Observing interaction

Second edition
Observing interaction:
An introduction to
sequential analysis
Second edition

ROGER BAKEMAN
Georgia State University

JOHN M. GOTTMAN
University of Washington

CAMBRIDGE
UNIVERSITY PRESS
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 2RU, UK

Published in the United States of America by Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9780521450089
© Cambridge University Press 1997

This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.

First published 1986

Second edition 1997

A catalogue record for this publication is available from the British Library

Library of Congress Cataloguing in Publication data

Bakeman, Roger.
Observing interaction: an introduction to sequential analysis /
Roger Bakeman, John M. Gottman. - 2nd ed.
p. cm.
Includes bibliographical references and index.
ISBN 0-521-45008-X (hardcover). - ISBN 0-521-57427-7 (pbk.)
1. Social sciences — Statistical methods. 2. Sequential analysis. 3. Social
interaction — Statistical methods. I. Gottman, John Mordechai. II. Title.
H62.B284 1997 96-26084
CIP

ISBN-13 978-0-521-45008-9 hardback

ISBN-10 0-521-45008-X hardback

ISBN-13 978-0-521-57427-3 paperback

ISBN-10 0-521-57427-7 paperback

Transferred to digital printing 2006

 Contents

Preface to the second edition page ix

Preface to the first edition xi

1. Introduction 1
1.1. Interaction and behavior sequences 1
1.2. Alternatives to systematic observation 2
1.3. Systematic observation defined 3
1.4. A nonsequential example: Parten 's study of children's play 4
1.5. Social process and sequential analysis 6
1.6. Another nonsequential example: Smith's study of parallel
play 1
1.7. A sequential example: Bakeman and Brownlee's study of
parallel play 8
1.8. Hypothesis-generating research 12
1.9. Summary: Systematic is not always sequential 13

2. Developing a coding scheme 15

2.1. Introduction 15
2.2. What is the question? 16
2.3. Physically versus socially based coding schemes 17
2.4. Detectors versus informants 22
2.5. The fallacy of equating objectivity with physically based
schemes 22
2.6. Keeping it simple 23
2.7. Splitting and lumping 24
2.8. Mutually exclusive and exhaustive codes 26
vi Contents

2.9. The evolution of a coding system 27

2.10. Example 1: Interaction ofprehatched chickens 28
2.11. Example 2: Children }s conversations 30
2.12. Example 3: Baby behavior codes 33
2.13. Example 4: Children's object struggles 34
2.14. Example 5: Monkeys' activities 35
2.15. Summary 36
3. Recording behavioral sequences 38
3.1. Recording units: Events vs. intervals 38
3.2. Momentary versus duration events 38
3.3. Continuous versus intermittent recording 39
3.4. Coding events 40
3.5. Recording onset and offset times 43
3.6. Timing pattern changes 45
3.7. Coding intervals 46
3.8. Cross-classifying events 49
3.9. Nonsequential considerations: Time sampling 50
3.10. The pleasures of pencil and paper 52
3.11. Why use electronics ? 52
3.12. Summary 55
4. Assessing observer agreement 56
4.1. Why bother? 56
4.2. Reliability versus agreement 59
4.3. The problem with agreement percentages 60
4.4. The advantages of Cohen's kappa 62
4.5. Agreement about unitizing 68
4.6. Agreement about codes: Examples using Cohen fs kappa 71
4.7. Generalizability theory 75
4.8. Unreliability as a research variable 79
4.9. Summary 80
5. Representing observational data 81
5.1. Representing versus recording 81
5.2. Event sequences 82
 Contents vii

5.3. State sequences 83

5.4. Timed-event sequences 84
5.5. Interval sequences 85
5.6. Cross-classified events 87
5.7. Transforming representations 88
5.8. Summary 90

6. Analyzing sequential data: First steps 91

6.1. Describing versus modeling 91
6.2. /toes and frequencies 92
6.3. Probabilities and percentages 93
6.4. Mean event durations 94
6.5. Transitional probabilities: An introduction 95
6.6. Summary 99

7. Analyzing event sequences 100

7.1. Describing particular sequences: Basic methods 100
7.2. Determining significance of particular chains 101
7.3. Transitional probabilities revisited 103
7.4. Computing z scores and testing significance 108
7.5. Classic lag sequential methods 111
7.6. LogAinear approaches to lag-sequential analysis 116
7.7. Computing Yule's Q or phi and testing for individual
differences 127
7.8. Summary 132

8. Issues in sequential analysis 136

8.1. Independence 136
8.2. Stationarity 138
8.3. Describing general orderliness 139
8.4. Individual versus pooled data 141
8.5. How many data points are enough? \AA
8.6. The type I error problem 147
8.7. Summary 148
viii Contents

9. Analyzing time sequences 150

9.1. The tyranny of time 150
9.2. Taking time into account 151
9.3. Micro to macro 153
9.4. Time-series analysis 154
9.5. Autocorrelation and time-series analysis 167
9.6. Summary 175

10. Analyzing cross-classified events 177

10.1. Describing cross-classified events 111
10.2. Log-linear models: A simple example 179
10.3. Summary 182

11. Epilogue 184

11.1. Kepler and Brahe 184
11.2. Soskin and John on marital interaction 185
11.3. Marital interaction research since 1963 188

Appendix 194
A Pascal program to compute kappa and weighted kappa 194

References 198

Index 205
 Preface to the second edition

Since the first edition of Observing Interaction appeared in 1986, the tech-
nology supporting recording and systematic coding of behavior has become
less expensive, more reliable, and considerably less exotic (Bakeman, in
press; revised chapter 3, this volume). Cumbersome videotape and cas-
sette recorders have given way to video camcorders. Visual time codes
are routinely recorded as part of the picture, and equipment to write and
read machine-readable time codes is readily available at reasonable cost.
Increasingly, computers assist coding, making routine what once was labor
intensive and time-consuming. Even physiological recording devices can
be added to the computer's net (Gottman & Bakeman, in press). Thus
an increasing circle of investigators can avail themselves of the methods
detailed in this book without mortgaging their careers, their lives, or the
lives of their associates.
At the same time, the way we think about sequential data has devel-
oped. This is reflected in a standard format for sequential data (Bakeman
& Quera, 1992, 1995a; revised chapter 5, this volume). SDIS - the Se-
quential Data Interchange Standard - has greatly facilitated the analysis of
sequential data. Again, an enterprise that formerly was time-consuming
and cumbersome has yielded to appropriately designed computer tools, as
described in my and Quera's Analyzing Interaction (1995), which should
be regarded as a companion to this volume. This revised version of Observ-
ing Interaction still explains how to conceptualize, code, record, organize,
and analyze sequential data, but now Analyzing Interaction provides the
tools to do so easily.
Another area of considerable development, and one responsible for many
of the differences between the first and second editions of Observing In-
teraction, concerns techniques for analyzing sequential data (chapters 7-9,
this volume; these chapters are extensively modified versions of chapters
7-8 from the first edition). Formerly many of the analytic techniques
proposed for sequential analysis were somewhat piecemeal and post hoc,
yet, waiting in the wings, log-linear analysis promises a coherent analytic
view for sequential phenomena (Bakeman & Quera, 1995b). This revised
 Preface to the second edition

edition moves log-linear techniques (which can be thought of as a multidi-

mensional extension of chi-square tests; see Bakeman & Robinson, 1994)
center stage. This simplifies matters and, at the same time, integrates
lag-sequential analysis with an established and well-supported statistical
tradition.
In the preface to thefirstedition, I suggested that far more people deserve
thanks than can be named explicitly. That is still so. Nonetheless, I would
like to thank three colleagues from the University of Barcelona: Maria
Teresa Anguera and Angel Blanco, who translated the first edition of Ob-
serving Interaction into Spanish, and Vicen? Quera, who has collaborated
with me these past several years in thinking about sequential analysis and
in developing the SDIS and the sequential analysis program we call the
Generalized Sequential Querier (GSEQ). I would also like to thank Debora
Gray, who emerged from her mountain fastness to redraw one last figure
for this second edition, and all those students in my observational methods
class, winter quarter 1996, who offered comments. Finally, I would like
to correct an error from the first edition's preface. Mildred Parten did not
disappear. She lived her life as a bureaucrat in Washington, D.C., at a time
in our history when such jobs were among the few open to women who
wanted both to work and to use their brains.

ROGER BAKEMAN

References

Bakeman, R. (in press). Behavioral Observations and Coding, In H. T.

Reis & C. K. Judd (Eds.), Handbook of research methods in social
psychology. New York: Cambridge University Press.
Gottman, J. M., & Bakeman, R. (in press). Systematic observational tech-
niques or building life as an observational researcher. In D. Gilbert,
S. Fiske, & G. Lindzey (Eds.), Handbook of social psychology (4th
Ed.). New York: McGraw-Hill.
 Preface to the first edition

Sometimes even a rather lengthy series of thoughts can be associated in

memory with a single, dominant image. For me (RB), when I reflect on
the writing of this book, the image of Mount Rainier, seemingly floating
on the late afternoon haze, is never far from my mind. This is a personal
image, of course, but it is understandable if I explain that John Gottman
and I first met at a conference, organized by Jim Sackett, held at Lake
Wilderness in Washington State. Thus some of the conversations that laid
the groundwork for this book took place against the backdrop of Mount
Rainier, dominating the horizon at the other end of the lake.
The conference was concerned with the Application of Observational/
Ethological Methods to the study of Mental Retardation, a title that suggests
both some of the research traditions that have influenced our writing and
some of the kinds of readers who might find this book useful. Throughout
this century, some of the most systematic and productive observers of social
interaction have been ethologists, especially those concerned with primates,
and developmental psychologists, especially those concerned with infants
and young children.
Although students of primate and children's behavior have been largely
responsible for the development of systematic observational methods, they
are not the only kinds of researchers who want to study social behavior
scientifically. Among others, this book should interest investigators in all
branches of animal behavior, in anthropology, in education (including those
concerned with classroom evaluation as well as early childhood education),
in management, in nursing, and in several branches of psychology (includ-
ing child, community, developmental, health, organizational, and social),
as well as investigators concerned with mental retardation.
As the title implies, this book is a primer. Our intent is to provide a clear
and straightforward introduction to scientific methods of observing social
behavior, of interest to the general practitioner. Avoided are arcane byways
of interest primarily to statistical specialists. Because the dynamics of peo-
ple (and other animals) interacting unfold in time, sequential approaches
to observing and understanding social behavior are emphasized.
xii Preface to the first edition

We assume that most readers of this book will be researchers (either

advanced undergraduates, graduate students, or established investigators
already seasoned in other methods) who want to know more about sys-
tematic observational methods and sequential analysis. Still, we hope that
those readers who are already experienced in observational methods will
recognize what they do clearly described here and may even gain additional
insights.
We further assume that most readers will have a rudimentary knowledge
of basic statistics, but, we should emphasize, an advanced knowledge of
statistical analysis is not required. Instead, throughout the book, conceptual
fundamentals are emphasized and are presented along with considerable
practical advice. Our hope is that readers of this book will be equipped to
carry out studies that make use of systematic observational methods and
sequential analysis.

Writers of acknowledgments frequently admit that far more people deserve

thanks than can be explicitly named. In this, I am no exception. I am
very much in debt, for example, to students and other researchers who
have consulted with me regarding methodological issues, because their
questions often forced me to consider problems I might not otherwise
have considered. I am also in debt to those mothers, infants, and children
who served as subjects in my own research, because the studies in which
they participated provide many of the examples used throughout this book.
Finally, I greatly appreciate those graduate students in my observational
methods course who read and commented on an earlier draft of this book.
Debora Gray is responsible for many of the figures included here. I
enjoyed working with her and very much appreciate her vision of what the
careful students's notebook should look like. I also want to thank Melodie
Burford, Connie Smith, and Anne Walters, who diligently ferreted out
errors in a near-final copy of the manuscript.
Chief among the people whose encouragement and support I value is Jim
Sackett. Early on, even before the Lake Wilderness conference in 1976,
he urged me to pursue my interest in observational methods and sequential
analysis. So did Leon Yarrow and Bob Cairns, to whom I am similarly
indebted for encouraging comments at an early point in my career.
Sharon Landesman-Dwyer also helped us considerably. Along with
Kathy Barnard and Jim Sackett, she arranged for us to spend a month
at the University of Washington in July 1983, consulting and giving lec-
tures. Not only did this give us a chance to discuss this book (with Mount
Rainier again in the background), it gave us a chance to present some of
the material incorporated here before a live, and very lively, audience.
I have been fortunate with collaborators. Most obviously, I appreciate
 Preface to the first edition xiii

the collaboration with John Gottman that has resulted in this book, but
also I have benefited greatly from my collaboration, first with Josephine
Brown, and more recently with Lauren Adamson. Their contribution to
this book is, I hope, made evident by how often their names appear in the
references. I would also like to thank my department and its chair, Duane
M. Rumbaugh, for the support I have received over the years, as well as
the NIMH, the NIH, and the NSF for grants supporting my research with
J. V. Brown and with L. B. Adamson.
Finally, I would like to acknowledge the seminal work of Mildred Parten.
Her research at the University of Minnesota's Institute of Child Develop-
ment in the 1920s has served as a model for generations of researchers
and is still a paradigmatic application of observational methods, as our first
chapter indicates. How much of a model she was, she probably never knew.
She did her work and then disappeared. In spite of the best efforts of Bill
Hartup, who until recently was the director of the Institute, her subsequent
history remains unknown.
Parten left an important legacy, however - one to which I hope readers
of this book will contribute.
ROGER BAKEMAN

A small percentage of current research employs observational measurement

of any sort. This is true despite the recent increased availability of new
technologies such as electronic notepads and videotape recording. It may
always be the case because it is more costly to observe than to use other
methods such as questionnaires.
We were motivated to write this book because we believe that observa-
tional methods deserve a special role in our measurement systems. First, we
think the descriptive stage of the scientific enterprise is extremely produc-
tive of research hypotheses, models, and theory. This ethological tradition
is full of examples of this fact, such as Darwin's classic work on emotional
expression. Second, the time is ripe for a reconsideration of observational
techniques because we now know a lot more about what to observe, how to
construct reliable measurement networks, and how to analyze data to detect
interaction sequences. Recently, we have been making new headway on
old problems with these new technologies.
We are optimistic that this book will fill a gap and stimulate new research
that employs useful (and not superficial) observational systems.

I (JMG) would like to acknowledge grants MH29910, MH35997, and

RSDA K200257 and sabbatical release time during 1984-1985.

JOHN M. GOTTMAN
 1
Introduction

1.1 Interaction and behavior sequences

Birds courting, monkeys fighting, children playing, couples discussing,
mothers and infants exchanging gleeful vocalizations all have this in com-
mon: Their interaction with others reveals itself unfolded in time. This
statement should surprise no one, certainly not readers of this volume. What
is surprising, however, is how often in the past few decades researchers in-
terested in dynamic aspects of interactive behavior - in how behavior is
sequenced moment to moment - have settled for static measures of inter-
action instead. This need not be. In fact, our aim in writing this book is to
demonstrate just how simple it often is not just to record observation data
in a way that preserves sequential information, but also to analyze that data
in a way that makes use of - and illuminates - its sequential nature.
We assume that readers of this book may be interested in different
species, observed at various ages and in diverse settings, but that most will
be interested specifically in observing interactive social behavior. This is
because we think sequential methods are tailor-made for the study of so-
cial interaction. As noted, a defining characteristic of interaction is that it
unfolds in time. Indeed, it can hardly be thought of without reference to
a time dimension. Sometimes we are concerned with actual time units -
what happens in successive seconds, for example; at other times, we are
just concerned with what events followed what. In either case, we think
it is a sequential view that offers the best chance for illuminating dynamic
processes of social interaction.
For example, we might ask a married couple to fill out a questionnaire,
and from their responses we might assign them a "marital satisfaction"
score. This would at least let us try to relate marital satisfaction to other
aspects of the couple's life circumstances, like their own early experience
or their current work commitments, but such a static measure would not
tell us much about how the couple interacts with each other, or whether the
way in which they interact relates in any way to how satisfied they report
being with their relationship. In order to "unpack" the variable of marital
 Introduction

satisfaction, we would need to examine more closely just how the couple
related to each other - and, in order to describe and ultimately attempt
to understand the dynamics of how they relate to each other, a sequential
view is essential. Our hope is that readers of this book not only take such
a sequential view, but also will learn here how to describe effectively the
sequential nature of whatever interaction they observe.

1.2 Alternatives to systematic observation

There is a second assumption we make about readers of this book. We
assume that they have considered a variety of different methods of inquiry
and have settled on systematic observation. For a moment, however, let
us consider the alternatives. When studying humans, at least those able
to read and write, researchers often use questionnaires, like the marital
satisfaction inventory mentioned. These questionnaires, as well as tests
of various sorts, certainly have their uses, although capturing the dynamic
quality of behavior sequences is not one of their stronger points.
We do not mean to suggest, however, that investigators must choose
between observational and other methods. In our own work, we usually
employ a convergent measurement network that taps the constructs we
are interested in studying. This network usually includes questionnaire,
interview, and other measurement operations (e.g., sociometric measures).
Still, we think there is something captured by observational procedures
that eludes these other measurement procedures. Nonetheless, there are at
least two time-honored alternatives to systematic observation for capturing
something of the sequential aspect of interaction. These are (a) narrative
descriptions, and (b) properly designed rating scales.
If we were forced to choose between the adjectives "humanistic" and
"scientific" to describe narrative descriptions, we would have to choose
humanistic. We do so, not to demean the writing of narrative reports - a
process for which we have considerable respect - but simply to distinguish
it from systematic observation. After all, not only have humanistic methods
of inquiry been used far longer than the upstart methods we characterize as
scientific, but also the preparation of narrative reports, or something akin
to it, is an important part of code development, a process that must precede
systematic observation.
Still, narrative reports depend mightily on the individual human doing
them, and judgments about the worth of the reports are inextricably bound
up with judgments about the personal qualities of their author. In fact, we
would be surprised if two reports from different authors were identical.
With systematic observation, on the other hand, the goal is for properly
 Systematic observation defined

trained observers to produce identical protocols, given that they observed

the same stream of behavior. The personal qualities of the observers (as-
suming some talent and proper training) should not matter. It is primarily
this drive for replicability that we think earns systematic observation the ap-
pellation "scientific" and that distinguishes it from the writing of narratives.
Rating scales, on the other hand, allow for every bit as much replicability
as systematic observation does. In fact, if we defined systematic observa-
tion more broadly than we do (see next section), the use of rating scales
could easily be regarded simply as another instance of systematic observa-
tion. For present purposes, however, we prefer to regard them separately
and to point out some of the differences between them. (For a discussion of
some of these differences, see Cairns & Green, 1979.) Imagine that we are
interested in the responsivity of two individuals to each other, for example, a
husband and wife or a mother and baby. We could define behavioral codes,
code the stream of behavior, and then note how and how often each was "re-
sponsive" to the other (systematic observation), or we could train observers
to rate the level of responsivity that characterized the interaction observed.
For many purposes, a rating-scale approach might be preferable. For
example, imagine an intervention or training study in which an investiga-
tor hopes to change maternal responsivity to infant cues. In this case, what
needs to be assessed is clearly known and is a relatively coherent concept
which can be clearly defined for raters. Then the far less stringent time
demands of a rating-scale approach would probably make it the methodol-
ogy of choice. On the other hand, if a researcher wants to describe exactly
how mothers are responsive to their infants and exactly how this respon-
sivity changes with infant development, then the more detailed methods of
systematic observation are required.

