A and into
B.
Proofs 409
Suppose that not every collider on U has a descendant in Z. Let R be the collider on U closest
to A that does not have a descendant in Z, and W be the collider on U closest to A. R A and*
R *B because A and B are not colliders on U.
Suppose first that R = W. There is a directed path from R to B that does not contain A, because
otherwise there is a cycle in G. R is not in Z because R has no descendant in Z. B is not on
U(A,R). By lemma 3.3.2, U(A,R) d-connects A and R given Z, and is out of A. By lemma
3.3.3 there is a d-connecting path C between A and B given Z that is out of A and into B.
Suppose then that R * W. Because U is out of A, W is a descendant of A. W has a descendant

in Z by definition of R. It follows that every collider on U that is an ancestor of A has a
descendant in Z. Hence R is an ancestor of B, and not of A. B is not on U(A,R). By lemma
3.3.2, U(A,R) d-connects A and R given Z and is out of A. By hypothesis, there is a directed
path D from R to B that does not contain A or any member of Z. By lemma 3.3.3, there is a
path that d-connects A and B given Z that is out of A and into B ..:.

V (or an inducing path graph over 0), there is an inducing path U over 0 between A and B that
is into A and into B then for every subset Z of O\{A,B} there is an undirected path path C that
d-connects A and B given Z that is into A and into B.
Proof. If every collider on U has a descendant in Z, then U is a d-connecting path between A
and B given Z that is into A and into B. Suppose then that there is a collider that does not have
a descendant in Z. Let W be the collider on U closest to A that does not have a descendant in Z.
Suppose that W the source of a directed path D to B that does not contain A. B is not on
U(A,W). By lemma 3.3.2, U(A,W) is a path that d-connects A and W given Z, and is into A.
By lemma 3.3.3, there is an undirected path C that d-connects A and B given Z and is into A
and into B. Similarly, if the first collider Won U after B that does not have a descendant in Z is
the source of a directed path D to A that does not contain B, then by lemma 3.3.3, A and B are
d-connected given Z by an undirected path into A and into B.
Suppose then that the collider Won U closest to A that does not have a descendant in Z is not
the source of a directed path to B that does not contain A, and that the collider R on U closest to
B that does not have a descendant in Z is not the source of a directed path to A that does not
contain B. It follows that there is a directed path D from W to A that does not contain B or any
member of Z, and there is a directed path D' from R to B that does not contain A or any
member of Z. By lemma 3.3.2, U(W,R) d-connects Rand W given Z = z\{R,W}. D' d-
connects R and B given Z = z\{R,B} and D d-connects Wand A given Z = Z\{A,W}. By
lemma 3.3.1 there is an undirected path that d-connects A and B given Z that is into A and into
B. :.
In a graph G, Let A(A,B) be the union of the ancestors of A or B.

V (or an inducing path graph over 0) and an undirected path U in G d-connects A and B given
(A(A,B) II O)\{A,B} then U is an inducing path between A and B over O.
Proof. If there is a path U that d-connects A and B given (A(A,B) II O)\{A,B} then every
collider on U is an ancestor of a member of (A(A,B) II O)\{A,B}, and hence an ancestor of A
or B. Every vertex on U is an ancestor of either A or B or a collider on U, and hence every
vertex on U is an ancestor of A or B. If U d-connects A and B given (A(A,B) II O)\{A,B}'
then every member of (A(A,B) II O)\{A,B} that is on U, except for the endpoints, is a
collider. Since every vertex on U is in A(A,B), every member of 0 that is on U, except for the
endpoints, is a collider. Hence U is an inducing path between A and B over O. :.
The following pair of lemmas state some basic properties of inducing paths.
Lemma 6.1.4: If G is a directed acyclic graph over V, 0 is a subset of V that contains A and
B, and G contains an inducing path over 0 between A and B that is out of A, then there is a
directed path from A to BinG.
Proof. Let U be an inducing path between A and B relative to 0 that is out of A. If U does not
contain a collider, then U is a directed path from A to B. If U does contain a collider, let C be
the fIrst collider after A. By defInition of inducing path, there is a directed path from C to B or
C to A. There is no path from C to A because there is no cycle in G; hence there is a directed
path from C to B. Because U is out of A, and C is the fIrst collider after A, there is a directed
path from A to C. Hence there is a directed path from A to B. :.

V (or an inducing path graph over 0) that contains an inducing path relative to 0 between A
and B that is out of A, then every inducing path relative to 0 between A and B is into B.
Proof. By lemma 6.1.4, if there an inducing path out of A, and an inducing path out of B,
there is a cycle in G. :.
Theorem 6.1: (Verma and Pearl): If V is a set of vertices, 0 is a subset of V containing A

and B, G is a directed acyclic graph over V (or an inducing path graph over 0) A and B are not
Proofs 411
d-separated by any subset of O\{ A,B} if and only if there is an inducing path over the subset 0
between A and B.
Proof. This follows from lemmas 6.1.1, 6.1.2, 6.1.3, and 6.1.5. :.
13.17 Theorem 6.2
Theorem 6.2: In an inducing path graph G' over 0, if A is not an ancestor of B, and A and
B are not adjacent then A and B are d-separated given D-SEP(A,B).
If G' is an inducing path graph over 0 and A *" B, let V E D-SEP(A,B) if and only if A *" V
and there is an undirected path U between A and V such that every vertex on U is an ancestor of
A or B, and (except for the endpoints) is a collider on U.
Lemma 6.2.1: If G' is the inducing path graph for Gover 0 and there is a directed path from
A to B in G', then there is a directed path from A to Bin G.
Proof. Suppose there is a directed path D from A to B in G'. Let X and Y be any two vertices
adjacent on the directed path and that occur in that order. There is a directed edge from X to Y
in G'. By the definition of inducing path graph, there is an inducing path between X and Yin G
that is out of X. Hence by lemma 6.1.4, there is a directed path from X to Yin G.
In G, the concatenation of the directed paths between vertices that are adjacent on D contains a
subpath that is a directed path from A to B. :.
Lemma 6.2.2: If G' is the inducing path graph for Gover 0, and there is a path U d-
connecting A and B given ZinG' then there is a path d-connecting A and B given Z in G.
Proof. Suppose that U d-connects A and B in G'. If there are vertices R, S, and Ton U such
that R and S are adjacent on U, and S and T are adjacent on U, and S is in Z, then S is a
collider on U. By the definition of inducing path graph, in G there are inducing paths over 0
between R and S, and S and T, such that each of them is into S. By lemmas 6.1.1 and 6.1.2, in
G there is a d-connecting path given z\{R,s} between Rand S, and a d-connecting path given
Z\{S,T} between S and T, such that each of them is into S.
If there are vertices R, S, and Ton U such that Rand S are adjacent on U, and S and T are
adjacent on U, and S is a collider on U, then S has a descendant in ZinG'. By the definition of
inducing path graph, in G there are inducing paths between R and S, and Sand T, that are both
into S. By lemmas 6.1.1 and 6.1.2, in G there is a d-connecting path given z\{R,S} between
R and S, and a d-connecting path given Z\{S,T} between S and T, and both are into S. If S has
a descendant in Z in GO then by lemma 6.2.1 it has a descendant in ZinG.
By lemma 3.3.1, there is a path in G that d-connects A and B given Z. :.
Lemma 6.2.3: If GO is the inducing path graph for directed acyclic graph Gover 0 and there
is an inducing path U over 0 between A and C in GO, then there is an edge between A and C in
GO.
Proof. Suppose there is an inducing path over 0 between A and C in GO. By lemmas 6.1.1
and 6.1.2, in GO there is an undirected path d-connecting A and C given A(A,C) (') O\{A,C}.
Hence by lemma 6.2.2 there is an undirected path in G such that A and C are d-connected given
A(A,C) (') O\{A,C} in G. By lemma 6.1.3 there is an inducing path over 0 between A and C
in G. It follows by defmition that there is an edge between A and C in GO. :.
Let a total order Ord of variables in an inducing path graph or directed acyclic graphGo be
acceptable if and only if whenever A *" B and there is a directed path from A to B in GO, A
precedes B in Ord. In a graph G, vertex X is after vertex Y if and only if there is a directed
path from Y to X inG, and it is before vertex Y if and only if there is a directed path from X to
Y inG. For inducing path graph GO and acceptable total ordering 0 rd, let
Predecessors(Ord,V) equal the set of all variables that precede V (not including V) according
to Ord. For inducing path graph GO and acceptable total ordering Ord, W is in SP(Ord,Go,V)
*"
(separating predecessors of V in GO for ordering Ord) if and only if W V and there is an
undirected path U between Wand V such that each vertex on U except for V precedes V in Ord
and every vertex on U except for the endpoints is a collider on U. Notice that by this definition
each parent of Vis in SP(Ord,Go,V). For example in figure 2, if Ord = <X,S,T,R,M,z,Q,Y>,
then SP(Ord,Go,y) = {Q,T,S} and if Ord =<X,S,T,R,M,z,Y,Q> then SP(Ord,Go,y) = 0.
Proofs 413
z
Figure 2
Lemma 6.2.4: If G' is an inducing path graph and Ord an acceptable total ordering then
Predecessors(Ord,x)'SP(Ord,G',x) is d-separated from X given SP(Ord,G',x).
Proof. Suppose on the contrary that there is a path U that d-connects some V in
Predecessors(Ord,x)'SP(Ord,G',x) to X given SP(Ord,G',x). There are three cases.
First suppose U has an edge into X that is not a double-headed arrow. (By a dou'ble-headed
arrow we mean e.g. A <-> B.) Then some parent R of X is on U, and is not a collider on U. R
is in SP(Ord,G',x) and hence is not equal to V. Because R is not a collider on U, U does not
d-connect V to X given SP(Ord,G',x), contrary to our assumption.
Next suppose U has an edge out of X. Since V is in Predecessors(Ord,x)'SP(Ord,G',x) it

precedes X in Ord; hence there is no directed path from X to V. It follows that U contains a
collider. Let the first collider after X on U be R. R is a descendant of X, and the descendants of
R are descendants of X. It follows that no descendant of R (including R itself) is in
SP(Ord,G',x), and hence U does not d-connect V and X, contrary to our assumption.
Suppose fmally that U contains a double-arrow into X. Because U d-connects X and V given
SP(Ord,G',x), each collider along U has a descendant in SP(Ord,G',x) and hence precedes
X in Ord; it follows that every ancestor of a collider on U precedes X in Ord. Let W be the
vertex on U closest to X not in SP(Ord,G,x), and R be the vertex adjacent to Won U and
between Wand X. If R is not a collider on U, then U does not d-connect V and X given
SP(Ord,G',x). If R is a collider on U, then W *-> R on U. W is either an ancestor of V or of
a collider on U, in which case it precedes X, and is a member of SP(Ord,G',x), contrary to
our assumption. :.
Theorem 6.2: In an inducing path graph G' over 0, if A is not an ancestor of B, and A and
B are not adjacent then A and B are d-separated given D-SEP(A,B).
Proof. Suppose that A and B are not adjacent, and A is not an ancestor of B. Let the total
order Ord on the variables in G' be such that all ancestors of A and all ancestors of B except for
A are prior to A, and all other vertices are after A. Then SP(Ord,G',A) = D.SEP(A,B). Hence
by lemma 6.2.4, if B is not in D.SEP(A,B) then D·SEP(A,B) d-separates A from B in G. B
is in D·SEP(A,B) if and only if there is a path from A to B in which each vertex except the
endpoints is a collider on the path, and each vertex on the path is an ancestor of A or B. But
then there is an inducing path between A and B, and by lemma 6.2.3 A and B are adjacent,
contrary to our assumption. :.
13.18 Theorem 6.3
Theorem 6.3: H the input to the CI algorithm is data over 0 that is faithful to G, the output is
a partially oriented inducing path graph of Gover O.
In an inducing path graph G', U is a discriminating path for B if and only if U is an

undirected path between X and Y containing B, B :F- X, B :F- Y, and
(i) if V and V' are adjacent on U, and V' is between V and B on U, then V *-> V' on U,
(ii) if V is between X and B on U and V is a collider on U then V -> Yin G', else V <-*
Yin G',
(iii) if V is between Y and Bon U and V is a collider on U then V -> X in G', else V <-*
X in G',
(iv) X and Yare not adjacent in G'.
B is a definite non·collider on undirected path U if and only if either B is an endpoint of U,

or there exist vertices A and C such that U contains one of the subpaths A <- B *-* C, A *-* B
-> C, or A *-!..B..!.-* C.
In a partially oriented inducing path graph 1&, U is a definite discriminating path for B if
and only if U is an undirected path between X and Y containing B, B :F- X, B :F- Y, every vertex
on U except for B and the endpoints is a collider or a definite non-collider on U, and
(ii) if V is between X and B on U and V is a collider on U then V -> Yin 1&, else V <-* Y
in 1&,
Proofs 415
(iii) if V is between Y and B on U and V is a collider on U then V -> X in 1C, else V <-* X
in 1C,
(iv) X and Yare not adjacent in 1C.
Figure 3: <E,F,G,A,C,B> is a definite discriminating path for C
Lemma 6.3.1: If G' is an inducing path graph, U is a discriminating path for B between X
and Y, and X and Yare d-separated given S, then for every vertex Von U not equal to B, V is
in S if and only if V is a collider on U.
Proof. First we will prove for each vertex V on U between X and B that V is in S if and only
if V is a collider on U. The proof is by induction on the number of vertices between X and V on
U.
Base Case: Let A be the first vertex on U after X. If A =B, then trivially for every vertex V
between X and A, V is in S if and only if V is a collider on U. Suppose then that A"* B. If A
is a collider on U then there is an edge from A to Y. A is not a collider on the concatenation of
U(X,A) and the edge between A and Y, and hence that path d-connects X and Y given Sunless
A is in S. If A is not a collider on U then there is an edge between Y and A that is into A. By
defmition of discriminating path, the edge between X and A is into A. Hence A is a collider on
the concatenation of U(X,A) and the edge between A and Y. Hence that path d-connects X and
Y given S unless A is not in S.
Induction Case: Suppose that if there are n or fewer vertices between X and V on U, then V is
in S if and only if V is a collider on U. If there are only n vertices between X and B then we are
done. Otherwise let A be the vertex such that there are n+l vertices between X and A on U.
Except for the endpoints, if V is on U(X,A) then V is a collider on U if and only if U is in S. If
A is a collider on U, then there is a directed edge from A to Y. A is not a collider on the
concatenation of U(X,A) and the edge from A to Y, so that path d-connects X and Y given S
unless A is in S. If A is not a collider on U, then there is an edge between A and Y that is into
A. Hence A is a collider on the concatenation of U(X,A) and the edge from A to Y, so that path
d-connects X and Y given S unless A is not in S.
Similarly, if V is between Yand B, V is in S if and only if V is a collider on U. :.
Lemma 6.3.2: If G' is an inducing path graph, U is a discriminating path for B between X
and Y, and X and Y are d-separated given S, then B is in S if and only if B is not a collider on
U.
Proof. By lemma 6.3.1, for every vertex Von U not equal to B, V is a collider on U if and
only if V is in S. If B is a collider and in S, then U d-connects X and Y given S, contrary to
our assumption. If B is not a collider and not in S, then U d-connects X and Y given S,
contrary to our assumption. Hence B is in S if and only if B is not a collider on U. :.
Theorem 6.3: If the input to the CI algorithm is data over 0 that is faithful to G, the output is
a partially oriented inducing path graph of Gover O.
Proof. The proof is by induction on the number of applications of orientation rules in the
repeat loop of the Causal Inference Algorithm. Let G' be the inducing path graph of G. Let the
object constructed by the algorithm after the nth iteration of the repeat loop be 1rn.
Base Case: Suppose that the only orientation rule that has been applied is that if A *-* B *-* C
in F, but A and C are not adjacent in F, A *-* B *-* C is oriented as A *-> B <-* C if B is not a
member of Sepset(A,C) and as A *-!JL!..-* C if B is a member of Sepset(A,C). Suppose A
*-> B <-* C in 1Il:J, but not in G'. It follows that B is not a member of Sepset(A,C), and either
B is a parent of A or a parent of C in G'. If B is a parent of either A or C in G', then there is an
undirected path between A and C that does not collide at B, and except for the endpoints
contains only B. For any subset S, if that path in G' does not d-connect A and C given S, then
S contains B. It follows that Sepset(A,C) contains B, which is a contradiction.
Suppose that A *-:!:.JL!.-* C in 1Il:J, but the edges between A and B, and Band C collide at B in
G'. It follows that Sepset(A,C) does contain B but every set that d-separates A and C in G'
does not contain B. Hence Sepset(A,C) does not contain B, which is a contradiction.
Induction Case: Suppose 1rn is a partially oriented inducing path graph of G. We will now
show that1rn+l is a partially oriented inducing path graph of G.
Case 1: There is a directed path from A to B and an edge A *-* B in 1rn, so A *-* B is oriented
as A *-> B. By the induction hypothesis if there is an edge R -> S in 1rn, then there is an edge
Proofs 417
R -> Sin G'. It follows that if there is a directed path from A to B in 1rn, then there is a directed
path from A to B in G'. Because G' is acyclic, A *-> BinG'.
Case 2: If B is a collider along <.A,B,C> in 1rn, B is adjacent to D, and A and C are not d-
connected given D, then orientB *-* D asB <-* D. By the induction hypothesis, B is a collider
along <.A,B,C> and D is adjacent to BinG'. If in G' A and C are not d-connected given D by
<.A,B,C> then B has no descendant in {D}. Hence D *-> Bin G'.
Case 3: If U is a definite discriminating path between A and B for M in 1rn, and P and R are
adjacent to M on U, and P-M-R is a triangle, then
if M is in Sepset(A,B) then mark M as a non-collider on subpath P *-!M.!.-* R
else orient P *-* M *-* R as P *-> M <- * R.
By the induction hypothesis, if U is a definite discriminating path for M in 1rn, then it is a

discriminating path for M in G'. By lemma 6.3.2, in G', if U is a discriminating path for M ,
then M is a collider on <P,Q.R> if and only if M is not in Sepset(A,B).
Case 4: If P *-~-* R then the orientation is changed to P *-> M -> R. By the induction
hypothesis, if P *-~-* R in 1rn, then in G' the edge from P to M is into M, but M is not a
collider on P *-> M *-* R. It follows that P *-> M -> RinG'. :.
13.19 Theorem 6.4
Theorem 6.4: If the input to the FC! algorithm is data over 0 that is faithful to G, the output
is a partially oriented inducing path graph of Gover O.
If A ># B in partially oriented inducing path graph 1r, V is in Possible-D-Sep(A,B) in 1r if and

only if V ># A, and there is an undirected path U between A and V in 1r such that for every
subpath <X,y,z> of U either Y is a collider on the subpath, or Y is not a definite non-collider
on U, and X, Y, and Z form a triangle in 1r.
Lemma 6.4.1: If G' is the inducing path graph of directed acyclic graph Gover 0, and P' is
the partially oriented graph constructed in step C) of Fast Causal Inference Algorithm for G
over 0, A and B are in 0, and A is not an ancestor of BinG', then every vertex in D-
SEP(A,B) in G' is in Possible-D-SEP(A,B) in F.
Proof. Suppose that A is not an ancestor of B. IT V is in D-SEP(A,B) in G', then there is an
undirected path U from A to V in which every vertex except the endpoints is a collider. It
follows that in G' for every subpath <X,Y,2> of U, Y is a collider on the subpath. Hence in 1C,
Yis either a collider, or X, Y, and Z form a triangle in 1C and Y is not a definite non-collider. :.
Theorem 6.4: IT the input to the FCl algorithm is data over 0 that is faithful to G, the output
is a partially oriented inducing path graph of Gover O.
Proof. This follows immediately from Theorem 6.3 and lemma 6.4.1. :.
13.20 Theorem 6.5
Theorem 6.5: IT 1C is a partially oriented inducing path graph of directed acyclic graph Gover
0, and there is a directed path U from A to B in 1C, then there is a directed path from A to B in
G.
Lemma 6.5.1: IT 1C is a partially oriented inducing path graph of directed acyclic graph G
over 0, and A -> B in 1C, then there is a directed path from A to BinG.
Proof. Let G' be the inducing path graph of G. If A -> Bin 1C, then A -> Bin G'. IT A -> Bin
G', then in G there is an inducing path from A to B that is not into A. Hence by lemma 6.1.4
there is a directed path from A to BinG. :.
Theorem 6.5: IT 1C is a partially oriented inducing path graph of directed acyclic graph Gover
0, and there is a directed path U from A to B in 1C, then there is a directed path from A to B in
G.
Proof. By lemma 6.5.1, for each edge between R and S in U there is a directed path from R to
Sin G. The concatenation of the directed paths in G contains a subpath that is a directed path
from A toB in G. :.
Proofs 419
13.21 Theorem 6.6
Theorem 6.6: If 1& is the CI partially oriented inducing path graph of directed acyclic graph G
over 0, and there is no semi-directed path from A to Bin 1&, then there is no directed path from
A toB in G.
Lemma 6.6.1: Suppose that G is a directed acyclic graph, and in G there is a sequence of
vertices M starting with A and ending with C, and a set of paths F such that for every pair of
vertices I and J adjacent in M there is exactly one inducing path W over 0 between I and J in F.
'*
Suppose further that if J C then W is into J, and if 1,* A then W is into I, and I and J are
ancestors of either A or C. Then in G there is an inducing path T over 0 between A and C such
that if the path in F between A and its successor in M is into A then U is into A, and if the path
in F between C and its predecessor in M is into C then U is into C.
Proof. Suppose that in G there is a sequence M of vertices in 0 starting with A and ending
with C, and a set of paths F such that for every pair of vertices I and J adjacent in M there is
'*
exactly one inducing path W over 0 between I and J, and if J C then W is into J, and if 1,* A
then W is into A, and I and J are ancestors of either A or C. Let T be the concatenation of the
paths in F. T may not be an acyclic undirected path because it might contain undirected cycles.
Let T be an acyclic undirected subpath of T between A and C.' We will now show that except
for the endpoints, every vertex in 0 on T is a collider, and every collider on T is an ancestor of
A orC.
If V is a vertex in 0 that is on T but that is not equal to A or C, every edge on every path in F is
into V. Hence, every edge on T that contains V is into V because the edges on T are a subset of
the edges on inducing paths in F.
Let Rand S be the endpoints of W. We will now show that every vertex on W is either an
ancestor of A or an ancestor of C. By hypothesis, R is an ancestor of either A or C, and S is an
ancestor of either A or C. Because W is an inducing path over 0, every collider on W is an
ancestor of either R or S, and hence an ancestor of either A or C. Every non-collider on W is
either an ancestor of R or S, or an ancestor of a collider on W. Hence every vertex on W is an
ancestor of either A or C. It follows that every collider on T is an ancestor of A or C, because
the vertices on T are a subset of the vertices on paths in F.
By definition, T is an inducing path between A and Cover O. Suppose the path in F between A
and its successor is into A. If the edge on T with endpoint A is on path in F on which A is an
endpoint, then T is into A because by hypothesis that inducing path is into A. If the edge on T
with endpoint A is on an inducing path over ° in which A is not an endpoint of the path, then T
is into A because A is in 0, and hence a collider on every inducing path for which it is not an
endpoint Similarly, T is into C if in F the path between C and its predecessor is into A . ...
In an inducing path or directed acyclic graph G that contains an undirected path U between X
and Y, the the edge between V and W is substitutable for U(V,W) in U if and only if V and
W are on U, V is between X and W on U, G contains an edge between V and W, V is a collider
on the concatenation of U(X, V) and the edge between V and W if and only if it is a collider on
U, and W is a collider on the concatenation of U(Y,W) and the edge between V and W if and
only if it is a collider on U.
Lemma 6.6.2: If G' is an inducing path graph for directed acyclic graph Gover 0, C is a
descendant of BinG, and U is an undirected path in G' between X and R containing subpath A
*-> B <-> C where A is between X and B, then in G' there is a vertex E on U between X and
A inclusive such that the edge between E and C is substitutable for U(E,C) in U. Furthermore
the concatenation of U(X,E) and the edge between E and C is into C, and if U is into X, then
the concatenation of U(X,E) and the edge between E and C is into X.
Proof. Suppose G' is an inducing path graph for directed acyclic graph Gover 0, C is a
descendant of BinG, and U is an undirected path in G' between X and R containing subpath A
*-> B <-> C where A is between X and B. If E and F are on U, we will say that F is the
successor of E on U if and only if there is an edge between E and F on U and E is between X
and F or E =X. Let Y be the successor of X on U.
First we consider the case where there is no vertex V on U between X and A inclusive such that
the edge from V to C is substitutable for U(V,C) in U, but each vertex on U between Y and A
inclusive is adjacent to C in G'. We will show that there is a directed path from Y to B.
Suppose that U(Y,B) is not a directed path from Y to B. Let E be the vertex on U closest to B
such that U(E,B) is not a directed path from E to B. Let F be the successor of E on U. F is an
ancestor of BinG', not a collider on U, and by assumption F is adjacent to C. The edge
between C and F is not out of C and into F, because G' is acyclic. Hence it is into C. If F = B,
then A <-> B <-> C in G'. It follows that in G there is an inducing path betweeen A and C that
is into A and C, and hence A <-> C in G', and the edge between A and C is substitutable for
the subpath of U between A and C. Suppose then that F"F B. U(F,B) is a directed path from F
to BinG'. Because the edge between F and C is not substitutable for U(F,C) in U it follows
that F is a collider on the concatenation of U(X,F') with the edge between F and C. Hence the
Proofs 421
edge between F and C is into F and into C, and the edge between E and Fan U is into F. It
follows that the edge between E and F is also into E because E is not an ancestor of B, and F is.
Hence G' contains the path E <-> F <-> C. Because F is an ancestor of Bin G', it is an
ancestor of B in G. Since F is an ancestor of B in G, and B is an ancestor of C in G, F is an
ancestor of BinG. It follows by lemma 6.6.1 that there is an inducing path between E and C in
G relative to 0 that is into E and into C. But then in G' the edge between E and C is
substitutable for U(E,C) in U, which is a contradiction.
We have shown that U(Y,B) is a directed path from Y to B. It follows that Y is an ancestor of
Bin G, and because B is an ancestor of C in G, Y is an ancestor of C in G. We have shown
that the edge between Yand its successor on U is out of Y. Hence Y is not a collider on U. By
assumption there is an edge between Y and C in G'. If the edge between Y and C is not
substitutable for U(Y,C) in U, then the edge between Yand C is into Y, and because G' is
acyclic (i.e. there is no directed cycle in G'), the edge between Y and C is also into C. Because
the edge between Yand C is not substitutable for U(Y,C) in U, and the edge between Yand C
is into Y, it follows that the edge between X and Y is into Y. Hence G' contains the path X *->
Y <-> C, and Yis an ancestor of C in G. It follows that there is an inducing path between X
and C in G relative to 0 that is into C, and if U is into X, also into X. But then the edge
between X and C is substitutable for U(X,C) in U, which is a contradiction.
Next we consider the case where there is no vertex Von U between X and A inclusive such that
the edge from V to C is substitutable for U(V,C) in U, but some vertex on U between Y and A
inclusive is not adjacent to C. Let E be the vertex on U closest to C and between X and C that is
not adjacent to C, and let F be the successor of Eon U. E '" A, because by lemma 6.6.1 there is
an inducing path between A and C in G, and hence A is adjacent to C in G'. From the previous
case, it follows that either there is an edge between Von U(E,C) and C that is substitutable for
U(V,C) in U(E,C) or F is an ancestor of BinG'. Suppose fIrst that there is an edge between V
on U(E,C) and C that is substitutable for U(V,C) in U(E,C). E is not adjacent to C, so V", E,
and V lies on U(F,C). If the edge between V and C is substitutable for U(V,C) in U(E,C), then
it is also substitutable for U(V,C) in U, which is a contradiction. Hence F is an ancestor of B in
G'. By the defInition of E, F is adjacent to C in G'. The edge between F and C is not out of C
and into F, because G' is acyclic. The edge between F and C is not out of F and into C because
the edge between F and C is not substitutable for U(F,C) in U(E,C), and U(F,B) is a directed
path from F to B. Hence the edge between F and C is into F and C. If the edge E <- F is on U,
then the F <-> C edge is substitutable for U(F,C) in U. If E *-> F in G' then G' contains the
path E *-> F <-> C, and F is an ancestor of C in G' and hence in G; it follows that there is an
inducing path between E and C relative to 0 in G, and E is adjacent to C in G'. This is a

contradiction.
It follows that for some vertex E on U between X and A inclusive there is an edge from E to C
that is substitutable for U(E,C) in U and is into C. If E = X then there is an inducing path
between X and C that contains the edge on U with X as endpoint. If E *- X then there is some
vertex E *-X on U such that there is an edge between E and C that is substitutable for U(E,C)
in U. In the fIrst case, the inducing path is into X if U is into X and hence the edge between C
and X is into X. In the second case the path consisting of the concatenation of U(X, V) and the
edge between V and C contains the edge on U with X as endpoint, and hence is into X if U is.
Lemma 6.6.4: If 1C is the CI partially oriented inducing path graph of graph Gover 0, and A
*-> Bin 1C, then every inducing path in G between A and B is into B.
Proof. We will prove that each orientation rule in the Causal Inference Algorithm is such that
if the rule orients the edge between A and B as A *-> B, then every inducing path between A
and B over 0 in G is into B. Let G' be the inducing path graph of G.
Case 1: By lemma 6.5.1 any of the rules that orients the edge between A and B as A -> B in 1C
entails that there is a directed path from A to B in G. If there is an inducing path over 0
between A and BinG that is out of B, and there is a directed path from B to A in G. But Gis
not cyclic, so there is no inducing path between A and BinG that is not into B.
Case 2: Suppose the edge between A and B is oriented as A *-> B in order to avoid a cycle in 1C
because there is a directed path from A to B in 1C. By Theorem 6.5 there is a directed path from
A to BinG. If there is an inducing path over 0 between A and BinG that is out of B, then
there is a directed path from B to A in G. But G is not cyclic, so there is no inducing path over
o between A and BinG that is out of B.
Case 3: Suppose that the edge between A and B is oriented as A *-> B because there is a vertex
C such that A and B are adjacent in 1C, B and C are adjacent in 1C, A and C are not adjacent in 1C,
and B is not in Sepset(A,C). It follows that A *-> B <-* C in G'. By the construction of G' it
follows that in G there is an inducing path over 0 between A and B into B, and an inducing
path over 0 between B and C into B. Suppose contrary to the theorem that there is another
inducing path over 0 between A and B in G that is out of B. By lemma 6.1.4, A is a
descendant of Bin G. By lemma 6.6.1 there is an inducing path over 0 between A and C. But
Proofs 423
if there is an inducing path over 0 between A and C in G, then A and C are adjacent in 11:,
contrary to our assumption.
Case 4: Suppose that the edge between A and B is oriented as A *-> B because B is a collider
along <C,B,D> in 11:, B is adjacent to A, and C and D are not d-connected given A. Suppose,
contrary to the theorem, that in G there is an inducing path over 0 between A and B that is out
of B. It follows that A is a descendant of BinG. Because there is an edge between C and B
that is into B in 11:, there is an edge between C and B that is into BinG'. The edge between C
and BinG' d-connects C and B given A and is into B. By lemmas 6.1.1 and 6.1.2 there is a
path in G that d-connects C and B given A that is into B. Similarly, there is a path in G that d-
connectsD andB given A that is intoB. By lemma 3.3.1, C andD are d-connected given A in
G. This is a contradiction.
Case 5: Suppose the edge between A and B in 11: is oriented as A *-> B because in 11: U is a
definite discriminating path for B between X and Y, B is in a triangle on U, and B is not in
Sepset(X,Y). Let A and C be the vertices adjacent to B on U . If U is a definite discriminating
path for B in 11:, then by the induction hypothesis, the corresponding path U' in G' is a
discriminating path for B. In G', X and Y are d-separated given Sepset(X,Y) because by
definition of definite discriminating path they are not adjacent If X and Y are d-separated given
Sepset(X,Y) in G', then by lemma 6.3.1 every collider on 'u' except for B is in
Sepset(X,Y), and every non-collider on U' is not in Sepset(X,y).
Suppose that there is an inducing path over 0 between B and A in G that is out of B. It follows
that there is a directed path from B to A in G and that A <-> B in G'. By definition of
discriminating path it follows that A is a collider on U'or A = X. By lemma 6.3.1 A is in
Sepset(X,y). Hence B is a collider on U'in G', and B has a descendant in Sepset(X,Y) in
G.
If some vertex Z on U is in Sepset(X,Y) then Z is a collider on U. Let Rand T be the vertices

on U' that are adjacent to Z on U'. By the definition of inducing path graph, in G there are
inducing paths over 0 between R and Z, and Z and T, such that each of them is into Z. By
lemmas 6.1.1 and 6.1.2, in G there is a d-connecting path given s\{R,z} between Rand Z,
and a d-connecting path given S\{ Z,T} between Z and T, such that each of them is into Z.
If there are vertices R, Z, and T on U' such that R and Z are adjacent on U, and Z and T are
adjacent on U', and Z is a collider on U', then either Z is in Sepset(X,Y) (if Z ¢' B), or Z has a
descendant in Sepset(X,y) in G (if Z = B). In either case Z has a descendant in Sepset(X,Y)
in G. By the defInition of inducing path graph, in G there are inducing paths over 0 between R
and Z, and Z and T, that are both into Z. By lemmas 6.1.1 and 6.1.2, in G there is a d-
connecting path given Sepset(X,y)\{R,z} between Rand Z, and a d-connecting path given
Sepset(X,y)\{Z,T} between Z and T, that are both into Z. By lemma 3.3.1, there is a path in
G that d-connects X and Y given Sepset(X,y). But this contradicts the assumption that X and
Yare d-separated given Sepset(X,y)' Hence there is no inducing path in G that is out of B. :.
A semi-directed path from A to B in partially oriented inducing path graph 11: is an

undirected path U from A to B in which no edge contains an arrowhead pointing towards A,
that is, there is no arrowhead at A on U, and if X and Yare adjacent on the path, and X is
between A and Y on the path, then there is no arrowhead at the X end of the edge between X
and Y.
Theorem 6.6: If 1Cis the CI partially oriented inducing path graph of directed acyclic graph G
over 0, and there is no semi-directed path from A to B in 11:, then there is no directed path from
AtoBinG.
Proof. Suppose there is a directed path P from A to BinG. Let P' in 11: be the sequence of
vertices in 0 along P in the order in which they occur. P' is an undirected path in 11: because for
each pair of vertices X and Y adjacent in P' for which X is between A and Y or X =A there is
an inducing path over 0 in G that is out of X. P' is a semi-directed path from X to Yin 11:
because by lemma 6.6.4, there is no arrowhead into X on P'. :.
13.22 Theorem 6.7
Theorem 6.7: If 11: is a partially oriented inducing path graph of directed acyclic graph Gover
0, A and B are adjacent in 11:, and there is no undirected path between A and B in 11: except for
the edge between A and B, then in G there is a trek between A and B that contains no variables
in 0 other than A or B.
Proof. Suppose that every trek between A and BinG contains some member of 0 other than
A or B. Because there is an edge between A and B in 11:, there is an inducing path between A
and B in G. Hence, A and Bared-connected given the empty set in G, and there is a trek T
between A and B. Let U be the sequence of observed vertices on T. Each subpath of T between
Proofs 425
variables adjacent in U is an inducing path relative to O. Hence U is an undirected path in 1C that

contains a member of 0 other than A or B. :.
13.23 Theorem 6.8
Theorem 6.8: If 1C is the CI partially oriented inducing path graph of directed acyclic graph G
over 0, and every semi-directed path from A to B contains some member of C in 1C, then every
directed path from A to B in G contains some member of C.
Proof. Suppose that U is a directed path in G from A to B that does not contain a member of
C. Let the sequence of observed variables on U in G be U'. Let X and Y be two adjacent
vertices in U', where X is between A and Y. U(X,Y) is a directed subpath of U that contains
no observed variables except for the endpoints. Hence U(X,Y) is an inducing path between X
and Y given 0 that is out of X. It follows that there is an edge between X and Yin 1C, and by
lemma 6.6.4 the edge between X and Y is not into X. Hence U' is a semi-directed path from A
to B in 1C that does not contain any member of C. :.
13.24 Theorem 6.9
Theorem 6.9: If 1C is a partially oriented inducing path graph of directed acyclic graph Gover
0, and A <-> Bin 1C, then there is a latent common cause of A and Bin G.
Proof. By Theorem 6.6, every inducing path over 0 in G between A and B is into B and into
A. By lemma 6.1.2, there is in Gad-connecting path U between A and B given the empty set
that is into A and into B in G. Because U d-connects A and B given the empty set in G it
contains no colliders, and hence no members of 0 except A and B. Because U contains an edge
into A and an edge into B, U is not a single edge between A and B. Hence there is some vertex
C not in 0 on U that is a common cause of A and B. :.
13.25 Theorem 6.10 (Tetrad Representation Theorem)
Tetrad Representation Tbeorem 6.10: In an acyclic LCF G, there exists an

LJ(T(JJ),T(K ,L),T(J,L),T(J ,K» choke point or an JK(T(J J),T(K,L),T(J,L),T(J ,K» choke
point iff G linearly implies PIJPKL - PILPJK = O.
In a graph G, the lengtb of a path equals the number of vertices in the path minus one. In a
graph G, a path U of length n is an initial segment of path V of length m iff m ~ n, and for 1
:5; i:5; n+l, the i th vertex of V equals the z1h vertex of U. In a graph G, path U of length n is a
final segment of path V of length m, iff m ~ n, and for 1 :5; i :5; n+ 1, the i th vertex of U equals
the (m-n+i)th vertex of V. A path U of length n is a proper initial segment of path V of
length miff U is an initial segment of V and U *- V. A path U of length n is a proper final
segment of path V of length miff U is a final segment of V and U *- V.
The proofs of the following lemma is obvious.
Lemma 6.10.1: In a directed graph G, if R(U,l) is an acyclic path, and X is a vertex on

R(UJ), then there is a unique initial segment of R(U /) from U to X.
Because the proofs refer to many different paths, we will usually designate a directed path by
R(X,Y) where X and Yare the endpoints of the path. When there is a path R(U/) in a proof,
and a vertex X on R(U /), R(U,}() will refer to the unique initial segment of R(U /) from U to
J, and R(X/) will refer to the unique final segment of R(U /) from X to J.
In a directed acyclic graph G, the last point of intersection of directed path R(U J) with
directed path R(VJ) is the last vertex on R(U J) that is also on R(V J). Note that if G is a
directed acyclic graph, the last point of intersection of directed path R(UJ) with directed path
R(VJ) equals the last point of intersection of R(VJ) with R(U /); this is not true of directed
cyclic paths.
Lemma 6.10.2: If G is a directed acyclic graph, for all variables Y and ZinG, if Y *- Z and
R and R' are two intersecting directed paths with sinks Y and Z respectively then there is a trek
between Y and Z that consists of subpaths of R and R'.
Proofs 427
proof. Since R and R' intersect, they have a last point of intersection X. Let the source of the
trek to be constructed be X. R(X,y) and R(XZ) do not intersect anywhere except at X. Since
*
Y Z, one of R(X,y) and R(X,z) is not empty. Hence (R(X,y),R(X,z)} is a trek. :.
In a directed acyclic graph, directed paths R(U/) and R(U,J) contain trek Tiff J(T(J,J) is a
fmal segment of R(U /) and J(T(J,J) is a final segment of R(U,J).
Lemma 6.10.3: In a directed acyclic graph, if R(U,/) and R(U,J) are directed paths that
contain both T(I,J) and T(J,J), then T(J,J) = T(J,J).
Proof. In a directed acyclic graph, there is a unique last point of intersection of R(U /) and
R(UJ), and unique final segments of R and R' whose source is the last point of intersection of
R(U/) and R(U,J). :.
If G is a directed acyclic graph, let Let Pxy be the set of all directed paths in G from X to Y. In
an LCF S, the path form of a product of covariances 'YIJ'YKL is the distributed form of
[L[L
UeU/J RePU/ R'ePUJ
LL(R)L(K)a?;lI L[L
VeUKL R"ePVK R"'ePVL
LL(R")L(R"')~ll
'YI.f'IKL - 'YIL'YlK is in path form iff both terms are in path form.
Henceforth, we will assume that all variances, covariances, products of covariances, and tetrad
differences are in path form unless otherwise stated.
We will adopt the following terminology. Suppose that m is a term in the path form of a
product of covariances rlJrK /. By definition, m is of the form
L(R(U/)L(R(U,J)L(R(V,K»L(R(V,L»a?;~. Let the paths associated with m be the
ordered quadruple <R(U /),R(U ,J),R(V,K),R(V,L». There is a one-to-one correspondence
between terms in the path form of a product of covariances, and such ordered quadruples. We
will consider terms m and m' to be distinct terms if their associated paths are different (Le. the
terms may contain the same number of occurrences of the same edge labels, but in different
orders.) Note that under this criterion of identity of terms, no term appears twice in the path
form of a product of covariances or tetrad difference. Henceforth when we consider sets of
terms appearing in some expression, we will do so under the assumption that each term occurs
at most once in the expression (although distinct terms that have identically equal values may
occur in the expression). We will say that a term m contains a path or trek X if its associated
quadruple contains X.
Lemma 6.10.4: A tetrad difference rlJ'YKL - 'YILrJK is not linearly implied to vanish by an
LCF S if there is a term m in the path form of 'YIJ'YKL such that every term m' in the path form of
'YILrJK contains an edge not in m.
Proof. Suppose that there is a term m in the path form of 'YIJ'YKI such that every term m' in the
path form of 'YIL'YlK contains an edge not in m. Set every variable not in m to be zero. Then
'YILrJK is zero since every term in 'YILrJK contains a variable not in m. Set every variable in m to
be positive. Then every non-zero term in the path form of 'YIJ'YKL is positive, since the e.c2.f of
each non-zero term is positive, and the c.f. of each non-zero term is positive. 'YIJ'YKL is not zero
since every term in it is either 0 or positive, and some are positive. Hence the tetrad difference
is not linearly implied to vanish. :.
Lemma 6.10.5: In an LCF S, if the paths in a term m in the path form of a tetrad difference
have different sources than the paths in a term m', then m contains some variable not in m'.
Proof. Each of the sources of the paths in m and m' are independent random variables, and it
is not the case that all of the paths in m or m' are empty. Let {I,J} be the sources of the paths in
m, and {K,z} be the sources of the paths in m' and suppose that {I,J} *' {K,Z}. Suppose
w.l.g. that J *' K. Since J, K, and Z are independent J does not occur on any paths with source
K or Z. m contains at least one edge X out of J. Since J does not occur on any path with source
K or Z, X does not occur on any path with source K or Z. Hence m contains a variable (the
label of X) that does not occur in m'. :.
In an LCF F, e(S) is equal to S if S is an independent variable, and it is equal to the error

variable into S if S is not an independent variable.
Lemma 6.10.6: In an LCF S, if there exist T(J,J) E T(J,J) and T(K,L) E T(K,L) such that
J(T(J,J) (') K(T(K,L» =0, J(T(J,J) n L(T(K,L» =0, and J(T(J,J) (') L(T(K,L» = 0, then
there exists a term m in 'YI.tYKL such that every term m' in 'YIL rJK contains an edge not in m.
Proof. Let S be the source of T(J,J) and S' be the source of T(K,L). (Note that since J(T(J,J)
does not intersect L(T(K,L», the source of T(J,J) does not equal the source of T(K,L), and
hence e(S) does not equal e(S'). See figure 4.
Let m =L(R(e(S)I»L(R(e(s),J)L(R(e(S').K»L(R(e(S'),L». m is the coefficient of a term in

'YIJ'YKL (the full term also cqntains a factor equal to the product of the variances of the sources of
paths in m.)
Proofs 429
Figure 4
Suppose there is a term m' in YILYJK whose associated paths contain only edges in m. m'
contains the product of the labels of edges in a trek T(I,L). Let the source of T(I,L) be S". If
S" "" Sand S" "" S', then e(S") "" e(S) and e(S") "" e(S'). Since e(S") is an i!ldependent
variable, and the only independent variables in m are e(S) and e(S'), if e(S") "" e(S) and e(S")
"" e(S'), then T(I,L) contains an edge label not in m. Suppose then w.l.g. that S" = S. There is
a path R(S,L) containing edge labels only in m. Since J(T(IJ) n L(T(K,L» =0, and I(T(IJ)
n L(T(K,L» = 0, the only path in m that contains L is L(T(K ,L». Hence R(S,L) intersects
L(T(K,L» at some vertex. The only two paths in m with Source S are I(T(IJ) and J(T(IJ),
and neither of them intersects L(T(K,L». Hence one of them intersects some other paths that in
turn intersects L(T(K,L». The only other path in m that intersects L(T(K,L» is K(T(K,L». So
R(S,L) intersects K(T(K,L». Since the last point of intersection of L(T(K,L» and K(T(K,L»
is S', R(S,L) intersects K(T(K,L» at or before S'. But the only paths with source S in m are
J(T(I,./) and I(T(I,./), and neither of them intersects K(T(K,L» at or before S'. Hence, there
is no path from S to L containing only edge labels in m. Similarly it can be shown that there is
no path from S' to I containing only edge labels in m. Hence m' contains an edge label not in
m. :.
Lemma 6.10.7: In an LCF S, if there exists a T(l,./) e T(l,./) and T(K,L) e T(K,L) such
that I(T(I,./) n K(T(K,L» = 0, and L(T(K,L» n J(T(IJ» = 0, or there exists a T(I,L) e
T(I,L) and T(J,K) e T(J,K) such that I(T(I,L» n K(T(J,K) = 0, and L(T(I,L» n J(T(J,K)
=0, then S does not linearly imply that YrJ'YK1- YrLYJK vanishes.
Proof. Suppose w.l.g. that I(T(I,./) n K(T(K,L» = 0, and L(T(K,L» n J(T(IJ» = 0.
There are four cases: either (i) I(T(I,./) n L(T(K,L» = 0 and J(T(I,./) n K(T(K,L» = 0, or
(li) I(T(IJ) n L(T(K,L» = 0 and J(T(IJ) n K(T(K,L» "" 0, or (iii) I(T(I,./) n L(T(K,L»
"# 0 and J(T(/,J» n K(T(K,L» = 0, or (iv) /(T(/,J» n L(T(K,L» "# 0 and J(T(/,J) n
K(T(K,L» "# 0.
In the first three cases, by lemma 6.lO.6 there exists a term m in rlJ'YKL such that every m' in
rlLrJK contains an edge label not in m.
In the fourth case, let X be the last point of intersection of /(T(/,./) and L(T(K,L», and Y be
the last point of intersection of J(T(/,./) and K(T(K,L». X is not the source of either trek, since
otherwise /(T(/,./) n K(T(K,L»"# 0 or J(T(/,./) n L(T(K,L» "# 0. Similarly, Y is not the
source of either trek. (R(X,I),R(X,L)} is a trek T(/,L) between / and L, by lemma 6.lO.2.
Similarly, (R(Y,./),R(Y,K)} form a trek T(J,K). (See figure 5.)
/ L K J
T(/,J) --+ T(K,L)--+
Figure 5
Now we will show that T(/,L) n T(J,K) = 0. /(T(/ ,L» n J(T(J,K) = 0 since /(T(/ ,L» is a
proper subpath of /(T(/,./) and J(T(J,K» is a proper subpath of J(T(/,./), and the last point of
intersection of /(T(/,./) and J(T(/,./) is the source of T(/,./). /(T(/ ,L» n K(T(J,K» = 0, since
/(T(/,L» is a subpath of /(T(/,./) and K(T(J,K» is a subpath of K(T(K,L», and /(T(/,./) n
K(T(K,L» = 0 by hypothesis. For similar reasons, L(T(/,L» n J(T(J,K» = 0, and L(T(/,L»
n K(T(J,K) = 0. It follows from lemma 6.lO.7 there exists a term m in 'YlL'YJK such that every
m' in 'YlJ'YKL contains an edge label not in m.
Proofs 431
Since there exists a term min 'YlLYlK such that every m' in 'YlfYKL contains an edge not in m, by
lemma 6.10.4 'YlJ'lKL - 'YlLYJK is not linearly implied. :.
A vanishing tetrad difference is a constraint upon the covariances of four pairs of variables:
<l,./>, <K ,L>, <l,L> and <.I,K>. Roughly speaking, a choke point for such a foursome of
variable pairs is a point where all of the treks between / and J intersect all of the treks between
K and L, and all of the treks between J and L intersect all of the treks between J and K. (A more
precise definition is given later.) In this section, we will prove that in an LCF G, the existence
of such a choke point is a necessary condition for the corresponding tetrad difference to vanish
in distributions perfectly represented by G. We will prove this by showing that the existence of
a choke point in G is equivalent to a condition that has already been proved to be a necessary
condition for S to linearly imply a vanishing tetrad difference; namely, the trek intersection
condition described in lemma 6.10.7. Unfortunately, this proof is long and tedious because
there are many different ways in which a choke point can fail to exist, depending upon which
treks are assumed to intersect and which treks are assumed not to intersect. In each case we
show that the non-existence of a choke point implies the violation of the necessary condition
described in lemma 6.10.7.
Two strategies are employed in the proofs. The fIrst is to show that the assumptions about
which treks intersect and don't intersect lead to contradictions. The second is to show that it is
possible to construct a pair of treks T'(J,J) and T'(K,L) such that J(T'(J,J) and K(T'(K,L»
don't intersect, and J(T'(J,./» and L(T'(K,L» don't intersect, or to construct a pair of treks
T'(J,L) and T'(J,K) such that J(T'(J ,L» and K(T'(J,K» don't intersect, and J(T'(J ,K» and
L(T(J,L» don't intersect. In either case, by lemma 6.10.7, it follows that PIJPKL- PlLPJK is
not linearly implied by G.
In general, when constructing a trek T(J,J) we will speak as if it suffices to show how to
construct a pair of (acyclic) directed paths Rand R' from a common source S to sinks / and J
respectively, without showing that the pair of directed paths constructed do not intersect. This
is because even if Rand R' do not form a trek because they intersect each other at some vertex
other than S, we have shown in lemma 6.10.2 that directed subpaths of Rand R' do form a
trek, and the existence of the directed subpaths of R and R' is enough for our purposes. We are
generally interested in showing that particular pairs of trek branches fail to intersect. If Rl and
R2 fail to intersect, then directed subpaths of Rl and R2 also fail to intersect. Hence, if the goal
is to show that trek branches T and T fail to intersect, it suffices to show that Rl and R2 fail to
intersect, even if T and T are actually equal to directed subpaths of Rl and R2 respectively.
Let S be a set of vertices, and RK(S) be the set of all directed paths with sink K and a source in
S. Let R(S,l) be a directed path from S in S to I. Let Xn be the nth vertex on R(S,l) such that
some directed path in RK(S) intersects it. Let the set of sources of directed paths in RK(S)
whose first point of intersection with R(S,l) is Xn be So. Let the last vertex in R(S,l) that is the
first intersection of some directed path in RK(S) with R(S,l) be Xmax. Note that Xmax. is not
necessarily the last point of intersection of some directed path in RK(S) with R(S,l); it is
merely the last of the first points of intersection. See figure 6.
S1 = {S I'S V
S 2 = (S 3)
Figure 6
Lemma 6.10.8: In a directed acyclic graph G, if R(M,l) is a directed path, and RK(S) is the
set of all directed paths to K from a given set of sources S, and there does not exist a vertex Z
such that all of the directed paths in RK(S) intersect R(M,l) at Z, then there is a pair of directed
paths, Rand R', with the following properties: M is the source of R, R' has a source in S,
Proofs 433
either R has sink I and R' has sink K or R has sink K and R' has sink I, and R does not
intersect R',
Proof. If there is a path R' in RK(S) that does not intersect R(M,I) the proof is done. Assume
then that every path in RK(S) intersects R(M,I). Let S" be the source of a path in Smax (the
set of all sources of paths in RK(S) whose fIrst intersection with R(M,I) isXmax.) The proof is
by induction on the number of distinct vertices in which the paths in RK(S) intersect R(M,I).
Base Case: Suppose the antecedent in the statement of the lemma is true. The paths in RK(S)
intersect R(M,I) in two distinct vertices. There is a path R(S',K) that does not intersect R(M,I)
at X2 (= Xmax), since otherwise all paths in RK(S) would intersect X2, contrary to our
hypothesis. In addition, R(S',K) does not intersect R(M,I) at any vertex prior to Xl, since
otherwise the paths in RK(S) would intersect R(MJ) at more than two distinct vertices,
contrary to our hypothesis. Similarly, there is a path R(S".K) that intersects R(M,I) only at
X2.
Let R(XI,K) be a fInal segment of R(S',K) and R(S",X2) an initial segment of R(S",K).
1. R(XI,K) does not intersect R(S",X2). See fIgure 7. Let R(M,XI) be an initial segment of
R(M,l), R(X2,1) be a fInal segment of R(MJ), R =R(M,XI) &R(XI,K) and R' = R(S",
X2)&R(X2,l). R and R' do not intersect for the following reasons.
S' S'
I
R(M.I ) ~ R(S', K ) .... R(S", K ).
R
'. I
Figure 7
R(M,xl) does not intersect R(S",x2). R(S",x2) is a subpath of R(S",K), which, by

hypothesis intersects R(M,l) only at X2. Since X2 occurs after Xl on R(M J), X2 does not
occur on R(M,xl). R(M,xl) does not intersect R(X2J). R(M,xI) and R(X2J) are both
subpaths of R(M,l), G is acyclic, and by hypothesis X I occurs before X2. R(X 1,K) does not
intersect R(S" ,(2) by hypothesis. R(X I,K) does not intersect R(X2,l). R(X 1,K) is a subpath
of R(S',K) and R(X2,l) is a subpath of R(M,l); by hypothesis R(S',K) intersects R(M,J) only
at X I, which does not occur on R(X2,l).
2. R(X 1,K) does intersect R(S",x 2) at Y. See figure 8. Let R(S" ,f) be an initial segment of
R(S",K), R(Y,K) be a final segment of R(S',K), R = R(M J) and R' = R(S",f)&R(Y,K). R
and R' do not intersect for the following reasons.
Proofs 435
S' S'
\/y
XI
S"
R(M /) --.- R(S', K ) - . R(S" ,K ) " ,
Figure 8
First we will show that R(M J) does not intersect R(S",y). Y;/: X2 since R(X t.K) intersect
R(MJ) only atXI. Also, G is acyclic, Yisprior toX2 on R(S",K), andX2 is the frrstpoint of
intersection of R(S",K) with R(M,!). Next we will show that R(M,l) does not intersect
R(Y,K). Y is on R(S",K) which does not contain X I; hence Y is not equal to X I. It follows
that R(Y,K) does not contain X \, since Y occurs after XI on R(S',K), and R(S',K). By
hypothesis R'(M,K) intersects R(MJ) only at XI. so that R(Y,K) does not intersect R(MJ) at
all.
Induction Case: Assume that the antecedent is true, and that the theorem is true for all m < n. If
there is a path in RK (S) that does not intersect R(M J), the proof is done. Suppose then that
every path in RK(S) intersects R(M J) and that the set of paths in RK(S) intersects R(MJ) at
exactly n distinct vertices. Let R(XmaxJ) be a final segment of R(MJ). Since not every path in
RK(S) intersects R(M J) at X max , there is a point of intersection prior to Xmax on R(M,/).
Hence the number of distinct points of intersection of the paths in RK(S) with R(XmaxJ) is
less than n. By the induction hypothesis, there is a path R 1 with source Xmax and a path R l'
with a source S' in the sources of RK(S), such that one of Rl and Rl' has a sink /, the other
has sink K, and Rl and Rt' do not intersect. Suppose w.l.g. that Rl has sink / and Rt' has
sink K. Since R l' does not contain X max , its fIrst point of intersection with R(M,J) is some
vertex X r , which occurs on R(M,I) before Xmax (by defInition of Xmax.) Let Rl'(Xr,K) be a
final segment of Rl', R(S",K) be a path in RK(S) whose first point of intersection with
R(M,J) is X max , and R(S", Xmax) an initial segment of R(S",K). There are two cases.
1. Assume that R(X,K) does not intersect R(S", Xmax). Let R =R(M,xr)&R 1'(Xr,K) and R'
= R(S", X max )&Rl. Rand R' do not intersect for reasons analogous to those in case I of the
base case (with Xr substituted for XI, and Xmax substituted for Xz; see figure 9.)
2. Assume that Rl '(Xr,K) does intersect R(S",xmax), and the last point of intersection is Y. Y
;c Xmax because it lies on Rl'(Xr,K) and Rt'(Xr,K) does not contain X max . Let Rl'(Y,K) be a
fInal segment of Rl'(Xr,K). There are two cases.
a. Assume that Rl '(Y,K) intersects R(M ,xmax) and the fIrst point of intersection is Z. Let
R(S",Y) be an initial segment of R(S",xmax), R(Y,Z) an initial segment of Rt'(Y,K), and
R(M,2) an initial segment of R(M,I). Z;c Xmax because Rl'(Y,K) does not intersect X max .
(See fIgure 9).
We will now prove Z is not after X max . Consider the path R(S",Y)&R(Y,2). R(S",Y) does
not intersect R(M,J) because Y occurs before X max , R(S",Y) is an initial segment of R(S",K)
and the fIrst point of intersection of R(M,I) and R(S",K) is X max . The fIrst point of
intersection of R(Y,2) and R(M,J) is Z, since R(Y,2) is an initial segment of Rl'(Y,K) and Z is
the fIrst point of intersection of Rl'(Y,K) and R(M,J). Hence the fIrst point of intersection of
R(S",Y)&R(Y,Z) with R(M,J) is Z. R(S",Y)&R(Y,2) is an initial segment of a path from S"
to K that is in RK(S), It follows that there is a path in RK(S) whose fIrst point of intersection
with R(M,I) is Z. If Z is after X max , then there is a path in RK(S) whose fIrst point of
intersection with R(M,J) is after X max, contrary to the defmition of X max .
Proofs 437
S· s·
R(M.!) ----.- R(S·. K)~ R(S",K) +
Figure 9
Let R = R(M ,Z)&R J '(Z,K) and R' = R(S",xmax)&R I. R(M ,Z) does not intersect
R(S",xmax) since R(S",xmax) is an initial segment of R(S",K) and R(M,Z) is an initial
segment of R(M /) and the first point of intersection of R(M /) and R(S",K) is X max . R(M,2)
does not intersect Rl (which has source Xmax) since Z occurs before Xmax and the directed
graph is acyclic. R J '(Z,K) does not intersect R 1 since R !'(ZX) is a subpath of R!, that does
not intersect RI by construction. RJ '(Z,K) does not intersect R(S",xmax) since RJ '(Z,K) is a
final segment of RI'(XrX), Z is after Y, and Y is the last point of intersection of R!'(XrX)
and R(S",xmax).
b. Assume that R 1'(Y,K) does not intersect R (M ,xmax). (This is similar to part 2 of the Base
case, with Xmax substituted for X2. See figure 8.) Let R' = R(S",y)&RJ'(Y,K) and R =
R(M,xmax)&RI. We have already shown that R(S",y) does not intersect R(M,!) and
R(M,xmax) is an initial segment of R(M/). R(S",y) does not intersectRI because Y is before
X max , and the directed graph is acyclic. Rl'(Y,K) does not intersect R(M,xmax) by
hypothesis, and Rl'(Y,K) does not intersect Rl because it is a subpath of R( that does not
intersect R 1 by construction....
In an directed acyclic graph G, if all L(T(K,L» and allJ(T(/,J) intersect at a vertex Q, then Q
is an U(T(/,J),T(K,L» choke point. Similarly, if all L(T(K,L» and all J(T(/,J) intersect at a
vertex Q, and all L(T(I,L» and all J(T(J,K» also intersect at Q, then Q is a
LJ(T(/,J),T(K,L),T(/,L),T(J,K» choke point.
Lemma 6.10.9: In a directed acyclic graph G, if there is no U(T(/,J),T(K,L» choke point,