1.3 Systematic observation defined

For present purposes, we define systematic observation as a particular ap-
proach to quantifying behavior. This approach typically is concerned with
naturally occurring behavior observed in naturalistic contexts. The aim
is to define beforehand various forms of behavior - behavioral codes -
and then ask observers to record whenever behavior corresponding to the
predefined codes occurs. A major concern is to train observers so that all
of them will produce an essentially similar protocol, given that they have
observed the same stream of behavior.
The heart and foundation of any research using systematic observation is
the catalog of behavior codes developed for a particular project (see chapter
2). As inventories of questions are to personality or marital satisfaction
 Introduction

research, as IQ tests are to cognitive development research, so are code

catalogs (or coding schemes) to systematic observation. They are the mea-
suring instruments of observational research; they specify which behavior
is to be selected from the passing stream and recorded for subsequent study.
In many ways, systematic observation is not very different from other
approaches to behavioral research. Here, too, investigators need to say
what they hope to find out; they need to define what seems important
conceptually, they need to find ways to measure those concepts, and they
need to establish the reliability of their measuring instruments. However,
because human observers are such an important part of the instrumentation,
reliability issues loom especially large in observational research, a matter
which we discuss further in chapter 4.
In sum, the twin hallmarks of systematic observation are (a) the use of
predefined catalogs of behavioral codes, (b) by observers of demonstrated
reliability. The entire process of defining and developing coding schemes
followed by training observers to acceptable levels of agreement can be
both time-consuming and demanding. But without such an effort, the
investigator who goes no further than only telling others what he or she
sees runs the risk of having skeptical colleagues dismiss such narrative
reports as just one person's tale spinning.

1.4 A nonsequential example: Par ten's study of

children's play
An early and well-known example of systematic observation is Mildred
Patten's (1932) study of social participation among preschool children,
conducted at the University of Minnesota's Institute of Child Welfare in
the late 1920s. There are in fact many excellent observational studies of
children's behavior which were done in the 1920s and 1930s, and many
of the basic techniques still in use were first articulated then. We discuss
Parten's study here as an exemplar of that early work and as a way of
defining by example what we mean by "systematic observation." At the
same time, Parten's study was not sequential, as we use the term, and so
describing both what she did and what she did not do should clarify what
we mean by "sequential."
During the school year of 1926-1927, some 42 children whose ages
ranged from not quite 2 to almost 5 years were observed during indoor
free play. Parten was interested in the development of social behavior in
young children, and to that end defined six levels or categories of social
participation as follows:
1. Unoccupied. The child does not appear to be engaged with any-
thing specific; rather, his behavior seems somewhat aimless. He
 Parten ys study of children's play

might watch something of momentary interest, play with his own

body, just wander around, or perhaps stand or sit in one place.
2. Onlooker. The child watches other children play, but does not
enter into their play. This differs from Unoccupied because the
child is definitely watching particular children, not just anything
that happens to be exciting.
3. Solitary Independent Play. The child plays alone and indepen-
dently with whatever toys are of interest. The child's activity does
not appear affected by what others are doing.
4. Parallel Activity. The child still plays independently, but his activ-
ity "naturally brings him among other children." He plays beside
them, not with them, but with toys that are similar to those the
children around him are using. There is no attempt to control the
coming or going of children in the group.
5. Associative Play. The child plays with other children. There may
be some sharing of play material and mild attempts to control which
children are in the group. However, there is no division of labor
or assigning of roles: Most children engage in essentially similar
activity. Although each child acts pretty much as he or she wishes,
the sense is that the child's interest lies more with the association
with others than with the particular activity.
6. Cooperative or Organized Supplementary Play. The child plays in
a group that is organized for some purpose. The purpose might be
to dramatize a situation - for example, playing house - or to play a
formal game, or to attain some competitive goal. There is a sense
of belonging or not to the group. There is also a division of labor,
a taking of roles, and an organization of activity so that the efforts
of one child are supplemented by those of another. (The above
definitions are paraphrased from Parten, 1932, pp. 250-251.)

Each child was observed for 1 minute each day. The order of observation
was determined beforehand and was varied systematically so that the 1-
minute samples for any one child would be distributed more or less evenly
throughout the hour-long free-play period. On the average, children were
observed about 70 different times, and each time they were observed, their
degree of social participation was characterized using one of the six codes
defined above.
Florence Goodenough (1928) called this the method of repeated short
samples. Today it is often called "time sampling," but its purpose remains
the same. A number of relatively brief, nonsuccessive time intervals are
categorized, and the percentage of time intervals assigned a particular code
is used to estimate the proportion of time an individual devotes to that
kind of activity. For example, one 3-year-old child in Parten's study was
 Introduction

observed 100 times. None of the 1-minute time samples was coded Unoc-
cupied, 18 were coded Solitary, 5 Onlooking, 51 Parallel, 18 Associative,
and 8 Cooperative. It seems reasonable to assume that had Parten observed
this child continuously hour after hour, day after day, that about 51% of
that child's time would have been spent in parallel play.
The method of repeated short samples, or time sampling, is a way of
recording data, but it is only one of several different ways that could be
used in an observational study. What makes Parten's study an example of
systematic observation is not the recording strategy she used but the coding
scheme she developed, along with her concern that observers apply that
scheme reliably.
Parten was primarily concerned with describing the level of social par-
ticipation among children of different ages, and with how the level of social
participation was affected by children's age, IQ, and family composition.
For such purposes, her coding scheme and her method of data collection
were completely satisfactory. After all, for each child she could compute
six percentages representing amount of time devoted to each of her six lev-
els of social participation. Further, she could have assigned, and did assign,
weights to each code (—3 to Unoccupied, —2 to Solitary, —1 to Onlooker,
1 to Parallel, 2 to Associative, and 3 to Cooperative), multiplied a child's
percent scores by the corresponding weights, and summed the resulting
products, which yielded a single composite social participation score for
each child - scores that were then correlated with the child's age and IQ.
Knowing that older children are likely to spend a greater amount of time
in associative and cooperative play than younger ones, however, does not
tell us much about moment-by-moment social process or how Parten's par-
ticipation codes might be sequenced in the stream of behavior. This is not
because her codes are inadequate to the task, but because her way of record-
ing data did not capture behavior sequences. There is no reason, of course,
why she should have collected sequential data - her research questions
did not require examining how behavior is sequenced on a moment-by-
moment basis. However, there are interesting questions to ask about the
sort of children's behavior Parten observed that do require a sequential
view. An example of such a question is presented below.

1.5 Social process and sequential analysis

The purpose of this book is to emphasize sequential analyses of sequentially
recorded data, but we should not let this emphasis obscure how useful and
interesting nonsequential data (or the nonsequential analysis of sequential
data) can be. At the same time, we want to argue that sequential analyses
 Smith's study of parallel play

can provide an additional level of information about whatever behavior we

are observing, a level that is not accessible to nonsequential analyses.
In many ways, Patten's study typifies the sort of "time-budget" informa-
tion that nonsequential analyses of observational data can provide. Indeed,
it is often very useful to know how children, or mothers with infants, or
animals in the wild, or office workers distribute their time among various
possible activities. Nor is time the only thing that can be "distributed." We
could, for example, observe married couples in conversation, code each
"utterance" made, and then report percent scores for each utterance code.
Computing such percentages is a nonsequential use of the data, to be sure,
but it does allow us to determine, for example, whether disagreements are
more common among "distressed" as opposed to "nondistressed" couples.
There are, however, additional questions that can be asked. When utter-
ances are recorded sequentially, we can go on to ask what happens after one
spouse disagrees or after one spouse complains. Are there characteristic
ways the other spouse responds? Are these ways different for husbands
and wives? Are they different for distressed and nondistressed couples?
(For answers to these questions, see Gottman, 1979a.) At this point, we are
beginning to probe social process in a way that only sequential analyses
make possible.
In general, when we want to know how behavior works, or functions,
within an ongoing interaction, some form of sequential analysis is prob-
ably required. For example, a nonsequential analysis could tell us that
distressed husbands and wives complain more than nondistressed ones do,
but only a sequential analysis could tell us that distressed couples, but not
nondistressed ones, tend to react to each other's complaints with additional
complaints. Similarly, a nonsequential analysis can tell us that 3-year-olds
engage in less parallel play than 2-year-olds, but only a sequential analysis
can tell us if, in the moment-by-moment stream of activity, young children
use parallel play as a bridge into group activity. An example of such a
sequential analysis will be discussed later, but first we present a second
nonsequential example.

1.6 Another nonsequential example: Smith's study

of parallel play
Parten believed that her study established a relationship between children's
age and their degree of participation in social groups: As children became
older, they participated more. Her cross-sectional study is often interpreted
as suggesting a developmental progression; thus parallel play is seen as a
"stage" through which children pass as they develop from solitary to social
 Introduction

group players. This idea found its way into textbooks but was not tested
empirically until Peter Smith did so in the late 1970s (Smith, 1978). In the
present context, Smith's study is interesting for at least three reasons: for
what he found out, for the way he both made use of and modified Parten's
coding scheme, and for his method, which only appears sequential, as we
define the term.
For simplicity, Smith reduced Parten's six categories to three:
1. Alone, which lumped together Parten's Unoccupied, Onlooker, and
Solitary
2. Parallel, as defined by Parten
3. Group, which lumped together Parten's Associative and Coopera-
tive
After all, because he wanted to test the notion that Parallel play character-
izes an intermediate stage of social development, finer distinctions within
Alone and within Group play were not necessary. Smith then used these
codes and a time-sampling recording strategy to develop time-budget in-
formation for each of the 48 children in his study. However, Smith did not
compute percent scores for the entire period of the study, as Parten did, but
instead computed them separately for each of six successive 5-week peri-
ods (the entire study took 9 months). These percent scores were then used
to code the 5-week periods: Whichever of the three participation categories
occurred most frequently became the category assigned to a time period.
Smith's method is interesting, in part because it forces us to define exactly
what we mean by a sequential approach. Certainly his method has in com-
mon with sequential approaches that successive "units" (in his case, 5-week
periods) are categorized, that is, are matched up with one of the codes from
the coding scheme. However, what Smith did does not satisfy our sense of
what we usually mean by "sequential." It is only a matter of definition, of
course, but for the purpose of this book we would prefer to reserve the word
"sequential" for those approaches that examine the way discrete sequences
of behavior occur. Normally this means that sequential approaches are
concerned with the way behavior unfolds in time, as a sequence of rela-
tively discrete events, usually on a moment-by-moment or event-by-event
basis. In contrast, Smith's 5-week periods are not at all discrete, and thus
his approach is not sequential - as we use the term here - but is a reasonable
data reduction technique, given the question he sought to answer.

1.7 A sequential example: Bakeman and

Brownlee's study of parallel play
What Smith reported is that many children moved directly from a 5-week
period in which Alone play predominated, to one in which Group play
 Bakeman and Brownlee 's study of parallel play

Table 1.1. Three coding schemes for participation in social

groups
Bakeman &
Parten (1932) Smith (1978) Brownlee (1980)

Together
Unoccupied
Unoccupied
Onlooker Alone

Solitary Solitary

Parallel Parallel Parallel

Associative
Group Group
Cooperative

Note: A coding scheme is an investigator's attempt to cleave an often intractable

world "at the joints." Given here are coding schemes used by the three stud-
ies discussed in this chapter. The dashed lines indicate that what Parten coded
Unoccupied, Bakeman and Brownlee might have coded either Together or Un-
occupied. Similarly, what Bakeman and Brownlee coded Unoccupied, Parten
might have coded either Unoccupied or Onlooker. Smith would have coded all
of these Alone, as well as what both Parten, and Bakeman and Brownlee, coded
Solitary.

did, without an intervening period during which Parallel play was most
frequent. He concluded that a period during development characterized
by parallel play may be optional, a stage that children may or may not go
through, instead of obligatory, as Parten seems to have suggested. Still,
Smith's children engaged in parallel play about a quarter of the time, on the
average, and therefore, although it was seldom the most frequent mode of
play, it was nonetheless a common occurrence. This caused Bakeman and
Brownlee (1980) to think that perhaps parallel play might be more fruitfully
regarded, not as the hallmark of a developmental stage, but as a type of play
important because of the way it is positioned in the stream of children's
play behavior. Thus Bakeman and Brownlee raised a uniquely sequential
question about parallel play, one quite different from the question Parten
and Smith pursued.
Like Smith, Bakeman and Brownlee modified Parten's coding scheme
somewhat (see Table 1.1). They defined five codes as follows:
10 Introduction

1. Unoccupied, which lumped together Parten's Unoccupied and On-

looker.
2. Solitary. Unlike Smith, Bakeman and Brownlee chose to keep
Unoccupied and Solitary separate. Because they were interested
in how these "play states" are sequenced, and because both Solitary
and Parallel play involve objects, whereas Unoccupied does not,
they thought the distinction worth preserving.
3. Together. As far as we know, this code has not been used in other
published studies. It appears to be a particularly social way of
being unoccupied and is characterized by children clearly being
with others - there seems to be an awareness of their association -
but without the kind of focus on objects or activities required for
Parallel or Group play.
4. Parallel as defined by Parten.
5. Group. Like Smith, Bakeman and Brownlee lumped together
Parten's Associative and Cooperative.
The source material for this study consisted of videotapes, made during
indoor free play. Forty-one 3-year-olds were taped for about 100 min-
utes each. Observers then viewed these tapes and decided which of the
five codes best characterized each successive 15-second interval. This
method of recording data represents something of a compromise. It would
have been more accurate if observers had simply noted when a different
"play state" started. That way, not only would an accurate sequencing of
states have been preserved, but accurate time-budget information (percent-
age of time spent in the various play states) would have been available
as well.
This raises an interesting question. Is continuous recording (noting times
when different codable events begin and end) better than interval recording
(assigning codes to successive time intervals)? We shall have considerably
more to say about this matter later. For now, let us simply say that Bakeman
and Brownlee were able to extract from their data a reasonably accurate
sequence of events, that is, a record of the way different play states followed
each other in time.
Viewing their data as a sequence of play states, Bakeman and Brownlee
first counted how often each code followed the other codes (for example,
they determined how often Group followed Parallel, followed Together,
etc.). Then, using methods described in chapter 7, they compared ob-
served counts to their expected values. This was done separately for each
possible transition for each child, which means, for example, that if the
Parallel to Group transition occurred at greater than expected levels for a
particular child, the expected levels were based on how often that child
engaged in Group play.
 Bakeman and Brownlee 's study of parallel play 11

Among other things, Bakeman and Brownlee wanted to know if certain

transitions were especially characteristic of the children they observed. Of
particular interest was the Parallel to Group transition, one they thought
should be frequent if parallel play functions as a bridge into group play.
Now just by chance alone, observed values for this transition should exceed
expected values for about half of the children. In fact, observed values for
the Parallel to Group transition exceeded chance for 32 of the 41 children
observed, a deviation from chance that was significant at better than the .01
level (determined by a two-tailed sign test). Thus, Bakeman and Brownlee
concluded, the movement from parallel to group play may be more a matter
of moments than of months, and parallel play may often serve as a brief in-
terlude during which young children have both an increased opportunity to
socialize as well as a chance to "size up" those to whom they are proximal,
before plunging into more involved group activity.
The point of this example is that, given a sequential view, coupled with
what are really quite simple statistics, Bakeman and Brownlee were able to
learn a fair amount about young children's experience in free-play groups.
For one thing, it appears that children changed the focus of their activity
in quite systematic ways, "one step" at a time. Some transitions were
"probable," meaning that observed exceeded expected values for signifi-
cantly more than half of the children. Other transitions were "improbable,"
meaning that observed exceeded expected values for significantly less than
half of the children (for example, observed exceeded expected values for
the Unoccupied to Parallel transition for only 2 of the 41 children). The
remaining transitions were neither probable nor improbable; Bakeman and
Brownlee called them "possible" or "chance" transitions.
The probable transitions all involved either remaining alone (moving
between Unoccupied and Solitary) or else remaining with others (moving
between Together, Parallel, and Group). What is interesting, however, is
how any movement at all occurred between being Alone and being To-
gether. Thus transitions from Unoccupied to Together (adding a social
focus to no focus) and from Solitary to Parallel (adding a combined object
and social focus to an existing object focus) were possible, whereas tran-
sitions from Unoccupied to Parallel (which would require simultaneously
adding both an object and a social focus) or from Solitary to Together
(which would require simultaneously dropping an object focus and adding
a social one) were improbable. Theoreticians can now argue about what
this "one step at a time" model means, but for present purposes, we would
just like to emphasize again that without a sequential view and some sim-
ple sequential techniques, this bit of social process would have remained
undescribed.
12 Introduction

1.8 Hypothesis-generating research

Although we have defined systematic observation in terms of the use of
predetermined categories of behavioral codes, we do not think that an
investigator need have a coding system before collecting data of interest.
For example, Gottman began his research on acquaintanceship in children
(1983) by collecting tape recordings of children becoming acquainted. He
had no idea at first which situations and experimental arrangements were
best for collecting the data. Although the literature suggested some social
processes to study, such as communication clarity, conflict resolution, and
self-disclosure, he had little idea how to operationalize these constructs.
A great deal of work was necessary before a useful coding system was
devised. He also found that several different coding systems were necessary
to capture different aspects of the children's interaction.
Furthermore, we have found in our own research that as an investigator
engages in programmatic research in an area, across a series of studies
much is learned about how the initial coding system operates. This leads
to revisions of the coding system, and, in some cases, to simplifications. For
example, consider conflict resolution in preschool children. An example
of a disagreement chain is as follows (Gottman, 1983):
Host (H): This is stretchy.
Guest (G): No, it's not.
H: Uh huh.
G: Yes.
H: Uh huh.
G: It's dirty.
H: Uh uh.
G: Uh huh.
H: Uh uh.
G: Uh huh.
H: Uh uh.
G: Uh huh.
H: Uh uh. It's not dirty, (p. 27)
These disagreement chains can have fairly neutral affect. An escalated
form of conflict involves negative affect (e.g., anger, crying) and is called
"squabbling." Preschoolers getting acquainted at home do not squabble
very much. Instead, they manage conflict so that it does not escalate. One
of the ways that they do this is by giving a reason for disagreeing. This is a
sequence in which a child disagrees and then gives a reason. For example:
"No, I don't wanna play house. 'Cause I'm not finished coloring." It turns
out that this sequence is very powerful in keeping conflict from escalating,
compared to interactions in which the disagreement is not as likely to be
followed by a reason for disagreeing. This fact was discovered in an initial
study on acquaintanceship; this sequence (disagreement is followed by
 Summary: Systematic is not always sequential 13

giving a reason for disagreeing) then became an operational definition of

a salient social process.
It is perfectly legitimate, in our view, to begin the process of systematic
observation with the simple goal of description. As we gain experience
with the phenomena we are investigating, we learn which variables are
important to us. We can begin our investigation with a search for order.
Usually we have some hypotheses about what we might expect to find.
The wonderful thing about observational research is that it maximizes the
possibility of being surprised.
There is a danger in this hypothesis-generating approach, and this has to
do with the temptation of not thinking very much about what one might ex-
pect, and instead looking at everything. Our experience in consulting leads
us to recognize the danger in this approach. Usually investigators who do
not generate hypotheses at all will be overwhelmed by their data. A delicate
balance must be worked out that is consistent with the researcher's style.
A basic requirement of this kind of exploratory research is that it is essen-
tial to replicate and search for consistency across studies. Our experience
with this approach is positive: We typically find that confusing results do
not replicate, and that interesting, interpretable results do replicate.
To summarize, hypothesis-generating research can play a vital role in
the process of description and in the identification of phenomena. This
kind of observational research is essential in new areas of investigation.
However, it needs to be carefully done by incorporating it in programmatic
research that builds in replication.