then either there is a trek T (K,L) such that there is no vertex V' that occurs in the intersection
of all J(T(/,.I» with L(T(K ,L», or there is a trek T(/,J) such that there is no vertex V' that
occurs in the intersection of all L(T(K,L» with J(T(/,J)
Proof. Suppose that the lemma is false. Then, for each trek T(K,L) there is a non-empty set
of points P(T(K,L» such that every point in P(T(K,L» is in the intersection of all J(T(/,J)
with L(T'(K,L». Similarly, for each trek T'(/,J) there is a non-empty set of points P(T(/,.I)
such that every point in P(T(/,J) is in the intersection of all L(T(K,L» with J(T'(/,J). Every
J(T(/,J) contains every vertex in UP(T(K,L» (since every J(T(/,.I) intersects each
T(K,L)eT(K,L)
L(T(K,L» at some vertex in P(T(K,L»), and every vertex in UP(T(K,L» occurs on
T(K,L)eT(K,L)
some trek L(T(K,L». Similarly, every L(T(K,L» contains every vertex in UP(T(I,J».
T(l ,J)eT(l ,J)
Furthermore, for every vertex in UP(T(K,L» there is some L(T'(K,L» that does not
T(K,L)eT(K,L)
contain it (else all J(T(I,J» and all L(T(K,L» intersect at a single vertex), and some
L(T'(K,L» that does contain it. Similarly, for every vertex in UP(T(I,J» there is some
T(l ,J)eT(l,J)
J(T(/,J) that does not contain it and someJ(T'(/,J) that does contain it.
Since every vertex in UP(T(K,L» occurs on every J(T(/,J), they can be ordered by the
T(K,L)eT(K,L)
order of their occurrence on someJ(T(/,J); similarly every vertex in UP(T(I,J» can be
T(l ,J)eT(l,J)
ordered. By the antecedent of the lemma, there are at least two vertices in each of
UP(T(K,L» and UP(T(I,J».
T(K,L)eT(K,L) T(l,J)eT(l,J)
Proofs 439
L(T'(K,L)) - - . J(T'(I.J)) -+ J(TU(I,J)) - - .
Figure 10
See figure 10. Let A be theJirst vertex in UP(T(I,J» and B be the first vertex in
T(l,J)eT(l,J) ,
UP(T(K,L». Suppose w.l.g. that A is before B. There exists an L(T(K,L» that
T(K,L)eT(K,L)
contains A (since every L(T(K,L» contains A), that does not contain B, but that does contain
some vertex C (*B) in UP(T(K,L».
T(K,L)eT(K,L)
There is also a J(T(/,.!» that contains A. Let S be the source of T(/,J), R(S,A) an initial
segment of J(T(I,J», R(A,C) a segment of L(T(K,L», and R(C,J) a final segment of
J(T(/,J». Let J(T"(/,f) = R(S,A)&R(A,C)&R(C,.!), and I(T"(/,.!» = I(T(/,f). J(T"(/,J»
does not contain B for the following reasons. R(S,A) does not contain B because A occurs
before B. R(A,C) does not contain B because it is a segment of L(T(K,L» which does not
contain B. R(C,.!) does not contain B because it is a segment of J(T(/,.!), and since B is the
first vertex in UP(T(K,L» it occurs before C on J(T(/,.!).
T(K,L)eT(K,L)
But this contradicts the fact that for every T(/,.!), J(T(/,.!) contains B. :.
Lemma 6.10.10. In a directed acyclic graph G, if there is no lK(T(/ j),T(K,L)) choke point,
then either there is a trek T(K,L) such that there is no vertex V' that occurs in the intersection
of all /(T(/ j)) with K(T(K ,L)), or there is a trek T(/ j) such that there is no vertex V' that
occurs in the intersection of all K(T(KL)) with /(T(/j))
Proof. The proof of lemma 6.10.10 is the same as that of lemma 6.10.9 with /, J, K, L
permuted ....
Lemma 6.10.11: In an acyclic LCF G, if there is a trek T(K,L) such that there is no vertex
V' that occurs in the intersection of all J(T(/ j)) with L(T(K,L)), then either there are treks
T"(/,J) and T"(K,L) such that J(T"(/,J)) does not intersect L(T"(K,L)) or PUPKI- P1LPJK is
not linearly implied by G.
Proof. Let S be the source of T(K,L), and S be the set of sources of treks between / and J.
By lemma 6.10.8 it is possible to construct a pair of paths Rand R', with sources S and S' (in
S), and sinks J and L, such that R and R' do not intersect. There are two cases.
1. If R is a path from S to L, and R' is a path from S' to J, then the following treks can be
formed from subpaths of Rand R'. (See figure 11.) J(T"(/ j)) =R', /(T"(/ ,1)) = /(T(/ ,1)),
K(T"(K,L)) = K(T(K,L)), and L(T"(K,L)) = R. By construction R does not intersect R';
hence J(T"(/j)) does not intersect L(T'(K,L)).
S' S'
KI J L K
T'(J,J) .... T '(K ,L ) ---+ T"(I,J ) . . . . T "( K ,L ) ......
Figure II
Proofs 441
2. If R is a path from S to I, and R' is a path from S' to L, there are two cases.
a. K(T(KL» intersects I(T(J,J), and the first vertex of intersection is Y. Let R(S,y) be an
initial segment of K(T'(KL», R(Y,K) a final segment of K(T(KL», R(S',Y) an initial
segment of J(T'(J ,]», R(Y J) a final segment of J(T'(I,]», I(TH(I,I» = R, I(TH(J,]» =
R(S,Y)&R(Y,J), K(T"(KL» =R(S',y)&R(Y,K), and L(T"(KL» = R'. (See figure 12.) By
construction, I(T"(J,J) and L(T"(KL» do not intersect.
K J K
I(T'(I,J )) --.. K (T'( K ,L )) + T"(I,J) . . . T"( K , L ) ~
I
R' --.. R ...
Figure 12
b. If K(T(KL» does not intersect I(T(I,]», the following treks can be formed. (See figure
13.) J(T'(I,L» = J(T(I,I», L(T(J,L» = R', I(T(I,K» =R, and K(T'(I,K» = K(T(K,L».
By hypothesis, K(T(I,K» does not intersect I(T(JL». By construction, L(T(JL» does not
intersectJ(T(I,K). Hence by lemma 6.10.7, PIJPKL - PILPJK is not linearly implied by G. :.
L j K
T(I.}) ---+- T(K.L) - . T'(I.L) ---+- T(}.K) . . . .
Figure 13
Lemma 6.10.12: In an acyclic LCF G, if there is a trek T(J,1) such that there is no vertex V'
that occurs in the intersection of all L(T(K,L» with J(T'(J,J), then either there are treks
T"(J,1) and T"(K,L) such thatJ(T'(J,J) does not intersect L(T'(K,L» or PIJPKL - P1LPJK =
ois not linearly implied by G.
Lemma 6.10.13: In an acyclic LCF G, if there is a trek T(J,1) such that there is no vertex V'
that occurs in the intersection of all K(T(K,L» with J(T'(J,1), then either there are treks
T"(J,J) and t"(K,L) such that J(T"(J,1) does not intersect K(T'(K ,L» or PIJPKL - P1LPJK =
Lemma 6.10.14: In an acyclic LCF G, if there is a trek T(K,L) such that there is no vertex
V' that occurs in the intersection of all J(T(J,1) with K(T(K,L», then either there are treks
T"(J,1) and T"(K,L) such that J(T"(J,1) does not intersect K(T"(K ,L» or PIJPKL - PILPJK =
The proofs of lemmas 6.10.12, 6.10.13, and 6.10.14 can all be obtained from the proof of
lemma 6.10.11 by permuting J,J, K, and 1.
Proofs 443
Lemma 6.10.15: In an acyclic LCF G, if there is no LJ(T(/,J),T(K,L» choke point, and

there is no /K(T(/,J),T(K,L» choke vertex, then there exist treks T(/,J) ,T(K,L), T"(/,J),
and T"(K,L) such that /(T(/,J» does not intersect K(T(K,L» and J(T"(/,J» does not
intersect L(T'(K,L», or PIJPKL - P1LPJK = 0 is not linearly implied by G.
Proof. This follows directly from lemmas 6.10.9 through 6.10.14. :.
Lemma 6.10.16: In an acyclic LCF G, if there is no U(T(/,J),T(K,L» choke point, and

there is no /K(T(l,J),T(K,L» choke point, then PIJPKL - P1LPJK = 0 is not linearly implied by
G.
Proof. Assume that there is no LJ(T(I,J),T(K,L» choke point, and there is no
/K(T(/,J),T(K,L» choke point. By lemma 6.10.15 either PUPKL - PILPJK = 0 is not linearly
implied by G or there exist treks T(/,J), T(K,L), T"(/,J), and T"(K,L) such that /(T(/,J»
does not intersect K(T(K,L» and J(T"(/,J» does not intersect L(T"(K,L». If PIJPKL -
P1LPJK = 0 is not linearly implied by G, the proof is done. Assume then that there exist treks
T(/,J), T(K,L), T"(/,J), and T'(K,L) such that /(T(/,J» does not intersect K(T(~,L» and
J(T"(/,J) does not intersectL(T"(K,L». There are three cases.
1. Suppose for all T(/,J),'J(T(/,J) intersects L(T(K,L» at each vertex in a non-empty set of
vertices pI, and all L(T(K,L» intersects J(T(/,J) at each vertex in a non-empty set of vertices
P. Hence, all L(T(K,L» contain every vertex in P and all J(T(/,J) contain every vertex in P'.
Since there is no U(T(/,J),T(K,L» choke point, there is no vertex Z such that for all T(/,!) and
all T(K,L), Z occurs in the intersection of L(T(/,J) and J(T(/,J». Hence P and pI do not
intersect.
Let A be the first vertex in P, and B be the first vertex in P'. Suppose w.l.g. that A occurs
before B. Let S'(/,J) be the source of T(/,J), S'(K,L) the source of T(K,L) and S"(/,J) the
source of T"(/,J), and S"(K,L) the source of T"(K,L). L(T"(K,L» contains A (since all
L(T(K,L» contain A), and J(T'(/,J) contains B, (since all J(T(I,J) contain B.) There are two
cases.
a. Suppose K(T"(K,L» does not intersect /(T"(/,J». Then, since K(T"(K,L» does not
intersect /(T'(/,J) and J(T'(K,L» does not intersect L(T'(K,L», by lemma 6.10.7, PIJPKL -
P1LPJK = 0 is not linearly implied by G.
b. Suppose K(T"(K,L» does intersect /(T"(I,J» at a vertex X. (See figure 14.) Let
R(S"(/,J),x) be an initial segment of /(T'(/,J), R(X,K) a final segment of L(T"(K,L». Let
R(S"(/,J),B) be an initial segment of J(T'(/,J) and R(B,L) be a final segment of L(T(K,L».
Form the trek K(T'''(K,L» = R(S"(l,J),X)&R(X,K), and L(T"'(K,L)) =

R(S"(J,J),B)&R(B,L). R(S"(/,J),B) does not contain A, since it is a subpath of J(T"(/,J))
which does not intersect L(t"(K,L)), which does contain A. R(B,L) does not contain A, since
A occurs before B. Hence L(T"(K,L» does not contain A; but this is a contradiction.
R(S"(I.J ),X)
S"(I,1)
J L 1 L
T'(I ,J) ~ T'(K,L) ~
+-
T'''(K,L) . . . .
T"(I.J ) T"(K,L) +-
Figurel4
2. All L(T(K,L)) intersect J(T'(J,J)) , but not at a single vertex, or all J(T(J,J)) intersect
L(T(K,L)) but not at a single vertex. Assume w.l.g. that the latter is the case. Let S' be the
source of T(J,J) and S be the source of T(K ,L). Let S be the set of sources of treks between J
and J . By lemma 6.10.8, it is possible to form two paths R(S",L) and R(S,J) or R(S",J) and
R(S,L) that don't intersect, where S" is in S. Assume that it is possible to form the paths
R(S",L) and R(S,J) that don't intersect. (If the paths that don't intersect are R(S",J) and
R(S,L) the proof is the same except that the indices are permuted.) Let T'(J,J) be a trek with
Proofs 445
source S" (See figure 15.) Let the first point of intersection of J(T'(J,f) with J(T(J,f) be M.
a. Assume that J(T"(J,J» does not intersect K(T(K,L» before it intersects J(T'(J,f) at M.
(See figure 15.) Let R(MJ) be a final segment of J(T(J,f) and R(S",M) be an initial segment
of J(T"(J,J). Let J(T'(J,L» =R(S",M)&R(M,l), L(T'(J,L» =R(S",L), J(T'(J,K» =R(S,f)
and K(T'(J,K» = K(T'(K ,L». R(S",M) and R(M,J) do not intersect K(T'(K,L» by
hypothesis. By lemma 6.10.7 PIJPKL - P1LPJK =0 is not linearly implied by G.
S'
S
L J K
L
T'(U ) • T'(K ,L ) - . T"( /,l) " , T (J,K) . . .

T (/ ,L ) .
Figure IS
b. Assume that J(t"(J,f) does intersect K(T'(K,L» before it intersects J(T(J,f), and the first
point of intersection is Q. Let R(Q,K) be a final segment of K(T'(K,L» and R(S",Q) be an
initial segment of J(t"(J,f). Let Y be the first point of intersection of R(S,f) and J(T(J,f), and
R(S',y) be an initial segment of J(T(J ,f). There are two cases.
1. Assume that R(S",L) intersects R(S',y) and the first point of intersection is Z. Let R(S',z)
be an initial segment of J(T'(J,J», R(Z,L) be a final segment of R(S",L), L(T'(J ,L» =
R(S',Z)&R(Z,L), J(T'(J,L» = J(T'(J,J», J(T'(J,K» = R(S,f), and K(T(J,K» = K(T'(K,L».
(See figure 16.)
T'(I,J ) T '(J,L ) . . T'(} ,K ) +

J(T"( I ,} )) + R(S" ,L ) +
Figure 16
K(T(J,K) does not intersect J(T(J,L» by hypothesis. J(T(J,K) does not intersect L(T'(J,L»
for the following reasons. R(S',z) does not intersect R(S,J) because R(S',z) is a subpath of
J(T'(I,J), Z is before Y, and the first point of intersection of J(T'(J,J) and R(S,]) is Y.
R(Z,L) does not intersect R(S,]) because it is a subpath of R(S",L) which does not intersect
R(S,]) by construction. By lemma 6.10.7. PIJPKL - PILPJK = 0 is not linearly implied by G.
2. Assume that R(S",L) does not intersect R(S',Y). Let L(T"(K,L» = R(S",L), K(T"(K,L»
= R(S",Q)&R(Q,K), J(T"'(J,]» = J(T'(J,]), and J(T"'(J,]) = R(S',Y)&R(Y,J). (See figure
17.) 'K(T"(K,L» does not intersect J(T'''(I,]) for the following reasons. R(S",Q) does not
intersect J(T(J,]) since R(S",Q) is an initial segment of J(T'(J,]), and Q occurs before the
first point of intersection of J(T"(I,J» and J(T'(J,]». R(Q,K) does not intersect J(T(J,]»
because it is a final segment of K(T(K,L», which does not intersect J(T(J,]) by hypothesis.
L(T'(K,L» does not intersect J(T"(I,]) for the following reasons. R(S',y) does not intersect
R(S",L) by hypothesis, and R(Y,]) is a subpath of R(S,]) which does not intersect R(S",L)
by construction. By lemma 6.10.7 PIJPKL - PILPJK = 0 is not linearly implied by G.
Proofs 447
T'(I,J) • R(S,J) ~ T"(K,l..) . T"'(I,J) +

K(T'(K ,l..» ~ R(S", l..) +
I (T"(I ,J» +
Figure 17
3. Either there is an L(T'(K,L)) that does not intersect J(T(l,./)) or there is a J(T'(/,./)) that
does not intersect L(T'(K,L)). Assume w.l.g. that J(T"(/,J)) with source S"(/,1) does not
intersect L(T(K,L)). There are two cases.
a. Suppose that /(T'(/,./)) does not intersect K(T(K,L)) before it intersects /(T(/,./)) at vertex
X. See figure 18.
S'(I ,J) S'(K,L) S'(U) S'(K .L)
S"(I.J ) S"(I.J)
T"(/,1) -~• • Y'(K.L) ---1~~ Y'''(I.1 )
T'(/,J) •
Figure 18
Let R(X,/) be a final segment of /(T'(/,J)) and R(S"(I,J),x) be an initial segment of

/(T'(/..I)) . The trek T"(/..I) can be formed as follows. J(T"(/,J)) = J(T"(I,J)) and /(t"'(I,J))
= R(S"(/,J),x)&R(XJ). R(S"(/,J),x) does not intersect K(T(K,L)) because by hypothesis
X occurs on /(T"(/,J)) before it intersects K(T(K,L)). R(X,I) does not intersect K(T(K,L))
because it is a subpath of /(T(/..I)) which does not intersect K(T(K,L)) by hypothesis. Hence
/(T'''(/,1)) does not intersect K (T'(K ,L )). J(T"'(I,))) = J(T"(/ ,))) does not intersect
L(T(K,L)) by hypothesis. By lemma 6.10.7, PlJPKL - PILPJK =0 is not linearly imp,lied by
G.
b. Suppose /(T'(/,J)) intersects K(T(I,J)) at Y before it intersects /(T(/,J)) at X. Let Z be the

fIrst point of intersection of J(T(/..I)) and L(T(K,L)). (If no such vertex exists, then J(T(/,J))
and L(T(K,L)) do not intersect, /(T(/,J)) and K(T(K,L)) do not intersect by hypothesis, and
by lemma 6.10.7 PIJPKL - PILPJK =0 is not linearly implied by G .). LetR(S'(/,J),Z) be an
initial segment of /(T(/ ..I)), and R(Z,L) be a fInal segment of L(T(K,L)). There are two cases.
I. Suppose that J(T'(/..I)) does not intersect R(S'(/..1),2). See fIgure 19.
Proofs 449
T"(/,J) - . . T'(K,L)
.... T'(JX) _...... T'(/ ,L) _ .....~~
T (/ ,J ) ..
Figure 19
Let R(Y,K) be a final segment of KCT(K,L)) and R(S"(J,J),Y) be an initial segment of

J(T"(J,J)). Let J(T(J,K)) =J(T"(J,J)), K(T'(J,K)) = R(S"(J,J),Y)&R(Y,K), J(T'(J,L)) =
J(T'(J,.!)), LCT'CJ,L)) =R(S'(J,.!),Z) & R(Z,L).JCT(J,L)) and K(T(J,K)) do not intersect for
the following reasons. J(TCJ,L)) does not intersect R(S"CJ,J),Y) because by hypothesis,
J(T'(J,.!)) intersects K(T(K,L)) at Y before it intersects J(T(I,.!)). I(T(I,L)) does not intersect
R(Y,K) because J(T(J,L)) = J(T(J,J)) and R(Y,K) is a subpath of K(T(K,L)), which does
not intersect J(T(J,.!)) by hypothesis. J(TCJ,K)) does not intersect L(T(J,L)) for the following
reasons. J(T(J,K)) does not intersect R(S'CJ,.!),Z) because J(T'(J,K)) =J(T"(J,.!)), which
does not intersect R(S'(J "!),7.). by hypothesis. J(T(J,K)) does not intersect R(Z,L) because
J(T(JX)) = JCT"CJ,J)) which does not intersect L(TCK,L)) (which contains R(Z,L)) by
hypothesis. By lemma 6.10.7, PIJPKL - PILPJK = 0 is not linearly implied by G.
2. Suppose that J(T"(J,.!» does intersect R(S'(J,.I),z) and the fust point of intersection is M.
(See figure 20.) M '" Z because J(T'(J,.I) does not intersect L(T(K,L» which contains Z. Let
R(S'(J,.I),M) be an initial segment of J(T'(J,.I) and R(M,.I) be a final segment of J(T'(J ,.I).
Let J(T'''(J,]» = J(T'(J,J» and J(T'''(J,J» = R(S'(J,J),M)&R(M,J). J(T"'(J,J» does not
intersect K(T(K,L» by hypothesis. J(T"'(J,.I) does not intersect L(T(K,L» for the following
reasons. R(S'(J,.I),M) does not intersect L(T'(K,L» since M is before Z on J(T'(J,.I), and the
first point of intersection of J(T'(J,J» with L(T'(K,L» is Z. R(M,J) does not intersect
L(T'(K,L» because it is a subpath of J(T"(J,J» which does not intersect L(T'(K,L» by
hypothesis. By lemma 6.10.7, PlJPKL - PILPJK = 0 is not linearly implied by G.
S'(I.J )
S'(I,J)
R(S'(I,J) ,M) R(S'(I,J),M)
M M
S"(I.J)
S"(I,J)
T"(I.1) ---+ T'(K,L) - - . . T"'(I.)) ~
T'(I.J) ~
Figure 2 0
Lemma 6.10.17: In an acyclic LCF G, if there is no U(T(J,L),T(J,K» choke point, and

there is no IK(T(J,L),T(J,K) choke point, then PIJPKL - PILPJK = 0 is not linearly implied by
G.
Proof. The proof is the same as that of lemma 6.10.16, with the indices permuted. :.
Proofs 451
Lemma 6.10.18: In an acyclic LCF G, if G linearly implies PIJPKL - PILPJK = 0, then either
there is an U(T(/,J),T(K,L)) choke point and an U(T(/,L),T(J,K)) choke point, or there is an
/K(T(/,J),T(K,L)) choke point and an /K(T(/,L),T(J,K)) choke point.
Proof. Assume that G linearly implies PIJPKL - PILPJK = O. By lemmas 6.10.16 and 6.10.17,
if G linearly implies PIJPKL - PILPJK = 0 then either there is an U(T(/,J),T(K,L)) choke point
or an /K(T(/ ,J),T(K,L)) choke point, and there is either an U(T(/,L),T(J,K)) choke point or an
/K(T(I,L),T(J,K)) choke point. If there is an LJ(T(I,J),T(K ,L)) choke point and an
LJ(T(/ ,L),T(J,K)) choke point, or there is an /K(T(I,J),T(K,L)) choke point and an
/K(T(I,L),T(J,K)) choke point, the proof is done. Suppose then that there is an
LJ(T(I,J),T(K,L)) choke point and an /K(T(I,L),T(J,K)) choke point, but no
/K(T(/,J),T(K,L)) choke point and no U(T(/,L),T(J,K)) choke point. (The case where there is
an LJ(T(I,L),T(J,K)) choke point and an /K(T(I,J),T(K,L)) choke point, but no
U(T(/,J),T(K,L)) choke point and no /K(T(/,L),T(J,K)) choke point is essentially the same,
with the indices permuted.)
By lemmas 6.10.9 through 6.10.14, if there is no U(T(/,L),T(J,K)) choke point, then either
there is a pair of treks T(/,L) and T(J,K) such that L(T'(/,L)) does not intersect J(T(J,K)) or
PIJPKL - PILPJK = 0 is not linearly implied by G. Since the latter possibility contradicts our
hypothesis, assume that there is a pair of treks T(/,L) and T(J,K) such that L(T(/ ,L)) does
not intersect J(T(J,K)). There are two cases.
If /(T(/,L)) does not intersect K(T'(J.K)) then by lemma 6.10.7, G does not linearly imply
PIJPKL - PILPJK = 0, contrary to our hypothesis. Suppose then that /(T'(/ ,L)) does intersect
K(T(J,K)) at a vertex Y. (See figure 21.)
S S
S'
L K I J L K I J
T'(I,L) --+ T(J,K)--+ T(I,J) --+ T(K,L)--+
Figure 21
Let S be the source of T(/,L), S' the source of T'(J,K), R(S,Y) an initial segment of
I(T'(I,L», R(Y,K) a final segment of K(T'(J,K», R(S',Y) an initial segment of K(T'(J,K»,
R(Y,l) a final segment of I(T'(I,L», I(T'(I,J» = R(S',Y)&R(Y,I), J(T'(I,J» = J(T'(J,K»,
K(T'(K,L» = R(S,Y)&R(Y,K), and L(T'(K,L» = L(T'(/,L». But since J(T'(/,J» =
J(T'(J,K» does not intersect L(T'(K,L» =L(T'(I,L», there is no U(T(/,J),T(K,L» choke
point, contrary to our hypothesis. :.
Lemma 6.10.19: In an acyclic LCF G, if G linearly implies PIJPKL - PILPJK = 0, then either
there is an LJ(T(I,J),T(K ,L),T(I,L),T(J,K» choke point, or there is an
IK(T(I,J),T(K,L),T(I,L ),T(J,K» choke point.
Proof. Assume that G linearly implies PIJPKL - PILPJK = O. By lemma 6.10.18, either there is
an LJ(T(I ,J),T(K ,L» choke point and an LJ(T(I,L),T(J,K» choke point, or there is an
IK(T(/,J),T(K,L» choke point and an IK(T(I,L),T(J,K» choke point. Suppose w.l.g. that the
former is the case. If some U(T(/,l),T(K,L» choke point is also an U(T(I,L),T(J,K) choke
point, the proof is done. Suppose then that no U(T(/,J),T(K,L» choke point is also an
U(T(I,L),T(J,K» choke point. Let C be an U(T(/,J),T(K,L» choke point. By hypothesis C
is not an U(T(I,L),T(J,K) choke point, so there exist a pair of treks T(I,L) and T(J,K) with
sources Sand S' respectively, such that L(T'(I,L» and J(T(J,K) do not intersect at C. (See
figure 22.)
Proofs 453
l
S S' S S'
y)
Y
R(~(Y'L)
J J L K J J L K
T'(J,L)--. T'(J,K) ~
T'(U) ...... T'(K,L) ~
Figure 22
Hence there is at most one occurrence of C in the pair of paths L(T(J,L» and J(T(J,K». Since
there is an U(T(J,L),T(J,K» choke point, L(T'(J,L» and J(T'(J,K» intersect at a point Y. Let
R(S,Y) be an initial segment of L(T'(J,L», R(Y,J) be a final segment of J(T'(J,K», R(S',Y)
an initial segment of J(T'(J,K», R(Y,L) a final segment of L(T'(J,L», J(T'(J,1) = J(T(J ,L»,
J(T'(J,J» = R(S,Y)&R(Y,J), K(T'(K,L» = K(T'(J,K» and L(T'(K,L» = R(S',Y)&R(Y,L).
Since L(T'(K ,L» and J(T'(J,J» are rearrangements of the vertices in J(T'(J,K» and
L(T'(l,L», the number of occurrences of any vertex in L(T(K,L» and J(T(J,1) is less than or
equal to the number of occurrences of that vertex in J(T(J,K» and L(T(J,L». Since C occurs
at most once in J(T'(J,K» and L(T'(J,L», it occurs at most once in L(T'(K,L» and J(T'(J,1).
Hence L(T(K,L» and J(T'(J,1) do not intersect at C, contrary to the hypothesis that C is an
U(T(l,1),T(K,L» choke point. :.
Lemma 6.10.20: For any probability distribution over a set of random variables W, if there
exists a subset P of V such that PIJ.PPKL.P - PIL.PPJK.P = 0, and for all variables U in P and
all subsets V of P not containing U, either PIU.V =0 and PKU.V = 0, or PJu.V = 0 and PLU.V
= 0, then PIJPKL- PILPJK = O.
Proof. The proof is by induction on the cardinality of P.
Base Case: Suppose the cardinality of P is zero. Then PIJPKI - PILPJK = 0 is equivalent
tOPIJ.PPKL.P - PIL.PPJK.P = O.
Induction Case: Suppose that the lemma is true for all sets of cardinality n or less. Let P have
cardinality n+ l. Assume that PIJ.PPKL.P - PIL.PPJK.P = O.
Let Y be a variable in P, and P' =P - {Y}. Since PIJ.PPKL.P - PIL.PPJK.P, by the recursion
formula for partial correlation,
The denominator of the l.h.s. equals the denominator of the r.h.s., so the numerator of the
l.h.s. equals the numerator of the r.h.s. Expanding the numerators of each side,
PIJ.P·PKL.P' - PIJ.P'PKY.P'PLY.P' - PKL.P'PIY.P'PJY.P' - PIY.P'PJY.P'PKY.P'PLY.P' =

PIL.P'PJK.P' - PIL.P'PJY.P'PKY.P' - PJK.P'PIY.P'PLY.P' - PIY.P'PJY.P'PKY.P'PLY.P'
The fourth terms on both sides are equal. By hypothesis, either PIY.P' = PKY.P' = 0, or PJY.p·
= PLY.P· = O. In either case, the second and third terms on each side are equal to zero. It
follows that PIJ.P·PKL.P· - PIL.P·PJK.P· = O. Since P' has one less member than P, by the
induction hypothesis, PIJPKL - PILPJK = O. :.
Lemma 6.10.21: In an acyclic LCF G, if there exists an U(T(J,J),T(K,L),T(J,L),T(J,K»

choke point or an JK(T(J,J),T(K,L),T(J,L),T(J,K» choke point, then G linearly implies
PIJPKL - PILPJK = O.
Proof. Suppose w.l.g. that X is the last U(T(J,J),T(K,L),T(J,L),T(J,K» choke point. There
are two cases.
First consider the case where there is no trek between at least one of the pairs J and J, and K
and L, and there is no trek between at least one of the pairs J and L, and J and K. It follows that
at least one of PIJ and PKL equals 0, and at least one of PIL and PJK is equal to zero. Hence
PIJPKL - PILPJK = O.
Next suppose w.l.g. that there are treks T'(J,J) and T'(K,L). We will prove that PIJPKL-
PILPJK = 0 by proving that there exists a set Q' of variables such that PIJ.Q'PKL.Q'-
Proofs 455
PIL.Q'PJK.Q' = 0, and for all variables U in Q' and all subsets V of Q' not containing U, either
pW.V =0 and PKU.V =0, or PJU.V =0 and Pw.V =0, and applying lemma 6.10.20.
Let Q = {sources of treks between X and J or X and L}. Since X is on J(T'(J,J» and
L(T(K,L)), and by defmition the sink of J(T'(J,J) is J, and the sink of L(T(K,L)) is L, there
are directed paths R(X,J) and R(X,L); hence X is in Q. We will now demonstrate that J ll.J1Q
by showing that J and J are d-separated given Q. We will show that J and J are d-separated
given Q by showing that every undirected path between J and J either contains a vertex V that is
a collider that is not the source of a directed path from V to any vertex in Q, or it contains some
vertex in Q that is not a collider.
Consider first the undirected paths between J and J without colliders. If there is an undirected
path with no collider between J and J that does not contain X, there is a trek between J and J
that does not contain X. But, every T(J,J) contains X, since X is a choke point. Hence, there
does not exist an undirected path between J and J without colliders that does not contain X.
Since X is in Q, every undirected path that does not contain a collider contains a vertex in Q.
Consider now undirected paths between J and J that contain colliders. If some vertex W is a
collider and is not the source of a directed path from W to some vertex in Q, the proof is done.
Suppose then that every vertex W that is a collider is the source of a directed path from W to
some vertex in Q. Consider w.l.g. an arbitrary undirected path R(JI) from J to J. Let Z be the
first vertex on R(JI) that is a collider. By hypothesis, there is a directed path R(Z,U) where U
is a vertex in Q. Since the undirected path from J to Z does not contain any colliders, there is a
vertex S that is the source of a pair of directed paths R(S,J) and R(S,z). Since Z has an edge
directed into it, S *' Z. There are two cases.
a. S = J. See figure 23. There is a directed path R(J,z). There is a directed path R(Z,U). Since
U is the source of a trek between X and J, there is a directed path R(U ,x). We have already
shown that there is a directed path R (X ,]). Hence there is a cyclic path
R(J,Z)&R(Z,U)&R(U ,x)&R(X ..I).
I K
Figure 23
b. S '¢' J. See figure 24. There is a directed path R (S ,1), and a directed path
R(S,z)&R(Z,U)&R(U)(). By lemma 6.10.2 there is a trek T(J)() with source M, where M
is the last point of intersection of R(S,.!) and R(S,z)&R(Z,U)&R(U )(), and J(T(J )(» is a
subpath of R(S,.!). Since M is on R(S,.!), and S occurs before Z on R(J,/), M occurs before Z
on R(J,/). Hence there is no collision at M in R(J,/). Also, M is in Q, since it is the source of a
trek between X and J. The undirected path R(J,/) contains a vertex in Q that is not a collider.
Proofs 457
l~
Figure 24
In either case Q d-separates X and Y, so IllJIQ. Similarly, it can be shown that

K liLIQJ liLIQJ liKIQ. It follows that PIJ.Q= 0, PKL.Q =0, PIL.Q= 0, and PJK.Q = O. Let
Q' =Q - {X}. By the recursion formula for partial correlation, PIJ.Q' =PIX.Q'PlX.Q', PKL.Q' =
PKX.Q'PLX.Q', PIL.Q' = PIX.Q'PLX.Q', and PJK.Q' = PJX.Q'PKX.Q'. Hence PIJ.Q' PKL.Q' =
PIX.Q'PJX.Q'PKX.Q'PLX.Q' =PIX.Q'PLX.Q'PlX.Q'PKX.Q' =PIL.Q' PJK.Q' .
We will next demonstrate that for each variable U in Q', and each subset V of Q' not
containing U,I llUIV, by showing that I and U are d-separated given V. We will show that I
and U are d-separated given V by showing that every undirected path between I and U either
contains a vertex W that is a collider that is not the source of a directed path from W to any
vertex in V, or it contains some vertex in V that is not a collider.
For U in Q', consider an arbitrary undirected path R(I,U) that contains colliders. Let Z be the
first point of R(/,U) after I that is a collider, and R(I,'Z) be an initial segment of R(/,U). If Z is
not the source of a path to some vertex M in Q', then I and U are d-separated given Q', and the
proof is done. Suppose then that there is a directed path R(Z,M) to some M in Q'. Since R(I,'Z)
contains no colliders, there is a vertex S on R(/,z) that is the source of directed paths R(SJ)
and R(S,Z). Hence S is the source of directed paths to I and M, R(SJ) and R(S,M) =
R(S,'Z)&R(Z,M) respectively. (If R(/,U) is an undirected path that contains no colliders, then it
still follows that there is a vertex S on R(/,U) that is the source of directed paths R(SJ) and
R(S,U).) M is either the source of a trek between X and J or X and L. Suppose w.l.g. that M is
the source of a trek between X and J. Then M is the source of a directed path R(M,J) and a
directed path R(M,x). M does not equal X by hypothesis. Hence R(M,J) does not contain X,
since R(M,J) is a branch of a trek between J and X, and the two branches of the trek intersect
only at M. R(S,M) does not contain X, else there is a cycle. Either R(SJ) contains X or it
doesn't.
a. If R (S ,I) does not contain X, then there are a pair of paths, R (S,I) and
R(S,7:)&R(Z,M)&R(M,J) that do not contain X. See figure 25. Hence there is a trek between I
and J that does not contain X. This contradicts the assumption that X is an
Ll(T(I,J),T(K,L),T(I,L),T(J,K) choke point.
Figure 25
b. Suppose R(SJ) does contain X. See figure 26. Then there is a directed path R(S,}() that is a
subpath of R(SJ). There is also a directed path R(S,M) =R(S,z)&R(Z,M)&R(M,J). Hence
there is a trek between X and J whose source S' lies on R(S,}(), which is a subpath of R(SJ).
S' is in Q' since it is the source of a trek between X and J. S' lies on R(S,x) which is a
Proofs 459
subpath of R(SJ). S' is not a collier on R(SJ), since S' occurs before Z, which is the fIrst
collider on R(J,U). Hence R(S}) contains a vertex in Q' that is not a collider.
Figure 26
By lemma 6.10.20, PIJPKL - PILPJK = O. :.
Tetrad Representation Theorem 6.10: In an acyclic LCF G, there exists an

LJ(T(J,J),T(K,L),T(/,L),T(J,K» choke point or an JK(T(J ,J),T(K,L),T(J,L),T(J,K» choke
point iff G linearly implies PIJPKL - PILPJK = O.
Proof. This follows directly from lemma 6.10.19 and lemma 6.10.21. :.
Corollary 6.10.1: If an acyclic LCF G' is a subgraph of an acyclic LCF G, and G linearly
implies PIJPKL - PILPJK = 0, then G' linearly implies PIJPKL - PILPJK = 0.
Proof. If G linearly implies PIJPKL - PILPJK = 0, then by lemma 6.10.21 G has either an
U(T(I.J),T(K,L),T(I,L),T(J,K» choke point or an IK(T(I.J),T(K,L),T(I,L),T(J,K» choke
point. If G has either an LJ(T(l,J),T(K,L),T(l,L),T(J,K» choke point or an
IK(T(l,J),T(K,L),T(l,L),T(J,K» choke point,then G' has either an
point. By lemma 6.10.21, G' linearly implies PIJPKL - PILPJK = 0. :.
13.26 Theorem 6.11
Theorem 6.11: An acyclic LCF G linearly implies PIJPKL - PILPJK = ° only if either pIJ or
PKL= 0, and PIL or PJr- 0, or there exists a (possible empty) set Q of random variables in G
such that PIJ.Q =PKL.Q =PIL.Q =PJK.Q =0.
Proof. By theorem 6.10, if G linearly implies PIJPKL - PILPJK = 0, then there is either an
point in G. In the proof of lemma 6.10.21 we demonstrated that the existence of an
point then either pIJ or PKL = 0, and PIL or PJK =0, or there exists a set Q of random variables
such that PIJ.Q =0, PKL.Q =0, PIL.Q= 0, and PJK.Q =0. :.
13.27 Theorem 7.1
If G is a directed acyclic graph over a set of variables V u W, W is exogenous with respect to

V in G, Y and Z are disjoint subsets of V, P(V u W) is a distribution that satisfies the Markov
Proofs 461
condition for G, and Manipulated(W) = X, then P(YIZ) is invariant under direct

manipulation of X in G by changing W from WI to W2 if and only if P(YIZ,W = WI) =
P(YIZ,W = W2) wherever they are both defined ..
Theorem 7.1: If GComb is a directed acyclic graph over V u W, W is exogenous with

respect to V in GComb, Y and Z are disjoint subsets of V, P(V u W) is a distribution that
satisfies the Markov condition for GComb, no member of X II Z is a member of IP(Y,Z) in
GUnman, and no member of)0:Z is a member of IV(Y,Z) in GUnman, then P(YIZ) is invariant
under a direct manipulation of X in GComb by changing W from WI to W2 ..
Proof. Suppose that GComb is a directed acyclic graph over V u W, W is exogenous with
respect to V, GUnman is the subgraph of GComb over V, P(V u W) is a distribution that
satisfies the Markov condition for GComb, X = Manipulated(W), P(YIZ,W= WI) '"
P(YIZ,W= W2) when GComb is manipulated by changing the value of W from ~I to W2, Y
and Z are disjoint subsets of V, no member of X II Z is a member of IP(Y,Z) in GUnman, and
no member of)0:Z is a member of IV(Y,Z) in GUnman, but P(YIZ) is not invariant when X is
manipulated. Hence there is an undirected path U in GComb that d-connects some R in W to
some Yin Y given Z. Let W be the vertex on U closest to Y that is in W. By lemma 3.3.2,
U(W,y) d-connects Wand Y given Z\{W,Y} = Z. Because U(W,y) contains no member of
W except W, every subpath of U(W,Y) that does not contain W is an undirected path in
GUnman. Because U(W,y) is an undirected path between Wand Y, it contains some variable X
in Manipulated(W). There are two cases: either X is in Z or it is not in Z.
If X is in Z then X is a collider on U in GUnman, and the vertex T adjacent to X on U and

between X and Y is a parent of X, and hence not a collider on U. Because T is not a collider on
U, T is not in Z, and Z\{T} = Z. If T is in Y, then X is in IP(Y,Z), contrary to our
assumption. If T is not in Y, then U(T,y) d-connects T and Y given Z\{T,y} = Z in GUnman.
T has a descendant (X) in Z in GUnman, and hence T is in IV(Y,Z) in GUnman. But then X is
in IP(Y,Z) in G, contrary to our assumption.
If X is not in Z, then U(X,y) d-connects Y and X given Z\{X} = Z in GUnman. If X is a

collider on U then X has a descendant in Z in GUnman. If X is not a collider on U then U(X,y)
is out of X because X is a child of W. Either X is an ancestor of a collider on U(X,Y), in which
case it is an ancestor of some member of Z in GComb, or U(X,y) is a directed path to Y, in
which case it is an ancestor of some member of Y in GComb. If X has a descendant in Z u Y in
GComb, then X has a descendant in Z u Y in GUnman, because W is exogenous with respect to
V. Hence X has a descendant in Y v Z in GUnman. It follows that X is in IV(Y,Z) in GUnman,

contrary to our assumption..'.
13.28 Theorem 7.2
Theorem 7.2: If P(O) is the marginal of a distribution faithful to G over V, 1r is a partially

oriented inducing path graph of Gover 0, and Ord is an ordering of variables in 0 acceptable
for some inducing path graph over 0 with partially oriented inducing path graph 1r, then there is
a minimal I-map GMin of P(O) in which Definite-SP(Ord,X) in 1r is included in
Parents(GMin,x) which is included in Possihle-SP(Ord,x) in 1r.
Proof. Suppose that GIP is an inducing path graph over 0 with partially oriented inducing path
graph 1r. By lemma 6.2.4 if GlP is an inducing path graph over 0 and Ord an acceptable total
ordering of variables for GlP, then Predecessors(Ord,x)'SP(Ord,GlP,x) is d-separated from
X given SP(Ord,GlP,x). Hence, if Parents(GMin,x) = SP(Ord,GIP,x) then GMin is an 1-
map ofP(O).
We will now show that no subgraph of GMin is an I-map of P(O). Suppose in GSub
Parents(GSub,x) is properly included in Parents(GMin,x) and hence properly included in
SP(Ord,GIP,x). Let V be some variable in Parents(GMin,x)\Parents(GSub,x). Because V
is in SP(Ord,GlP,x) there is an undirected path U in GlP between V and X on which all of the
vertices except the endpoints are colliders, and precede X in Ord. Let W be the vertex on U
closest to X but not equal to X that is in Parents(GMin,x)\Parents(GSub,x). It follows that
U(W,x) is an undirected path in GIP between Wand X such that every vertex on U(W,x)
except for the endpoints is a collider and in Parents(G Sub,x). Hence W is in
Predecessors(Ord,x)\Parents(GSub,x) and is d-connected to X given Parents(GSub,x) in
GIP. Hence W is d-connected to X given Parents(GSub,x) in G, and because P(V) is faithful
to G, Wand X are dependent given Parents(GSub,x). Hence P(O) does not satisfy the
Markov Condition for GSub.
For a partially oriented inducing path graph 1r and ordering Ord acceptable for 1r, V is in
Possihle-SP(Ord,x) if and only if V *" X and there is an undirected path U in 1r between V
and X such that every vertex on U except for X is a predecessor of X in Ord, and no vertex on
U except for the endpoints is a definite-non-collider on U. For a partially oriented inducing path
graph 1r and ordering Ord acceptable for 1r, V is in Definite-SP(Ord,x) if and only if V *" X
Proofs 463
and there is an undirected path U in 7C between V and X such that every vertex on U except for
X is a predecessor of X in Ord, and every vertex on U except for the endpoints is a collider on
U. From these definitions and the definition of partially oriented inducing path graph it follows
that Definite-SP(Ord,x) is included in Parents(GMin,x) which is included in Possible-
SP(Ord,x). :.
13.29 Theorem 7.3
Theorem 7.3: If G is a directed acyclic graph over V v W, W is exogenous with respect to

V in G, 0 is included in V, GUnman is the subgraph of G over V, 7C is the FeI partially
oriented inducing path graph over 0 of GUnman, Y and Z are included in 0, X is included in
Z, Y and Z are disjoint, and no X in X is in Possibly-IP(Y,Z) in 7C, then P(YIZ) is invariant
under direct manipulation of X in G by changing the value of W from WI to wz.
If A and B are not in Z, and A *- B, then an undirected path U between A and B in a partially
oriented inducing path graph 7C over 0 is a possibly d-connecting path of A and B given Z
if and only if every collider on U is the source of a semi-directed path to a member of Z, and
every definite non-collider is not in Z.
Lemma 7.3.1: If G is a directed acyclic graph, U is a path that d-connects V and Y given Z,
X is in Z, and X is on U, then there is a path that d-connects X and Y given z\{X} that is into
X and that contains only edges that lie on a directed path to X, and a subpath of U(X,Y).
Proof. Suppose that G is a directed acyclic graph, U is a path that d-connects V and Y given
Z, X is in Z, and X is on U. Because X is in Z and on U, it follows that X is a collider on U,
and hence U(X,Y) is into X. No non-collider on U(X,Y) except for the endpoints is in Z, so
no non-collider on U(X,Y) except for the endpoints is in z\{X}. Every collider on U(X,Y) has
a descendant in Z. If every collider on U(X,Y) has a descendant in Z\{X} then U(X,Y) d-
connects X and Y given Z\{X}. Suppose then that some collider on U(X,Y) has X as a
descendant but no other member of Z as a descendant, and let C be the closest such collider on
U to Y. U(C,Y) d-connects C and Y given z\{X} because C is not in z\{X},every collider on
U(C,Y) has a descendant in z\{X}, and no non-collider on U(C,Y) is in z\{X}. There is a
directed path from C to X that contains no member of z\{X}. Hence by lemma 3.3.3 X is d-
connected to Y given z\{X} by a path that is into X, and that contains only edges that lie on a
directed path to X and a subpath of U(X,Y).
Lemma 7.3.2: If G' is the inducing path graph for Gover 0, X and Y are in 0, Z is
included in 0, and there is a path U d-connecting X and Y given Z in G, then there is a path T
d-connecting X and Y given Z in G' such that if U is into X in G, then T is into X in G' and if
U is into Yin G then T is into Yin G'.
Proof. Suppose that in G with inducing path graph G' that U is a path d-connecting X and Y
given Z. We will use the following algorithm to construct two sequences of vertices, Ancestor,
and D-Path. (yVe are actually interested only in the undirected path D-path; Ancestor is used
solely as a device to constructD-path.) The vertices in D-Path are always observed (i.e. vertices
in 0), but might not be on U; vertices in Ancestor are always on the path U, but might not be
observed. For any sequence of vertices R of vertices, R(n) refers to the nth vertex in R. We will
say that for any pair of variables V and Won U that W is after Von U if V is between Wand X
on U or V=X.
Algorithm D·Path
Ancestor(O) = <X>.
D-path(O) =<X>.
n =0.
repeat
if Ancestor(n) = D-path(n) then
if there is no collider between Ancestor(n) and the next observed variable V on U,
Ancestor(n+ 1) =D-path(n+ 1) =V;
else Ancestor(n+ 1) =first collider on U after Ancestor(n) and D-path(n+l) =first
observed variable on a path from Ancestor(n+ 1) to a member of Z;
else if Ancestor(n) *D-path(n) then
if on U there is no collider C after Ancestor(n) that has D-path(n) as the first
observed variable on a directed path from C to a member of Z, then Ancestor(n+ 1)
= D-path(n+ 1) = first observed variable on U after Ancestor (n)
else
let C2 be the collider closest to Y that has D-path(n) as the first observed variable
on a directed path from C2 to a member of Z;
if there is no collider between C2 and the ftrst observed variable after C2 on U
then Ancestor(n+ 1) =D-path(n+ 1) =ftrst observed variable after C2 on U;
else let Cl be the ftrst collider after C2, let Ancestor(n+l) = Cl and D-path(n+l)
= the ftrst observed variable on a directed path from Cl to a member of Z;
elsen=n+1.
until Y is in D·path.
Proofs 465
Figure 27
For example, when the algorithm is applied to the graph in figure 27 (where the circled vertices
are not observed), for U = <X,RS,T,Q,Y>, and the result is Ancestor = <X,R,Q,Y> and D-
path = <X,M,Q,Y>.
We will now show that either D-path d-connectsX and Y given ZinG', or some other path in
G' d-connects X and Y given Z.
All of the vertices in D-path are observed variables, and hence in G'. By the way that D-path is
constructed, each adjacent pair of vertices A and BinD-path is connected in G by a trek T(A,B)
that contains no observed variables, except for the endpoints. If A and B are both on U then
T(A,B) contains the edges in U(A,B); if A is on U and B is not then T(A,B) contains the edges
in U(A,Ancestor(B» and a directed path from Ancestor(B) to B; if A is not on U and B is, then
T(A,B) consists of a directed path from Ancestor(A) to A and U(Ancestor(A),B); and if neither
is on U, then T(A,B) contains the edges in a directed path from Ancestor(A) to A,
U(Ancestor(A),Ancestor(B», and a directed path from Ancestor(B) to B. T(A,B) is constructed
out of subpaths of U, and subpaths of directed paths from colliders on U to vertices in Z.
T(A,B) is an inducing path in G, and hence each adjacent pair of vertices in D-path is adjacent
in G'. The method of construction of D-path makes D-path acyclic. It follows that D-path is an
acyclic undirected path from X to Y in G'.
If W is on D-path, but is not a collider on D-path, then W is on U in G, and is not a collider on

U. It follows that W is not in Z.
We will now show that we can transform D-path into a path D-path' in G' such that every
collider Bon D-path' has a descendant in ZinG. LetB be the vertex on D-Path closest to X
that is a collider on D-path but that in G does not have a descendant in Z, and A be the
predecessor of B on D-path, and C be the successor of B on D-path. If in G T(A,B) and
T(B,C) are both into B, then by the construction of D-path, B has a descendant in Z in G.
Hence at least one of T(A,B) and T(B,C) is out of BinG. Suppose without loss of generality
that T(B ,C) is out of BinG, and B is between X and C on D-path. It follows that B is an
ancestor of C in G. In addition since there is an arrowhead at BinG', there is an inducing path
between B and C that is into B and C. By lemma 6.6.2, there is a vertex Von D-path(X,C)
such that there is an edge between V and C in G' that is substitutable for D-path(V,C). Let D-
path' be the concatenation of D-path(X, V) with the edge between V and C. By lemma 6.6.2, D-
path' is into X if D-path is. Every collider on D-path' is a collider on D-path, and every non-
collider on D-path' is a non-collider on D-path. Furthermore, D-path' does not contain the
vertex B which in G does not have a descendant in Z. Repeat this process until every vertex on
the modified D-path that in G does not have a descendant in Z has been removed from the path.
Call the result D-path'.
Suppose now that some collider B on D-path' has a descendant in ZinG but not in G'. We
will show how to transform D-path' into a path in G' in which every collider has a descendant
in ZinG'. Let P be a directed path in G from B to some Z that is a member of Z. In G', let P'
be the undirected path from B to Z that consists of the observed variables on P in the order in
which they occur. P' is an undirected path in G' because in G the directed path between any
two observed variables on P is an inducing path. Let S be the vertex on P' closest to Z such that
there is no directed path from B to SinG'. Let R be the predecessor of Son P'. If P'(B,R) is
not a directed path from B to R then form P" by substituting some directed path from B to R in
G' for P'(B,R) in P'. There is an inducing path between R and SinG that is into S, so in G'
the edge between Rand S is into S. Because P"(B,S) is not a directed path from B to S, but
P"(B,R) is a directed path from B to R, it follows that R <-> S in G'.
We will now demonstrate that there is an edge B <-> SinG'. If B = R, it follows from what
we have just shown. Suppose then that R '# B. In that case let Q be the predecessor of R on
P". Because P"(B,R) is a directed path from B to R, Q -> RinG'. By lemma 6.6.2, there is a
vertex E on P"(B,R) such that there is an edge between E and S that is into S and is
substitutable for P"(E,S) in P"(B,S). If the edge between E and S is out of E, then there is a
directed path from B to SinG', contrary to our assumption. It follows that the edge between E
and S is into E. But because P"(B,R) is a directed path from B to R, if the edge between E and
Proofs 467
S is into E, the edge between E and S is not substitutable for P"(E,S) in P"(B,S) unless E =
B. It follows then that B <-> S in G'.
We will now form a path D-path" between X and Y by the following iteration, where at each
stage of the iteration the vertices Band S are defined as above. Let the Oth stage D-path" equal
D-path'. If S is on the n-l th stage D-path"(X,B) let the nth stage D-path"(X,S) equal the n-l th
stage D-path"(X,S). If S is not on the n-l th stage D-path"(X,B) let V equal the concatenation
of the n-l th stage D-path"(X,B) and B <-> S. By lemma 6.6.2 there is a vertex E on V that is
not equal to B and not equal to S such that there is an edge from E to S that is into S, and is a
collider on V if and only if it is a collider on the concatenation of V(X,E) with the edge between
E and S. Let the nth stage D-path"(X,S) equal the concatenation of V(X,E) and the edge
between E and S. Similarly, form the nth stage D-path"(Y,S). The nth stage D-path"(X,S)
does not intersect the nth stage D-path"(Y,S) except at S because except for the edges
containing S, they are subpaths of paths that do not intersect except possibly at S. Let the nth
stage D-path" be the concatenation of D-path"(X,S) and D-path"(Y,S). If S does not have a
descendant in Z in G', repeat this process until some vertex M on P' that does have a
descendant in Z in G' is on D-path". See figure 28, where D-path' is <X,E,B,F,Y> and D-
path" consists of the edges in boldface.
X --+E .....1--- Y
Figure 28
The nth stage D-path" is into X if the n-l th stage D-path" is, and into Y if the n-l th stage D-
path" is. Moreover, the Olh stage D-path" (D-path') is into X if U is, and into Y if U is. Every
non-collider on the nth stage D-path" is a non-collider on the n-lIh stage D-path". Because
every non-collider on D-path' is not in Z, every non-collider on the nth stage D-path" is not in
Z. Every collider on the nth stage D-path" with the possible exception of M is a collider on the
n-l th stage D-path", and hence a collider on D-path'. M is a collider on the nth stage D-path",
but it has a descendant in Z. There is at least one fewer collider on nth stage D-path" that does
not have a descendant in Z than there is on D-path' (because D-path' contains B, and the nth
stage D-path" does not) This process can be repeated until every collider on D-path" has a
descendant in Z. The resulting path d-connects X and Y given Z in G', is into X if U is, and
into Y if U is. :.
Lemma 7.3.3: If G is a directed acyclic graph over V, 1r is the FCI partially oriented
inducing path graph of Gover 0, and some path U in G d-connects X and Y given Z, then
there is a path U" in 1r that possibly d-connects X and Y given Z. Furthermore if U is into X,
then U" is not out of X.
Proof. Suppose that some path U in G d-connects X and Y given Z. Let G' be the inducing
path graph of G. By lemma 7.3.2, there is a path U' in G' that d-connects X and Y given Z,
and if U is into X then U' is into X. Let U" be the path in 1r that corresponds to U' in G'. If R
is a collider on U", then by the definition of partially oriented inducing path graph R is a
collider on U'. Because R is a collider on U', and U' d-connects X and Y given Z, R has a
descendant in Z in G'. By Theorem 6.6, there is a semi-directed path from R to a member of Z
in 1r. If R is a definite non-collider on 0", then by definition of partially oriented inducing path
graph R is a non-collider on U'. Because R is a non-collider on 0', and U' d-connects X and Y
given Z, R is not in Z. Hence U" is a possibly d-connecting path between X and Y given Z.
Furthermore, if 0' is into X, then by definition of partially oriented inducing path graph U" is
not out of X. :.
If 1ris a partially oriented inducing path graph of Gover 0, then X is in Possibly-IV(Y,Z) if

and only if X is not in Z, there is a possibly d-connecting path between X and some Yin Y
given Z, and there is a semi-directed path from X to a member of Y u Z. If 1r is a partially
oriented inducing path graph of Gover 0, then X is in Possibly-IP(Y,Z) if and only if Y
and Z are disjoint, X is in Z, and there is a possibly d-connecting path between X and some Y
in Y given z\{X} that is not out of X. If 1ris the FCI partially oriented inducing path graph of
Gover 0, then X is in Definite-Non-Descendants(Y) if and only if there is no semi-
directed path from any member of Y to X in 1r.
Lemma 7.3.4: If X is in IV(Y,Z) in directed acyclic graph G, Y and Z are disjoint subsets
of 0, X is in 0, and 1r is the FC! partially oriented inducing path graph of Gover 0, then X is
in Possibly-IV(Y,Z) in 1C.
Proof. Suppose that X is in IV(Y,Z) in G, Y and Z are disjoint subsets of 0, X is in 0, and

1Cis the FCI partially oriented inducing path graph of Gover 0. Because X is in IV(Y,Z)in G,
X has a descendant in Y u ZinG. Hence, by Theorem 6.6, there is a semi-directed path from
Proofs 469
X to a member of Y u Z in 11:. Also, there is a path that d-connects X and some member Y of Y
given ZinG. Hence, by lemma 7.3.3 there is a path that possibly d-connects X and some
member Y of Y given Z in 11:. By defInition X is in Possibly-IV(Y,Z) in 11:. :.
Lemma 7.3.5: H X is in IP(Y,Z) in directed acyclic graph G, Y and Z are disjoint subsets
of 0, and 11: is the FeI partially oriented inducing path graph of Gover 0, then X is in
Possibly-IP(Y,Z) in 11:.
Proof. Suppose that X is in IP(Y,Z) in G, Y and Z are disjoint subsets of 0, and 11: is the
FeI partially oriented inducing path graph of Gover O. Because X is in IP(Y,Z) in G, some
v¢able Tin G is a parent of X and in IV(Y,Z) or Y. H T is in Y then there isa directed path
from a member TofY toX that d-connects T and X given z\{X}. HTis in IV(Y,Z) then Tis
d-connected to some Yin Y given Z by some path U. H X is on U then X is a collider on U
and U(X,y) is into X; furthermore, by lemma 7.3.1 there is an undirected path that d-connects
X and Y given Z\{X} that is into X. H X is not on U then the concatenation of the edge from T
to X and U is a path that d-connects X and Y given Z\{ X} and is into X. Hence, by lenuna
7.3.3 there is a path that possibly d-connects X and Y given z\{X} in 11: that is not out of X.
By defInition X is in Possibly-IP(Y,Z) in 11:. :.
Theorem 7.3: H G is a directed acyclic graph over V u W, W is exogenous with respect to

V in G, 0 is included in V, GUnman is the subgraph of G over V, 11: is the FeI partially
oriented inducing path graph over 0 of GUnman, Y and Z are included in 0, X is included in
Z, Y and Z are disjoint, and no X in X is in Possibly-IP(Y,Z) in 11:, then P(YIZ) is invariant
under direct manipulation of X in G by changing the value of W from WI to Wl.
Proof. Suppose that G is a directed acyclic graph over V u W, 0 is included in V, W is
exogenous with respect to V in G, GUnman is the subgraph of G over V, 11: is the FCI partially
oriented inducing path over 0 of GUnman, Y and Z are included in 0, X is included in Z, Y
and Z are disjoint, and no X in X is in Possibly-IP(Y,Z) in 11:. H P(YIZ) is not invariant
when X is manipulated by changing the value of W from WI to Wl then W is d-connected to
Y given ZinG. Suppose that W is d-connected to Y given ZinG. Let W be a member of W
that is d-connected to some Yin Y by an undirected path U in G that contains no other member
of W. No non-collider on U is in Z, and every collider on U has a descendant in Z.
Note that if R and N are in V and R is a descendant of N in G, then R is a descendant of N in

GUnman, because there is no edge from any member of V into a member of W. In G,U
contains some X in X. Because X is in Z,X is a collider on U, and U(X,Y) is into X. By
lImu:na 7.3.1 in G there is an undirected path M that d-connects X and Y given Z\{X}, is into
X, and contains only edges that lie on a directed path to X and a subpath of U(X,y). Hence M
is an undirected path in GUnman, no non-collider on M is in Z\{X}, and every collider on M

has a descendant in z\{X} in G, and hence in GUnman. It follows that M d-connects X and Y
given z\{X} in GUnman. Let T be the vertex adjacent to X on M. If T = Y then X is in IP(Y,Z)
*
in GUnman. If T Y then T has a descendant in Z (namely X) in GUnman. Also T is not a
collider on U(X,Y), and hence not in Z. By lemma 3.3.2 Tis d-connected to Y given z\{T} =
Z in GUnman. It follows that T is in IV(Y,Z) in GUnman, and hence X is in IP(Y,Z) in
GUnman. In either case X is in IP(Y,Z) in GUnman and by lemma 7.3.5, X is in Possibly-
IP(Y,Z) in 1T:, contrary to our assumption. :.
13.30 Theorem 7.4
Theorem 7.4: If G is a directed acyclic graph over V u W, W is exogenous with respect to

V in G, 0 is included in V, GUnman is the subgraph of G over V, 1T: is the FCI partially
oriented inducing path graph over 0 of GUnman, X, Y and Z are included in 0, X, Y and Z
are pairwise disjoint, and no X in X is in Possibly-IV(Y,Z) in 1T:, then P(YIZ) is invariant
under direct manipulation of X in G by changing the value of W from WI to wz.
Proof. Suppose G is a directed acyclic graph over V u W, W is exogenous with respect to V
in G, 0 is included in V, GUnman is the subgraph of G over V, 1T: is the FCI partially oriented
inducing path over 0 of GUnman, Y and Z are included in 0, X, Y and Z are pairwise disjoint,
and no X in X is in Possibly-IV(Y,Z). If P(YIZ) is not invariant when X is manipulated by
changing the value of W from WI to Wz then W is d-connected to Y given Z in G. Let W be a
member of W that is d-connected to some Y in Y given Z by an undirected path U in G that
contains no other member of W.
Because U d-connects Wand Y given Z, no non-collider on U is in Z, and every collider on U

has a descendant in Z. U contains some X in X. By lemma 3.3.2 U(X,Y) is an undirected path
that d-connects X and Y given ZinG. There is a path U'(X,Y) in GUnman with the same edges
as U(X,Y) in G, because U(X,Y) contains no member ofW. No non-collider on U'(X,Y) is in
Z. In G, every collider on U(X,Y) has a descendant in Z; hence every collider on U'(X,Y) has
a descendant in Z in GUnman. Hence U(X,Y) d-connects X and Y given Z in GUnman. By
lemma 7.3.3 there is a possibly d-connecting path between X and some Yin Y given Z in 1T:.
Now we will show that X has a descendant in Y u Z in GUnman. If X is a collider on U, then

X has a descendant in Z in G, and hence in GUnman. Suppose then that X is not a collider on
Proofs 471
U. The edge from W to X on U is into X, so the edge containing X on U(X,Y) is out of X. If

U(X,Y) contains no colliders then Y is a descendant of X. If U(X,Y) contains a collider, then
the collider on U(X,Y) closest to X is a descendant of X, and an ancestor of a member of Z.
Hence X is an ancestor of a member of Z. In either case, X has a descendant in Y u Z in G,
and hence in GUnman.
It follows that X is in IV(Y,Z) in GUnman, and hence by lemma 7.3.4 X is in in Possibly.