1.9 Summary: Systematic is not always sequential

We make two assumptions about readers of this book. First, we as-
sume that they are already convinced that systematic observation is an
important method for measuring behavior. Second, we assume that they
want to explore dynamic, process-oriented aspects of the behavior they
observe. This requires that sequential techniques be added to systematic
observation.
Systematic observation has two major defining characteristics: the use
of predefined behavior codes and a concern with observer reliability. The
method is time consuming but offers a degree of certainty and replica-
bility that narrative reports cannot. Even when systematic observation is
not sequential, much can be learned, as the work of Parten and of Smith
described above demonstrates. In particular, nonsequential systematic ob-
servation can be used to answer questions about how individuals distribute
their time among various activities, or distribute their utterances among
14 Introduction

different utterance categories. The data derived from such studies can
then be used to ask the usual questions regard-ing how different groups of
individuals vary, or how individuals change with age.
Sequential techniques, added to systematic observation, allow a whole
new set of questions to be addressed. In particular, sequential techniques
can be used to answer questions as to how behavior is sequenced in time,
which in turn should help us understand how behavior functions moment
to moment. In fact, for purposes of this book, we use the word "sequen-
tial" to refer to relatively momentary phenomena, not for developmental
phenomena, which are expressed over months or years.
The purpose of this introductory chapter has been to suggest, both in
words and by example, what sequential analysis is, why it is useful, and
what it can do. In the following chapters, we discuss the various compo-
nents required of a study invoking sequential analysis.
 Developing a coding scheme

2.1 Introduction
The first step in observational research is developing a coding scheme. It
is a step that deserves a good deal of time and attention. Put simply, the
success of observational studies depends on those distinctions that early
on become enshrined in the coding scheme. Later on, it will be the job of
observers to note when the behaviors defined in the code catalog occur in
the stream of behavior. What the investigator is saying, in effect, is: This is
what I think important; this is what I want extracted from the passing stream.
Yet sometimes the development of coding schemes is approached almost
casually, and so we sometimes hear people ask: Do you have a coding
scheme I can borrow? This seems to us a little like wearing someone else's
underwear. Developing a coding scheme is very much a theoretical act,
one that should begin in the privacy of one's own study, and the coding
scheme itself represents an hypothesis, even if it is rarely treated as such.
After all, it embodies the behaviors and distinctions that the investigator
thinks important for exploring the problem at hand. It is, very simply, the
lens with which he or she has chosen to view the world.
Now if that lens is thoughtfully constructed and well formed (and aimed
in the right direction), a clearer view of the world should emerge. But
if not, no amount of corrective action will bring things into focus later.
That is, no amount of technical virtuosity, no mathematical geniuses or
statistical saviors, can wrest understanding from ill-conceived or wrong-
headed coding schemes.
How does one go about constructing a well-formed coding scheme?
This may be a little like asking how one formulates a good research ques-
tion, and although no mechanical prescriptions guaranteeing success are
possible, either for coding schemes or for research questions, still some
general guidelines may prove helpful. The rest of this chapter discusses
various issues that need to be considered when coding schemes are being
developed.

15
16 Developing a coding scheme

2.2 What is the question?

Perhaps the single most important piece of advice for those just beginning
to develop a coding scheme is, begin with a clear question. For any child,
mother-infant pair, couple, animal, or group one might want to study, there
are an infinite number of behaviors that could be coded. Without the focus
provided by a clear question, it is hard to know just where to look, and
it is very easy to be overwhelmed. We have all probably experienced at
one time or another how operational goals can take over an activity. For
example, a consultant we knew once found the employees of a large college
athletic complex diligently recording who used what facility for how long
in ways that had become ever more elaborate over the years. It seemed all
too easy to think of more complicated ways to encode the passing stream of
behavior, but that encoding seemed completely unguided by any purpose.
No doubt once there had been some purpose, but apparently it had been
forgotten long ago.
Similarly, many investigators, ourselves included, seem tempted to in-
clude more and more separate codes, and make ever finer distinctions,
simply because it is possible to get observers to record data using such
schemes. There is an argument to be made for such a "wide-net" strategy,
and it usually goes something like this: Because we are not really sure what
will turn out to be important, we need to record everything - or at least lots
- and then scan our voluminous data archive for important features. Some-
how, it is hoped, coherence will emerge from the multitude. However, we
suspect that this happens very rarely, and so when asked, "But what do I
do with all this data that I collected so laboriously?" our first suspicion is
that a clear question, and not statistical expertise, was lacking. Typically,
"categorical overkill" seems to inundate investigators in tedious and not
very fruitful detail, whereas studies involving clearly stated questions and
tightly focused coding schemes seem far more productive.
For example, consider the following question. Among monkeys, whose
early days are spent clinging to their mothers, is it the infants whofirstleave
their mothers as they begin to explore the world, or is it the mothers who first
push their reluctant infants out into the world? Given this question, most of
us would probably agree on the behavior to be recorded and how it should
be categorized. We would want to record separations between mother
and infant and categorize each one as being either initiated by the mother
or initiated by the infant. Thus our coding scheme would contain two
codes - Infant-Initiated Separation and Mother-Initiated Separation - and
recording would be "activated" whenever the event of interest - separation
- occurred. About the only way to complicate this simple example would
be to add a third code: Mutual Disengagement.
 Physically vs. socially based coding schemes 17

With such a simple coding scheme, the progression from data collection
to analysis to interpretation would be simple and straightforward. We might
find, for example, that over the course of the first several months of life the
number of separation episodes gradually increased and then decreased. At
first, there would be few separations because the infant is almost always
clinging to the mother, whereas later on there might be few because the
infant is almost always off the mother, but in between there would be
more movement from and to the mother. Further, the data could show
that when the number of separations is first increasing, the mother initiated
considerably more of them than her infant, whereas later, when the number
of separations begins to decrease, the infant initiated more, leading us to
conclude that it is the mothers who first push their presumably reluctant
infants out into the world.
The point is, developing a coding scheme is theoretical, not mechanical
work. In order to work well, a coding scheme has to fit a researcher's ideas
and questions. As a result, only rarely can a coding scheme be borrowed
from someone else. However, when research questions are clearly stated, it
is a much easier matter to determine which distinctions the coding scheme
should make. Without clear questions, code development is an unguided
affair.

2.3 Physically versus socially based coding schemes

Regardless of whether coding schemes should be borrowed or not - and
our previous paragraph objects only to mindless borrowing, not borrowing
per se - the fact of the matter is that only rarely are coding schemes used in
more than one laboratory. This might seem an undesirable state of affairs
to those who believe that scientific work should be replicable. After all, all
laboratories interested in temperature use thermometers.
The reason for the commonality of thermometers and the multiplicity of
coding schemes has to do, we think, with the nature of the "stuff" being
measured - temperature or social behavior. Now it may be that some
aspects of social behavior can be measured as precisely and commonly as
temperature. Whatever the case, we think it is helpful, when developing a
coding scheme, to distinguish between codes that are physically based, on
the one hand, and socially based, on the other.
At the risk of precipitating a lengthy discussion with the philosophically
inclined, let us state that it seems possible to locate coding schemes along
a continuum. One end would be anchored by physically based schemes
- schemes that classify behavior with clear and well-understood roots in
the organism's physiology - whereas the other end would be anchored
18 Developing a coding scheme

by socially based schemes - schemes that deal with behavior whose very
classification depends far more on ideas in the mind of the investigator (and
others) than on mechanisms in the body. We have called these "socially
based," not because they necessarily deal with social behaviors - even
though they typically do - but because both their specification and their
use depend on social processes. Instead of following quite clearly from
physical features or physical mechanisms in a way that causes almost no
disagreement, socially based schemes follow from cultural tradition or
simply negotiation among people as to a meaningful way to view and
categorize the behavior under discussion. Moreover, their use typically
requires the observer to make some inference about the individual observed.
For example, some people are paid to determine the sex of baby chicks.
The "coding scheme" in this case is simple and obvious: male or female.
This is not an easy discrimination to make, and chicken sexers require a fair
amount of training, but few people would suggest that the categories exist
mainly as ideas in the observers' heads. Their connection with something
"seeable," even if difficult to see, is obvious.
Other people (therapists and students influenced by Eric Berne) go about
detecting, counting, and giving "strokes" - statements of encouragement
or support offered in the course of interaction. In effect, their "coding
scheme" categorizes responses made to others as strokes or nonstrokes. For
some purposes, therapeutic and otherwise, this may turn out to be a useful
construct, but few would argue that "strokes" are a feature of the natural
world. Instead, they are a product of the social world and "exist" among
those who find the construct useful. Moreover, coding a given behavior as
a "stroke" requires making an inference about another's intentions.
Other examples of physically and socially based coding schemes could
be drawn from the study of emotion. For example, Ekman and Friesen's
(1978) Facial Action Coding System scores facial movement in terms of
visible changes in the face brought about by the motion of specific muscle
groups (called action units or "AUs"). The muscles that raise the inner
corner of the eyebrows receive the code "AU1." The muscles that draw the
brows down and together receive the code "AU4." When these muscles
act together, they result in a particular configuration of the brow called
"AU1+4." The brows wrinkle in specific ways for each of these three
action units (see Figure 2.1).
The brow configuration AU1 +4 has been called "Darwin's grief muscle"
as a result of Darwin's 1873 book on the expression of emotion. This
leads us to the point that AU1 and AU1+4 are both brow configurations
typically associated with distress and sadness. In fact, there is not always
a one-to-one correspondence between sadness or distress and these brow
configurations. For example, Woody Allen uses AU1+4 as an underliner
 Physically vs. socially based coding schemes 19

Figure 2.1. Examples of action units from Ekman and Friesen's (1978) Facial
Action Coding System.

after he tells a joke. But in most social interaction there are additional cues
of sadness or distress.
In a physically based coding scheme we would be coding such things
as specific brow configuration, but in a socially based coding scheme we
would be coding such things as sadness. The socially based coding scheme
requires considerably more inference, and probably requires sensitive ob-
servers. However, it is not necessarily less "real" than the system that
records brow action. It is simply a different level of description.
Not all researchers would agree with this last statement. Some, espe-
cially those trained in animal behavior and ethology, might argue that if
the problem is analyzed properly, then any socially based scheme can and
should be replaced with a physically based one. We disagree. In fact, we
think there are often very good reasons for using socially based coding
schemes.
First, it is often the case that physically based schemes, like Ekman and
Friesen's mentioned above, are time consuming to learn and to apply, and
therefore, as a practical matter, it may be much easier to use a socially based
alternative. Even if it is not, a socially based scheme may more faithfully
reflect the level of description appropriate for a given research issue. In any
given case, of course, investigators' decisions are influenced by the problem
at hand and by the audience they want to address, but it seems worth asking,
before embarking on an ambitious observational study, whether something
simpler, and perhaps more in tune with the research question, would not
do as well. Some people will be unhappy with any coding scheme that is
not clearly grounded in the physical world, but others, ourselves included,
will be tolerant of almost any kind of coding scheme, even one that is quite
20 Developing a coding scheme

inferential, as long as it fits the research concerns and as long as trained

observers acting independently can agree (see chapter 4).
Second, often various aspects of the socially created world are exactly
what we want to study, and in those cases socially based coding schemes
are the appropriate "filter" to capture the behavior of interest. Sometimes
investigators are advised to avoid coding schemes that require observers
to make any judgments about intentionality. In one sense, this is sound
advice. Just from the standpoint of observer agreement, there is no reason
to ask observers to make more complex judgments (was the baby trying
to get her mother's attention?) when simpler ones will do (did the baby
approach within 1 meter of her mother and look at her?). But in another
sense, if this advice were followed rigorously, we would end up not study-
ing some very interesting behavior, or else defending some possibly quite
silly coding schemes.
For example, consider the Parten-like coding scheme for play state, de-
scribed above, that Peter Smith used. When using this scheme, observers
need to discriminate between children playing alone and playing with oth-
ers, whether in parallel or in a more engaged state. Now one crucial dis-
tinction is between being alone and being "with" others; however, whereas
observers usually have little trouble agreeing when children are with oth-
ers, it is not an easy matter to define "withness" in a mechanical way. Is
it a matter of proximity? If so, then exactly how near? Is it a matter
of visual regard? If so, then how frequently and/or how long? Or does
it also include elements of posture, orientation, activity, etc.? We have
two choices. Either we define "Alone," "Parallel," and "Group" play in a
mechanical way, dependent solely on physical features, and run the risk
of occasionally violating our own and our observers' intuitions about the
"true" state of affairs; or we view the difference between being alone and
being with others as determined by the children's intent and experienced
by them subjectively, and we make use of the ability that humans (in this
case, our observers) have to judge, more or less accurately (as determined
by agreement or consensus), the intentions of others in their culture.
The question is, should we deny or make use of these very human abilities
(see Shotter, 1978)? In general, we think we should use them, and that it is
especially appropriate to do so when studying social behavior in our fellow
humans. (We recognize that this argument begins to break down when
animals are being studied.) This is not a new issue in psychological studies,
and there is little hope that it will be settled here. Even so, when we argue for
more "mentalistic" categories than some traditional psychologists or animal
behaviorists would admit, we do not intend this as a license for overly rich
and undisciplined speculation and inference. Other things being equal,
we much prefer coding categories that are defined in terms of observable
 Physically vs. socially based coding schemes 21

and concrete features, but we are not willing to let this one consideration
override all others, especially if the meaningfulness of the data collected
or the ability of those data to answer the question at hand might suffer. It
is possible, for example, to record accurately how many times an infant
approaches within 1 meter of his or her mother and what proportion of time
was spent in contact with the mother, within 1 meter of the mother, and
looking at the mother, and still not be able to gauge validly the quality of
the mother-infant relationship. That task might be accomplished better, as
Mary Ainsworth, Alan Sroufe, and others have argued, by asking observers
to rate, on 7-point scales, how much the infant seeks to maintain contact
with the mother, resists contact with the mother, etc., and then assessing the
relationship on the basis of the pattern of the rating scales (see Ainsworth,
Blehar, Waters, & Wall, 1978; Sroufe & Waters, 1977).
Still, when one is developing coding schemes (or rating scales, for that
matter), it is a very useful exercise to describe each behavior (or points on
the rating scale) in as specific a way as possible. For example, Bakeman
and Brownlee (1980), in their parallel play study, required observers to
distinguish between children who were playing alone and who were playing
in parallel with others. First, Bakeman and Brownlee viewed videotapes
of 3-year-olds in a free play situation. They continually asked each other,
is this child in solitary or parallel play? - and other questions, too - and
even when there was consensus, they tried to state in as specific terms as
possible what cues prompted them to make the judgments they did. They
were thus able to make a list of features that distinguished parallel from
solitary play; when engaged in "prototypic" parallel play, children glanced
at others at least once every 30 seconds, were engaged in similar activities,
were within 1 meter of each other, and were no more than 90 degrees away
from facing each other directly.
These features were all described in the code catalog (the written de-
scription of the coding scheme), but the observers were instructed to treat
these defining features as somewhat flexible guidelines and not as rigid
mandates. Once they had thoroughly discussed with their observers what
most writers mean by parallel play, and had described parallel play in their
code catalog, they were willing to let observers decide individual cases
on the basis of "family resemblance" to parallel play. We agree with this
procedure and would argue as follows: By not insisting that observers
slavishly adhere to the letter of the rules, we then make use of, instead of
denying, their human inferential abilities. However, those abilities need to
be disciplined by discussion, by training, and - perhaps most importantly -
by convincing documentation of observer agreement. The result should be
more accurate data and data that can "speak" to the complex questions that
often arise when social behavior and social development are being studied.
22 Developing a coding scheme

2.4 Detectors versus informants

In the previous section we suggested that it is useful to distinguish be-
tween physically based and socially based coding schemes and mentioned
a number of implications of this view. We would now like to suggest one
final implication of this view, one that affects how the observers' work is
regarded and, ultimately, how observers are trained. It is fairly common
to regard observers merely as "detectors," as instruments like any other
whose job is passively to record what is "there," much as a thermometer
records temperature. For physically based coding schemes, this seems like
an appropriate view. In such cases, it is not the human ability to make
inferences about fellow humans that is being exploited by the investigator,
but simply the quite incredible ability of the visual-perceptual system to
"see what is there."
For socially based coding schemes, however, it makes sense to regard
the observer more as a cultural informant than as a detector. When socially
based distinctions are being made, distinctions that typically require some
inference about others' intentions, this strikes us as a more accurate way
for both the researchers and the observers to think about their task. At
least one code catalog that we know of, which was developed during the
course of a study of citizen-police interaction, presents the observers' task
to them exactly in this way (Wallen & Sykes, 1974). We think that this
is a good idea, and that observers using socially based coding schemes do
a better job if they appreciate the "participant-observer" aspects of their
observation task.