IV(Y,Z), contrary to our assumption. :.
13.31 Theorem 7.5
Theorem 7.5: If G is a directed acyclic graph over V u W, W is exogenous with respect to

V in G, GUnman is the subgraph of G over V, PUnman(W)(V) = P(VIW = WI) is faithful to
GUnman, and changing the value of W from WI to W2 is a direct manipulation of X in G, then
the Prediction Algorithm is correct
Proof. Let GMan be the manipulated graph, and F the minimal I-map of PUnman(W)(V)
constructed by the algorithm for the given ordering of varIables Ord. Step A) is trivial. Step B)
is correct by Theorem 6.4. Step Cl) is correct by Theorem 7.2. In step C2, by lemma 3.3.5,
for all values of V for which the conditional distributions in the factorization are defined,
-7
L n
PUnman(W) (VI Parents(F, V»
PUnman(W) (YIZ) = IV~,Z)VEIV(Y,Z)UIP(Y,Z)UY
L n PUnman(W) (VIParents(F, V»
IV(Y,Z)uYVEIV(Y,Z)uIP(Y,Z)uY
for all values z of Z such that PMan(Z) 0# O.
Because GMan is a subgraph of GUnman, if F is an I-map of PUnman(w)(V)then F is an I-map

of PMan(W)(V). Hence PMan(W)(V) satisfies the Markov condition for F, and by lemma 3.3.5
L
~
n PMan(W) (VIParents(F, V»
(1) PMan(W) (YIZ) = --=-IV.:...:~...::y..!:,z:.<).:...V.:::eI:....;V-,-(Y::..:,=Z)<.::U:..::IP:...;(>..::Y..!:,Z:.<)U=Y_ _ _ _ __
L n
PMan(W)(VIParents(F, V»
IV(Y,Z)uYVeIV(Y,Z)uIP(Y,Z)uY
for all values z of Z such that P Man(Z) *" 0, and for all values for which the conditional
distributions in the factorization exist
PMan(W)(V) satisfies the Markov condition for GMan by hypothesis. Hence in PMan(W)(V) X is
independent of its non-parental non-descendants in GMan given Parents(GMan,x). The
predecessors of X in Ord by hypothesis are either in Definite-Non-Descendants(1r,x), in
which case they are in Non-Descendants(GUnman,x) or they are in Parents(GMan,x).
GMan is a subgraph of GUnman, so any vertex that is a non-descendant of X in GUnman is a
non-descendant of X in GMan. Hence each predecessor of X in Ord is a non-de~endant of X in
GMan. The algorithm guarantees that Parents(GMan,x) is included in Predecessors(Ord,K).
It follows that Parents(GMan,X) is a subset of Predecessors(Ord,x) such that
Predecessors(Ord,x)\Parents(GMan,x) is independent of X given Parents(GMan,x) in
PMan(W)(V). Hence, if Parents(GMan,K) is substituted for Parents(F,x) in F, the resulting
graph is still an I-map of P Man(W)(V), by lemma '3.7.1. So in (1) we can substitute
P(XIParents(GMan,x» for P(XIParents(F,x» By assumption the algorithm returns a value
only if PMan(W)(VlParents(F, V» = PUnman(W)(VlParents(F, V» for each V *" X, so we can
substitute PUnman(W)(VlParents(F,V» for PMan(W)(VlParents(F,V» in (1). :.
13.32 Theorem 9.1
Theorem 9.1 If P(S) is faithful to G(S), and X and Y are sets of variables in G(S) not
containing S, then P(YIX) = P(YIX,s) if and only if X d-separates Y and Sin G(S).
13.33 Theorem 9.2
Theorem 9.2: For a joint distribution, P ,faithful to graph G, exactly one of <Y 11 XIZ; Y
11 XIZ u IS} > is true in P if and only if the corresponding member and only that member of
<Z d-separates X; Y; Z u IS} d-separates X, Y> is true in G.
Proofs 473
13.34 Theorem 10.1
Theorem 10.1: If G is an almost pure latent variable graph over V u T u C, T is causally

sufficient, and each latent variable in T has at least two measured indicators, then latent
variables Tl and T3, whose measured indicators include J and L respectively, are d-separated
given latent variable T2, whose measured indicators include I and K, if and only if G linearly
implies PJIPLK = PJLPKI = PJKPIL·
Figure 29
We say that a measurement model is almost pure if the only kind of impurities among the
measured variables are common cause impurities. An almost pure latent variable graph is
one in which the measurement model is almost pure.
Lemma 10.1.1: If G' is an almost pure latent variable graph over V u T u C, T is causally
sufficient, and each latent variable in T has at least two measured indicators, and latent variables
Tl and T3, whose measured indicators include J and L respectively, are d-separated given latent
variable T2, whose measured indicators include I and K, then G' linearly implies PJIPLK =
PJLPKI = PJKPIL.
Proof. Let G be a pure latent variable subgraph of G', formed by removing the sources of all
treks creating common cause impurities. IfTl and T2 are d-separated given T2 in G' then they
are d-separated given T2 in G. Because/ and K are pure indicators of T2 in G, and thus children
only of T2, T2 is a non-collider on all undirected paths between / and any other indicator or K
and any other indicator. Therefore J and / are d-separated given T2, K and L are d-separated
given T2, and K and / are d-separated given T2.
Since Tl and T3 are d-separated given T2, and again J and L are children only of Tl and T3
respectively, then J and L are d-separated given T2. X and Z are d-separated given Y if and and
only if G linearly implies PXZ.Y = O. Hence G linearly implies PIJ.T2 = 0, and
PIJ = P1T2 X P JT2 . Similarly, G linearly implies PKL =PKT2 X PLT2' P JL =P JT2 X PLT2 and
PIK =P1T2 xPKT2' Hence G linearly implies PJlPLK = PJT2 xP1T2 xPLT2 XPKT2 =
P JT2 X PLT2 X PKT2 X P1T2 = PJLPKI. G linearly implies the same vanishing tetrad differences
as G', so G' linearly implies PJIPLK = PJLPKI. The proof that PJLPKI = PJKPIL is linearly
implied by G' is essentially the same. :.
Lemma 10.1.2: If G is an almost pure latent variable graph over V u T u C, T is causally

sufficient, and each latent variable in T has at least two measured indicators, 'then latent
variables Tl and T3, whose measure indicators respectively include J and L, are d-separated
given latent variable T2, whose measured indicators include / and K, if G linearly implies
PJIPLK = PJLPKI •
Proof. Suppose that G linearly implies PJIPLK = PJLPKI but Tl and T3 are not d-separated
given T2.
By the Tetrad Representation Theorem, if G linearly implies PJI PLK = PJLPKI then either there
is an /L(T(I,J),T(L,K),T(L,J),T(I,K» choke point, or there is a
JK(T(/ ,J),T(L,K),T(L,J),T(/,K» choke point.
Let T(/,K) be the trek consisting of the edges from T2 to / and T2 to K. Suppose fIrst that there
is an /L(T(/,J),T(L,K),T(L,J),T(/,K) choke point. The choke point is either / or T2 because
those are the only vertices in /(T(/,K). / is not the choke point because it does not lie on any
trek between Land K. Hence T2 is the choke point. Similarly, if there is a
JK(T(/,J),T(L,K),T(L,J),T(/,K» choke point it is T2. Hence, in either case T2 is a choke
point.
There are two ways that Tl and T3 might fail to be d-separated given T2. Either there is a trek
between Tl and T3 that does not contain T2, or there is some undirected path U between Tl and
T3 such that T2 is a descendent of every collider on U, and T2 is not a non-collider on U.
Proofs 475
First assume that there is some trek between Tl and T3 that does not contain T2. Then there is a
trek between 1 and L that does not contain T2. But then T2 is not a choke point, contrary to
what we have just proved.
Now assume that there is some undirected path U between Tl and T3 such that T2 is a
descendent of every collider on U, and T2 is not a non-collider on U. In that case U d-connects
Tl and T3 given T2. Again there are two cases.
Suppose fIrst that T2 is an IL(T(I.», T(L,K), T(L.», T(I,K) choke point. Let C be the collider
on the undirected path U that is closest to T3. (See fIgure 30.)
~IZJ
Figure 30
U(T3,C) does not contain any colliders on U except C because C is the closest collider to T3 on
U; hence U(T3,C) is a trek between T3 and C. There is a vertex Won U(T3,C) that is the
source of a trek between T3 and C. W"# C because W is not a collider on U, but C is. Hence
U(W,T3) contains no colliders on U. It follows that U(W,T3) does not contain T2, because T2
is not a non-collider on U. Hence there is a trek T(K,L) between K and L whose K branch
consists of the concatenation of U(W,C), a directed path from C to T2, and the edge from T2 to
K, and whose L branch consists of the concatenation of U(W,T3) and the edge from T3 to L.
Because neither U(W,T3) nor the edge from T3 to L contains T2, T2 is not in L(T(K,L», and
hence is not an IL(T(/.»,T(L,K), T(L.»,T(I,K) choke point, contrary to our hypothesis.
A similar argument shows that if there is some undirected path U between Tl and T3 such that
T2 is a descendent of every collider on U and T2 is not a non-collider on U, then there is no
lK(T(/.», T(L,K), T(L,.!), T(I,K) choke point.
Therefore Tl and T3 are d-separated given T2. :.


sufficient, each latent variable in T has at least two measured indicators, then latent variables Tl
and T3, whose measured indicators include J and L respectively, are d-separated given latent
variable T2, whose measured indicators include I and K, if and only if G linearly implies
PJIPLK =PJLPKI =PJKPIL·
Proof. The theorem follows from lemmas 10.1.1 and 10.1.2.
13.35 Theorem 10.2

sufficient, each variable in T has at least two measured indicators, the input to MIMBuild is a
list of all vanishing zero and first order correlations among the latent variables linear\)' implied
n
by G, and is the output of MIMBuild then
A-I) If X and Yare not adjacent in n, then they are not adjacent in G.
A-2) If X and Yare adjacent in n and the edge is not labelled with a "7", then X and Y
are adjacent in G.
0-1) If X -> Y is in n, then every trek in G between X and Y is into Y.
0-2) If X -> Y is in n and the edge between X and Y is not labelled with a "7", then X ->
YisinG.

list of all vanishing zero and first order correlations among the latent variables linearly implied
by G, n is the output of MIMBuild, and X and Y are not adjacent in n, then they are not
adjacent in G.
Proof. This follows directly from Theorem 3.4.

by G, n is the output of MIMBuild, and X -> Y is in n, then every trek in G between X and Y
is into Y.
Proofs 477
Proof. Suppose X -> Y is in n. The proof is by induction on the number of iterations of the
repeat loop in step 0) in the PC Algorithm.
Base Case: There is a trek between X and Y in G, because otherwise X and Y are d-separated
given the empty set and therefore not adjacent in n. Suppose that X -> Yis oriented asX -> Y
<- Z by step C) of the PC Algorithm (i.e. X and Z are d-separated by some set not containing
Y.) If in G, there is a trek between X and Y, and a trek between Y and Z that are not both into
Y, then there is a trek between X and Z and hence X and Z are not d-separated given the empty
set. Suppose then that X and Z are d-separated by some W'* Y in G. Because X and Y are
adjacent in IT, W does not d-separate X and Yin G. Similarly, W does not d-separate Y and Z.
If there is a trek in G between X and Y that is out of Y then there is a directed path U from Y to
X in G. If U does not contain W then U d-connects X and Y given Win G. There is also a path
V in G that d-connects Y and Z given W. Because U is out of Y, U and V do not collide at Y in
G. Hence by lemma 3.3.1 X and Z are d-connected given WinG, contrary to our assumption.
If U does contain W, then W is a descendant of Y, and by lemma 3.3.1 X and Z are d-
connected given W, contrary to our assumption. Hence no trek in G between X and Y is out of
Y.
Induction Case: Suppose after n-1 iterations of the repeat loop in step 0) of the PC Algorithm,
if Z -> X in IT, then every trek between Z and X in G is into X. Suppose that the X -> Y edge
is oriented because there is some vertex Z such that Z -> X - Y in IT and Z is not adjacent to Y
in n. Because the edge between X anf! Y in IT was not oriented into Y, X and Z are d-separated
given Y. There are treks between X and Y, and between Y and Z in G, because they are
adjacent in n. If there is a trek between Y and X that is into X, then by lemma 3.3.1, X and Z
are d-connected given Y, contrary to our assumption. :.
Y is a definite non-collider on an undirected path U in pattern IT if and only if either X *-*

Y -> Z, or X <- Y *-* Z are subpaths of U, or X and Z are not adjacent and not X -> Y <- Z on
U.

by G, IT is the output of MIMBuild, and Y is a definite non-collider on undirected path U in P,
and the corresponding path U' exists in G, then Y is a non-collider on U'.
Proof. If U contains X *-* Y -> Z in IT, then by lemma 10.2.2, if the corresponding path U'
exists in G, then the edge between Y and Z in G is out of Y; hence Y is not a collider on U'.
Similarly, if X <- Y *-* Z in n, then Y is not a collider on U'. Suppose then that X and Z are
not adjacent and not X -> Y <- Z on U in n. It follows that X and Z are d-separated given Yin
G. Hence if the edges between X and Y and between Y and Z exist in G, they do not collide at
Y.

by G, n is the output of MIMBuild, and X - Y or X -> Y is in n, and the edge is not labelled
by a "?", then X and Y are adjacent in G.
Proof. Suppose that X - Y or X -> Y is in n, the edge is not labelled by a"?", but that X and
Y are not adjacent in G. Then there is some set S that d-separates X and Yin G. Let P be the
set of undirected paths in P between X and Y of length ~ 2. Any such S has cardinality ~ 2,
because otherwise MIMBuild would have found it with some test of vanishing zero or first
order partial correlations. X - Y or X -> Y was not labelled with a "?" so either (i) P is empty,
or (ii) every path in P contains a collider, or (iii) there is some vertex Z that is a definite non-
collider on every path in P, or (iv) every path in P contains some subpath <A,B,C>.
Suppose P is empty. Because by lemma 10.2.1 non-adjacencies in n are non-adjacencies in G,

the adjacencies in n are a superset of those in G, and thus the set of undirected paths in n is a
superset of the undirected paths in G. It follows that there is no undirected path of length ~ 2 in
G. If in G there is also no edge between X and Y, then X and Y are d-separated given the
empty set in G. But since there is an edge between X and Yin n, X and Y are not d-separated
given the empty set in G. Hence there is an edge between X and Yin G.
Suppose every path in P contains a collider and there is no edge between X and Yin G. By
lemmas 10.2.1 and 10.2.2 every path in G between X and Y contains a collider. Hence there is
no trek between X and Yin G. But then there is no edge between X and Yin n, contrary to our
assumption.
Suppose there is some vertex Z that is a definite non-collider on every path in P. It follows
from lemma 10.2.1, 10.2.2, and 10.2.3 that if there is no edge between X and Yin G, then Z
is a non-collider on every undirected path between X and Y in G. Hence X and Y are d-
separated by Z. It follows that there is no edge between X and Y in n, contrary to our
assumption.
Proofs 479
Suppose every path in P contains some subpath <A,B,C>. If there is no edge between X and Y
in G, then every undirected path in G between X and Y contains <A,B,C>. It follows that B is
either a collider on every path between X and Yin G, in which case X and Yare d-separated
given the empty set, or B is a non-collider on every path between X and Yin G, in which case,
X and Yare d-separated given Bin G. In either case, there is no edge between X and Y in II,
contrary to our assumption. :.

by G, II is the output of MIMBuild, and X -> Y is in II, and the edge is not labelled by a"?",
then X -> Yis in G.
Proof. This follows from lemmas 10.2.2 and 10.2.4. :.

by G, and II is the output of MIMBuild then
A-I) If X and Yare not adjacent in II, then they are not adjacent in G.
A-2) If X and Yare adjacent in II and the edge is not labelled with a "?", then X and Y
are adjacent in G.
0-1) If X -> Y is in II, then every trek in G between X and Y is into Y.
0-2) If X -> Y is in II and the edge between X and Y is not labelled with a "?", then X ->
Y is in G.
Proof. This follows from lemmas 10.2.1 through 10.2.5. :.
13.36 Theorem 11.1
Theorem 11.1. If G is a subgraph of directed acyclic graph G', than the set of tetrad
equations among variables of G that are linearly implied by G' is a subset of those linearly
implied by G.
Proof. If G is a subgraph of directed acyclic graph G', then the treks in G are a subset of the
treks in G'. Hence if there is a choke point in G', there is a choke point in G. By the Tetrad
Representation Theorem, if G' linearly implies that a tetrad difference t vanishes, then G
linearly implies t vanishes. :.
Glossary
A: In a graph G, Let A(A,B) be the union of the ancestors of A or B.
Acceptable: Let a total order Ord of variables in a graph G' be acceptable for G if and only
*
if whenever A B and there is a directed path from A to B in G', A precedes B in Ord.
After: In a graph G, vertex X is after vertex Y if and only if there is a directed path from Y to
X in G.
Almost Pure: We say that a measurement model is almost pure if the only kind of
impurities among the measured variables are common cause impurities. An almost pure
latent variable graph is one in which the measurement model is almost pure.
Before: In a graph G, vertex X is before vertex Y if and only if there is a directed path from
XtoYinG.
C.F: See constant factor.
Choke point: In a directed acyclic graph G, if for all T(KL) in T(K,L) and all T(J,J) in
T(J,J), L(T(KL» and J(T(J,J) intersect at a vertex Q, then Q is an LJ(T(J,J),T(KL» choke
point. Similarly, if for all T(K L) in T(K,L) and all T(J,J) in T(J,J), L(T(K L» and all
J(T(J,J) intersect at a vertex Q, and for all T(JL) in T(JL) and all T(J,K) in T(J,K), L(T(JL»
and J(T(J,K» also intersect at Q, then Q is an LJ(T(J,J),T(KL),T(JL),T(J,K» choke point.
Also see the defmition of trek.
Combined graph: See manipulation.
Constant factor: In an LCF or LCf T, if an expression is equal to ce, where c is a non-zero

constant, and e is a product of equation coefficients raised to positive integral powers, then c is
the constant factor of (c.f.) ceo
Contains: In a directed acyclic graph, directed paths R(U J) and R(U,J) contain trek T iff
J(T(J,J) is a fInal segment of R(UJ) and J(T(J,J) is a fmal segment of R(U,J).
D: Given a directed acyclic graph G, D(Xi,xj) is the set of all directed paths from Xi to Xj.
D-connection: See D-separation.
Definite discriminating path: In a partially oriented inducing path graph 1C, U is a

definite discriminating path for B if and only if U is an undirected path between X and Y
*" *"
containing B, B X, B Y, every vertex on U except for B and the endpoints is a collider or a
defInite non-collider on U, and
(ii) if V is between X and B on U and V is a collider on U then V -> Y in 1C, else V <-* Y
in 1C,
(iii) if V is between Y and B on U and V is a collider on U then V -> X in 1C, elsCi V <-* X
in 1C,
(iv) X and Y are not adjacent in 1C.
Definite non-collider: A vertex B is a definite non-collider on undirected path U if and

only if either B is an endpoint of U, or there exist vertices A 'and C such that U contains one of
the subpaths A <- B *-* C, A *-* B -> C, or A *-!..JL!.-* C.
Definite non-descendant: If 1C is the FCI partially oriented inducing path graph of Gover
0, then X is in Definite-Non-Descendants(Y) if and only if there is no semi-directed path
from any member of Y to X in 1C.
Definite-SP: For a partially oriented inducing path graph 1C over 0 and ordering Ord
acceptable for 1C, *"
V is in Definite-SP(Ord,x) if and only if V X and there is an undirected
path U in 1C between V and X such that every vertex on U except for X is a predecessor of X in
Ord, and every vertex on U except for the endpoints is a collider on U.
Dependent: In an LCT or LCFS , a variable Xi is dependent iff Xi does not have zero
indegree.
Det: Det(Z) is the set of variables determined by any subset of Z.

Glossary 483
Determines: A set of variables Z determines the set of variables A, when every variable in
function of any proper subset of Z.
Det-connected: See Det-separation.
Det-separated: If G is a directed acyclic graph over V, Z is a subset of V that does not

*
contain X or Y, and X Y, then X and Y are det-separated given Z and Deterministic(V)
if and only if either X and Yare d-separated given Z u Det(Z) in some Mod(G) relative to
Deterministic(V) and Z, or X or Y is in Det(Z); otherwise if X * Y and X and Y are not in
Z, then X and Y are det-connected given Z and Deterministic(V). If X, Y and Z are
disjoint sets of variables in V, and X and Y are non-empty, then X and Y are det-separated
given Z if and only if every member X of X and every member Y of Y are det-separated given
Z; otherise if X, Y and Z are disjoint sets of variables in V, and X and Y are non-empty, then
X and Y are det-connected given Z and Deterministic(V).
Discriminating path: In an inducing path graph G', U is a discriminating path for B if

*
and only if U is an undirected path between X and Y containing B, B X, B Y, and *
(ii) if V is between X and B on U and V is a collider on U then V -> Yin G', else V <-*
Yin G',
(iii) if V is between Y and Bon U and V is a collider on U then V -> X in G', else V <-*
XinG',
(iv) X and Y are not adjacent in G'.
Distributed form: The distributed form of an expression or equation E is the result of

carrying out every multiplication, but no additions, subtractions, or divisions in E. If there are
no divisions in an equation then its distributed form is a sum of terms. For example, the
distributed form of the equation u = (a + b)(c + d)v is u = acv + adv + bcv + bdv.
D-map: An acyclic graph G over V is a D-map of probability distribution P(V) iff for every
X, Y, and Z that are disjoint sets of random variables in V, if X is not d-separated from Y
given ZinG then X is not independent of Y given Z in P(V). However, when D-map is
applied to the graph in an LCf, the quantifiers in the definitions apply only to sets of non-error
variables.
D-Sep: If G' is an inducing path graph over 0 and A *- B, let V E D-SEP(A,B) if and only if
A *- V and there is an undirected path U between A and V such that every vertex on U is an
ancestor of A or B, and (except for the endpoints) is a collider on U.
D-separated: If G is a directed acyclic graph with vertex set V, Z is a set of vertices not
containing X or Y, X *- Y, and X and Y are not in Z, then X and Y are D-separated given Z
and Deterministic(V) if and only if there is no undirected path U in G between X and Y such
that each collider on U has a descendant in Z, and no other vertex on U is in Det(Z); otherwise
if X*- Y and X and Y are not in Z, then X and Y are D-connected given Z and
Deterministic(V). Similarly, if X, Y, and Z are disjoint sets of variables, and X and Yare
non-empty, then X and Y are D-separated given Z and Deterministic(V) if and only if each
pair <X ,Y> in the Cartesian product of X and Y are D-separated given Z and
Deterministic(V); otherwise if X, Y, and Z are disjoint, and X and Yare non-empty, then
X and Y are D-connected given Z and Deterministic(V).(Note that this is different from d-
separation, which begins with a lowercase "d", and d-connection, which also'begins with a
lowercase "d".)
e: In an LCF F, e(S) is equal to S if S is an independent variable, and it is equal to the error

variable into S if S is not an independent variable.
E: If X is a random variable, E(X) is the expected value of X.
Equiv(G'): If G' is an inducing path graph over 0, Equiv(G') is the set of inducing path
graphs over the same vertices with the same d-connections as G.
E.C.F: See equation coefficient factor.
Equation coefficient: See linear causal theory, linear causal form.
Equation coefficient factor: In an LCF or LCT T, if an expression is equal to ce, where c

is a non-zero constant, and e is a product of equation coefficients raised to positive integral
powers, then e is the equation coefficient factor(e.c.f.) of ce ..
Equivalent to a polynomial: In an LCF, a quantity (e.g. a covariance) X is equivalent

to a polynomial in tbe coefficients and variances of exogenous variables if and
only if for each LCF F = «R,M,E>, C, V, EQ,L,Err> and in every LCT S =
«R',M',E'>, (nf'p), EQ',L',Err'> that is an instance of F, there is a polynomial in the
Glossary 485
variables in C and V such that X is equal to the result of substituting the linear coefficients of S
in as values for the corresponding variables in C, and the variances of the exogenous variables
in S as values for the corresponding variables in V.
Error variable: See linear causal theory, linear causal form.
Exogenous: If G is a directed acyclic graph over a set of variables V u W, and V (') W = 0,

then W is exogenous with respect to V in G if and only if there is no directed edge from
any member of V to any member of W.
Faithfully indistinguishable: We will say that two directed acyclic graphs, G, G' are
faithfully indistinguishable (f.L) if and only if every distribution faithful to G is faithful to
G' and vice-versa.
F.I.: See faithfully indistinguishable.
Final segment: In a graph G, a path U of length n is a final segment of path V of length m

iff m ;:: n, and for 1 S; is; n+ 1, the zth vertex of V equals the (m-n+O th vertex of U.
I-Map: An acyclic directed graph Gover V is an I-map of probability distribution P(V) iff for
every X, Y, and Z that are disjoint sets of random variables in V, if X is d-separated from Y
given ZinG then X is independent of Y given Z in P(V). However, when I-map is applied to
the graph in an LCf, the quantifiers iii the definitions apply only to sets of non-e"or variables.
Ind: For a directed acyclic graph G, Ind is the set of independent variables in G.
Ind aIJ : IndaIJ is the coefficient of J in the independent equational for I. See also
independent equational.
Independent: In an LCf or LCFS , a variable Xi is independent iff Xi has zero indegree