2.5 The fallacy of equating objectivity with physically

based schemes
Sometimes physically based coding schemes are mistakenly labeled as
more "objective" than socially based systems. Let us consider why this is
a fallacy. In the study of affect, one of our colleagues, Mary Lynn Fletcher,
pointed out that if you employed observers to study the conversation of
Chinese people, it would be absurd to hire observers who could not un-
derstand Chinese, claiming that this makes them more "objective." In an
analogous way, one can argue that to code complex social processes such
as emotions, it is often necessary for observers to be able to understand
the "language" of affect. This requires that coders be expert cultural infor-
mants about emotion and also be able to respond subjectively. The key to
clarifying the issue of "objectivity" is to stress, not objectivity as such, but
replicability instead. Independent observers should agree.
 Keeping it simple 23

There is a second issue inherent in the physically based versus socially

based systems, namely, the level of description that is of interest to the in-
vestigator. For example, expert cultural informants have a complex catalog
of physical cues that they can call upon to decide that a husband is angry in
a marital discussion. If an investigator is interested in emotional categories
such as angry, sad, contemptuous, etc., then it may be fruitless to attempt to
construct a complete list of cues for each emotion judgment. Furthermore,
it may not even be possible because the kinds of cues that go into a judgment
of anger could be infinite, varying with context, the words spoken, pause,
stress, rhythm, amplitude, major or minor key, contours of speech, facial
expressions, gestures, and the nonadditive interaction of these features. We
may not be interested in a description such as the following:
At 5 :19 the husband lowered his brows, compressed his lips, lowered the pitch of
his voice, disagreed with his wife, while folding his arms in the akimbo position.
Instead, we may simply wish to describe the husband as angry. As long as
we know what we mean by "angry," which will be required if we are ever to
achieve reliability between observers, incidents such as the one previously
described illustrate the construct of anger. They need not exhaust it.
In other words, competent observers can make extremely complex so-
cial judgments reliably and need not be viewed as inherently undesirable
measuring instruments. At times they are exactly what we need, and their
subjective observational skills can be precise and the best scientific choice.

2.6 Keeping it simple

Some readers, because of their intellectual convictions, the nature of the
problems they study, or both, may not have found the foregoing discussion
very helpful. Still, they could probably agree with advice to keep coding
schemes simple. Like much good advice, this is so familiar that it seems
trite, yet like maintaining liberty, maintaining simplicity also requires con-
stant vigilance. The "natural" press of events seems against simplicity, at
least with respect to coding schemes. The investigator's question may be
only vaguely apprehended, as discussed earlier, a coding scheme may have
been borrowed from someone else inappropriately, the conceptual analysis
on which a coding scheme is based may not have been worked through
sufficiently, or a proposed scheme may not have been adequately refined
through pilot observations and critical discussion. All of these circum-
stances can confuse the issue and result in overly elaborated and unwieldy
coding schemes.
Several points may be useful for keeping the coding system simple. First,
it is important to have clear conceptual categories that are essentially at the
24 Developing a coding scheme

same level of description. For example, it is not desirable to ask coders to

decide whether an utterance was a question and at the same time whether
the utterance was responsive to the previous speaker. These judgments are
made on two different levels of conceptual complexity. They should not
be combined in one coding scheme. Second, codes ought to be reasonably
distinct; that is, behavior categories should be homogeneous, which means
that even when acts appear somewhat similar, they should not be put in
the same category if there is good reason to think that either their causes
or their functions are different. Third, it is better to "split" than to "lump"
codes. Behavior categories can always be lumped together during data
analysis, if that seems useful, but behaviors lumped together by the coding
scheme cannot be split out later (unless we have a taped record and recode
specific moments).

2.7 Splitting and lumping

At some point in discussions of this sort, the "level of analysis" problem
is frequently broached (see, e.g., Hartup, 1979; Suomi, 1979). We can
conceive of almost any phenomenon as consisting of levels, hierarchically
arranged, with larger and more inclusive or more molar concepts occupy-
ing each higher level, and smaller and more detailed or more molecular
concepts occupying each lower level. Then the question arises, what level
should our coding categories occupy? In the abstract, without the context
of a particular question, an absolute answer to this question is hardly pos-
sible, but a relative one is. First we need to decide what conceptual level
seems appropriate for the question at hand (which is easier said than done);
then we should choose coding categories no higher than that level. In fact,
there is considerable merit in locating at least some of the categories one
level below, on a slightly more detailed level that seems required.
There are at least three reasons why using coding categories that represent
a somewhat more molecular level than the level planned for analysis may be
a desirable strategy. For example, imagine that our question concerned how
often 2-, 3-, and 4-year-old children accompany vocalizations with gestures
when directing their attention to others. The question requires only that
"bids" for attention be tallied and the presence or absence of vocalizations
and of gestures for each bid be recorded. However, we could ask our ob-
servers to record whether the bid was directed to an adult or a peer, whether
a vocalization involved language or not, whether a nonverbal vocalization
was of high or low intensity, whether a verbalization was a question, a
command, or a comment, etc. Dropping to a coding level more molecular
than required for the question might seem to place additional burdens on
 Splitting and lumping 25

the observers, but paradoxically we think that this strategy increases the
chances of collecting reliable data. Just as it is often easier to remember
three elements, say, instead of one, if those three are structured in some way,
so too observers are often more likely to see and record events accurately
when those events are broken up into a number of more specific pieces (as
long as that number is not too great, of course). This seems to provide the
passing stream of behavior with more "hooks" for the observers to grab.
Further, when data are collected at a somewhat more detailed level than
required, we are in a position to justify empirically our later lumping. Given
the coding scheme presented in the previous paragraph, for example, a
critic might object that the different kinds of vocalization we coded are so
different that they should not be dealt with as a single class. Yet if we can
show that the frequency of gesture use was not different for the different
kinds of vocalizations in the different age groups, then there would be no
reason, for these purposes, to use anything other than the lumped category.
Moreover, the decision to lump would then be based on something more
than our initial hunches.
Finally, and this is the third reason, more detailed data may reveal some-
thing of interest to others whose concerns may differ from ours, and at the
same time may suggest something unanticipated to us. For example, given
the coding scheme described above, we might find out that how gestures
and vocalizations are coordinated depends on whether the other person
involved is a peer or an adult, even if initially we had not been much in-
terested in, nor had even expected, effects associated with the interactive
partner.
We should note that often the issue of level of analysis is not the same
issue as whether to code data at a detailed or global level. We may have
a set of research questions that call for more than one coding system.
For example, Gottman and Levenson are currently employing a socially
based coding scheme to describe emotional moments as angry, sad, etc.
Observers also note if there was facial movement during each emotional
moment. These facial movements are then coded with a detailed, physically
based coding system, Ekman and Friesen's Facial Action Coding System
(FACS). Gottman and Levenson collected psychophysiological data while
married couples interacted. One research question concerns whether there
are specific physiological profiles for specific categories of facial expres-
sions. The FACS coding is needed to address this question, but in the
Gottman and Levenson study a decision had to be made about sampling
moments for FACS coding because detailed FACS coding is so costly. The
socially based system is thus used as an aid to employing a more detailed,
physically based coding system. Coding schemes at different levels of
analysis can thus be used in tandem within the same study.
26 Developing a coding scheme

2.8 Mutually exclusive and exhaustive codes

Almost all the examples of coding schemes presented so far consist of mu-
tually exclusive and exhaustive codes. This means that only one code can be
associated with a particular event (mutually exclusive) but that there is some
code for every event (exhaustive). For example, with respect to Parten's
six social participation codes, only one category was appropriate for each
1-minute time sample, but all time samples could be categorized. Obser-
vational studies do not require that all coding schemes consist of mutually
exclusive and exhaustive codes, but in fact such schemes have several de-
sirable features - their construction requires a certain amount of conceptual
analysis, for example, and their use can simplify data analysis - and as a re-
sult such schemes are frequently encountered. Some writers even state that
coding schemes must consist of mutually exclusive and exhaustive codes,
but there are other possibilities, as discussed in the following paragraphs.
In principle, of course, codes for any behavior can be defined in a way that
makes them mutually exclusive and exhaustive (ME&E; see S. Altmann,
1965). For example, if we were interested in the coordination of vocal
and visual behavior during face-to-face interaction, we might record (a)
when person A was looking at, and (b) when person A was talking to his
or her partner. Now these two behavioral codes can cooccur and so are
not mutually exclusive, but if we regarded their coocurrence as a new or
different code, then we could construct an ME&E coding scheme consisting
of four codes: (a) A looks, (b) A talks, (c) A both looks and talks, and (d)
the "null" code, A neither looks nor talks.
The coding scheme consisting of the two nonmutually exclusive be-
haviors may offer certain advantages. For one thing, observers have to
remember only two behavioral codes and not three (or four, counting the
"null" code). When only two nonmutually exclusive "base behaviors" are
under consideration, as in the present example, the difference between their
number and the number of ME&E "behavior patterns" they generate is not
great, but with only a few more base behaviors the number of possible
patterns becomes huge. For example, with four base codes there are 16
patterns, withfivebase codes, 32 patterns, etc., which could rapidly result in
observer overload. Moreover, if the times of onsets and offsets for the base
behaviors were recorded, data consisting of ME&E categories could always
be generated later, if such data were required for subsequent analyses.
It should be noted that coding time may be increased in some ME&E
systems because of the nature of the decisions the coder has to make. This
applies to schemes that include codes of the sort (a) event A, (b) event B,
(c) both event A and B coocurring. If coding time is an issue, an alternative
is to have coders use a checklist of items that can cooccur and need not be
exhaustive, or a rating system. This is like having coders fill out a brief
 The evolution of a coding system 27

"questionnaire" after every interactive unit occurs. Coders decide about

each "item" of the "questionnaire" independently of every other item. Each
item can still be precisely defined.

2.9 The evolution of a coding system

Rosenblum (1978) described the initial stages involved in the creation of
a coding system. First, he discussed the importance of establishing the
conditions of observation. In particular, the situation selected will affect
the diversity of behavior displayed, which will, in turn, determine the
complexity of the coding system. Second, he suggested beginning with
informal observation of behavior. He wrote,
[I]t is best to begin in the most unstructured fashion possible. There is great
advantage to beginning such observations with only a pencil and blank pad for
recording, putting aside the spectres of anthropomorphism, adultomorphism, or
any of the other rigidifying constraints that must be imposed in separating wheat
from chaff later on in the development of the research program; it is vital to
begin by using the incredible synthesizing and integrative functions of the human
mind. . . . At the beginning, the observer must simply watch, allowing individual
subjects to arise as separate entities in the group and behavior patterns to emerge
as figures against the background of initially amorphous activity, (pp. 16-17)

We suggest that writing narrative summaries is very helpful at this stage.

From the narrative a set of codes is generated, ideally an ME&E set.
As research experience is obtained using a particular coding scheme,
it can be modified. For example, Patterson and Moore (1979) discussed
using interactive patterns as units of behavior. They analyzed the behavior
of Tina using the Family Interaction Coding System (FICS) and found an
organization of FICS codes in time. These could later become units of
observation. They wrote:
Examination of Tina's aversive behavior showed that they tended to occur in
bursts, followed by periods of neutral or prosocial behavior.... It seemed, then,
that the bursts, themselves, would be a likely area to search next for structure in
social interaction, (p. 83)

Next they noted what FICS behavior initiated and maintained these bursts
and identified a set of sequences characteristic of the bursts. Using the
larger interaction units, they discovered evidence for what they called an
"interactional ripple effect," by which they meant the increased likelihood
of the initiating event of the chain occurring once a chain has been run off.
There are many consequences of employing a coding system in a se-
ries of studies. New codes may appear, new distinctions may be made, or
28 Developing a coding scheme

distinctions may disappear as a new lumping scheme is derived. Gottman

(1979a) used sequential analysis of a 28-code system for coding marital
interaction to create eight summary codes. Two codes were lumped to-
gether only if they were functionally similar and sequentially related. For
example, the "yes but" code functioned in a similar way to simple disagree-
ment in the sense that both led to disagreement by the spouse, so they were
lumped together. Gottman argued against the logic of using factor analysis
to lump codes. He wrote:
The problem with factor analysis in this application is that it lumps two codes on
the basis of high correlations between these two codes across subjects. However,
just because subjects who do a lot of code A also do a lot of code B does not
imply that these two codes are functionally equivalent, (p. 91)

A coding system can evolve as it is being used by intelligent coders. To

make this systematic, require observers to write a narrative summary of
each observation session, keeping a log of any new behavior that occurs
and seems important. Then ask observers to note the contexts in which the
new behavior occurs, its antecedents and consequences. The new behavior
may be part of a functional behavior set already described, or it may require
a category of its own.

2.10 Example 1: Interaction of prehatched chickens

In this chapter and the previous one, we have given several examples of
coding schemes. For the most part, these have been quite simple, coding
just one kind of behavior, like social participation, with just a few mutually
exclusive and exhaustive codes. We would like to end this chapter by
describing five somewhat more complex coding schemes. Thefirstinvolves
chickens. In order to study interactions between not-yet-hatched chickens
and their setting hens, Tuculescu and Griswold (1983) defined 16 codes,
organized as follows:
Embryonic behaviors
Distress-type calls
1. Phioo
2. Soft Peep
3. Peep
4. Screech
Pleasure-type calls
5. Twitter
 Example 1: Interaction of prehatched chickens 29

6. Food Call
7. Huddling Call
Maternal behaviors
Body movements
8. Undetermined Move
9. Egg Turn
10. Resettle
Head movements
11. Peck
12. Beak Clap
Vocalizations
13. Cluck
14. Intermediate Call
15. Food Call
16. Mild Alarm Call
To those familiar with the behavior of chickens, these codes appear
"natural" and discrete. Trained observers apparently have no trouble dis-
criminating, for example, between a Phioo, a Soft Peep, a Peep, and a
Screech, each of which in fact appears somewhat different on a spectro-
graphic recording. Thus "physical reality" may undergird these codes, but
human observers are still asked to make the determinations.
These codes are also clearly organized. There are three levels to this
particular hierarchy. On the first level, embryonic and maternal behavior
are distinguished; on the second, different kinds of embryonic (distress
and pleasure calls) and different kinds of maternal (body movements, head
movements, vocalizations) behavior are differentiated; and on the third, the
codes themselves are defined. Within each "second-level" category, codes
are mutually exclusive, but codes across different second-level categories
can cooccur. Indeed, cooccurrence of certain kinds of behavior, like em-
bryonic distress calls and maternal body movements, was very much of
interest to the investigators.
There are at least three reasons why we think organizing codes in this
hierarchical fashion is often desirable. First, it both ensures and reflects a
certain amount of conceptual analysis. Second, it makes the codes easier to
explain to others and easier for observers to memorize. Third, it facilitates
analysis. For example, for some analyses all embryonic distress calls and
all maternal vocalizations were lumped together, which is an example of a
practice we recommended earlier - analyzing on a more molar level than
that used for data collection.
30 Developing a coding scheme

2.11 Example 2: Children's conversations

The second example is derived from a study conducted by John Gottman
(1983) on how children become friends. Working from audiotapes, ob-
servers categorized each successive "thought unit" in the stream of talk
according to a catalog of 42 mutually exclusive and exhaustive content
codes. These 42 codes were grouped into seven superordinate categories:
(a) demands for the other child, (b) demands for the pair, (c) you and
me statements, (d) self-focus statements, (e) emotive statements, (f) social
rules, and (g) information exchange and message clarification statements.
Here we reproduce definitions and examples just for the first 16 codes (from
Gottman, 1983, p. 13):
Demands for the other child
1. Command (Gimme that.)
2. Polite Requests (That one, please.)
3. Polite Request in Question Form (Would you gimme that?)
4. Suggestion (You could make that black.)
5. Suggestion in Question Form (Why don't you make that black?)
6. Asking Permission (Can I play with that now?)
7. Demands in the Form of an Information Statement (I think my crayons are
next to you.)
8. Demands in the Form of a Question for Information (Have you got any
sixes?)
9. Wanna (I wanna play house.)
10. Question Wanna (Do you wanna play house?)
11. Requirements for the Other Child (You should stay in the lines.)
12. Asks Help (Would you tie this for me?)
We demands (demands for the pair)
13. Hafta Wanna (We have to take a nap.)
14. Let's (Let's play house.)
15. Let's in Question Form (How about drawing now?)
16. Roles to Both (You be the cop and I'll be the robber.)

Unlike with the codes for chicken behavior described above, it is hard to
claim that any physical reality undergrids Gottman's content codes. For that
very reason, he took considerable pains to demonstrate observer reliability
for his codes - which Tuculescu and Griswold did not. Gottman also made
finer distinctions when defining his codes than he found useful for later
analyses - which is natural tendency when the cleavage between codes is
not all that clear. Still, like Tuculescu and Griswold if for slightly different
reasons, he found it useful to lump codes for analysis. In fact, the initial 42
content codes were reduced to 20 for his analyses of friendship formation.
The three codes derived from the 16 initial codes list above were (a) Weak
demands - numbers 2, 3, 5, 6, 7, 8, and 10 above; (b) Strong Demands -
 Example 2: Children's conversations 31

numbers 1,4,9,11, and 12 above; and (c) Demands for the Pair - numbers
13, 14, 15, and 16 above.
Using the result of sequential analysis from a detailed coding system,
Gottman devised a "macro" coding system. The macro system was de-
signed so that it would be faster to use (2 hours per hour of tape instead of
30) and would code for the sequences previously identified as important. In
the process of building the macro system, new codes were added because
in moving to a larger interaction unit, he noticed new phenomena that had
never been noticed before. For example, the codes escalation and deesca-
lation of a common-ground activity were created. Gottman (1983) wrote:
Escalation and deescalation of common-ground activity were included as cate-
gories because it appeared that the children often initially established a relatively
simple common-ground activity (such as coloring side by side) that made low
demands of each child for social responsiveness. For example, in coloring side
by side, each child would narrate his or her own activity (e.g., "I'm coloring mine
green"). This involved extensive use of the ME codes. Piaget (1930) described
this as collective monologue, though such conversation is clearly an acceptable
form of dialogue. However, in the present investigation the common-ground
activity was usually escalated after a while. This anecdotal observation is consis-
tent with Bakeman and Brownlee's (1980) recent report that parallel play among
preschool children is usually the temporal precursor of group play. However, the
extent of this process of escalation was far greater than Bakeman and Brownlee
(1980) imagined. An example of this escalation is the following: Both children
begin narrating their own activity; then one child may introduce INX codes (nar-
ration of the other child's activity - e.g., "You're coloring in the lines"); next, a
child may begin giving suggestions or other commands to the other child (e.g.,
"Use blue. That'd be nice"). The activity escalates in each case in terms of the
responsiveness demand it places on the children. A joint activity is then suggested
and the complexity of this activity will be escalated from time to time.
This escalation process was sometimes smooth, but sometimes it introduced
conflict. When it did introduce conflict, the children often deescalated that ac-
tivity, either returning to a previous activity that they had been able to maintain
or moving to information exchange. While many investigators have called at-
tention to individual differences in the complexity of children's dialogue during
play (e.g., Garvey, 1974; Garvey & Berndt, 1977), the anecdotal observation here
is that a dyad will escalate the complexity of the play (with complexity defined
in terms of the responsiveness demand) and manage this complexity as the play
proceeds. I had not noticed this complex process until I designed this coding
system. However, I do not mean to suggest that these processes are subtle or hard
to notice, but only that they have until now been overlooked. An example will
help clarify this point. D, the host, is 4-0; and J, the guest, is 4-2. They begin
playing in parallel, but note that their dialogue is connected.

18. J: I got a fruit cutter plate.

19. D: Mine's white.
32 Developing a coding scheme

20. J: You got white Play-Doh and this color and that color.
21. D: Every color. That's the colors we got.
They continue playing, escalating the responsiveness demand by using strong
forms of demands.
29. D: I'm putting pink in the blue.
30. J: Mix pink.
31. D: Pass the blue.
32. J: I think I'll pass the blue.
They next move toward doing the same thing together (common-ground activity).
35. D: And you make those for after we get it together, OK?
36. J: 'Kay.
37. D: Have to make these.
38. J: Pretend like those little roll cookies, too, OK?
39. D: And make, um, make a, um, pancake, too.
40. J: Oh rats. This is a little pancake.
41. D: OK. Make, make me, um, make 2 flat cookies. Cause I'm, I'm cutting
any, I'm cutting this. My snake.
nex escalation includes offers.
The next
54. J: You want all my blue?
55. D: Yes. To make cookies. Just to make cookies, but we can't mess the
cookies all up.
56. J: Nope.
They then introduce a joint activity and begin using "we" terms in describing
what the activity is:
57. D: Put this the right way, OK? We're making supper, huh?
58. J: We 're making supper. Maybe we could use, if you get white, we could
use that too, maybe.
59. D: I don't have any white. Yes, we, yes I do.
60. J: If you got some white, we could have some, y'know.
As they continue the play, they employ occasional contextual reminders that this
is a joint activity:
72. D: Oh, we've got to have our dinner. Trying to make some.
D then tries to escalate the play by introducing some fantasy. This escalation is
not successful. J is first allocated a low-status role (baby), then a higher-status
role (sister), then a higher-status (but still not an equal-status) role (big sister).
76. D: I'm the mommy.
77. J: Who am I?
78. D: Um, the baby.
79. J: Daddy.
80. D: Sister.
81. J: I wanna be the daddy.
82. D: You're the sister.
 Example 3: Baby behavior codes 33

83. J: Daddy.
84. D: You're the big sister!
85. J: Don't play house. I don't want to play house.
The escalation failure leads to a deescalation.
87. J: Just play eat-eat. We can play eat-eat. We have to play that way.
However, in this case, the successful deescalation was not accomplished without
some conflict:
89. J: Look hungry!
90. D: Huh?
91. J: I said look hungry!
92. D: Look hungry? This is dumb.
93. J: Look hungry!
94. D: No!
The children then successfully returned to the previous level of common ground
activity, preparing a meal together. Common ground activity is thus viewed in
this coding system as a hierarchy in terms of the responsiveness it demands of
each child and in terms of the fun it promises, (pp. 55-57)

2.12 Example 3: Baby behavior codes

Our third example describes the behavior of young infants during times
when they are being administered a neonatal assessment and was devel-
oped by Sharon Landesman-Dwyer (1975). The scheme contains 50 codes
grouped into five superordinate categories. The five categories are not mu-
tually exclusive. In fact, they code different possible kinds of behavior,
namely, (a) the kind of external stimulation being provided and what the
infant is doing with his or her (b) eyes, (c) face (including vocalizations),
(d) head, and (e) body. The ten codes within each of these superordinate
categories, however, are mutually exclusive and exhaustive. For example,
the 10 codes for eyes are:
0. Can't See
1. Closed
2. Slow Roll
3. Squint
4. Open-Close
5. Daze
6. Bright
7. Tracking
8. REM
9. Other
34 Developing a coding scheme

and the 10 codes for face and vocalization are

0. Repose
1. Small Move
2. Mouth/Suck
3. Grimace
4. Smile
5. Tremor
6. Yawn
7. Whimper
8. Cry
9. Cry/Tremor

2.13 Example 4: Children's object struggles

The fourth example is taken from a study of the social rules governing
object conflicts in toddlers and preschoolers, conducted by Bakeman and
Brownlee (1982). They defined six codes, organized into three superor-
dinate categories. Each superordinate category can be cast as a question
about an object struggle as follows:
Prior possession: Did the child now attempting to take an object
from another child play with the contested object at any point
in the previous minute?
1. Yes
2. No
Resistance: Does the child now holding the object resist the at-
tempted take?
1. Yes
2. No
Success: Is the child who is attempting to take the object successful
in getting it?
1. Yes
2. No
The four schemes described above are alike in at least one sense. In each
case, the codes defined can be grouped into a relatively small number of
superordinate categories. They are also alike in that all codes within a su-
perordinate category are at least mutually exclusive and, with the possible
exception of Tuculescu and Griswold, exhaustive as well. Gottman's super-
ordinate categories are themselves mutually exclusive, however, whereas
that is not the case for the other three schemes. In fact, the superordinate
 Example 5: Monkeys' activities 35

categories for the other three schemes represent different modes or kinds of
behavior or different questions about a particular behavioral event and are
clearly not mutually exclusive (with the exception of embryonic distress
and pleasure calls).
The Landesman-Dwyer and the Bakeman and Brownlee schemes are
formally identical. Both consist of several sets of mutually exclusive and
exhaustive codes. This is a useful structure for codes because it ensures that
answers to a number of different questions will be answered: What is the
baby doing with his eyes? With his mouth? Did the taker have prior posses-
sion? They differ, however, in when the questions are asked. Landesman-
Dwyer's scheme is used to characterize each successive moment in time,
whereas Bakeman and Brownlee's scheme is used to characterize a partic-
ular event and is "activated" only when the event of interest occurs.

2.14 Example 5: Monkeys' activities

The fifth example is from the work of Suomi and his coworkers concerning
the social development of rhesus monkeys (Suomi, Mineka, & DeLizio,
1983). They defined 14 codes for monkeys' activities that, although not
necessarily mutually exclusive, were designed to be exhaustive. Observers
were asked to record frequency and duration information for each. The
14 codes (along with brief definitions) were as follows (paraphrased from
Suomi et al., 1983, p. 774):
1. Mother-Infant Ventral (mutual ventral and/or nipple contact)
2. Mother-Infant Reject (any successful or unsuccessful break of
Mother-Infant Ventral, or rejection of an infant-initiated Mother-
Infant Ventral)
3. Ventral Cling (ventral contact with an animal other than the mother)
4. Self-Clasp (clasping any part of own body)
5. Self-Mouth (oral contact with own body)
6. Environmental Exploration (tactual or oral manipulation of inani-
mate objects)
7. Passive (absence of all social, exploratory, and locomotor activity;
could cooccur with self categories and vocalizations)
8. Stereotypy (patterned movements maintained in a rhythmic and
repetitive fashion)
9. Locomotion (any self-induced change in location of self, exclusive
of stereotypy)
10. Vocalization (any sound emitted by subject)
11. Huddle (a maintained, self-enclosed, fetal-like position)
36 Developing a coding scheme

12. Play and Sex (any type of play and/or sexual posturing, exclusive
of locomotion)
13. Aggression (vigorous and/or prolonged biting, hair pulling, clasp-
ing, accompanied by one or more of threat, barking, piloerection,
or strutting)
14. Social Contact (contact and/or proximity with another subject, ex-
clusive of Mother-Infant Ventral, Ventral Cling, Aggression, or
Play and Sex)
Except for Passive (which can cooccur with Self-Clasp, Self-Mouth, and
Vocalization), these codes appear to be mutually exclusive. In some cases,
activities that could cooccur have been made mutually exclusive by defini-
tion. For example, if an activity involves both Stereotypy and Locomotion,
then Stereotypy is coded. Similarly, if what appears to be Social Contact
involves a more specific activity for which a code is defined (like Play
and Sex), then the specific code takes precedence. Defining such rules of
precedence is, in fact, a common way to make a set of codes mutually
exclusive.

2.15 Summary
No other single element is as important to the success of an observational
study as the coding scheme. Yet developing an appropriate scheme (or
schemes) is often an arduous task. It should be assumed that it will in-
volve a fair number of hours of informal observation (either "live" or
using videotapes), a fair amount of occasionally heated discussion, and
several successively refined versions of the coding scheme. Throughout
this process, participants should continually ask themselves, exactly what
questions do we want to answer, and how will this way of coding behavior
help us answer those questions?
There is no reason to expect this process to be easy. After all, quite apart
from our current research traditions, developing "coding schemes" (making
distinctions, categorizing, developing taxonomies) is an ancient, perhaps
even fundamental, intellectual activity. It seems reasonable to view one
product of this activity, the coding scheme, as an informal hypothesis, and
the entire study in which the coding scheme is embedded as a "test" of that
hypothesis. If the "hypothesis" has merit, if it is clearly focused and makes
proper distinctions, then sensible and interpretable results should emerge.
When results seem confused and inconclusive, however, this state of affairs
should not automatically be blamed on a lack of proper data-analytic tools
for observational data. First we should ask, are questions clearly stated,
and do the coding schemes fit the questions? We hope that a consideration
 Summary 37

of the various issues raised in this chapter will make affirmative answers
to these questions more likely.
In this chapter, we have confined our discussion to coding schemes. Five
examples of developed schemes were presented, and an additional four are
detailed in the companion volume (Bakeman and Quera, 1995a, chapter
2). We have stressed in particular how coding schemes can be organized
or structured and have left for the next chapter a discussion of how coding
schemes are put into use. This separation is somewhat artificial. How
behavioral sequences are to be recorded can and often does affect how codes
are defined and organized in the first place. This is especially evident when
the Landesman-Dwyer and the Bakeman and Brownlee schemes discussed
earlier are compared. Still, a scheme like Landesman-Dwyer's could be
recorded in two quite different ways, as we discuss in the next chapter. It
is the task of that chapter to describe the different ways behavior can be
recorded, once behavioral codes have been defined.
 Recording behavioral
sequences

3.1 Recording units: Events versus intervals

The title of this chapter, Recording behavioral sequences, suggests two
different but somewhat related topics. The chapter could deal primarily
with mechanical matters and describe devices used to record data, or it
could deal just with conceptual matters and describe different strategies for
collecting data about behavior sequences. Actually, this chapter will be a
mixture of both, although conceptual and strategic matters will be stressed.
One important issue has already been raised by the example of the Bake-
man and Brownlee (1980) study of parallel play in chapter 1, and that is
the issue of "units." Before selecting a particular recording strategy, an
investigator needs first to decide what "units" are to be used for recording.
As we use the term, the recording unit identifies what prompts the ob-
server to record, and usually is either an interval or an event. For example,
an investigator might choose to code time intervals, as Bakeman and Brown-
lee did, assigning codes to successive time intervals. This is a common
strategy, but for many purposes more accurate data result when the events
themselves are coded instead. In such cases, observers wait for an event
of interest to occur. When one occurs, they code it (i.e., note what kind of
event it was) and perhaps record onset and offset times for the event as well.
Which is better, to code events or to code intervals? That depends on a
number of factors, including the kind and complexity of the coding scheme,
the desired accuracy for the data, and the kind of recording equipment
available. These issues are discussed later on in this chapter in the context of
talking about applications of specific recording strategies, but first we want
to make another distinction, one between momentary and duration events.

3.2 Momentary versus duration events

Sometimes investigators are concerned only with how often certain events
occur, or in what order they occur, and are not much concerned with how
38
 Continuous vs. intermittent recording 39

long they last. At other times, duration - the mean amount of time a par-
ticular kind of event lasts or the proportion of time devoted to a particular
kind of event - is very much of concern. As a result, many writers have
found it convenient to distinguish between "momentary events" (or fre-
quency behaviors) on the one hand, and "behavioral states" (or duration
behaviors) on the other (J. Altmann, 1974; Sackett, 1978). The distinction
is not absolute, of course, but examples of relatively brief and discrete, mo-
mentary events could include baby burps, dog yelps, child points, or any of
Gottman's thought unit codes described in section 2.11, whereas examples
of duration events could include baby asleep, dog hunting, child engaged in
parallel play, or any of Landesman-Dwyer's baby behavior codes described
in section 2.12.
One particular way of conceptualizing duration events is both so common
and so useful it deserves comment. Often researchers view the events they
code as "behavioral states." Typically, the assumption is that the behav-
ior observers see reflects some underlying "organization," and that at any
given time the infant, animal, dyad, etc., will be "in" a particular state. The
observers' task then is to segment the stream of behavior into mutually ex-
clusive and exhaustive behavioral states, such as the arousal states often de-
scribed for young infants (REM sleep, quiet alert, fussy, etc.; Wolff, 1966).
The distinction between momentary and duration events (or between dis-
crete events and behavioral states) seems worth making to us, partly because
of the implications it may have for how data are recorded. When the investi-
gator wants to know only the order of events (for example, Gottman's study
of friendship formation) or how behavioral states are sequenced (Bakeman
and Brownlee's study of parallel play), then the recording system need not
preserve time. However, if the investigator wants also to report proportion
of time devoted to the different behavioral states, then time information
of course needs to be recorded. In general, when duration matters, the
recording system must somehow preserve elapsed time for each of the
coded events. Moreover, when occurrences of different kinds of events
are to be related, beginning times for these events need to be preserved as
well. (Examples are provided by Landesman-Dwyer and by Tuculescu and
Griswold; see section 3.5, Recording onset and offset times.)

3.3 Continuous versus intermittent recording

Before discussing particular recording strategies as such, there is one more
distinction that we would like to make. Almost all strategies we describe
here are examples of continuous, not intermittent, recording. The phrase
"continuous recording" evokes an event recorder, with its continuously
40 Recording behavioral sequences

Figure 3.1. An example of continuous recording: A deflected-pen event recorder.

moving roll of paper and its pens, ready to record events by their deflection
(see Figure 3.1). It is a rather cumbersome device, rarely used in observa-
tional research. Almost always, researchers prefer pencil, paper, and some
sort of clock, or else (increasingly) an electronic recording device. Still,
the phrase "continuous recording" seems appropriate for the strategies we
describe here, not because the paper rolls on, but because the observers are
continuously alert, paying attention, ready to record whenever an event of
interest occurs, whenever a behavioral state changes, or whenever a specific
time interval elapses.
Given that this is a book about sequential analysis in particular, and
not just systematic observation in general, the emphasis on continuous
recording is understandable. After all, for sequential analysis to make
much sense, the record of the passing stream of behavior captured by the
coding/recording system needs to be essentially continuous, free of gaps.
However, we do discuss intermittent recording in section 3.9 (Nonsequen-
tial considerations: time sampling).
The purpose of the following sections is to describe different ways of col-
lecting observational data, including recording strategies that code events
and ones that code intervals. For each way of collecting data, we note
what sort of time information is preserved, as well as other advantages and
disadvantages.

3.4 Coding events

In this section we shall discuss event-based coding. The basic aspect to
take note of is the observer's task, and, in particular what gets the observer
 Coding events 41

Figure 3.2. Tallying events with a simple checklist.

to record a particular code. When the events of interest, not a time interval
running out, are what stir an observer into action, we would say that an event
coding strategy is being used to record observational data. The simplest
example of event coding occurs when observers are asked just to code
events, making no note of time. For example, an investigator might be
interested in how often preschool children try to hit each other, how often
they quarrel, and how often they ask for an adult's aid. The observer's task
then is simply to make a tally whenever one of these codable events occurs.
Such data are often collected with a "checklist." The behavioral codes are
written across a page, at the top of columns. Then when a codable event
occurs, a tally mark is made in the appropriate column. No doubt our
readers are already quite aware of this simple way of collecting data. Still,
it is useful whenever investigators want only to know how often events of
interest occur (frequency information) or at what rate they occur (relative
frequency information) (see Figure 3.2).
Such data can be important. However, of more immediate concern to
us, given the focus of this book, are event coding strategies that result
in sequential data. For example, Gottman segmented the stream of talk
into successive thought units. Each of these events was then coded, pro-
viding a continuous record of how different kinds of thought units were
sequenced in the conversations Gottman tape-recorded. Similarly, Bake-
man and Brownlee (who actually used an interval coding strategy) could
have asked observers to note instead whenever the play state of the child
they were observing changed. Each new play state would then have been
coded, resulting in a record of how different kinds of play states were
sequenced during free play (see Figure 3.3).
42 Recording behavioral sequences

A P A P ft A frP & P A

Figure 3.3. Recording the sequence of events.

In these two examples, the basic requirement for sequential data - con-
tinuity between successive coded units - is assured because the stream of
talk or the stream of behavior is segmented into successive events (units) in
a way that leaves no gaps. However, sequential data may also result when
observers simply report that this happened, then that happened, then that
happened next, recording the order of codable events. Whether such data
are regarded as sequential or not depends on how plausible the assumption
of continuity between successive events is, which in turn depends on the
coding scheme and the local circumstances surrounding the observation.
However, rather than become involved in questions of plausibility, we think
it better if codes are defined so as to be mutually exclusive and exhaustive in
the first place. Then it is easy to argue that the data consist of a continuous
record of successive events or behavioral states.
Whether behavior is observed "live" or viewed on videotape does not
matter. For example, observers could be instructed to sit in front of a cage
from 10 to 11 a.m. on Monday, 3 to 4 p.m. on Tuesday, etc., and record
whenever an infant monkey changed his activity, or observers could be
instructed to watch several segments of videotape and to record whenever
the "play state" of the "focal child" changed, perhaps using Smith's social
participation coding scheme (Alone, Parallel, Group). In both cases, ob-
servers would record the number and sequence of codable events. An obvi-
ous advantage of working from videotapes is that events can be played and
replayed until observers feel sure about how to code a particular sequence.
 Recording onset and offset times 43

Still, both result in a complete record of the codable events that occurred
during some specified time period.