(Le. there are no edges directed into it). Note that the property of independence is completely
distinct from the relation of statistical independence. The context will make clear in which of
these senses the term is used.
Independent equational: In an LCF «R,M,E>, C, V, EQ,L,S> an equation is an

independent equational for a dependent variable Xj if and only if it is implied by EQ
and the variables in R which appear on the r.h.s. are independent and occur at most once
Inducing path: If G is a directed acyclic graph over a set of variables V, 0 is a subset of V

containing A and B, and A "* B, then an undirected path U between A and B is an inducing
path relative to 0 if and only if every member of 0 on U except for the endpoints is a
collider on U, and every collider on U is an ancestor of either A or B. We will sometimes refer
to members of 0 as observed variables.
Inducing path graph: G' is an inducing path graph over 0 for directed acyclic
graph G if and only if 0 is a subset of the vertices in G, there is an edge between variables A
and B with an arrowhead at A if and only if A and B are in 0, and there is an inducing path in
G between A and B relative to 0 that is into A. (Using the notation of Chapter 2, the set of
marks in an inducing path graph is (>, EM}.)
Initial segment: In a graph G, a path U of length n is an initial segment of path V of

length m iff m ~ n, and for I :;:; i :;:; n+ I, the Ith vertex of V equals the ith vertex of U.
Into: In a graph G, an edge between A and B is into A if and only if the mark at the A end of
the edge is an ">". If an undirected path U between A and B contains an edge into A we will
say that U is into A.
Invariant: If G is a directed acyclic graph over a set 'of variables V u W, W is exogenous

with respect to V in G, Y and Z are disjoint subsets of V, P(V u W) is a distribution that
satisfies the Markov condition for G, and Manipulated(W) = X, then P(YIZ) is invariant
under direct manipulation of X in G by changing W from WI to W2 if and only if P(YIZ,W =
WI) =P(YIZ,W = W2) wherever they are both defmed.
Instance: An LCT S is an instance of an LCF F if and only if the graph of S is isomorphic

to the graph of F.
IP: In a directed acyclic graph G, if Y (') Z = 0, W is in IP(Y,Z) (W has a parent that is an

informative variable for Y given Z) if and only if W is a member of Z, and W has a parent in
IV(Y,Z) u Y.
IV: In a directed acyclic graph G, ifY (') Z = 0, then Vis in IV(Y,Z) (informative variables
for Y given Z) if and only if V is d-connected to Y given Z, and V is not in ND(YZ). (This
entails that V is not in Y u Z.)
Glossary 487
Label: See linear causal theory, linear causal form.
Length: In a graph G, the length of a path equals the number of vertices in the path minus
one.
Last point of intersection: In a directed acyclic graph G, the last point of intersection
of directed path R(U J) with directed path R(V.I) is the last vertex on R(U J) that is also on
R(V.I). Note that if G is a directed acyclic graph, the last point of intersection of directed path
R(U J) with directed path R(V.I) equals the last point of intersection of R(V.I) with R(U J);
this is not true of directed cyclic paths.
LCF: See linear causal form.
LCT: See linear causal theory.
Linear causal form: A linear causal form is an unestimated LCT in which the linear
coefficients and the variances of the exogenous variables are real variables instead of constants.
This entails that an edge label in an LCF is a real variable instead of a constant (except that the
label of an edge from an error variable is fIxed at one.) More formally, let a linear causal form
(LCF) be <<R,M,E>, C, V, EQ.L.Err> where
i. <R,M,E> is a directed acyclic graph. Err is a subset of R called the error variables.
Each error variable is of indegree 0 and outdegree 1. For every Xi in R of indegree '* 0
there is exactly one error variable with an edge into Xi.
ii. Cij is a unique real variable associated with an edge from Xj to Xi • and C is the set of
Cij.V is the set of variables a;,where Xi is an exogenous variable in <R,M,E> and a;
is
a variable that ranges over the positive real numbers.
iii. L is a function with domain E such that for each e in E, L(e) = Cij iff head(e) =Xi and
tail(e) = Xi. L(e) will be called the label of e. By extension, the product of labels of edges
of U. The label of an empty path is fIxed at 1.
iv. EQ is a consistent set of independent homogeneous linear equationals in variables in R.
For each Xi in R of positive indegree there is an equation in EQ of the form
Xi = I,cij Xj
XjEParents(Xj)
where each Cij is a real variable in C and each Xi is in R. There are no other equations in
EQ. Cji is the equation coefficient of Xj in the equation for Xi.
Linear causal tbeory: Let a linear causal tbeory be (LCT) be «R,M,E>, (DI,P),
EQ,L,Err> where
i. (DI,P) is a probability space, where D is the sample space,fis a sigma-field over D,
and P is a probability distribution overf
ii. <R,M,E> is a directed acyclic graph. R is a set of random variables over (Dj,P).
iii. The variables in R have a joint distribution. Every variable in R has a non-zero variance.
E is a set of directed edges between variables in R. (M is the set of marks that occur in a
directed graph, i.e. {EM, >}.
iv. EQ is a consistent set of independent homogeneous linear equations in random
variables in R. For each Xi in R of positive indegree there is an equation in EQ of the form
Xi = Laij Xj
XjEParents(Xj)
where each aij is a non-zero real number and each Xi is in R. This implies that each vertex
Xi in R of positive indegree can be expressed as a linear. function of all and only its parents.
There are no other equations in EQ. A non-zero value of aij is the equation coefficient
of Xj in the equation for Xi.
v. If vertices (random variables) Xi and Xj are exogenous, then Xi and Xj are pairwise
statistically independent
vi. L is a function with domain E such that for each e in E, L(e) = aij iff bead (e) =Xi and
tail (e) = Xj. L(e) will be called the label of e. By extension, the product of labels of edges
vii. There is a subset of S of R called the error variables, each of indegree 0 and
outdegree Note that the variance of any endogenous variable I conditional on any set of
variables that does not contain the error variable of I is not equal to zero.
Linear Representation: A directed acyclic graph G over V linearly represents a

distribution P(V) if and only if there exists a a directed acyclic graph G' over V' and a
distribution P"(V') such that
(i) V is included in V';
Glossary 489
(ii) for each endogenous (that is, with positive indegree) variable X in V, there is a unique
variable ex in V,\V with zero indegree, positive variance, outdegree equal to one, and a
directed edge from ex to X;
(iii) G is the subgraph of G' over V;
(iv) each endogenous variable in G is a linear function of its parents in G';
(v) in P"(V') the correlation between any two exogenous variables in G' is zero;
(vi) P(V) is the marginal of P"(V') over V.
The members of V,\V are called error variables and we call G' the expanded graph.
Linearly implies: A directed acyclic graph G linearly implies PAB.H = 0 if and only if
PAB.H = 0 in all distributions linearly represented by G. (We assume all partial correlations are
defined for the distribution.)
Manipulate: See manipulation.
Manipulated graph: See manipulation.
Manipulation: If G is a directed acyclic graph over a set of variables V u W, and V II W =

0, then W is exogenous with respect to V in G if and only if there is no directed edge
from any member of V to any member of W. If GCornb is a directed acyclic graph over a set of
variables V u W, and P(V u W) satisfies the Markov condition for GCornb, then changing the
value of W from WI to wz is a manipulation of GComb with respect to V if and only if W is
exogenous with respect to V, and P(VIW = WI) *- P(VIW = wz).We define PUnrruJn(W)(V) =
P(VIW = WI), and P Man(W)(V) = P(VIW = wz), and similarly for various marginal and
conditional distributions formed from P(V). We refer to GCornb as the combined graph, and
the subgraph of GComb over V as the un manipulated graph GUnman. V is in
Manipulated(W) (that is, V is a variable directly influenced by one of the manipulation
variables) if and only if V is in Children(W) II V; we will also say that the variables in
policy variables. The manipulated graph, GMan is a subgraph of GUnman for which
PMan(W)(V) satisfies the Markov Condition and which differs from GUnman in at most the
parents of members of Manipulated(W).
Minimal I-map: An acyclic graph G is a minimal I-map of probability distribution P iff

G is an I-map of P, and no subgraph of G is an I-map of P. However, when minimal I-map is
applied to the graph in an LCI', the quantifiers in the definitions apply only to sets of IWn-e"or
variables.
Mod: If G is a directed acyclic graph over V, and Z is included in V, then G' is in Mod(G)
relative to Deterrninistic(V) and Z if and only if for each V in V
(i) if there exists a set of vertices included in Z that are non-descendants of V in G and
that determine V, then Parents(G',V)= X, where X is some set of vertices included in
Z that are non-descendants of V in G and that determine V;
(ii) if there is no set X of vertices included in Z that are non-descendants of V in G and
that determine V, then Parents(G',V) = Parents(G,V).
ND: In a directed acyclic graph G, ND(Y) is the set of all vertices that do not have a
descendant in Y.
Non-Descendants: In a directed acyclic graph G, X is in Non-Descendants(Y) if and

only if there is no directed path from any member of Y to X in G.
Observed: See inducing path graph, inducing path.
Out of: In a graph G, an edge between A and B is out of A if and only if the mark at the A
endpoint is the empty mark. If an undirected path U between A and B contains an edge out of A
we will say that U is out of A.
Parallel embedding: Directed acyclic graphs G1 and G2 with common vertex set 0 have a
parallel embedding in directed acyclic graphsH 1and H2 having a common set U of vertices
that includes 0 if and only if
(i) G1 is the subgraph of H1 over 0 and Gz is the subgraph of H2 over 0;.
(ii) every directed edge in H1 but not in G1 is in H2 and every directed edge in H2 but not in
G2 is in H1.
Path form: If G is a directed acyclic graph, let Let Pxy be the set of all directed paths in G
from X to Y. In an LCF S, the path form of a product of covariances rlJ'YKL is the
distributed form of
[ L (L LL(R)L(R')dfjlI L [L
UeU/J RePU/R'ePUJ
LL(R")L(R"')~ll
R"ePYKR"'ePYL VeUKL
Glossary 491
YIJ'YKL - YILrJK is in path form iff both tenns are in path fonn. YIIYKL - YILrJK is in path form
iff both tenns are in path fonn.
Policy variables: See manipulate.
*'
Possible-D-SEP(A,B): If A B in partially oriented inducing path graph tr, V is in
*'
Possible-D-Sep(A,B) in tr if and only if V A, and there is an undirected path U between A
and V in tr such that for every subpath <X,y.;z> of U either Y is a collider on the subpath, or Y
is not a definite non-collider on U, and X, Y, and Z fonn a triangle in tr.
Possibly d-connecting: If A and B are not in Z, and A*, B, then an undirected path U
between A and B in a partially oriented inducing path graph tr over 0 is a possibly d-
connecting path of A and B given Z if and only if every collider on U is the source of a semi-
directed path to a member of Z, and every definite non-collider is not in Z.
Possibly-IP: If tr is a partially oriented inducing path graph of Gover 0, then X is in

Possibly-IP(Y,Z) if and only if Y and Z are disjoint, X is in Z, and there is a possibly d-
connecting path between X and some Yin Y given z\{X} that is not out of X.
Possibly-IV: If tr is a partially oriented inducing path graph of Gover 0, then X is in

Possibly-IV(Y,Z) if and only if X is not in Z, there is a possibly d-connecting path between
X and some Y in Y given Z, and there is a semi-directed path from X to a member of Y u Z.
Possible-SP: For a partially oriented inducing path graph tr and ordering Ord acceptable for
tr, let V be in Possible-SP(Ord,x) if and only if V *' X and there is an undirected path U in tr
between V and X such that every vertex on U except for X is a predecessor of X in Ord, and no
vertex on U except for the endpoints is a definite-non-collider on U.
Predecessors: For inducing path graph G' and acceptable total ordering Ord, let
Predecessors(Ord,V) equal the set of all variables that precede V (not including V) according
toOrd.
Proper final segment: A path U of length n is a proper final segment of path V of

length m iff U is a final segment of V and U V. *'
Proper initial segment: A path U of length n is a proper initial segment of path V of

length m iff U is an initial segment of V and U ¢ V.
PMan(W)(V): See manipulate.
PUnflllln(W)(V): See manipulate.
Pure Latent Variable Grapb: A pure latent variable grapb is a directed acyclic graph
in which each measured variable is a child of exactly one latent variable, and a parent of no
other variable.
Random coefficient linear causal tbeory: The definition of a random coefficient

linear causal tbeory is the same as that of a linear causal theory except that each linear
coefficient is a random variable independent of the set of all other random variables in the
model.
Rigidly statistically indistinguisbable: If directed acyclic graphs G and G' are strongly
statistically indistinguishable and every parallel embedding of G and G' is strongly statistically
indistinguishable then structures G and G' are rigidly statistically indistinguisbable
(r .s.i.).
R.S.I.: See rigidly statistically indistinguishable.
Semi-directed: A semi-directed patb from A to B in partially oriented inducing path

graph 1& is an undirected path U from A to B in which no edge contains an arrowhead pointing
towards A, that is, there is no arrowhead at A on U, and if X and Yare adjacent on the path,
and X is between A and Y on the path, then there is no arrowhead at the X end of the edge
between X and Y.
Source: See trek.
SP: For inducing path graph G' and acceptable total ordering Ord, W is in SP(Ord,G',V)
(separating predecessors of V in G' for ordering Ord) if and only if W ¢ V and there is an
undirected path U between Wand V such that each vertex on U except for V precedes V in Ord
and every vertex on U except for the endpoints is a collider on U.
S.S.L: See strongly statistically indistinguisable.

Glossary 493
Strongly statistically indistinguishable: Two directed acyclic graphs G, G' are

strongly statistically indistinguishable if and only if they have the same vertex set V
and every distribution P on V satisfying the Minirnality and Markov Conditions for G satisfies
those conditions for G', and vice-versa.
Substituable: In an inducing path or directed acyclic graph G that contains an undirected path
U between X and Y, the the edge between V and W is substitutable for U(V,W) in U if and
only if V and Ware on U, V is between X and Won U, G contains an edge between V and W,
V is a collider on the concatenation of U(X, V) and the edge between V and W if and only if it is
a collider on U, and W is a collider on the concatenation of U(Y,W) and the edge between V
and W if and only if it is a collider on U.
T: See trek.
Termini: See trek.
Trek: A trek T(I,J) between two distinct vertices I and J is an unordered pair of acyclic
directed paths from some vertex K to I and J respectively that intersect only at K. The source of
the paths in the trek is called the source of the trek. I and J are called the termini of the
trek. Given a trek T(/,J) between I and J, I(T(I,]» will denote the path in T(/,J) from the
source of T(/,J) to I and J(T(/,J) will denote the path in T(/,J) from the source ofT(/,J) to J.
One of the paths in a trek may be an empty path. However, since the termini of a trek are
distinct, only one path in a trek can be empty. T(/,J) is the set of all treks between I and J.
T(/,J) will represent a trek in T(/,J). S(T(/,J) represents the source of the trek T(/,J).
Undirected: In a graph G, Let V be in Undirected(X,y) if and only if V lies on some

undirected path between X and Y.
Unmanipulated graph: See manipulation.
Ux: In an LCF S, Ux is the set of all independent variables that are the source of a directed
path to X. (Note that if X is independent then X e Ux since there is an empty path from every
vertex to itself.)
Uxy: In an LCF S, UXy is Ux () Uy.

Weakly faithfully indistinguishable: Two directed acyclic graphs are weakly

faithfully indistinguishable (w.f.i.) if and only if there exists a probability distribution
faithful to both of them.
Weakly statistically indistinguishable: Two directed acyclic graphs are weakly

statistically indistinguishable (w.s.i.) if and only if there exists a probability distribution
meeting the Minirnality and Markov Conditions for both of them.
W.F.I.: See weakly faithfully indistinguishable.
W.S.L: See weakly statistically indistinguishable.

Bibliography
Aigner, D. and Goldberger, A. (1977). Latent Variables in Socio-economic Models. North-

Holland, Amsterdam.
Aitkin, M. (1979). A simultaneous test procedure for contingency table models. Appl. Statist.
28, 233-242.
Allen, D. (1974). The relationship between variable selection and data augmentation and a
method for prediction. Technometrics 17, 125-127.
Anderson, 1., and Gerbing, D. (1982). Some methods for respecifying measurement models to
obtain unidemensional construct measurement. Journal of Marketing Research. 19, 453-60.
Anderson, T. (1984). An Introduction to Multivariate Statistical Analysis. Wiley, New York.
Asher, H. (1976). Causal Modeling. Sage Publications, Beverly Hills, CA.
Asmussen, S. and Edwards, D. (1983). Collapsibility and response variables in contingency

tables. Biometrika 70,567-578.
Bagozzi, Richard P. (1980). Causal Models in Marketing. Wiley, New York.
Bartlett, M. (1935). Contingency table interaction. J. Roy. Statist. Soc. Suppl. 2, 248-252.
Bartlett, M. (1954). A note on the multiplying factors for various chi-squared approximations.
J. Roy. Statist. Soc. Ser. B 16, 296-298.
Basmann, R. (1965). A note on the statistical testability of 'explicit causal chains' against the
class of 'interdependent' models. JASA, 60, 1080-1093.
Beale, E., Kendall, M., and Mann, D. (1967). The discarding of variables in multivariate
analysis. Biometrika 54,357-366.
Becker, O. (1964). Human Capital. National Bureau of Economic Research, New York.
Beinlich, I., Suerrnondt, H., Chavez, R., and Cooper, O. (1989). The ALARM monitoring
system: A case study with two probabilistic inference techniques for belief networks. Proc.
Second European Conference on Artificial Intelligence in Medicine, London, England. 247-
256.
Bentler, P. (1985). Theory and Implementation of EQS: A Structural Equations Program.

BMDP Statistical Software Inc., Los Angeles.
Bentler, P. (1980). Multivariate analysis with latent variables: causal modeling. Annual Review
of Psychology 31, 419-456.
Bentler, P. and Bonett, D. (1980). Significance tests and goodness of fit in the analysis of
covariance structures. Psychological Bulletin 88, 588-606.
Bentler, P., and Peeler, W. (1979). Models of female orgasm. Archives of Sexual Behavior 8,
405-423.
Birch, M. (1963). Maximum likelihood in three-way contingency tables. 1. Roy. Statist. Soc.
25, 220-223.
Bishop, Y., Fienberg, S., and Holland, P. (1975). Discrete Multivariate Analysis. MIT Press,
Cambridge, MA.
Blalock, H. (1961). Causal Inferences in Nonexperimental Research. University of North

Carolina Press, Chapel Hill, NC.
Blalock, H. (1969). Theory Construction: From Verbal to Mathematical Formulations.

Prentice-Hall, Englewood Cliffs, NJ.
Blalock, H. (1971). Causal Models in the Social Sciences. Aldine-Atherton, Chicago.
Blau, P. and Duncan, O. (1967). The American Occupational Structure. Wiley, New York.
Bibliography 497
Blum, R. (1984). Discovery, confirmation and incorporation of causal relationships from a

time-oriented clinical data base: The RX project. Readings in Medical Artificial Intelligence,
Clancey, W. and Shortliffe, E. (eds.). Addison-Wesley, Reading, MA.
Blyth, C. (1972). On Simpson's paradox and the sure-thing principle. JASA 67,364-366.
Bollen, K. (1989). Structural Equations with Latent Variables. Wiley, New York.
Bollen, K. (1990). Outlier screening and a distribution-free test for vanishing tetrads.
Sociological Methods and Research 19,80-92.
Bowden, R. and Turkington, D. (1984). Instrumental Variables. Cambridge University Press,

New York.
Breslow, N. and Day, N. (1980). Statistical Methods in Cancer Research, Vol 1: The Analysis
of Case-Control Studies. IARC, Lyon.
Brownlee, K. (1965). A review of "Smoking and Health." JASA 60, 722-739.
Bunker, J., Forrest, W., Mosteller, F. and Vandam, L. (1969). The National Halothane Study:
Report of the Subcommittee on the National Halothane Study of the Committee on
Anesthesia, Division of Medical Sciences, National Academy of Sciences. National
Research Council, Washington~ D.C. U.S. Government Printing Office.
Burch, P. (1978). Smoking and lung cancer: The problem of inferring cause (with discussion).
J. Roy. Statist. Soc. Ser. A 141,437-477.
Burch, P. (1983). The Surgeon General's "Epidemiologic Criteria for Causality." A critique.
Journal of Chronic Diseases 36, 821-836.
Burch, P. (1984). The Surgeon General's "Epidemiologic Criteria for Causality." Reply to
Lilienfeld. Journal of Chronic Diseases, 37, 148-157.
Byron, R. (1972). Testing for misspecification in econometric systems using full information.
International Economic Review 28, 138-151.
Callahan, J., and Sorensen, S., (1992) Using TETRAD IT as an automated exploratory tool.
Sociological Methods and Research. Fall
Campbell, D. and Stanley, J. (1963). Experimental and Quasi-Experimental Designs. Rand

McNally, Chicago.
Campbell, D., Schwartz, R., Sechrest, L., and Webb, E. (1966). Unobtrusive Measures:
Nonreactive Research in the Social Sciences. Rand McNally, Chicago.
Caramazza, A. (1986). On drawing inferences about the structure of normal cognitive

processes from patterns of impaired performance: the case for single patient studies. Brain
and Cognition, 5, 41-66.
Cartwright, N. (1983). How the Laws of Physics Lie. Oxford University Press, New York.
Cartwright, N. (1989). Nature's Capacities and Their Measurement. Clarendon Press, Oxford.
Cavallo, R. and Klir, G. (1979). Reconstructability analysis of multi-dimensional relations: A

theoretical basis for computer-aided determination of acceptable systems models.
International Journal of General Systems 5,143-171.
Cederlof, R., Friberg, L., Lundman, T. (1972). The interactions of smoking, environment and
heredity and their implications for disease etiology. Acta Med Scand. 612 (Suppl).
Chow, C. and Liu, C. (1968). Approximating discrete probability distributions with

dependence trees. IEEE Trans. on Info. Theory IT-14, 462-467.
Chow, C. and Wagner, T. (1973). Consistency of an estimate of tree-dependent probability

distributions. IEEE Trans. on Info. Theory IT-19, 369-371.
Christensen, R. (1990). Log-Linear Models. Springer-Verlag, New York.
Coleman, J. (1964). Introduction to Mathematical Sociology. Free Press, New York.
Cooper, G. and Herskovits, E. (1991). A Bayesian method for constructing Bayesian belief
networks from databases. Proceedings of the Conference on Uncertainty in Artificial
Intelligence. LA, CA. 86-94.
Bibliography 499
Cooper, G. and Herskovits, E. (1992). A Bayesian method for the induction of probabilistic
networks from data. Machine Learning (to appear).
Cooper, G. (1989). Current research in the development of expert systems based on belief
networks. Applied Stochastic Models and Data Analysis 5, 39-52.
Cornfield, J., Haensze1, W., Hammond, E., Lilienfe1d, A., Shimkin, M., and Wynder, E.
(1959). Smoking and lung cancer: Recent evidence and a discussion of some questions.
Journal of the National Cancer Institute 22, 173-203.
Costner, H. (1971). Theory, deduction and rules of correspondence. Causal Models in the
Social Sciences, Blalock, H. (ed.). Aldine, Chicago.
Costner, H. and Schoenberg, R. (1973). Diagnosing indicator ills in multiple indicator models,
in Structural Equation Models in the Social Sciences, Goldberger, A. and Duncan, O.
(eds.). Seminar Press, New York.
Costner, H. and Herting, J. (1985). Respecification in multiple indicator models, in Causal

Models in the Social Sciences, 2nd ed., Blalock, H. (ed.), 321-393. Aldine, New York.
Cox, D. (1958). Planning of Experimen~. Wiley, New York.
Crawford, S. and Fung, R. (1990). An Analysis of Two Probabilistic Model Induction

Techniques. Third International Workshop on Artificial Intelligence and Statistics, Fort
Lauderdale, FL.
Darroch, J., Lauritzen, S., and Speed, T. (1980). Markov fields and log linear interaction
models for contingency tables. Ann. Stat. 8, 522-539.
Davis, W. (1988). Probabilistic theories of causation. Probability and Causality, James Fetzer
(ed.). D. Reidel, Dordrecht.
Dawid, A. (1979). Conditional independence in statistical theory (with discussion). J. Roy.

Statist. Soc. Ser. B 41, 1-31.
Dempster, A. (1972). Covariance selection. Biometrics 28, 157-175.

Doll, R. and Hill, A. (1952). A study of the aetiology of carcinoma of the lung. Brit. Med. J.
2, 1271-1286.
Duncan, O. (1975). Introduction to Structural Equation Models. Academic Press, New York.
Duncan, 0., Featherman, D., and Duncan, B. (1972). Socioeconomic Background and
Achievement. Seminar Press, New York.
Edwards, A. (1976). An Introduction to Linear Regression and Correlation. W. H. Freeman,

New York.
Edwards, D. and Havranek, T. (1985). A fast procedure for model search in multi-dimensional
contingency tables. Biometrika 72,339-351.
Edwards, D. and Havranek, T. (1987). A fast model selection procedure for large families of
models. J. Amer. Statist. Assoc. 82, 205-211.
Edwards, D. and Kreiner, S. (1983). The analysis of contingency tables by graphical models.
Biometrika 70, 553-565.
Elby, A. (1992) Should we explain the EPR causally? Philosophy of Science 59,16-25.
Fienberg, S. (1977). The Analysis of Cross-classified Categorical Data. MIT Press,

Cambridge, MA.
Fine, A. (1982). Hidden variables, joint probability, and the Bell inequalities. Physical Review
Letters 48, 291-295.
Fine, T. (1973). Theories of Probability. Academic Press, New York.
Fisher, F. (1966). The Identification Problem in Economics. McGraw-Hill, New York.
Fisher, R. (1951). The Design of Experiments. Oliver and Boyd, Edinburgh.
Fisher, R. (1959). Smoking. The Cancer Controversy. Oliver and Boyd, Edinburgh.
Bibliography 501
Flack, V. and Chang, P. (1987). Frequency of selecting noise variables in subset regression
analysis: A simulation study. American Statistician 41,84-86.
Forbes, H. and Tufte, E. (1968). A note of caution in causal modeling. American Political
Science Review 62, 1258-1264.
Fox, J. (1984). Linear Statistical Models and Related Methods. Wiley, New York.
Freedman, D. (1983a). Structural-equation models: A case study. Report, Department of

Statistics, University of California, Berkeley.
Freedman, D. (1983b). A note on screening regression equations. American Statistician 37,

152-155.
Freedman, D., Navidi, W., and Peters, S. (1986). On the impact of variable selection in fitting
regression equations. On Model Uncertainty and its Statistical hnplications. Lecture Notes
in Economics and Mathematical Systems 307, Dijkstra, T. (ed.). Springer-Verlag, Berlin.
Friedman, M. (1957). A Theory of the Consumption Function. Princeton University Press,

Princeton, NJ.
Fung, R. and Crawford, S. (1990). Constructor: A system for the induction of probabilistic
models. Proceedings of the Eighth National Conference on AI, Boston, AAAI.
Furnival, G., and Wilson, R. (1974). Regression by leaps and bounds. Technometrics 16,
4990-5111
Geiger, D. (1990). Graphoids: A Qualitative Framework for Probabilistic Inference. Ph.D.

Thesis, University of California, Los Angeles.
Geiger, D., and Heckerman, D. (1991) Advances in probabilistic reasoning. Proc. Seventh
Conference on Uncertainty in AI, B. D'Ambrosio et al. (ed.). Morgan Kauman, Los
Angeles, California.
Geiger, D. and Pearl, J. (1989a). Logical and Algorithmic Properties of Conditional

Independence and Qualitative Independence. Report CSD 870056, R-97-IIL, Cognitive
Systems Laboratory, University of California, Los Angeles.
Geiger, D. and Pearl, J. (l989b). Axioms and Algorithms for Inferences Involving Conditional
Independence. Report CSD 890031, R-119-I, Cognitive Systems Laboratory, University
of California, Los Angeles.
Geiger, D., Verma, T., and Pearl, J. (1990) Identifying independence in Bayesian Networks.
Networks 20, 507-533.
Geisser, S. and Eddy, W. (1979). A predictive approach to model selection. JASA 74, 153-
160.
Geweke, J., Meese, R., and Dent, W. (1983). Comparing alternative tests of causality in
temporal systems. Journal of Econometrics 21,161-194.
Glymour, C. (1983). Social science and social physics. Behavioral Science 28, 126-133.
Glymour, C., Scheines, R., Spirtes, P., and Kelly, K. (1987). Discovering Causal Structure.
Acadernic Press, San Diego, CA.
Glymour, C., Scheines, R., and Spirtes, P. (1989). Why Aviators Leave the Navy:
Applications of Artificial Intelligence Procedures in Manpower Research. Report to the
Naval Personnel Research Development Center, January.
Glymour, C., Spirtes, P., and Scheines, R. (1991a). Independence relations produced by
parameter values. Philosophical Topics, v. 18, no. 2, Fall.
Glymour, C., Spirtes P., and Scheines, R. (1991b). From probability to causality.
Philosophical Studies, v. 64, no. 1, 1-36.
Gold, E. (1967). Language identification in the limit Information and Control 10, 447-474.
Gold, E. (1965). Limiting recursion. Journal of Symbolic Logic 30, 27- 48.
Goldberg, A., Duncan, O. (eds.) (1973). Structural Equation Models in the Social Sciences.
Seminar Press, New York.
Bibliography 503
Goodman, L. (1973a). Causal analysis of data from panel studies and other kinds of surveys.
Amer. J. Sociol. 78, 1135-1191.
Goodman, L. (1973b). The analysis of multidimensional contingency tables when some

variables are posterior to others: A modified path analysis approach. Biometrika 60, 179-
192.
Granger, C. (1969). Investigating causal relations by econometric models and cross-spectral

methods. Econometrica 37, 424-438.
Greenland, S. (1989). Modelling variable selection in epidemiologic analysis. American

Journal of Public Health 79, 340-349.
Griffiths, W., Hill, R., and Pope, P. (1987). Small sample properties of probit model
estimators. J ASA 82, 929-937.
Haberman, S.J. (1979). Analysis of Qualitative Data. Volume 2. Academic Press.
Harary, F., Norman R., and Cartwright, D. (1965). Structural Models: An Introduction to the
Theory of Directed Graphs. Wiley, New York.
Harary, F., and Palmer, E. (1973). Graphical Enumeration. Academic Press, New York.
Hausman, D. (1984). Causal priority. Nous 18,261-279.
Havranek, T. (1984). A procedure for model search in multi-dimensional contingency tables.