3.5 Recording onset and offset times

When time information is required, which usually means that the investi-
gator wants to report time-budget information or else wants to report how
different kinds of behavior are coordinated in time, observers can be asked
to record, not just that a codable event occurred, but its onset and offset
times as well. This is one way to preserve time information, but it is not
the only way, as we discuss in the next section.
The task is made even easier when codes are mutually exclusive and
exhaustive (or consist of sets of ME&E codes) because then offset times
do not need to be recorded. In such cases, the offset of a code is implied
by the onset of another mutually exhaustive code. As an example, con-
sider Landesman-Dwyer's codes for baby's eyes (Closed, Slow Roll, Daze,
Bright, etc.). She could have instructed observers to record the time when-
ever the "state" of the baby's eyes changed-for example, from Closed to
Slow Roll. Then the elapsed time or duration for this episode of Closed
would be the onset time of Slow Roll minus the onset time of Closed.
Moreover, because the times when the baby's eyes were closed is known,
she could also ask what kinds of face, head, and body movements occurred
when the baby's eyes were closed.
This strategy of recording onset times (or onset and offset times when
codes are not mutually exclusive and exhaustive) seems so simple, general,
and straightforward to us that we are surprised that it is not used more often.
The reason for this, we believe, is primarily technical. Before electronic
devices that automatically record time became available, we suspect that
observers found it distracting to write down times. Thus the older literature
especially contains many examples of a "lined paper" approach to recording
time information. Behavioral codes would head columns across the top of
the page. Then each row (the space between lines) would represent a period
of time. Observers would then make a check or draw a vertical line in the
appropriate column to indicate when that kind of event began and how long
it lasted. As we discuss later, such an "interval coding" strategy has the
merit of requiring only pencil and paper, but it remains an approximation
of what researchers often really want to do, which is to record the exact
times when the behavior of interest begins and ends.
Timing onsets and offsets does not require electronic recording devices,
of course, but such devices do make the task easy. In this paragraph and
the next, we describe two applications, one that uses such devices and one
44 Recording behavioral sequences

co&JL,

II 0 IS
if 18 5
11 23 19
If 42
16 56

Figure 3.4. An example of a hand-held electronic recording device: The OS-3

(Observational Systems, Redmond, WA 98052). Data production in the form
shown is just one of many options.

that does not. First, imagine that observers using the Landesman-Dwyer
baby's eyes code were equipped with electronic recording devices. They
would learn to push the correct keys by touch; thus they could observe
the baby continuously, entering codes to indicate when behavior changed.
(For example, an 11 might be used for Closed, a 12 for Slow Roll, a 16
for Bright, etc.) Times would be recorded automatically by a clock in the
device (see Figure 3.4). Later, the observer's record would be "dumped" to
a computer, and time budget (i.e., percentage scores for different behavioral
codes) and other information would be computed by a computer program.
(In fact, Landesman-Dwyer used such devices but a different recording
strategy, as described in the next section.)
The second application we want to describe involves pencil and paper
recording, but couples that with "time-stamped" videotapes. For a study
of preverbal communication development, Bakeman and Adamson (1984)
videotaped infants playing with mothers and with peers at different ages.
 Timing pattern changes 45

At the same time that the picture and sound were recorded, a time code
was also placed on the tape. Later, observers were instructed to segment
the tapes into different "engagement states" as defined by the investigators'
code catalog. In practice, observers would play and replay the tapes until
they felt certain about the point where the engagement state changed. They
would then record the code and onset time for the new engagement state.
In summary, one way to preserve a complete record of how behavior
unfolds in time (when such is desired) is to record onset (and if necessary
offset) times for all codable events. This is easy to do when one is working
from time-stamped videotapes. When coding live, this is probably best
done with electronic recording devices (either special purpose devices like
the one shown in Figure 3.4 or general-purpose handheld or notebook
computers, programmed appropriately, which increasingly are replacing
special purpose devices). When such devices are not available, the same
information can be obtained with pencil, paper, and some sort of clock,
but this complicates the observers' task. In such cases, investigators might
want to consider the approximate methods described in section 3.7, on
Coding intervals.

3.6 Timing pattern changes

There is a second way of preserving complete time information, but it
applies only to coding schemes structured like the Landesman-Dwyer Baby
Behavior Code described earlier. That is, the scheme must consist of groups
of mutually exclusive and exhaustive codes, with each group referring to a
different aspect of behavior. Paradoxically, this approach seems like more
work for the observers, but many observers (and investigators) prefer it to
the simple recording of onset times discussed above.
To describe this recording strategy, let us continue with the example
of the Baby Behavior Code. Recall that there were five groups of codes:
External Stimulation, Eyes, Face, Head, and Body. Observers are taught to
think of this as a 5-digit code: The first digit represents the kind of External
Stimulation, the second digit represents the Eyes, the third the Face, etc.
Thus the code 18440 means that external stimulation was a reflex (code 1),
REM movement was evident in the eyes (code 8), there was a smile on the
face (code 4), the head was up (also code 4), and the body was in repose
(code 0). If code 18440 were followed by code 18040, it would mean that
nothing had changed with respect to External Stimulation, Eyes, Head, and
Body but that the Face was now in repose (code 0), not smiling as before.
Each time there is a change in codable behavior, even if it involves only
one of the five superordinate groups, a complete 5-digit code is entered. On
46 Recording behavioral sequences

the surface of it, this seems like more work than necessary. After all, why
require observers to enter codes for external stimulation, eyes, head, and
body when there has been no change just because there has been a change
in facial behavior? In fact, observers who use this approach to recording
data report that, once they are trained, it does not seem like extra work at
all. Moreover, because the status of all five groups is noted whenever any
change occurs, investigators feel confident that changes are seldom missed,
which they might not be if observers were responsible for monitoring five
different kinds of behavior separately.
Paradoxically, then, more may sometimes be less, meaning that more
structure - that is, always entering a structured 5-digit code - may seem
like less work, less to remember. This is the approach that Landesman-
Dwyer actually uses for her Baby Behavior Code. We should add, however,
that her observers use an electronic recording device so that time is automat-
ically recorded everytime a 5-digit code is entered. In fact, we suspect that
the Baby Behavior Code would be next to impossible to use without such
a device, no matter which recording strategy (timing onsets or timing pat-
tern changes) were employed. Given proper instrumentation, however, the
same information (frequencies, mean duration, percents, cooccurrences,
etc.) would be available from data recorded using either of these strate-
gies. When a researcher's coding scheme is structured appropriately, then,
which strategy should be used? It probably depends in part on observer
preference and investigator taste, but we think that recording the timing of
pattern changes is a strategy worth considering.

3.7 Coding intervals

For the recording strategies just described - coding events, recording onset
and offset times, recording pattern change times - observers are "event
triggered," that is, they are stirred to record data whenever a codable event
occurs. When using an interval coding (or interval recording) strategy, on
the other hand, observers are "time triggered," that is, typically they record
at certain predetermined times.
The essence of this strategy is as follows. A child, a dyad, an animal, a
group, or whatever, is observed for a period of time. That period is divided
into a number of relatively brief intervals, typically on the order of 10 or 15
seconds or so. Observers then either categorize each interval or else note
which codable events, if any, occurred during each interval.
Such data can be collected easily with pencil and paper via a "checklist"
format. As mentioned briefly earlier, behavioral codes head columns across
the top of a page, whereas the rows down the page represent successive
 Coding intervals 47

T
Ik,
T
1
z y
3
4
5
6 y
7 y
8 y
q y
10 y
II y
12 y
13
14 y

Figure 3.5. Coding intervals using a checklist format.

intervals of time. When a codable event happens within an interval, ob-

servers place a check in the appropriate row and column. Even though
observers may actually first move pencil to paper when an event happens,
and so are "event-triggered" in that sense, the end result is to characterize
intervals as containing or not containing an event of interest, which is why
we call this an "interval coding" strategy.
An even simpler example of interval coding is provided by the Bakeman
and Browenlee study of parallel play, described in chapter 1. Their ob-
servers watched 12-15 7-minute video recordings for each of 41 children.
The tapes were played at normal speed without stopping, in effect simu-
lating live "real time" observation. Every 15 seconds, a simple electronic
device delivered a "click" to the ear, at which time observers wrote down
the social participation category that they felt best characterized the time
since the last click (see Figure 3.5).
Interval coding, or some variant of it, has been both widely used and
much criticized in the literature. We think the reason it has been so widely
48 Recording behavioral sequences

used has something to do with historical precedent, a human propensity to

impose clock units on passing time, the ready availability of lined paper, and
the unavailability until quite recently of electronic recording devices. It has
been criticized because it simply cannot provide as accurate information
as an event coding strategy. The major problem is that more than one code
may occur within an interval or near the boundaries of intervals. What
would a coder do in this case? If a hierarchical decision rule is employed,
sequential and other information within the interval is lost. This problem is
minimized, of course, if the time intervals are short relative to the behavior
observed.
However, we can see no good theoretical reason ever to use interval
coding; its merits are all practical. It requires only pencil, paper, and
some sort of simple timing device. No sophisticated electronic recording
devices or computers need be involved. In addition, as mentioned earlier,
sometimes observers find it easier to categorize intervals than to identify
when codable events began. (We suspect, however, that this is more of
a consideration when behavior is recorded live, and is much less of a
consideration when observers can play and replay videotaped behavior, as
for the Bakeman and Adamson study of infant engagement states described
in section 3.5.)
Clearly, the key consideration when an interval coding strategy is used is
the length chosen for the interval. If that interval is somewhat shorter than
the shortest duration typically encountered for a codable event, then little
distortion should be introduced into the data (Smith & Connolly, 1972; cf.
J. Altmann, 1974). No doubt, most investigators who use an interval coding
strategy understand perfectly well that the interval used should not be so
long as to mask onsets and offsets of the events being studied because if it
is, then not only will estimates of frequencies, durations, and percentages
be inaccurate, but behavioral sequences will be distorted as well.
For example, imagine that the codable event is baby within arm's length
of mother and that the interval used for recording is 10 seconds. We should
be able to assume that when a checked interval follows an unchecked
interval, the baby approached the mother. More importantly, we should
also be able to assume that when a string of checked intervals follow each
other, the baby did not leave the mother and return but rather stayed near
her. Clearly, an interval of 5 minutes would not be satisfactory, 10-second
intervals are probably acceptable, and 1-second intervals might be even
better. Whether the 1-second interval provides sufficiently more accurate
data to be worth the extra effort, however, is debatable, although the matter
could be determined with a small pilot study.
In sum, when simplicity and low cost of instrumentation matter more
than accuracy, or when the interval is shorter than most codable events
 Cross-classifying events 49

and observers prefer checking or categorizing intervals to recording onset

times, then an interval coding strategy probably makes sense. For all other
cases, some variant of event coding should probably be used.

3.8 Cross-classifying events

In sections 3.4, 3.5, and 3.6, we referred to event coding. By that we
meant a recording strategy that requires observers to detect codable events
whenever they occur in the passing stream of behavior. Once a codable
event has been detected, it is then the observers' task to classify the event,
that is, to assign to it one of the codes from the coding scheme. This strategy
becomes a way of recording behavioral sequences when continuity between
successive events can be assumed, as was discussed in section 3.4.
There is a second way of coding events, however, which does not require
any kind of continuity between successive events but which nonetheless
results in behavioral sequences being captured. This method does not
simply classify events (on a single dimension) but instead cross-classifies
them (on several dimensions). The key feature of this approach is the
coding scheme. Sequential data result when the superordinate categories
of the scheme represent logically sequential aspects of the event.
For example, imagine that observers were asked to note whenever chil-
dren quarreled. If observers did only this, the result would be a frequency
count for quarrels but no sequential information. However, observers could
also be asked to note what the children were doing just before the quarrel
began, what kind of quarrel it was, and how the quarrel was resolved. As-
suming that a mutually exclusive and exhaustive set of codes was defined
for each of these three questions, the observer would be cross-classifying
the quarrel. This is essentially the same kind of task as asking a child to
classify a set of objects by shape (circles, squares, and triangles), color (red,
blue, and green), and material (wood, metal, plastic). However, there is a
key difference between these two classification tasks. The three schemes
used to cross-classify quarrels have a natural temporal order - preceding
circumstance, quarrel, resolution - whereas the three schemes used to
cross-classify objects do not. Thus the strategy of cross-classifying events
is not necessarily sequential. Whether it is or not depends on how the
coding scheme is defined.
A second example of this strategy is provided by the Bakeman and
Brownlee study of object struggles. (Their coding scheme was described
in section 2.13.) In that study, the event of interest was an object struggle.
Whenever one occurred, observers recorded (a) whether the child attempt-
ing to take the object had had prior possession or not, (b) whether the
50 Recording behavioral sequences

current possessor resisted the take attempt or not, and (c) whether the child
attempting to take the object was successful or not. Note that the possible
answers to each of these three questions (yes/no) are mutually exclusive
and exhaustive and that the three questions have a natural temporal order.
The cross-classification of events is a recording strategy with many ad-
vantages. For one thing, techniques for analyzing cross-classified data
(contingency tables) have received a good deal of attention, both histori-
cally and currently, and they are relatively well worked out (see chapter
10). Also, clear and simple descriptive data typically result. For example,
in another study, Brownlee and Bakeman (1981) were interested in what
"hitting" might mean to very young children. They defined three kinds of
hits (Open, Hard, and Novelty) and then asked observers to record when-
ever one occurred and to note the consequence (classified as No Further
Interaction, Ensuing Negative Interaction, or Ensuing Positive Interaction).
They were able to report that open hits were followed by no further interac-
tion and novelty hits by ensuing positive interaction more often than chance
would suggest, but only for one of the age groups studied, whereas hard
hits were associated with negative consequences in all age groups.
A further advantage is that a coding scheme appropriate for cross-
classifying events (temporally ordered superordinate categories, mutually
exclusive and exhaustive codes within each superordinate category) im-
plies a certain amount of conceptual analysis and forethought. In general,
we think that this is desirable, but in some circumstances it could be a
liability. When cross-classifying events, observers do impose a certain
amount of structure on the passing stream of behavior, which could mean
that interesting sequences not accounted for in the coding scheme might
pass by unseen, like ships in the night. Certainly, cross-classifying events
is a useful and powerful way of recording data about behavioral sequences
when investigators have fairly specific questions in mind. It may be a less
useful strategy for more exploratory work.

3.9 Nonsequential considerations: Time sampling

The emphasis of this book is on the sequential analysis of data derived
from systematic observation. However, not all such data are appropriate
for sequential analyses. Recognizing that, we have stressed in this chapter
recording strategies that can yield sequential data. There exist other useful
and widely used recording strategies, however, which are worth mentioning
if only to distinguish them from the sequential strategies described here.
Perhaps the most widely used nonsequential approach to recording ob-
servational data is time sampling, or some variant of it. We have already
 Nonsequential considerations: Time sampling 51

discussed this approach in reference to Parten's classic study of social par-

ticipation described in chapter 1. The essence of time sampling is that
observing is intermittent, not continuous. Repeated noncontiguous brief
periods of time are sampled, and something about them is recorded. There
is no continuity between the separate samples, and thus the resulting data
are usually not appropriate candidates for sequential analysis.
A variant of time sampling, which at first glance might appear sequen-
tial, requires observers to intersperse periods of observing and periods of
recording. For example, an observer might watch a baby orangutan for
15 seconds, then record data of some sort for the next 15 seconds, then
return to a 15-second observation period, etc. Assuming that the behavior
of interest occurs essentially on a second-by-second basis, this might be
a reasonable time-sampling strategy, but it would not produce sequential
data. Even if the events being coded typically lasted longer than 15 sec-
onds, it would still be a dubious way to collect sequential data because of
all the gaps between observation periods.
For every general rule we might state, there are almost always excep-
tions. Imagine, for example, that we were interested in how adults change
position when sleeping. One obvious way to collect data would be to use
time-lapse photography, snapping a picture of the sleeping person every
minute or so. Now surely this fits the definition of a time-sampling strategy.
Yet assuming that sleeping adults rarely change position more than once
within the time period that separates samples, such data would be appro-
priate for sequential analysis. Logically, such data are not at all different
from the data produced for Bakeman and Brownlee's study of parallel play
described in chapter 1. In each case, the data consist of a string of num-
bers, with each number representing a particular code. The only difference
lies with the unit coded. For the sleeping study, the unit would be an es-
sentially instantaneous point in time, whereas for the parallel play study,
observers were asked to characterize an entire period of time, integrating
what they saw.
As the previous paragraph makes clear, time sampling can be regarded
as a somewhat degenerate form of interval coding (section 3.7 above).
Nonetheless, sequential data may result. Whether it does or not, however,
is a conceptual matter, one that cannot be mechanically decided. To restate,
whether data are regarded as sequential depends on whether it is reasonable
to assume some sort of continuity between adjacent coded units. Usually,
but not always, time-sampled data fail this test. This does not mean that
a time-sampling recording strategy is of little use, only that it is often
not appropriate when an investigator's concerns are primarily sequential.
However, when researchers want to know how individuals spend time, time
sampling may be the most efficient recording strategy available.
52 Recording behavioral sequences

3.10 The pleasures of pencil and paper

Throughout this chapter, we have made occasional comments about the
mechanics of data recording. Although there are other possibilities, mainly
we have mentioned just two: pencil and paper methods, on the one hand,
and the use of electronic recording devices, on the other. In this section, we
would like to argue that pencil and paper have their pleasures and should
not be shunned simply because they seem unsophisticated.
Pencil and paper methods have many advantages. For one thing, it
is difficult to imagine recording instruments that cost less or are easier
to replace. There are no batteries that may at a critical moment appear
mysteriously discharged. Also, pencil and paper are easy to transport and
use almost anywhere. Moreover, there is a satisfying physicality about
pencil marks on paper. The whole record can be seen easily and parts of it
modified with nothing more complicated than an eraser. Almost never does
the record of an entire observation session disappear while still apparently
is one's hands. Although paper can be lost, it almost never malfunctions.
Pencil and paper methods are equally usable when the researcher is ob-
serving live behavior, is viewing videotapes, or is working from a corpus
of written transcripts. For example, pencil and paper recording was used
for Brownlee and Bakeman's study of hitting, mentioned in section 3.8
(children were observed live for the study proper, although researchers
developed the codes while watching videotapes); for Bakeman and Adam-
son's study of communication development in infants, referred to in section
3.5 (observers worked with videotapes, stopping, reversing, and replaying
as appropriate); and for Gottman's study of friendship formation, described
in section 2.11 (observers worked with transcripts, segmented into thought
units).