Biometrics 40, 95-100.
Heise, D. (1975). Causal Analysis. Wiley, New York.
Herskovits, E., and Cooper, G. (1990). Kutato: An entropy-driven system for construction of
probabilistic expert systems from databases. Proc. Sixth Conf. Uncertainty in AI.
Association for Uncertainty in AI, Inc., Mountain View, CA.
Herskovits, E. (1992).Computer Based Probabilistic-Network Construction. Ph.D Thesis,

Departments of Computer Science and Medicine, Stanford University.
Herting, J. and Costner, J. (1985). Respecification in multiple indicator models. Causal

Models in the Social Sciences, Blalock, H. (ed.). Aldine, NY.
Hocking, R. and Leslie, R. (1967). Selection of the best subset in regression analysis.
Technometrics 9, 531-540.
Holland, P. (1986). Statistics and causal inference. JASA 81, 945-960.
Holland, P. and Rosenbaum, P. (1986). Conditional association and unidimensionality in

monotone latent variable models. Ann. Stat. 14, 1523-1543.
Hosmer, D. and Lemeshow, S. (1989). Applied Logistic Regression. Wiley, New York.
Howson, C., and Urbach, P. (1989). Scientific Reasoning: The Bayesian Approach. Open
Court, La Salle, illinois.
Hsiao, C. (1981). Autoregressive modelling and money-income causality deduction. Journal of

Monetary Economics 7, 85-106.
James, L., Mulaik, S., and Brett, J. (1982). Causal Analysis: Assumptions, Models and Data.
Sage Publications, Beverly Hills, CA.
Jeffreys, H. (1957). Scientific Inference. Cambridge University Press, New York.
Joreskog, K. (1973). A general method for estimating a linear structural equation. Structural
Equation Models in the Social Sciences, Goldberger, A., and Duncan, O. (eds.). Seminar
Press, New York.
Joreskog, K. (1978). Structural analysis of covariance and correlation matrices. Psychometrika

43,443-447.
Joreskog, K. (1981). Analysis of covariance structures. Scandinavian Journal of Statistics 8,

65-92.
Joreskog, K. and Sorbom, D. (1984). LISREL VI User's Guide. Scientific Software, Inc.,
Mooresville, IN.
Bibliography 505
Joreskog, K. and Sorbom, D. (1990). Model search with TETRAD II and LlSREL.
Sociological Methods and Research 19,93-106.
Kadane, J. and Sedransk, N. (1980). Toward a more ethical clinical trial. Bayesian Statistics:
Proceedings of the First International Meeting Held in Valencia (Spain), May 28 to June 2,
1979, edited by J. Bernardo et al., University Press, 229-238.
Kadane, J. and Sedransk, N. (1992). Results of the clinical trial. Toward a More Ethical
Clinical Trial, J. Kadane (ed.), John Wiley & Sons, NY, forthcoming.
Kadane, J. and Seidenfeld, T. (1990). Randomization in a Bayesian Perspective. Journal of

Statistical Planning and Inference, 25, North-Holland, 329-345.
Kadane, J. and Seidenfeld, T. (1992). Statistical issues in the analysis of data gathered in the
new designs. Toward a More Ethical Clinical Trial, J. Kadane (ed.), John Wiley & Sons,
NY, forthcoming.
Kelley, T. (1928). Crossroads in the Mind of Man. Stanford University Press, Stanford.
Kendall, M. (1948). The Advanced Theory of Statistics. Charles Griffm and Co., London.
Kenny, D. (1979). Correlation and Causality. Wiley, New York.
Kiiveri H. (1982). A Unified Approach to Causal Models. Ph.D. thesis, Univ. of Western
Australia, in preparation.
Kiiveri, H. and Speed, T. (1982). Structural analysis of multivariate data: A review.

Sociological Methodology, Leinhardt, S. (ed.). Jossey-Bass, San Francisco.
Kiiveri, H., Speed, T., and Carlin, J. (1984). Recursive causal models. Journal of the
Australian Mathematical Society 36, 30-52.
Klein, L., (1961). An Econometric Model of the United Kingdom. Oxford University, Institute
of Statistics, Oxford.
Kleinbaum, D, Kupper, L., and Morgenstern, H. (1982). Epidemiologic Research. Lifetime

Learning Publications, Belmont, CA.
Klepper, S. (1988). Regressor diagnostics for the classical errors in variables model. Journal
of Econometrics 37, 225-250.
Klir, G., and Parviz, B. (1986). General reconstruction characteristics of probabilistic and
possibilistic systems. International Journal of Man-Machine Studies 25, 367-397.
Kullback, S. (1959, 1968). Information Theory and Statistics. Wiley, New York.
Ku11back, S. (1967). A lower bound for discrimination information in terms of variation. IEEE
Trans. Information Theory 13, 126-127.
Kullback, S. (1968). Probability densities with given marginal. Ann. Math. Statist. 39,79-86.
Kohn, M. (1969). Class and Conformity. Dorsey Press, Homewood, IL.
Lauritzen, S., Speed, T., and Vijayan, K. (1978). Decomposable Graphs and Hypergraphs.
Preprint 9, Institute of Mathematical Statistics, University of Copenhagen.
Lauritzen, S. and Wermuth, N. (1984). Graphical models for associations between variables,
some of which are qualitative and some quantitative. Ann. Stat 17,31-57.
Lawley, D., and Maxwell, A. (1971). Factor Analysis as a Statistical Method. Butterworth,
London.
Lazarsfeld, P., and Henry, N. (1968). Latent Structure Analysis. Houghton Mifflin, Boston.
Leamer, E. (1978). Specification Searches: Ad Hoc Inference with Non-experimental Data,

Wiley Series in Probability and Mathematical Statistics, John Wiley and Sons, NY.
Lee, S. (1985). Analysis of covariance and correlation structures. Computational Statistics and
Data Analysis 2, 279-295.
Lee, S. (1987). Model Equivalence in Covariance Structure Modeling. Ph.D. Thesis,

Department of Psychology, Ohio State University.
Lewis, D. (1973). Counterfactuals. Harvard University Press, Cambridge, MA.

Bibliography 507
Lewis, D. (1973). Causation. Journal of Philosophy 70, 556-572.
Lilienfeld, A. (1983). The Surgeon General's "Epidemiologic Criteria for Causality." A

Criticism of Burch's Critique. 36, 837-845.
Linthurst, R. A. (1979). Aeration, nitrogen, pH and salinity as factors affecting Spartina

Alterniflora growth and dieback. Ph.D. thesis, North Carolina State University.
Lohrnoller, J. (1989). Latent Variable Path Modeling with Partial Least Squares. Physica-
Verlag, Heidelberg.
Long, J. (1983a). Qualitative Applications in the Social Sciences. Vol. 33: Confirmatory
Factory Analysis. Sage Publications, Beverly Hills, CA.
Long, J. (1983b). Qualitative Applications in the Social Sciences. Vol. 34: Covariance
Structure Models. Sage Publications, Beverly Hills, CA.
Luijben, T., Boomsma, A., and Molenaar, I. (1986). Modification of factor analysis models in
covariance structure analysis. A Monte Carlo study. On Model Uncertainty and its
Statistical Implications. Lecture Notes in Economics and Mathematical Systems 307,
Dijkstra, T. (ed.). Springer-Verlag, Berlin.
MacCallum, R. (1986). Specification searches in covariance structure modeling. Psychological

Bulletin 100, 107-120.
Mackie, J. (1974). The Cement of the Universe. Oxford University Press, New York.
Mallows, C. (1973). Some comments on Cpo Technometrics 15, 661-676.
Mardia, K., Kent, J., and Bibby, 1. (1979). Multivariate Analysis. Academic Press, New
York.
Maruyama, G. and McGarvey, B. (1980). Evaluating causal models: An application of

maximum likelihood analysis of structural equations. Psychological Bulletin 87, 502-512.
McPherson, J., Welch, S., and Clark, C. (1977). The stability and reliability of political
efficacy: Using path analysis to test alternative models. American Political Science Review
71: 509-21.
Miller, R. Jr. (1981). Simultaneous Statistical Inference, 2nd ed. McGraw-Hill, New York.
Miller, W. and Stokes, D. (1963). Constituency influence in Congress. American Political

Science Review 1963, 45-456.
Miller, J., Slomczynski, K., and Schoenberg, R. (1981). Assessing comparability of

measurement in cross-national research: Authoritarian-conservatism in different
sociocultural settings. Social Psychology Quarterly 44, 178-191.
Mitchell, T. and Beauchamp, J. (1988). Baysian variable selection in linear regression. JASA
83, 1023-1032.
Mitchell, T. (1977). Version spaces: A candidate elimination approach to rule learning.

Proceedings Fifth International Joint Conference on AI. DCAI - 77, Pittsburgh, PA.
Mosteller, F., and Tukey, J. (1977). Data Analysis and Regression, A Second Coure in
Regression. Addison-Wesley, Massachusetts.
MRFIT Research Group (1982). Multiple risk factor intervention trial; risk factor changes and
mortality results. JAMA 248, 1465-1477.
Neopolitan, R. (1990) Probabilistic Reasoning in Expert Systems, Wiley, New York.
Neyman, J. (1935) Statistical problems with agricultural experimentation. J. Roy. Stat. Soc.
Suppl. 2, 107-180.
Osherson, D., Stob, T., and Weinstein, S. (1986). Systems That Learn. MIT Press,
Cambridge, MA.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems. Morgan and Kaufman, San
Mateo.
Bibliography 509
Pearl, J. and Dechter, R. (1989). Learning structure from data: A survey. Proceedings COLT
'89, 30-244.
Pearl, J., Geiger, D. and Verma, T. (1990). The logic of influence diagrams. Influence
Diagrams, Belief Nets and Decision Analysis. R. Oliver and J. Smith, editors. John Wiley
& Sons Ltd.
Pearl, J. and Tarsi, M. (1986). Structuring causal trees. Journal of Complexity 2, 60-77.
Pearl, J. and Verma, T. (1987). The Logic of Representing Dependencies by Directed Graphs.
Report CSD 870004, R-79-IT, University of California at Los Angeles Cognitive Systems
Laboratory.
Pearl, J. and Verma, T. (1990). A Formal Theory ofInductive Causation. Technical Report R-
155, Cognitive Systems Labratory, Computer Science Dept. UCLA.
Pearl, J. and Verma, T. (1991). A theory of inferred causation. Principles of Knowledge

Representation and Reasoning: Proceedings of the Second International Conference,
Morgan Kaufmann, San Mateo, CA.
Pratt, J. and Schlaifer, R. (1988). On the interpretation and observation of laws. Journal of
Econometrics 39, 23-52.
Putnam, H. (1965). Trial and error predicates and a solution to a problem of Mostowski.
Journal of Symbolic Logic 30, 49-57.
Rawlings, J. (1988). Applied Regression Analysis. Wadsworth, Belmont, CA.
Reichenbach, H. (1956). The Direction of Time. Univ. of California Press, Berkeley, CA.
Reiss, 1., Banwart, A., and Forman, H. (1975). Premarital contraceptive usage: A study and
some theoretical explorations. J. Marriage and the Family 37, 619-630.
Risannen, J. (1987). Stochastic complexity. J. Roy. Statist. Soc. Ser. B 49, 223-239.
Rindfuss, R., Bumpass, L., and St. John, C. (1980). Education and fertility: Implications for
the roles women occupy. American Sociological Review 45, 431-447.
Rodgers, R. and Maranto, C. (1989). Causal models of publishing productivity in psychology.

Journal of Applied Psychology 74, 636-649.
Rose, G. Hamilton, P. Colwell, L., and Shipley, J. (1982). A randomised controlled trial of
anti-smoking advice: lO-year results. Journal of Epidemiology and Community Health 36,
102-108.
Rosenbaum, P. (1984). From association to causation in observational studies. JASA 79, 41-
48.
Rubin, D. (1974). Estimating causal effects of treatments in randomized and nonrandomized

studies. Journal of Educational Psychology 66, 688-701.
Rubin, D. (1977). Assignment to treatment group on the basis of a covariate. Journal of

Educational Statistics 2, 1-26.
Rubin, D. (1978). Bayesian inference for causal effects: The role of randomizations. Ann. Stat.
6,34-58.
Rubin, D. (1986). Comment: Which ifs have causal answers. JASA 81, 396.
Salmon, W. (1980). Probabilistic causality. Pacific Philosophical Quarterly 61, 50-74.
Salmon, W. (1984). Scientific Explanation and the Causal Structure of the World. Princeton
Univ. Press, Princeton, NJ.
Saris, W. and Stronkhorst, H. (1984). Causal Modeling in Nonexperimental Research.

Sociometric Research Foundation, Amsterdam.
Scheines, R. (1988). Automating creativity. Aspects of Artificial Intelligence, J. Fetzer (ed.).

Kluwer, Boston.
Scheines, R., and Spirtes, P. (1992). Finding latent variable models in large data bases.
International Journal of Intelligent Systems, G. Piatetski-Shapiro, (ed.).
Bibliography 511
Scheines, R., Spirtes, P., Glymour, G., and Sorensen, S. (1990). Causes of Success and
Satisfaction Among Naval Recruiters. Report to the Navy Personnel Research Development
Center, San Diego, CA
Scheines, R., Spirtes, P., and Glymour, C. (1990). A qualitative approach to causal
modeling. Qualitative Simulation Modeling and Analysis, Fishwick, P. and Luker, P.
(eds.). Advances in Simulation 5, 72-97. Springer-Verlag, New York.
Sclove, S. (undated). On Criteria for Choosing a Regression Equation for Prediction. Technical
Report 28, Department of Statistics, Carnegie-Mellon University.
Sewell, W. and Shah, V. (1968). Social class, parental encouragement, and educational
aspirations. American Journal of Sociology. 73, 559-572.
Simon, H. (1953). Causal ordering and identifiability. Studies in Econometric Methods. Hood
and Koopmans (eds). 49-74.Wiley, NY.
Simon, H. (1954).Spurious correlation: a causal interpretation. JASA. 49, 467-479.
Simon, H. (1977). Models of Discovery. D. Reidel, Dordrecht
Simpson, C. (1951). The interpretation of interaction in contingency tables. J. Roy. Statist.

Soc. Ser. B 13, 238-241.
Sims, C. (1972). Money, income, and causality. American Economic Review 62, 540-552.
Skyrrns, B. (1980). Causal Necessity: A Pragmatic Investigation of the Necessity of Laws.

Yale University Press, New Haven.
Sober, E. (1987). The principle of the common cause. Probability and Causality, Fetzer, J.
(ed.). D. Reidel, Dordrecht.
Sorbom., D. (1975). Detection of correlated errors in longitudinal data. British Journal of

Mathematical and Statistical Psychology 28,138-151.
Spearman, C. (1904). General intelligence objectively determined and measured. American

Journal of Psychology 15,201-293.
Spence, M. (1973). Job market signalling. Quarterly Journal of Economics 87, 355-379.
Spiegelhalter, D. (1986). Probabilistic reasoning in predictive expert systems. Uncertainty in

Artificial Intelligence, Kanal, K. and Lemmer, 1. (eds.). North-Holland, Amsterdam.
Spiegelhalter, D., and Knell-Jones, R. (1984). Statistical and knowledge-based approaches to

clinical decision-support systems. J. Royal Statist. Soc. Ser. A 147,35-77.
Spirtes, P. (1989a). A Necessary and Sufficient Condition for Conditional Independencies to

Imply a Vanishing Tetrad Difference. Technical Report CMU-LCL-89-3, Laboratory for
Computational Linguistics, Carnegie Mellon University, Pgh, PA.
Spirtes, P. (1989b). Fast Geometrical Calculations of Overidentifying Constraints. Technical

Report CMU-LCL-89-3, Laboratory for Computational Linguistics, Carnegie Mellon
University, Pgh, PA.
Spirtes, P. (1992). Building causal graphs from statistical data in the presence of latent
variables", forthcoming in Proceedings of the IX International Congress on Logic,
Methodology, and the Philosophy of Science, B. Skyrms, ed., Uppsala, Sweden, 1991.
Spirtes, P. and Glymour, C. (1988). Latent variables, causal models and overidentifying
constraints. Journal of Econometrics 39,175-198.
Spirtes, P. and Glymour, C. (1990). Causal Structure Among Measured Variables Preserved
with Unmeasured Variables. Technical Report CMU-LCL-90-5, Laboratory for
Computational Linguistics, Camegie Mellon University.
Spirtes, P., Glymour, C., and Scheines, R. (1990b). Causality from probability. Evolving
Knowledge in Natural Science and Artificial Intelligence, Tiles, J. et. al. (eds.). Pitman,
London, 181-199.
Spirtes P., Glymour C., and Scheines, R. (1990c). Causality from probability. Conference
Proceedings: Advanced Computing for the Social Sciences, Williamsburgh, V A.
Spirtes, P., Glymour, C., and Scheines, R. (1991a). An algorithm for fast recovery of sparse
causal graphs. Social Science Computer Review, vol. 9, 1991, pp. 62-72.
Bibliography 513
Spirtes P., Glymour C., and Scheines, R. (1991b). From probability to causality.
Philosophical Studies, 64, 1-36.
Spirtes P., Glymour C., Scheines, R., Meek, C., Fienberg, S., and Slate, E. (1992).
Prediction and Experimental Design with Graphical Causal Models. Technical Report
CMU-Phil-32, Philosophy Department, Carnegie-Mellon University.
Spirtes, P., Glymour, C., Scheines, R., and Sorensen, S. (1990). TETRAD Studies of Data
for Naval Air Traffic Controller Trainees. Report to the Navy Personnel Research
Development Center, San Diego, CA.
Spirtes, P., Scheines, R., and Glymour, C. (1990a). Simulation studies of the reliability of
computer aided specification using the TETRAD II, EQS, and LISREL Programs.
Sociological Methods and Research, 19,3-66.
Spirtes P., and Verma, T. (1992). Equivalence of Causal Models with Latent Variables.
Technical Report CMU-Phil-33, Philosophy Department, Carnegie-Mellon University.
Spohn, W. (1983). Determinisitc and probabilistic reasons and causes. Methodology,

Epistemology, and Philosophy of Science: Essays in Honour of Wolfgang Stegmuller on
the Occasion of his 60th Birthday, C. G. Hempel, H. Putnma, and W.K. Essler (eds.), D.
Reidel, Dordrecht, Holland, 371-396.
Spohn, W. (1990). Direct and indirect causes. Topoi, 9, 125-145.
Spohn, W. (1991). On Reicenbach's principle of the common cause. Proceedings of the First
Pittsburgh-Konstanz Colloquium, W. Salmon, G. Walters (eds.).
Spohn, W. (1992). Causal laws are objectifications of inductive schemes. Theory of

Probability, J. Dubucs (ed.), Kluwer, Dordrecht, Holland.
Stein, C. (1960). Multiple regression. Contributions to Probability and Statistics. Essays in

Honor of Harold Hotelling, aikin, I. (ed.). Stanford Univ. Press, Stanford, CA.
Stetzl, I. (1986). Changing causal relationships without changing the fit: Some rules for
generating equivalent LISREL models. Multivariate Behavior Research 21, 309-331.
Suppes, P. (1970). A Probabilistic Theory of Causality. North-Holland, Amsterdam.
Suppes, P. and Zanotti, M. (1981). When are probabilistic explanations possible. Synthese 48,
191-199.
Surgeon General of the United States (1964). Smoking and Health. U.S. Government Printing
Office.
Surgeon General of the United States (1979). Smoking and Health. U.S. Government Printing
Office.
Swamy, P. (1971). Statistical Inference in Random Coefficient Regression Models. Springer-

Verlag, Berlin.
Thomson, G. (1916). A hierarchy without a general factor. British Journal of Psy,cho1ogy 8,

271-28l.
Thomson, G. (1935). On complete families of correlation coefficients and their tendency to

zero tetrad-differences: Including a statement of the sampling theory of abilities. British
Journal of Psychology 26, 63-92.
Thurstone, L. (1935). The Vectors of Mind. Univ. of Chicago Press, Chicago.
Timberlake, M. and Williams, K. (1984). Dependence, political exclusion, and government

repression: Some cross-national evidence. American Sociological Review 49, 141-146.
Verma, T. (1987). Causal networks: semantics and expressiveness. Technical Report R-65-I,
Cognitive Systems Laboratory, University of California, Los Angeles.
Verma, T. and Pearl, J. (1990a). On Equivalence of Causal Models. Technical Report R-150,
Cognitive Systems Laboratory, University of California, Los Angeles.
Verma, T. and Pearl, J. (1990b). Equivalence and synthesis of causal models. Proc. Sixth
Conference on Uncertainty in AI. Association for Uncertainty in AI, Inc., Mountain View,
CA, 220-227.
Bibliography 515
Verma, T. and Pearl 1. (1990c). Causal networks: semantics and expressiveness. Uncertainty
in Artificial Intelligence 4, R. Shacter, T. Levitt, L. Kanal, J. Lerner (eds.) Elsevier Science
Publishers, North-Holland.
Verma, T. and Pearl, 1. (1991). Equivalence and synthesis of causal models. Technical Report
R-150, Cognitive Systems Laboratory, University of California, Los Angeles.
Wallace, C. and Freeman, P (1987). Estimation and inference by compact coding. J. Roy.
Statist. Soc. Ser. B 49, 240-265.
Weisberg, S. (1985). Applied Linear Regression, 2nd ed. Wiley, New York.
Wermuth, N. (1976). Model search among multiplicative models. Biometrika 32,253-363.
Wermuth, N. (1980). Linear recursive equations, covariance selection and path analysis. JASA
75, 963-972.
Wermuth, N. and Lauritzen, S. (1983). Graphical and recursive models for contingency tables.
Biometrika 72,537-552.
Wermuth, N. and Lauritzen, S. (1990). On substantive research hypotheses, conditional

independence graphs and graphical chain models. J. Roy. Statist. Soc. Ser. B 52, 21-50.
Wheaton, B., Muthen, B., Alwin, D., and Summers, G. (1977). Assessing Reliability and
Stability in Panel Models. Sociological Methodology 1977, Heise, D. (ed.). Jossey-Bass,
San Francisco.
Whitney, H. (1957). Elementary structures of real algebraic varieties. Ann. Math. 66,545-556.
Whittaker, J. (1990). Graphical Models in Applied Multivariate Statistics. Wiley, New York.
Wise, D. (1975). Academic achievement and job performance. American Economic Review 65,
350-366.
Wishart, J. (1928). Sampling errors in the theory of two factors. British Journal of
Psychology, Vol. 19, pp. 180-187.
Wright, S. (1934). The method of path coefficients. Ann. Math. Stat. 5, 161-215.
Younger, M. (1978). Handbook for Linear Regression. Wadsworth, California.
Yule, G. (1903). Notes on the theory of association of attributes in statistics. Biometrika 2,

121-134.
Yule, G. (1926). Why do we sometimes get nonsensical relations between time-series? -- A

study in sampling and the nature of time series. Journal of the Royal Statistical Socitey, 89,
1-64.
Index
2x2 Foursomes 312-314 Bootstrap Procedures 365

3xl Foursomes 314 Bouck, L. ii
Acceptable Ordering 221 Bowden, R. 2
Acyclic Path 29 Breslow, N. 164
Adjacency 29 Brownlee, K. 17,291,297-298,302
Adjacent 28 Bumpass, L. 139-140
Aitkin, M. 105 Bunker, J. 3
Almost Pure Latent Variable Graph 309 Burch, P. 291, 298-300, 301, 302
Almost Pure Measurement Model 309, 473 Byron, R. 104
Ancestor 31 Califano, 1. 291
Anderson, T. 128 Callahan, 1. ii
Asmussen, S. 4 Cambell, D. 338
Asymptotically Reliable iv Carlin, J. i, 40, 162
Axioms Cartwright, N. 96
Causal Markov Condition 54-55, 57-64 Causal Chain 44
Statement 54 Causal Graph 47
Causal Minirnality Condition 55~56 Causal Inference Algorithm 181-184
Statement 55 Statement 183
Faithfulness 35-36 Causal Markov Condition 54-55, 57-64
Faithfulness Condition Statement 54
Statement 56 Causal Mediary 43
Markov 33 Causal Minirnality Condition 55-56
Basmann, R. 5,101 Statement 55
Bayesianism 23, 68, 70, 277-286, 323 Causal Representation Convention 47
Search Strategies 109-111 Causal Structure 45
Bell, J. 63 Deterministic 49
Bentler, P. 104, 140-142,323 Generates Probability Distribution 50,
Bibby, J. 197 54
Birch, M. 105, 107 Generation 45
Bishop, E. 39 Genuinely Indeterministic 53
Bishop, Y. 4, 105 Indeterministic 51
Blalock, H. i, 6, 75, 86 Isomorphism 45
Blau, P. 142-144 Linear Deterministic 49
Blyth, C. 267 Linear Pseudo-indeterministic 52
Bollen, K. i, 107, 195,325 Causal System
Causally Sufficient 45 Cooper, G. i, II, 110, 146,323

Causally Connected 45 Cornfield, J. 291, 293-296
Causally Sufficient 45 Costner, H. i, 86, 338
Causation Covariance Structures 104
Reciprocal 354 Crawford,S. 108, 125
Cause Cross-Construct Foursomes 314
Boolean 43 Cross-Construct Impure 309
Common 44, 47 Cross-Validation 365
Direct 43 Cycle
Indirect 44 Chordless 361
Representation 47 Cyclic 29
Scaled Variable 44 Cyclic Directed Graph 354-360
Cederlof, R. 298 d-connection 36, 74,83
Child 28 D-SEP 176
Children 28 d-separation 36, 73-75,82,83,113
Choke Point 196 Cyclic Graph 359
Chordless cycle 361 Faithful 113
Christensen, R. ii, 23, 105, 150 Order 316
CI (see Causal Inference Algorithm) DAG-I~omorph 36
CI Partially Oriented Inducing Path Graph Darroch, J. 39
190 Data Sets
Clique 31 Abortion Opinions 150
Maximal 31 AFQT 243-244
Coleman, J. 147-149 Alarm Network 11, 145-147
Collapsibility 3-4 American Occupational Structure 142-
Collider 31, 181 145
Combined Graph 78 College Plans 149-150
Common Cause 44, 47 Education and Fertility 139-140
Common Cause Impure 309 Female Orgasm 140-142
Complete Graph 30 Leading Crowd 147-149
Concatenation 29 Mathematical Marks 197-200
Conditional 41-42 Political Exclusion 248-250
Conditional Independence 32 Publishing Productivity 13-14, 133-
Conditional Independence Tests 366 138, 166
Connected Graph 30 Rat Liver 252-257
Constraints Simulation Studies 161,250-251,320,
Non-independence 191-193 332,351
Index 519
Results 154-161 Discovery 103-104

Sample Generation Discovery Problems 103-104
Spartina grass 244-248 Doll,R.291
Daughter 29 Druzdel, M. ii
Davis, W. 61 Duncan, O. 142-144
Day, N. 164 Earman, J. ii
Definite d-connection 189 Edge 28, 29
Defmite Discriminating Path 181 Directed 28
Definite Non-Collider 180,224,319 Edge-end 28
Definite-Non-Descendant 221 Endpoint 28
Defmite-SP 223 Into 28
Degree 29 Outof28
Dempster, A. 104 Edge-end 28
Descendant 31 Edwards, A. 128
Descendants 33 Edwards, D. 4, 109
Det83 Empty Path 29
Deterministic 83 Endogenous Variable 49
Deterministic Causal Structure 49 Endpoint 28
Deterministic Graph 38 EQS 104, 109, 138, 143, 144,323
Deterministic System 38 Automatic Model Modification 329,
Deterministic Variable 82 332
Direct Cause 43 Boldness 346
Direct Manipulation 78 Lagrange Multiplier Statistic 331
Directed Acyclic Graph 30, 32 Reliability 346
DAG-Isomorph 36 Equiv 177
I-map 33 Error
Linear Implication 196 Edge Direction of Commision 146
Minimal I-map 34 Edge Direction of Omission 146
Perfect Map 36 Edge Existence of Commision 146
Represents Probability Distribution 34 Edge Existence of Omission 146
Directed Edge 28 Error Variable 37
Head 30 Estimation 132
Tail 30 Multinomial Distributions 132
Directed Graph 25, 26, 28 Ethics
Acyclic 30 Experimental Design 286-290
Directed Independence Graph 34-35, III Exogenous Variable 49,78
Directed Path 25, 29 Expanded Graph 37
Experimental design 17,23,277-286 Glymour, C. 86, 200

Ethical 286-290 Gold,M.162
Prospective 274 Granger Causality 363-365
Randomized 261 Graph 25, 28
Retrospective 274 Adjacent 28
F.I. (see Faithful Indistinguishability) Causal 47
Factor Analysis 107, 198 Clique 31
Faithful Indistinguishability 89-90, 361- Combined 78
363 Complete 30
Faithfulness 36 Connected 30
Faithfulness Condition 9, 22, 35, 66 Cyclic Directed 354-360
Statement 56 d-connection 74,83
Fast Causal Inference Algorithm 184 d-separation 73, 83
Statement 188 Deterministic 38
FCI (see Fast Causal Inference Algorithm) Directed 25, 26
FCI Partially Oriented Inducing Path Directed Acyclic 30,32
Graph 190 Directed Independence 34-35, III
Feedback 354-360 Edge 28
Fienberg, S. i, 4, 39, 86, 105, 106, 147, Expanded 37
237 Faithful Representation of D-
Fisher, F. 10 I separations 113
Fisher, R. iii, 17,259,260,291,292-293, Inducing Path 25,174-177
302 Manipulated 78
Forrest, W. 3 Over 28
Foursomes Parallel Embedding 94
Cross-Construct 310 Partially Oriented Inducing Path 25,
Intra-Construct 310 177, 181
Fox, 1. 238 Pattern 27, 90, 113
Freedman, D. 142, 365 Subgraph 30
Friberg, L. 298 Trek 97,196
Friedman, M. 126 Trek Sum 97
Functional Determination 82 Triangle 31
Fung, R. 108, 125 Underlying 37
Geiger, D. i, 48, 71,86,224 Undirected 25
Generation 45, 50, 54 Undirected Independence 37-38, 108
Genuinely Indeterministic Causal Structure Unmanipulated 78
53 Haenszel, W. 291
Index 521
Hammond, E. 291 Intersect 29

Hausman, D. ii Into 28, 181
Havranek, T. 109 Intra-Construct Foursomes 310-311
Head 30 Intra-Construct Impure 308, 311
Heckerman, D. 48 Invariance 208,461,486
Heise, D. 75, 355 1P208
Henry, N. 86 Isomophic Causal Structures 45
Herskovits, E. 11, 110, 146,323 N208
Hierarchical Models 106 Jacknife Procedures 365
Hill, A. 291 Jeffreys, H. 162
Hill, J. ii Joint Independence 32
Holland, P. i, 4, 8-9, 39, 105, 164 Joreskog, K. 104, 107, 132
Howson, C 260 Kadane,J. ii, 261,277-286
I-map 33 Kelly, K. 86,200
IG Algorithm 124-125 Kelly, T. 164
Implication Kendall, M. 64-66, 67
Linear 37 Kent, J. 197
Impure Indicators Kiiveri, H. i, 7, 39, 132, 162
Common Cause 309 Klein, L. 163
Cross-Construct 309 Kleinbaum, D. 164
Intra-Construct 308, 311 Klepper, S. ii, 126
Latent-Measured 308 Klir, G. 105
Indegree 29 Kohn, M. 307, 333
Independence 32 Kullback, S. 106
Indeterministic Causal Structure 51 Kupper, L. 164
Indeterministic System 38 Lagrange Multiplier Statistic 331
Indirect Cause 44 Latent Variable Graph
Indistinguishability Almost Pure 309
Faithful 361-363 Latent Variable Models
Indistinguishable Cross-Construct Foursome 310
Linearly Faithful 361 Intra-Construct Foursome 310
Inducing Path Graph 25, 26, 174-177 Latent Variables 21, 306
D-separability 176 Causal structure among 19
SP223 Latent-Measured Impure 308
Informative Parents 208 Lauritzen, S. i, iv, 7, 37, 39, 111
Informative Variables 208 Lazarsfeld, P. 86
Instrumental Variable 101 Leamer, E. 109
Lee, S.101 Marginal 32

Lilienfeld, A. 291, 293, 300-301, 302 Markov Condition 7,8,9,22,33,359-360
Linear Deterministic Causal Structure 49 Maruyama, G. 338
Linear Faithfulness 74 Mason's Theorem 355
Linear Implication 37, 196 Mason, S. 355
Linear Model 196 Maximal Clique 31
Linear Pseudo-Indeterministic 39 Maximally Oriented Partially Oriented
Linear Pseudo-indeterministic Causal Inducing Path Graph 180
Structure 52 Maximum Likelihood Estimation
Linear Pseudo-Indeterministic Models 307 Fitting Function 330
Linear Regression 101, 108 McGarvey, B. 338
Linear Representation 36 McPherson J., 333
Linear Statistical Indistinguishability 94, Measurement Model
99 Almost Pure 309, 473
Linearly Faithful Indistinguishable 361 Meek, C. i, 86
Linthurst, R. ii, 16,244-248 Meyer,M. ii
LISREL 104,107,109,143,323 MIMBuild Algorithm
Automatic Model Modification 329- Complexity 320
332 Reliability 320
Boldness 346 Simulation Study Results 320
Modification Indices 331 Minimal I-map 34
Reliability 346 MINITAB240
Log-linear models 4, 104, 105-107 Mitchell, T. 109
Hierarchical Models 106, 108 Mixture 58
Representing Colliders 106 Model selection 4
Lohmoller, J. 132 Model Specification 365
Lundman, T. 298 Modification Indices 331
Lung Cancer 291-302 Modified PC Algorithm 165
Manipulated 78 Statement 166
Manipulated Graph 78 Morgenstern, H. 164
Manipulation 7-9,22,78 Mosteller, F. 3, 202, 239
Direct 78 Multiple Risk Factor Intervention Tial
Invariance 208, 461, 486 Research Group 301
Manipulation Theorem 9, 75-81 National Halothane Study 3
Statement 79 Navidi, W. 365
Maranto, C. 13, 133-138, 166 ND208
Mardia, K. 197 Neyman, J. 7, 291
Index 523
No Descendants 208 Path

Node Acyclic 29
Visit 327 Adjacency 29
Noncollider 31 Collider on 31
Notational Conventions viii Concatenation of 29
Observed 173 Cyclic 29
Order Definite Discriminating 181
d-separability 316 Directed 25, 29
Osherson, D. 162 Empty 29
Out of 28, 181 Intersect 29
Outdegree 29 Into 29
Over 28 Noncollider on 31
Papineau, D. ii Out of 29
Parallel Embedding 94 Point of Intersection 29
Parameter Estimation 365 Undirected 25, 29
Parent 25, 28,181 Unshielded Collider on 31
Parents 28 Pattern
Partially Oriented Inducing Path Graph 25, Definite Non-Collider 319
26, 177, 181 Output 146
Acceptable Ordering 221 True 146
CI190 Pattern Graph 27, 90, 113
Collider 181 PC Algorithm 116-122
Definite d-connection 189 Applications 132
Definite Discriminating Path 181 Discrete Distributions 147
Definite Non-Collider 180, 224 Linear Structural Equation Models
Definite-SP 223 133
FCI190 Applied to Latent Variables 316
Into 181 Complexity 119-120
Maximally Oriented 180 Stability 120-122
Out Of 181 Statement 117,118
Parent 181 PC* Algorithm 122-123
Possible-D-SEP 187 Heuristics 123-124
Possible-SP 223 Statement 123
Possibly d-connecting 224 Pearl, J. i, iv, 36, 40, 68,71,86, 101, 124,
Possibly-IP 224 126,162,200,224,258,260
Possibly-IV 224 Pearson, K. 21
Semi-Directed Path 30, 190 Peeler, W. 140-142
Perfect Map 36 Recursive Diagram 111

Peters, S. 365 Regression 1-2,238-241,363
PMan 78 Logistic 105
Point of Intersection 29 Selection of Regressors 14-17
Policy Variable 78 Stepwise 105
Political Exclusion 248-250 Reichenbach, H. 6
Population Representation
Manipulated 76 Linear 36
Unmanipulated 76 Represents 34
Possible-D-SEP 187 Retrospective Design 274
Possible-SP 223 Rigid Statistical Indistinguishability 94
Possibly D-connecting 224 Rindfuss, R. 139-140
Possibly-IP 224 Rodgers, R. 13, 133-138, 166
Possibly-IV 224 Rose,G.300
Power 254 Rosenbaum, P. 164
Pratt, J. i, iv, 8-9, 22, 202, 203-212, 238 Rubin, D. i, iv, vi, 8,22,202,203-212,237
Predictable 213 S.S.I. (see Strong Statistical
Prediction Algorithm 227 Indistinguishability
Examples 227-237 Salmon, W. 62-63
Statement 221-222 Sampling 272-276
Probability 31 SAS 245
Conditional Independence 32 Scheines, R. 86, 200
Independence 32 Schlaifer, R. i, iv, 8-9, 22, 202, 203-212,
Interpretations 5 238
Probability Distribution Schoenberg, R. 338
Generated by Causal Structure 50, 54 Scott, D. ii
Prospective Design 274 Search Algorithms
Pseudo-Indeterministic Causal Structure Incorporating Background Knowledge
52 127
Pseudo-Indeterministic System 38, 39 Probabilities of Error 130-132
Pseudocorrelation Matrix 340 Statistical Decisions 128
PUnman 78 Variable Selections 125
Putnam, H. 162 Search Strategies 104-111
R.S.1. (see Rigid Statistical Sedransk,~.277-286
Indistinguishability) Seidenfeld, T. ii, 261, 277-286

Rawlings, J. 2, 16,244-248 Semi-Directed Path 30, 190
Reconstructability Analysis 105 Seneca, E. ii
Index 525
Sewell, W. 149-150 G2129

SGS Algorithm ll4-116 Vanishing Partial Correlation 128
Complexity ll5 Vanishing Tetrad Difference 325
Stability ll5 X 2 129
Statement ll4 Stepwise Regression 105
Shah, V. 149-150 Stetzl, I. 5, 101
Shimkin, M. 291 Stob, T. 162
Simon, H. i, 6, 75, 86 Strong Statistical Indistinguishability 88
Simpson's Paradox 64-68, 92, 267, 273 Structural Equation Models 307
Simpson, C. 64, 67-68, 267 Indistinguishability 5
Simultaneous Equation Model 101 Measurement Model 307
Skyrms, B. ii Structural Model 307
Slate, E. i, 86 Subgraph 30
Slezak, P. ii Suppes, P. 6, 101, 164
Smoking 291-302 Surgeon General's Report on Smoking and
Surgeon General's Report 291, 296- Health 291
297,298,302 Tai130
Sober, E. 63 Tetrad Differences 325
Sorbom, D. lO4, 107 Vanishing 194
Sorensen, S. ii Variance of Sampling Distribution 325
Source 34 TETRAD II 10, 127, 132, 138, 144, 146
SP223 Search Procedure 324
Spearman, C. i, 86,162,164,200 Boldness 346
Specification Searches 4 Implied_H 326
Error Probabilitites 252-257 Imp1iedH 326
Speed, T. i, iv, 7, 39, 132, 162 Limitations 351
Spirtes, P. 86, 200 Reliability 346
St. John,C. 139-140 Scoring Principles 324-326
Stationary 363 T-maxscore 327
Statistical Indistinguishability 21, 87 Tetrad-Score 326
Faithful 89-90 Weight 326
Linear 99 Tetrad Representation Theorem 22, 196-
Rigid 94 197
Strong 88 Statement 196, 315
Weak 90-93 Thomson, G. 200
Statistical Tests Thurstone, L. 107, 162, 198
Degrees of Freedom 129 Timberlake, M. 248-250
Time Series 356-358, 363-365 Verma, T. i, 71, 86, 101, 124, 162, 191,
Trek 34,97,196 200,224,258,260
Source 34 Version Spaces 109
Trek Sum 97 Vertex
Triangle 31 Ancestor 31
Tukey, J. 202, 239 Child 28
Turkington, D. 2 Children 28
Type 1 Error 254 Daughter 29
Type 2 Error 254 Degree 29
Underdetermination 91 Descendant 31
Underlying Graph 37 Descendants 33
Undirected Graph 25,28 Indegree29
Undirected Independence Graph 37-38, Outdegree 29
108 Parent 28
Algorithm 124 Parents 28
Undirected Path 25, 29 Vijayan, K. 39
Unmanipu1ated Graph 78 Visit 327
Unshielded Collider 31 W.F.I. (see Weak Faithful
Urbach, P. 260 IndistingJIishability)
Vandam,L.3 W.S.I. (see Weak Statistical
Vanishing Tetrad Differences 194, 310, Indistinguishability
314,315 Weak Faithful Indistinguishability 91
Linear Implication 196 Weak Statistical Indistinguishability 90-93
Variable Weinstein, S. 162
Determined 81 Weisberg, S. 252-257
Deterministic 82 Wermuth, N. i, iv, 7, 37, 40,104,111
Endogenous 49 Wermuth-Lauritzen Algorithm 111-112
Error 37 Wheaton, B. 333
Exogenous 49 Whittaker, J. 4,105,111,125,197
Functionally Determined 82 Williams, K. 248-250
Instrumental 101 Wishart, J. 325
Observed 173 Worrall, J. ii
Policy 78 Wright, S. 6, 75, 86
RedefIning 99-101 Wynder, E. 291
Variable Aggregation 125 Younger, M. 2
Variable RedefInition 99-10 1 Yule, G. 3, 21
Variable Selection 271-272 Zanotti,M. 101,164
Vol. 1: RA. Fisher: An Appreciation. Edited by S .E. Fienberg and Vol. 23: D.G. Saphire, Estimation of Victimization Prevalence
D.V. Hinkley. XI, 208 pages, 1980. Using Data from the National Crime Survey. V, 165 pages, 1984.
Vol. 2: Mathematical Statistics and Probability Theory. Proceed- Vol. 24: T.S. Rao, M.M. Gabr, An Introduction to Bispectral
ings 1978. Edited by W. Klonecki, A. Koze.k, and J. Rosinski. Analysis and Bilinear Time Series Models. VIII, 280pages, 1984.
XXIV, 373 pages, 1980.
Vol. 25: Time Series Analysis of Irregularly Observed Data.
Vol. 3: B.D. Spencer, Benefit-Cost Analysis of Data Used to Proceedings, 1983. Edited by E. Parzen. VII, 363 pages, 1984.
Allocate Funds. VIII, 296 pages, 1980.
Vol. 26: Robust and Nonlinear Time Series Analysis. Proceed-
Vol. 4: E.A. van Doom, Stochastic Monotonicity and Queueing ings, 1983. Edited by J. Franke, W.HiirdleandD. Martin. IX,286
Applications of Birth-Death Processes. VI, 118 pages, 1981. pages, 1984.
Vol. 5: T. Rolski, Stationary Random Processes Associated with Vol. 27: A. Janssen, H. Milbrodt, H. Strasser, InfInitely Divisible
Point Processes. VI, 139 pages, 1981. Statistical Experiments. VI, 163 pa8es, 1985.
Vol. 6: S.S. Gupta andD.-Y. Huang, Multiple Statistical Decision Vol. 28: S. Amari, Differential-Geometrical Methods in Statistics.
Theory: Recent Developments. VIII, 104 pages, 1981. V, 290 pages, 1985.
Vol. 7: M. Akahira and K. Takeuchi, Asymptotic Efficiency of Vol. 29: Statistics in Ornithology. Edited by B.J.T. Morgan and
Statistical Estimators. VIII, 242 pages, 1981. P.M North. XXV,418 pages, 1985.
Vol. 8: The First Pannonian Symposium on Mathematical Statis- Vol 30: J. Grandell, Stochastic Models ofAir PollutantConcentra-
tics. Edited by P. Reve.z, L. Schmetterer, and V.M. Zolotarev. VI, tion. V, 110 pages, 1985.
308 pages, 1981.
Vol. 31: J. pfanzagl, Asymptotic Expansions for General Statisti-
Vol. 9: B. Jllrgensen, Statistical Properties of the Generalized cal Models. VII, 505 pages, 1985.
Inverse Gaussian Distribution. VI, 188 pages, 1981.
Vol. 32: Genera1izedLinear Models. Proceedings, 1985. Edited by
Vol. 10: A.A. McIntosh, FittingLinear Models: An Application of R Gilchrist, B. Francis and J. Whittaker. VI, 178 pages, 1985.
Conjugate Gradient Algorithms. VI, 200 pages, 1982.
Vol. 33: M. CsBrgo, S. CsBrgo,L Horvath, AnAsymptoticTheory
Vol. 11: D.F. Nicholls and B.G. Quinn, Random Coefficient for Empirical Reliability and Concentratioo Processes. V, 171
Autoregressive Models: An Introduction. V, 154 pages, 1982 pages, 1986.
Vol. 12: M. Jacobsen, Statistical Analysis of Counting Processes. Vol. 34: D.E. Critchlow, Metric Methods for Analyzing Partially
VII, 226 pages, 1982 Ranked Data. X, 216 pages, 1985.
Vol. 13: J. pfanzagl (with the assistance of W. Wefe1meyer), Vol. 35: Linear Statistical Inference. Proceedings, 1984. Edited by
Contributions to a General Asymptotic Statistical Theory. VII, T. Calinski and W. Klooecki. VI, 318 pages, 1985.
315 pages, 1982
Vol. 36: B. Matern, Spatial Variation. Second Edition. 151 pages,
Vol. 14: GUM 82: Proceedings oftheIntemational Cooference on 1986.
Generalised Linear Models. Edited by R Gilchrist. V, 188 pages,
1982 Vol. 37: Advances in Order Restricted Statistical Inference. Pro-
ceedings, 1985. Edited by R. Dykstra, T. Robertson and F.T.
Vol. 15: K.R.W. Brewer and M. Hanif, Sampling with Unequal Wright. VIII, 295 pages, 1986.
Probabilities. IX, 164 pages, 1983.
Vol. 38: Survey Research Designs: Towards a Better Understand-
Vol. 16: Specifying Statistical Models: From Parametric to Non- ing of Their Costs and Benefits. Edited by R W. Pearson and R.F.
Parametric, Using Bayesian or Non-Bayesian Approaches. Edited Boruch. V, 129 pages, 1986.
by J.P. Florens, M Mouchart, J.P. Raoult, L. Simar, and A.F.M.
Smith, XI, 204 pages, 1983. Vo1. 39: J.D. Malley, Optimal Unbiased Estimation of Variance
Components. IX, 146 pages, 1986.
Vol. 17: I.V. Basawa and D.J. Scott, Asymptotic Optimal Infer-
enoe for Non-Ergodic Models. IX, 170 pages, 1983. Vo1. 40: H.R Lerche, Boundary Crossing of Brownian Motion. V,
142 pages, 1986.
Vol. 18: W. Britton, Conjugate Duality and the Exponential
Fourier Spectrum. V,226 pages, 1983. Vo1. 41: F. Baccelli, P. Bremaud, Palm Probabilities and Station-
ary Queues. VII, 106 pages, 1987.
Vol. 19:L Femholz, vonMises CalculusForStatisticalFunctionals.
VIII, 124 pages, 1983. Vol. 42: S. Kullback, J.C. Keegel, J.H. Kullback, Topics in
Statistical Information Theory. IX, 158 pages, 1987.
Vol. 20: Mathematical Learning Models - Theory and Algo-
rithms: Proceedings of a Conference. Edited by U. Herl<enrath, D. Vo1. 43: B.C. Arnold, Majorization and the Lorenz Order: A Brief
Kalin, W. Vogel. XIV, 226 pages, 1983. Introduction. VI, 122 pages, 1987.
Vol. 21: H. Tong, Threshold Models in Non-linear Time Series Vo1. 44: D.L McLeish, Christopher G. Small, The Theory and
Analysis. X, 323 pages, 1983. Applicatioos of Statistical Inference Functions. VI, 124 pages,
1987.
Vol. 22: S. Johansen, Functional Relations, Random Coefficients
and Nonlinear Regression with Application to Kinetic Data, VIII, Vo1. 45: J.K. Ghosh (Ed.), Statistical Information and Likelihood.
126 pages, 1984. 384 pages, 1988.
Vol. 46: H.-G. MUller, Nonpamnetric Regression Analysis of Vol. 65: A. Janssen, D.M. Mason, Non-Standard Rank Tests. VI,
Longitudinal Data. VI, 199 pages, 1988. 252 pages, 1990.
Vol. 47: A.J. Getson, F.C. Hsuan, {2J-Invemes and Their Statis- Vol. 66: T. Wright, Exact Confidence Bounds when Sampling
vm,
tical Application. 110 pages, 1988. from Small Finite Univemes. XVI, 431 pages, 1991.
Vol. 48: G.L. Bretthorst,Bayesian Spectrum Analysis andParam- Vol. 67: M.A. Tanner, Tools for Statisticallnference: Observed
eter Estimation. XII, 209 pages, 1988. Data and Data Augmentation Methods. VI, 110 pages, 1991.
Vol. 49: SL. Lauritzen, Extremal Families and Systems of Suffi- Vol. 68: M. Taniguchi,HigherOrder AsymptoticTheoryforTime
cient Statistics. XV, 268 pages, 1988. Series Analysis.vm, 160 pages, 1991.
Vol. 50: O.E. Bamdorff-Nie1sen, Pamnetrlc Statistical Models Vol. 69: NJD. Nage1kerke, Maximum Likelihood Estimation of
and Likelihood. VII, 276 pages, 1988. Functional Relationships. V, 110 pages, 1992.
Vol. 51: J. HUsler, R.-D. Rciaa (Eds.), Extreme Value Theory. Vol. 70: K. lida, Studies on the Optimal SCllICh Plan. vm, 130
Proceedings, 1987. X, 279 pages, 1989. pages, 1992.
Vol. 52: P.K. GoeI, T. Ramalingam, The Matching Methodology: Vol. 71: E.M.R.A. Engel, A Road to Randomness in Physical
vm,
Some Statistical Properties. 152 pages, 1989. Systems. IX, 155 pages, 1992.
Vol. 53: B.C. Arnold, N. Balakrishnan, Relations, Bounds and Vol. 72:J.K. Lindsey, The Analysis of Stochastic Processes using
Approximations for Order Statistics. IX, 173 pages, 1989. GUM. VI, 294 pages, 1992.
Vol. 54: K.R. Shah, B.K. Sinha, Theory of Optimal Designs. vm, Vol. 73: B.C. Arnold, E. Castillo, J.-M. Sarabia, Conditionally
171 pages, 1989. Specified Distributions. Xli, 151 pages, 1992.
Vol. 55: L. McDonald, B. Manly, J. Lockwood, J. Logan (Bds.), Vol. 74: P. Barone, A. Frigessi, M. Piccioni, Stochastic Models,
Estimation and Analysis ofInsectPopoiations. Proceedings, 1988. Statistical Methods, and Algorithms in Image Analysis. VI, 258
XIV, 492 pages, 1989. pages, 1992.
Vol. 56: J.K. Lindsey, The Analysis of Categorical Data Using Vol. 75: P.K. Goel, N.S. Iyengar (Eds.), Bayesian Analysis in
GUM. V, 168 pages, 1989. Statistics and Econometrics. XI, 410 pages, 1992.
Vol. 57: A. Decarli, B.J. Francis, R. Gilchrist, G.U.H. Seeber Vol. 76: L. Bondesson, Generalized Gamma Convolutions and
(Bds.), Statistical Modelling. Proceedings, 1989. IX, 343 pages, Related Classes of Distributions and Densities. vm,
173 pages,
1989. 1992.
Vol. 58: O.E. Bamdorff-Nielsen,P. Blzsild,P.S. Eriksen,Decom- Vol. 77: E. Mammen, When Does Bootstrap Work? Asymptotic
position and Invariance of Measures, and Statistical Tranaforma- Resolts and Simulations. VI, 196 pages, 1992.
tion Models. V, 147 pages, 1989.
Vol. 78: L. Fahnneir, B. Francis, R. Gilchrist, G. Tutz (Bds.),
Vol. 59: S. Gupta, R. Mukerjee,A CalcoIus for Factorial Anange- Advances in GUM and Statistical Modelling: Proceedings of the
ments. VI, 126 pages, 1989. GUM92 Conference and the 7th International Workshop on
Statistical Modelling, Munich, 13-17 Joly 1992. IX, 22S pages,
Vol. 60: L GyOrli, W. HlIrdle, P. Sarda, Ph. Vieu, Nonparametric 1992.
Curve Estimation from Time Series. vm, 153 pages, 1989.
Vol. 79: N. Schmitz, Optimal Sequentially Planned Decision
Vol. 61: 1. Breckling, The Analysis of Directional T1lIIe Series: Procedures. XII, 209 pages, 1992.
Applications to Wind Speed and Direction. vm, 238 pages, 1989.
Vol. 80: M. F1igner, J. Verduccl (Bds.), Probability Models and
Vol. 62: J.C. Akkerboom, Testing Problems with Linear or Angu- Statistical Analyses for Ranking Data. XXII, 306 pages, 1992.
lar Inequality Constraints. XII, 291 pages, 1990.
Vol. 81: P. Spirtes, C. Glymour, R. Scheines, Causation, Predic-
Vol. 63: 1. Pfanzagl, Estimation in Semiparametric Models: Some tion, and SCllICh. XXIII, 526 pages, 1993.
Recent Developments. ill, 112 pages, 1990.
Vol. 64: S. Gabler, Minimax Solutions in Sampling from Finite

PopoIations. V, 132 pages, 1990.
General Remarks
Lecture Notes are printed by photo-offset from the Series Editors:

master-copy delivered in camera-ready form by
Professor 1. Berger
the authors of monographs, resp. editors of pro-
Department of Statistics
ceedings volumes. For this purpose Springer-
Purdue University
Verlag provides technical instructions for the
West Lafayette, IN 47907
preparation of manuscripts. Volume editors are
USA
requested to distribute these to all contributing
authors of proceedings volumes. Some homo- Professor S. Fienberg
geneity in the presentation ofthe contributions in Office of the Vice President
a multi-author volume is desirable. York University
4700 Keele Street
Careful preparation of manuscripts will help keep
North York, Ontario M3J IP3
production time short and ensure a satisfactory
Canada
appearance of the finished book. The actual pro-
duction of a Lecture Notes volume normally takes Professor 1. Gani
approximately 8 weeks. Department of Statistics lAS
Australian National University
For monograph manuscripts typed or typeset
GPO Box 4
according to our instructions, Springer-Verlag can,
Canberra ACT 2601
if necessary, contribute towards the preparation
Australia
costs at a fixed rate.
Professor K. Krickeberg
Authors of monographs receive 50 free copies of
3 Rue de L 'Estrapade
their book. Editors of proceedings volumes simi-
75005 Paris
larly receive 50 copies ofthe book and are respon-
France
sible for redistributing these to authors etc. at their
discretion. No reprints of individual contributions Professor I. Olkin
can be supplied. No royalty is paid on Lecture Department of Statistics
Notes volumes. Stanford University
Stanford, CA 94305
Volume authors and editors are entitled to pur-
USA
chase further copies oftheir book for their personal
use at a discount of 33.3% and other Springer Professor B. Singer
mathematics books at a discount of 20% directly 60 College St., Room 210
from Springer-Verlag. Authors contributing to PO Box 3333
proceedings volumes may purchase the volume in Yale University
which their article appears at a discount of20%. New Haven, CT 06510
USA
Springer-Verlag secures the copyright for each
volume.

Spirt Es 1993

Uploaded by

Copyright:

Available Formats

Spirt Es 1993

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

Spirt Es 1993

Uploaded by

Copyright:

Available Formats

Lecture Notes in Statistics 81

Edited by J. Berger, S. Fienberg, J. Gani, K. Krickeberg,

Mathematics Subject Classification: 62-07

library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

© 1993 Springer-Verlag New Yorlc, Inc.

Camera ready copy provided by the authors.

In memory of Lucille Lynch Schwartz Watkins Speede Tindall Preston - C. O.

To Martha, for her support and love - R.S.

--G. U. Yule and M. G. Kendall 1950

The Theory of Estimation discusses the principles upon which

-- Sir Ronald Fisher, 1956

--G. Mosteller and J. Tukey, 1977

Causal inference is one of the most important, most subtle, and

sufficiently large. These procedures have been implemented in a computer program,

unmeasured variables in linear systems. Chapter 12 is a brief consideration of a variety of

Preface ................................................................................................................................ vii

Acknowledgments .............................................................................................................. xiii

1. Introduction and Advertisement •••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••.••..•••.••••.••• 1

2. Formal Preliminaries ..................................................................................................... 25

3. Causation and Prediction: Axioms and Explications .................................................41

3.3 Causality and Probability ................................................................................... 49

4. Statistical Indistinguishability ...................................................................................... 87

5. Discovery Algorithms for CausaUy Sufficient Structures •••••••••••••••••••••••••••••••••••••••••. 103

5.4.3 The IG (Independence Graph) Algorithm .......................................... 124

6. Discovery Algorithms without Causal Sufficiency ..................................................... 163

7. Prediction ........................................................................................................................ 201

7.3 Rubin-Holland-Pratt-Schlaifer Theory .............................................................. 203

8. Regression, Causation and Prediction ••••••••••••••••••••••••••••••••••••••••••.•••••••••••.•••••••••••••••••• 238

9. The Design of Empirical Studies ••••••••••••••...••••••.••••••••••.••••••••••••••••.•••••••••••.•.••••••••. _•.••• 259

10. The Structure of the Unobserved ............................................................................... 306

11. Elaborating Linear Tbeories witb Unmeasured Variables •••••••••••••••••••••••••••••••••••••. 323

12. Open Problems ......................................................................,'•••••••••••••••••••••••••••••••••••••• 354

13.8 Theorem 3.7 ..................................................................................................... 398

Glossary .............................................................................................................................. 481

Bibliography ....................................................................................................................... 495

Index .................................................................................................................................... 517

In the text, each technical term is written in boldface where it is defined.

Variables: capitalized, and in italics, e.g, X

If a description of a set of variables is a function of a graph G and variables in G, then we

L ' which has the following properties:

We sometimes use a special summation symbol,

LP(XIY = O,Z = 0) = P(X = OIY = O,Z = 0) + P(X = IIY = O,Z = 0)

However, if p(Y=o,z=o) = 0, then P(X=OIY=<i,z=O) and P(X=lIY=O,z=o) are not defined, so

Introduction and Advertisement

1.1 The Issue

"Model selection" or "specification search" is an area of research in which questions of causal

Of itself, the mathematical representation of causal influence by directed graphs is trivial.

the joint distribution must satisfy

P(X,Y,Z) =P(X) P(YIX) P(ZIY).

With an appropriate definition of "manipulation" the Markov Condition entails a generalization

1.2.1 Bayes Networks from the Data

The ALARM belief network

1.2.2 Structural Equation Models from the Data

GPQ =graduate program quality PUBS =publications since graduate school

1.2.3 Selection of Regressors

Y=aIXI +a2Xs +a3T+ey

1.2.4 Causal Inference without Experiment

Suppose B is unmeasured. Fisher's point is that non-experimental studies of A and C cannot

The common presumption is that non-experimental evidence cannot determine whether