3.11 Why use electronics?

As a general rule, users are attracted to more complex and sophisticated
devices of any kind because, once mastered, savings in time and labor re-
sult. The key phrase here is "once mastered," because almost inevitably,
more sophisticated devices - such as electronic recording devices instead
of pencil and paper - require that the user pay an "entry cost." This entry
cost includes, not just the initial cost of the device, but the time it takes
to learn how to use the device and to keep the device working properly.
Nonetheless, many researchers find that the entry costs of electronic record-
ing devices are well worth paying. (To our knowledge, such devices have
no common generic name or acronym; they are not, for example commonly
called ERDs.)
 Why use electronics? 53

Electronic recording devices can be either special purpose (see Fig-

ure 3.4) or, increasingly, general purpose laptop or notebook computers,
programmed to collect data for a particular study. A major advantage of
such devices is that data are recorded in machine-readable form from the
start. There is no need for data to pass through the hands of a keypuncher,
and hence keypunchers cannot introduce errors into data later. Such devices
work as follows. To indicate the sequence of codable events, observers de-
press keys corresponding to the appropriate behavioral codes. A record of
the keys "stroked" is stored in an internal memory. Once the observation
session is complete, the contents of memory can be transferred to a file on
the computer's hard disk or written to some external storage medium, like
a floppy disk or a magnetic backup tape. The stored data are then available
for whatever subsequent processing is desired.
A second major advantage of electronic recording devices is that they
usually contain an internal clock. This means that whenever an observer
enters a code, the time can be stored as well, automatically, without the
observer having to read the time. Reading the time and writing it down
are, of course, quite distracting for an observer. That is why we think
interval coding has been used so often in the past (a click every 10 seconds,
for example, can be delivered to the ear, leaving the eye free) and why,
given the availability of electronic recording devices, event coding (with
the automatic recording of onset and offset times or of pattern changes) is
becoming much more common than in the past.
A final advantage likewise has to do with minimizing distractions, leav-
ing observers free to devote their full attention to whatever or whomever
is being observed. Observers easily learn to push the appropriate keys
without looking at them, like a good typist. With these keyboard devices,
then, observers do not need to shift their eyes away from the recording task,
as they do when recording with pencil and paper. This matters most, of
course, when observing live.
Electronic devices for recording data, then, are recommended when a
machine-readable record is required, and it seems advantageous to avoid
a keypunching step during data preparation. They are also recommended
when observers need to keep their eye on the task without the distractions of
writing down codes, or worse, their times of occurrence. A code such as the
Baby Behavior Code described in section 3.6, which requires the frequent
entering of 5-digit codes along with the times they occurred, is probably
unworkable with anything other than an electronic recording device.
Other, more sophisticated possibilities for recording data electronically
should be mentioned. One common strategy involves time-stamped video-
tapes (i.e., videotapes on which the time, usually accurate to the nearest
second or some fraction of a second, forms part of the picture). Observers
54 Recording behavioral sequences

Table 3.1. Summary of recording schemes

Recording scheme Definition Advantages Disadvantages

Event recording Events activate May provide a Could lose time

coders realistic way of information
segmenting unless onsets
behaviors and offsets were
noted
Interval Time activates Easy to use 1. May
recording coders artificially
truncate
behavior
2. Need to
select interval
small enough or
could lose
information
Cross-classifying Events activate 1. Speed Could lose
events coders, but only 2. Can still valuable
a specific kind preserve information not
of event sequential accounted for
information by the highly
3. Statistics well selective scheme
worked out
4. Requires
conceptual
forethought
Time sampling Time activates Easy to use Coding is
coders intermittent so
sequential
information is
usually lost

view tapes, slowing down the speed and rewinding and reviewing as neces-
sary. The times when events occurred are read from the time displayed on
the screen. This time may be written down, using pencil and paper, or keyed
along with its accompanying code directly into a computer. An improve-
ment on this strategy involves recording time information on the videotape
in some machine-readable format and connecting a computer to the video
player. Then observers need only depress keys corresponding to particular
behaviors; the computer both reads the current time and stores it along
with the appropriate code. Moreover, the video player can be controlled
 Summary 55

directly from the computer keyboard. Then, with appropriate hardware

and software, coders can instruct the system to display all segments of tape
previously coded for a particular code or set of codes. Such systems are
very useful but, as you might expect, have a relatively high entry cost both
in terms of money and time (e.g., see Tapp & Walden, 1993).

3.12 Summary
Table 3.1 is a summary of the four major conceptual recording schemes
we have discussed in this chapter, together with their advantages and dis-
advantages. The particular recording scheme chosen clearly depends on
the research question. However, in general, we find event recording (with
or without timing of onsets and offsets or timing of pattern changes) and
cross-classifying events to be more useful for sequential analyses than ei-
ther interval recording or time sampling.
 Assessing observer agreement

4.1 Why bother?

Imagine that we want to study how objects are used in communication
between mothers and their 15-month-old infants, and have made a number
of videotapes of mothers and infants playing together. We might focus our
attention on those times when either the mother or the infant tries to engage
the other's interest in some object, and then describe what those attempts
look like and what their consequences are. After viewing the videotapes
several times, sometimes in slow motion, we might be convinced that we
had detected all such episodes, had accurately described them, and were
now ready to make statements about how mothers and infants go about
attempting to interest each other in objects.
We should not be surprised, however, if other investigators do not take
our conclusions as seriously as we do. After all, we have done nothing
to convince them that others viewing the videotapes would see the same
things, much less come to the same conclusions. We are probably all aware
how easy it is to see what we want to see, even given the best of intentions.
For that reason, we take elaborate precautions in scientific work to insulate
measuring procedures from the investigator's influence.
When measurements are recorded automatically and/or when there is lit-
tle ambiguity about the measurement (for example, the amount of sediment
in a standard sample of seawater), as is often the case in the "hard" sci-
ences, the problem of investigator bias is not so severe. But in observational
studies, especially when what we call "socially based" coding schemes are
being used, it becomes especially important to convince others that what
was observed does not unduly reflect either the investigator's desires or
some idiosyncratic worldview of the observer. The solution to the first
problem is to keep observers naive as to the hypotheses under investiga-
tion. This in fact is done by most investigators, judging from the reports
they write, and should be regarded as standard practice in observational
work. The solution to the second problem is to use more than one observer
and to assess how well they agree.

56
 Why bother? 57

Accuracy
The major conceptual reason for assessing interobserver agreement, then,
is to convince others as to the "accuracy" of the recorded data. The as-
sumption is, if two naive observers independently make essentially similar
codings of the same events, then the data they collect should reflect some-
thing more than a desire to please the "boss" by seeing what the boss wants,
and something more than one individual's unique and perhaps strange way
of seeing the world.
Some small-scale studies may require only one observer, but this does
not obviate the need for demonstrating agreement. For example, in one
study Brownlee and Bakeman (1981) were concerned with communicative
aspects of hitting in 1-, 2-, and 3-year-old children. After repeated viewings
of 9 hours of videotape collected in one day-care center, they developed
some hypotheses about hitting and a coding scheme they thought useful for
children of those ages. The next step was to have a single observer collect
data "live" in another day-care center. There were two reasons for this.
First, given a well-worked-out coding scheme, they thought observing live
would be more efficient (no videotapes to code later), and second, nursery
school personnel were concerned about the disruption to their program
that multiple observers and/or video equipment might entail. Two or more
observers, each observing at different times, could have been used, but
Brownlee and Bakeman thought that using one observer for the entire study
would result in more consistent data. Further, the amount of observation
required could easily be handled by one person. Nonetheless, two observers
were trained, and agreement between them was checked before the "main"
observer began collecting data. This was done so that the investigators,
and others, would be convinced that this observer did not have a unique
personal vision and that, on a few occasions at least, he and another person
independently reported seeing essentially the same events.

Calibration
Just as assuring accuracy is the major conceptual reason, so calibrating ob-
servers is probably the major practical reason for establishing interobserver
agreement. A study may involve a large number of separate observations
and/or extend over several months or years. Whatever the reason, when
different observers are used to collect the same kind of data, we need to
assure ourselves that the data collected do not vary as a function of the
observer. This means that we need to calibrate observers with each other
or, better yet, calibrate all observers against some standard protocol.
58 Assessing observer agreement

Reliability decay
Not only do we need to assure ourselves that different observers are coding
similar events in similar ways, we also need to be sure that an individual ob-
server's coding is consistent over time. Taplin and Reid (1973) conducted a
study of interobserver reliability as a function of observer's awareness that
their coding was being checked by an independent observer. There were
three groups - a group that was told that their work would not be checked, a
group that was told that their work would be spot-checked at regular inter-
vals, and a group that was told that their work would be randomly checked.
Actually the work of all three groups was checked for all seven sessions.
All groups showed a gradual decay in reliability from the 80% training
level. The no-check group showed the largest decay. The spot-check
group's reliability increased during sessions 3 and 6, when they thought
they were being checked. The random-check group performed the best over
all sessions, though lower than the spot-check group on session 3 and 6.
Reliability decay can be a serious problem when the coding process takes
a long time, which is often the case in a large study that employs a complex
coding scheme. One solution to the problem was reported by Gottman
(1979a). Gottman obtained a significant increment in reliability over time
by employing the following procedure in coding videotapes of marital
interaction. One employee was designated the "reliability checker"; the
reliability checker coded a random sample of every coder's work. A folder
was kept for each coder to assess consistent confusion in coding, so that
retraining could be conducted during the coder's periodic meetings with
the reliability checker. To test for the possibility that the checker changed
coding style for each coder, two procedures were employed. First, in one
study the checker did not know who had been assigned to any particular
tape until after it was coded. This procedure did not alter reliabilities.
Second, coders occasionally served as reliability checkers for one another
in another study. This procedure also did not alter reliabilities. Gottman
also conducted a few studies that varied the amount of interaction that the
checker coded. The reliabilities were essentially unaffected by sampling
larger segments, with one exception: The reliabilities of infrequent codes
are greatly affected by sampling smaller segments. It is thus necessary for
each coding system to determine the amount that the checker codes as a
function of the frequency of the least frequent codes.
What should be clear from the above is that investigators need to be
concerned not just with inter-, but also with intraobserver reliability. An
investigator who has dealt with the problems of inter- and intraobserver
agreement especially well is Gerald Patterson, currently of the Oregon
Social Learning Center. Over the past several years, Patterson and his co-
workers have trained a number of observers to use their coding schemes.
 Reliability vs. agreement 59

Although observers record data live, training and agreement assessments

depend on the use of videotapes. First, presumably "correct" codings or
"standard" versions were prepared for a number of videotaped sequences.
Then these standards were used to train new observers, who were not
regarded as trained until their coding reached a preset criterion of accuracy,
relative to the standard. Second, observers periodically recoded standard
versions, and their agreement both with the standard and with their own
previous coding was assessed. Such procedures require time and planning,
but there is probably no other way to ensure the continued accuracy of
human observers when a project requires more than one or two observers
and lasts for more than a month or two.

4.2 Reliability versus agreement

So far in this chapter we have used the term "observer agreement," yet the
term "observer reliability" often occurs in the literature. What, if any, is
the difference? Johnson and Bolstad (1973) make a nice distinction. They
argue that agreement is the more general term. It describes, as the word im-
plies, the extent to which two observers agree with each other. Reliability is
the more restrictive term. As used in psychometrics, it gauges how accurate
a measure is, how close it comes to "truth." Hence when two observers are
just compared with each other, only agreement can be reported. However,
when an observer is compared against a standard protocol assumed to be
"true," then observer reliability can be discussed.
Others would argue that reliability is the more general term. Inter-
observer agreement only addresses potential errors among observers and
ignores many other sources of potential errors, which in the context of ob-
servational research may be many (Pedhazur & Schmelkin, 1991, pp. 114-
115, 145-146). Yet, when two observers independently agree, the usual
presumption is that they are therefore accurate, even though it is possible, of
course, that they simply share a similar but nonetheless deviant worldview.
But this presumption is questionable because other sources of error may
be present. In observational research (and in this book as well), inter-
observer agreement is emphasized, but it is important to remember that,
although important, indices of interobserver agreement are not indices of
reliability, and that reliability could be low even when interobserver agree-
ment is high.
Not wishing to shortchange a complex topic (i.e., assessment of relia-
bility), we would nonetheless argue that there is some merit in preparing
standard protocols, presumed true, which can then be used as one, simple
index ofobserver reliability. If thefirstreason for assessing observer agree-
ment is to assure others that our observers are accurate and our procedures
60 Assessing observer agreement

replicable, and the second reason is to calibrate multiple observers, then a

third reason is to assure ourselves as investigators that observers are coding
what we want (i.e., are seeing the world as we do). And one way to test
this is to let observers code independently a sequence of events for which
we have already prepared a standard protocol. Clearly, this is easiest to do
when videotaped behavior or transcripts of conversations are coded, and
less easy to do when live behavior is coded. However, such procedures let
us speak, albeit relatively informally, of "observer reliability" (assuming of
course that investigators are relatively infallible), and they also give inves-
tigators additional confidence in their observers, a confidence that probably
becomes more important, the more socially based the coding scheme is.
In sum, "reliability" invokes a rich and complex psychometric tradition
and poses problems that lie well beyond the scope of this book. From this
point of view, "interobserver agreement" is the more limited and straight-
forward term. Moreover, it is the one that has been emphasized in obser-
vational research and, consequently, is addressed in the remainder of this
chapter. The question now is, how should observer agreement be assessed
and computed?

4.3 The problem with agreement percentages

Perhaps the most frequently encountered, and at the same time the most
misleading, index of observer agreement is a percentage of some sort. This
is usually referred to as a "percentage of agreement" and in its most general
form is defined as follows:
P A = — — — x 100
NA + ND
PA refers to the percentage of agreement, NA the number of agreements,
and No the number of disagreements. In any given application, the in-
vestigator would need to specify further the recording unit used (events
or intervals), which is after all the basis for determining agreement and
disagreement, and exactly how agreement and disagreement are defined.
For example, Tuculescu and Griswold (1983), in their study of pre-
hatched chickens (section 2.10), defined four kinds of embryonic distress
calls (Phioo, Soft Peep, Peep, and Screech). Observers coded events.
Whenever an event of interest occurred, they recorded which it was, when
it began, and when it ended. Given this coding scheme and this recording
strategy, observer agreement could have been computed as follows.
First, what constitutes an agreement needs to be defined. For example,
we might say that two observers agree if they record the same kind of
distress call at times that either overlap or are separated by no more than
 The problem with agreement percentages 61

two seconds. They disagree when one records a distress call and the other
does not, or when they agree that a distress call occurred but disagree
as to what kind it is. (Following an old classification system for sins,
some writers call these disagreements "omission errors" and "commission
errors," respectively.)
Once agreements and disagreements have been identified and tallied, the
percentage of agreement can be computed. For example, if two observers
both recorded eight Phioos at essentially the same time, but disagreed three
times (each observer recorded one Phioo that the other did not, and once one
observer recorded a Phioo that the other called a Soft Peep), the percentage
of agreement would be 73 (8 divided by 8 + 3 times 100). Percentage
agreement could also be reported, not just for Phioos in particular, but
for embryonic distress calls in general. For example, if the two observers
agreed as to type of distress call 35 times but disagreed 8 times, then the
percentage of agreement would be 81 (35 divided by 35 + 8 times 100).
Given a reasonable definition for agreement and for disagreement, the
percentage of agreement is easy enough to compute. However, it is not at
all clear what the number means. It is commonly thought that agreement
percentages are "good" if they are in the 90s, but there is no rational basis
for this belief. The problem is that too many factors can affect the per-
centage of agreement - including the number of codes in the code catalog
- so that comparability across studies is lost. One person's 91% can be
someone else's 78%.
Perhaps the most telling argument against agreement percentage scores
is this: Given a particular coding scheme and a particular recording strategy,
some agreement would occur just by chance alone, even with blindfolded
observers, and agreement percentage scores do not correct for this. This be-
comes most clear when an interval coding strategy is coupled with a simple
mutually exclusive and exhaustive scheme, as in the study of parallel play
described in section 1.7. Recall that for this study, Bakeman and Brownlee
had observers code each successive 15-second interval as either Unoccu-
pied, Solitary, Together, Parallel, or Group. If two observers had each
coded the same 100 intervals, the pattern of agreement might have been as
depicted in Figure 4.1. In this case, the percentage of agreement would be
87 (87 divided by 87 + 13 times 100). However, as we show in the next sec-
tion, an agreement of 22.5% would be expected, in this case, just by chance
alone. The problem with agreement percentages is that they do not take
into account the part of the observed agreement that is due just to chance.
Figure 4.1 is sometimes called a "confusion matrix," and it is useful
for monitoring areas of disagreement that are systematic or unsystematic.
After computing the frequencies of entries in the confusion matrix, the
reliability checker should scan for clusters off the diagonal. These indicate
62 Assessing observer agreement

UM.
i
1
LU, O i o o 8
i
i ffli 2 4 0 o o 2.5
I mrwr 11 w
i 1 z 23
wrttff I
wrr 21
0 o *i 25 1

^ A*. 0 o o I 15

25 21 2S 17 100

Figure 4.1. An agreement or "confusion" matrix. Tallies on the diagonal indicate

agreement between the two observers, whereas tallies off the diagonal pinpoint
disagreements.

confusions between specific codes. If many observers display the same

confusion, it may suggest retraining, or clarification of the coding manual,
or finding clear examples that sharpen distinctions between codes.

4.4 The advantages of Cohen's kappa

An agreement statistic that does correct for chance is Cohen's kappa (Co-
hen, 1960). As a result, it is almost always preferable to simple agreement
percentages. It is defined as follows:
~ *e
K =
exp
where Pobs is the proportion of agreement actually observed and Pexp is
the proportion expected by chance. Pobs is computed by summing up the
tallies representing agreement (those on the upper-left, lower-right diagonal
in Figure 4.1) and dividing by the total number of tallies. That is, it is
analogous to PA in the previous section, except that it is not multiplied by
 The advantages of Cohen's kappa 63

100. Symbolically:

Fobs —
N
where k is the number of codes (i.e., the order of the agreement matrix),
xu is the number of tallies for the /th row and column (i.e., the diagonal
cells), and N is the total number of tallies for the matrix. For the agreement
portrayed in Figure 4.1, this is:
7 + 2 4 + 1 7 + 25 + 14 on
Pobs = m = .87
Pexp is computed by summing up the chance agreement probabilities for
each category. For example, given the data in Figure 4.1, the probability
that an interval would be coded Unoccupied was .08 for the first observer
and .09 for the second. From basic probability theory, the probability of
two events occurring jointly (in this case, both observers coding an interval
Unoccupied), just due to chance, is the product of their simple probabilities.
Thus the probability that both observers would code an interval Unoccupied
just by chance is .0072 (.08 x .09). Similarly, the chance probability that
both would code an interval Solitary is .0625 (.25 x .25), Together is .0483
(.21 x .23), Parallel is .0812 (.28 x .29), and Group is .0255 (.17 x .15).
Summing the chance probabilities for each category gives the overall pro-
portion of agreement expected by chance (PeXp), which in this case is .2247.
A bit of algebraic manipulation suggests a somewhat simpler way to
compute PeXp- Multiply the first column by the first row total, add this to
the second column total multiplied by the second row total, etc., and then
divide the resulting sum of the column-row products by the total number
of tallies squared. Symbolically:

where JC+,- and JC,-+ are the sums for the /th column and row, respectively
(thus one row by column sum cross-product is computed for each diagonal
cell).
For the agreement given in Figure 4.1, this is:
_ 9 x 8 + 25 x 25 + 21 x 23 + 28 x 29 + 17 x 15
Pexp
" 100 x 100
= .2247
64 Assessing observer agreement

UM-. £K.

4^ 7 I

Z 12
-facto. <\ H\ too

Figure 4.2. An agreement matrix using a coding scheme with two codes. Although
there is 97% agreement, 84% would be expected just by chance alone.

Now we can compute kappa for our example data.

.87 - .2247
K= = .8323 (rounded)
As the reader can see, the amount of agreement corrected for chance (about
.83) is rather less than the uncorrected value (.87). In some cases, especially
when there are few coding categories and when the frequency with which
those codes occur is quite disproportionate, the difference can be quite
dramatic. Imagine, for example, that instead of the five categories listed
in Figure 4.1, only two had been used: Unengaged (meaning Unoccupied)
and Engaged. In that case,the data from Figure 4.1 could reduce to the
data shown in Figure 4.2. The proportion of agreement oberved here is
quite high, .97 (7 + 90 divided by 100), but so is the proportion of chance
agreement as well.
9 x 8 + 91 x 92
*exp — = .8444
100 x 100
As a result, the value of kappa, although still respectable, is considerably
lower than the level of agreement implied (misleadingly) by the .97 value:
.97-.8444
1.-.8444
The question now is, is a kappa of .8072 big enough? Fleiss, Cohen, and
Everitt (1969) have described the sampling distribution of kappa, and so
it is possible to determine if any given value of kappa differs significantly
from zero (see also Hubert, 1977). The way it works is as follows: First,
the population variance for kappa, assuming that kappa is zero, is estimated
 The advantages of Cohen fs kappa 65

from the sample data. Then the value of kappa estimated from the sample
data is divided by the square root of the estimated variance and the result
compared to the normal distribution. If the result were 2.58 or bigger, for
example, we would claim that kappa differed significantly from zero at the
.01 level or better.
In this paragraph, we show how to compute the estimated variance for
kappa, first defining the procedure generally and then illustrating it, using
the data from Figure 4.2. The formula incorporates the number of tallies
(N, in this case 100), the probability of chance agreement (Pexp in this case
.8444), and the row and column marginals: /?,-+ is the probability that a tally
will fall in the i"th row, whereas p+j is the probability that a tally will fall
in the jth column. In the present case, p\+ = .08, p2+ = .92, p+\ — .09,
and p+2 = -91. To estimate the variance of kappa, first compute
k
]T = Pi+ x p+i x [1 - (p+i + pi+)f
1=1
In the present case, this is:
.08 x .09 x [1 - (.09 + .08)]2 + .92 x .91 x [1 - (.91 + .92)]2
= .00496+ .57675 = .5817
Then add to it this sum:
k k

9
x x
TTiT^= Pt+ P+J (P+i + PJ+>
I— 1 J— 1

In the present case, this is:

.08 x .91 x (.09 + .92)2 + .92 x .09 x (.91 + .08)2
= .07426+ .08115 = .1554
Next subtract P*. In the present case this is:
2
.8444 = .7130
and the result is:
.5817+ .1554-.7130 =.0241
Finally, divide this result by Af x (1 — PeXp)2- This divisor is:
100 x (1 - .8444)2 = 2.421
and the final quotient is:
.0241/2.421 = .009955
This is the estimated variance for kappa, given the data in Figure 4.2. The z
score is 8.091, the estimated kappa (.8072) divided by the square root of the
variance (.09977). We would conclude that the agreement demonstrated
in Figure 4.2 is significantly better than chance.
66 Assessing observer agreement

For many investigators, this will not be stringent enough. Just as corre-
lation coefficients that account for little variance in absolute terms are often
significant, so too, quite low values of kappa often turn out to be signifi-
cant. This means only that the pattern of agreement observed is greater than
would be expected if the observers were guessing and not looking. This can
be unsatisfactory, however. We want from our observers not just better than
chance agreement; we want good agreement. Our own inclination, based
on using kappa with a number of different coding schemes, is to regard
kappas less than .7, even when significant, with some concern, but this is
only an informal rule of thumb. Fleiss (1981), for example, characterizes
kappas of .40 to .60 as fair, .60 to .75 as good, and over .75 as excellent.
The computation of kappa can be refined in at least three ways. The first
is fairly technical. Multiple observers may be used - not just two - and
investigators may want a generalized method for computing kappa across
the different pairs. In such cases, readers should consult Uebersax (1982);
a BASIC program that computes Uebersax's generalized kappa coefficient
has been written by Oud and Sattler (1984).
The second refinement is often useful, especially when codes are roughly
ordinal. Investigators may regard some disagreements (confusing Unoccu-
pied with Group Play, for example) as more serious than others (confusing
Together with Parallel Play, for example). Cohen (1968) has specified a
way of weighting different disagreements differently. Three kxk matrices
are involved: one for observed frequencies, one for expected frequencies,
and one for weights. Let x^, mtj, and W(j represent elements from these
three matrices, respectively: then ra,7 = (x+j x xt+) -f- N, and the wtj indi-
cate how seriously we choose to regard various disagreements. Usually the
diagonal elements of the weight matrix are 0, indicating agreement (i.e.,
wa = 0 for i = 1 through k)\ cells just off the diagonal are 1, indicating
some disagreement; cells farther off the diagonal are 2, indicating more
serious disagreement; etc. For the present example, we might enter 4 in
cells JC15 and x$u indicating that confusions between unoccupied and group
are given more weight than other disagreements. Then weighted kappa is
computed as follows:

K -1
Kwt l
-

If wn = 0 for / = 1 through k and 1 otherwise, then /c, as defined earlier,

and Kwt are identical. (To compute variance for weighted kappa, see Fleiss,
Cohen, & Everitt, 1969).
 The advantages of Cohen's kappa 67

A third way the computation of kappa can be refined involves PeXp- As

usually presented, Pexp is computed from the row and column totals (the
marginals), but in some cases investigators may believe they have better
ways to estimate the expected distribution of the codes, and could use them
to generate expected frequencies.
For example, imagine that an investigator has accumulated a consider-
able data archive using the coding scheme presented in Figure 4.1 and so
has a good idea of how often intervals are coded Unoccupied, etc., in gen-
eral. Imagine further that coders are subject to random agreement checks
and that, as luck would have it, a tape selected for one such check shows
an unusual child who spends almost all of his time Unoccupied with just
a little additional Solitary play. In this case, the kappa would be quite
low. What is really a 5-category scheme becomes for kappa computation
a 2-category scheme with a skewed distribution, which usually results in
low values. In effect, the observers have received no "credit" for knowing
that Together, Parallel, and Group Play did not occur. In such a case, it
may make sense to substitute the average values usually experienced for
the marginals rather than use the actual ones from such a deviant case.
As the above example makes clear yet again, assessing observer agree-
ment has more than one function. When our concern is to convince others
(especially journal editors and reviewers) that our observers are accurate,
then we might well pool tallies from several different agreement checks
into one kappa table, computing and reporting a single kappa value. This
has the merit of providing realistic marginals. When our concern is to
calibrate and train observers, however, we would want to compute kappas
separately for each agreement check, thus providing immediate feedback.
At the same time, we should counsel our observers not to be discouraged
when low kappas result simply from coding an unusual instance.
With respect to the training of observers, there is one final advantage of
Cohen's kappa that we should mention. The kappa table itself provides a
graphic display of disagreement. Casual inspection immediately reveals
which codes are often confused and which almost never are. An excessive
number of tallies in an off-diagonal cell would let us know, for example, that
intervals one observer codes together are regarded as parallel by the other
observer. Moreover, simple inspection can also reveal if one observer is
more "sensitive" than the other. If so, a pattern of considerably more tallies
above than below the diagonal (or vice versa) results. For example, if most
of the disagreements in Figure 4.1 had been above the diagonal (meaning
that what the first observer regarded as Unoccupied the second observer
often regarded as something more engaged but not vice versa), this would
indicate that the second observer was more sensitive, detecting engagement
when the first observer saw only an unengaged child. Such patterns have
68 Assessing observer agreement

clear implications for observer training. In the first case (code confusion),
further training would attempt to establish a consensual definition for the
two codes; and in the second (different sensitivities), further training would
attempt to establish consensual thresholds for all codes.

4.5 Agreement about unitizing

Constructing a kappa table like that shown in Figure 4.1 is easy enough.
For each "unit" coded, a tally is entered in the table. The codes assigned the
unit by the two observers determine in which cell the tally is placed. This
procedure, however, assumes previous agreement as to what constitutes a
unit. Sometimes unit boundaries are clear-cut, and may even be determined
by forces external to the observers. At other times, boundaries are not all
that distinct. In fact, determining unit boundaries may be part of the work
observers are asked to do. For example, observers may be asked to identify
and code relatively homogeneous stretches of talk, or to identify a particular
kind of episode (e.g., a negotiation or conflict episode).
In such cases, agreement needs to be demonstrated on two levels: first
with respect to "unitizing" (i.e., identifying the homogeneous stretches or
episodes), and second with respect to the codes assigned the homogeneous
stretch or episode (or perhaps events embedded within the episode itself).
When the unit being coded is a time interval, as in the example just given,
there is no problem. In that case, unit boundaries are determined by a clock
and not by observers. When the unit being coded is an event, however, the
matter becomes more difficult. For example, consider Gottman's study of
friendship formation described in section 2.11. In such cases, there are
two parts to the coding task. First, observers need to segment the stream
of recorded talk into thought units, whereas second, they need to code the
segmented thought units themselves.
Agreement with respect to the coding of thought units can be determined
using Cohen's kappa, as described in the last section. However, how should
agreement with respect to segmenting the stream of talk into thought units
be demonstrated? In the older literature, a percentage score has often been
used for this task.
If both observers worked from transcripts, marking thought unit bound-
aries on them, it should be a fairly simple matter to tally the boundaries
claimed by both observers and the ones noted only by one observer. But
in this case, there can be only omission disagreements. Moreover the
percentage agreement would not correct for chance; still the score may
have some descriptive value, albeit limited.
Alternatively, the investigator might regard every gap between adjacent
words as a potential boundary. In this case, the initial coding unit would
 Agreement about unitizing 69

be the word gap. Each gap would contribute one tally to a 2 x 2 kappa
table: (a) Both observers agree that this gap is a thought unit boundary;
(b) the first observer thinks it is, but the second does not; (c) the second
thinks it is, but the first disagrees; or (d) both agree that it is not a thought
unit boundary. Although the chance agreement would likely be high (two
coding categories, skewed distribution, assuming that most gaps would not
be boundaries), still the kappa computed would correct for that level of
chance agreement.
A better and more general procedure for determining agreement with re-
spect to unitizing (i.e., identifying homogeneous stretches of talk or partic-
ular kinds of episodes) requires that onset and offset times for the episodes
be available. Then the tallying unit for the kappa table becomes the unit
used for recording time. For example, imagine that times are recorded to
the nearest second and that observers are asked to identify conflict episodes,
recording their onset and offset times. Then kappa is computed on the ba-
sis of a simple 2 x 2 table like that shown in Figure 4.2, except that now
rows and columns are labeled yes/no, indicating whether or not a second
was coded for conflict. One further refinement is possible. When tally-
ing seconds, we could place a tally in the agreement (i.e., yes/yes) cell if
one observer claimed conflict for the second and the other observer claimed
conflict either for that second or an adjacent one, thereby counting 1-second
disagreement as agreements as often seems reasonable.
In the previous chapter, we described five general strategies for recording
observational data: (a) coding events, (b) timing onsets and offsets, (c) tim-
ing pattern changes, (d) coding intervals, and (e) cross-classifying events.
Now we would like to mention each in turn, describing the particular prob-
lems each strategy presents for determining agreement about unitizing.
Coding events, without any time information, is in some ways the most
problematic. If observers work from transcripts, marking event (thought
unit) boundaries, then the procedures outlined in the preceding paragraphs
can be applied. If observers note only the sequence of events, which
means that the recorded data consist of a string of numbers or symbols,
each representing a particular event or behavioral state, then determining
agreement as to unit boundaries is more difficult. The two protocols would
need to be aligned, which is relatively easy when agreement is high, and
much more difficult when it is not, and which requires some judgment in
any case. An example is presented in Figure 4.3.
When onset and offset or pattern-change times are recorded, however, the
matter is easier. Imagine, for example, that times are recorded to the nearest
second. Then the second can be the unit used for computing agreement both
for unitizing (identifying homogeneous stretches or episodes) and for the
individual codes themselves. Because second boundaries are determined
70 Assessing observer agreement

TVUdW

u u a.
s S a. a.
Gr a. a.
T ? A
G Gr a. a.
V
5 S a- a.
V T
U U a.
T
u
p F a-
& a. a.

Figure 4.3. Two methods for determining agreements ("a") and disagreements
("d") when two observers have independently coded the same sequence of events.
Method A ignores errors of omission. Method B counts both errors of commission
and errors of omission as disagreements.

by a clock external to the observers, there is no disagreement as to where

these boundaries fall (the only practical requirement is that the clocks used
by the two observers during an agreement check be synchronized in some
way). An example showing how second-by-second agreement would be
computed in such cases is given in the next section.
When time intervals are coded in the first place, the matter is similar.
Again, the underlying unitization is done by clocks, not by observers. When
cross-classifying events, however, we need to ask, to what extent are both
observers detecting the same events? In older literature, often a percentage
agreement was used. For example, in their study of social rules among
preschool children, Bakeman and Brownlee (1982) asked two observers
to cross-classify object struggles (see section 2.13). During an agreement
check, one observer recorded 50 such struggles, the other 44; however, all
44 of the latter had also been noted by the first observer, and hence their
percentage agreement was 88.0% (44 divided by 44 + 6). In a case like
this, there seems to be no obvious way to correct for chance agreement.
Our recommendation is as follows: Report the agreement-disagreement
tallies along with the percentage of agreement in such cases, but note
 Agreement about codes 71

their limitations. However, if at all possible, report time-based kappa

statistics to establish that observers detected the same events to cross-
classify. This is the same strategy we recommend to demonstrate that
observers are identifying the same homogeneous stretches or episodes,
which likewise are then subjected to further coding.

4.6 Agreement about codes: Examples using Cohen's kappa

In this section, we describe how Cohen's kappa can be used to document
observer agreement about codes for each of the recording strategies listed
in chapter 3. This is not the only way to determine observer agreement (or
reliability), as we discuss in the next section, but it may be among the most
stringent. This is because Cohen's kappa documents point-by-point agree-
ment, whereas many writers would argue that agreement does not need to be
determined for a level more detailed than that ultimately used for analysis.
We think that this argument has merit, but that there are at least two rea-
sons to favor a relatively stringent statistic like Cohen's kappa. First, as we
argued earlier in this chapter, determining observer agreement has more
than one function. For training observers and providing them feedback
on their performance, we favor an approach that demands point-by-point
agreement. We also like the graphic information about disagreement pro-
vided by the kappa table. Second, once agreement at a detailed level has
been established, we can safely assume agreement at less detailed levels,
and in any case a relatively detailed level is required for sequential analysis.
When observers code events, it is relatively straightforward to compute
agreement about how units are coded, once the units are identified. For
example, when thought units are coded from transcripts, the codes the two
observers assign each successive thought unit would determine the cell of a
26 x 26 kappa table in which a tally would be placed (assuming 26 possible
codes for thought units). What this example highlights is that kappa is
a summary statistic, describing agreement with respect to how a coding
scheme is used (not agreement about particular codes in the scheme), and
that the codes that define each kappa table must be mutually exclusive and
exhaustive.
As a second example, still assuming that events are being coded, imag-
ine that an investigator wants to segment the stream of behavior into the
five play states used by Bakeman and Brownlee (Unoccupied, Solitary,
Together, Parallel, Group). If observers are told where the segment bound-
aries are, their only task would be to code segments. In this case, it would
be easy to construct a kappa table. Telling observers where the segment
boundaries are, however, is not an easy matter. Other observers, working
72 Assessing observer agreement

with videotaped material, would need to have determined those boundaries

previously, and then the boundaries would need to be marked in some way,
perhaps with a brief tone dubbed on the soundtrack.
If observers are not told where the segment boundaries are, however, but
instead are asked to both segment and code at the same time, then protocols
like those shown in Figure 4.3 would result. The question now is, how do
we construct a kappa table from data like these? There are two choices.
We could ignore those parts of the protocols where observers did not agree
as to segment boundaries and tally only those segments whose boundaries
were agreed upon, in effect ignoring errors of omission. This would re-
sult in eight agreements and two disagreements, as shown in Figure 4.3
(Method A). Or we could assume that there "really" was a segment bound-
ary whenever one of the observers said there was. This would result in nine
agreements and four disagreements (Method B). Neither of these choices
seems completely satisfactory. Thefirstprobably overestimates agreement,
whereas the second probably underestimates it. Our preference is for com-
puting kappas using a time interval as the unit, but this requires timing
onsets and offsets, timing pattern changes, or coding intervals directly.
As an example, consider the recording of onset and offset times for dif-
ferent kinds of embryonic distress calls done by Tuculescu and Griswold
(section 2.10). Assume that these times were recorded to the nearest sec-
ond. Then each second can be categorized as "containing" (a) a Phioo,
(b) a Soft Peep, (c) a Peep, (d) a Screech, or (e) no distress call. If two
observers both code the same audiotape, agreement data could be like that
shown in Figure 4.4. Similarly, in addition to the kappa for embryonic dis-
tress calls, kappas could be computed for Tuculescu and Griswold's other
superordinate categories as well (embryonic pleasure calls, maternal body
movements, maternal head movements, and maternal vocalizations).
Occasionally, editors or colleagues ask for agreement statistics, not for
a coding scheme, as kappa provides, but for individual codes. We usu-
ally think that kappa coupled with the agreement matrix is sufficient, but
nonetheless kappas can be computed for individual codes by collapsing
the table appropriately. Consider the agreement matrix for the five distress
calls shown in Figure 4.4. From it we could derive five 2 x 2 matrices, first
collapsing the tallies into Phioo/not Phioo, then Soft Peep/not Soft Peep,
etc. Then a kappa could be computed separately for each table. In this case,
the individual kappas would be .79, .87, .43, .98, and .93 for Phioo, Soft
Peep, Peep, Screech, and None, respectively. Not surprisingly, the kappa
associated with Peep is relatively low; of the 10 noted by the first observer
and the 13 noted by the second observer, agreement occurred for only 5.
A second example of a time-based kappa is provided by Adamson and
Bakeman (1985), who asked observers to record whenever infants dis-
played heightened affectivity. These displays were typically quite brief,
 Agreement about codes 73

1W HZ 3 O O 8 53

i 13 7 o 5 \ob

z 3 5 o O \0

o 1 O 165-

7 1 i H3O HH3

5Z 13 4H5 777

Figure 4.4. An agreement matrix for coding of chicken embryonic distress calls.
Each tally represents a 1-second interval. The reader may want to verify that the
percentage of agreement observed is 94.2%, the percentage expected by chance
is 39.4%, the value of kappa is .904, its standard error is .0231, and the z score
comparing kappa to its standard error is 39.2.

lasting just a few seconds, relatively infrequent, and consisted of such

things as smiles, gleeful vocalizations, or excited arm waving. Assuming
a unit of 1 second, a 10-minute agreement check could produce a kappa
table like the one given on the left in Figure 4.5. What would happen, how-
ever, if a half-second unit had been used instead? The answer is essentially
nothing, as is demonstrated by the table on the right in Figure 4.5. By and
large, halving the length of the unit would result in tables that are roughly
proportional, except that one would have twice as many tallies as the other.
The kappa statistic (unlike chi-square) is not affected by this, however, as
the kappa computations in Figure 4.5 demonstrate.
In the preceding two paragraphs (and in Figures 4.4 and 4.5), we have
suggested how kappa can be computed when onset times for mutually
exclusive and exhaustive codes are recorded. The principle is exactly
the same when timing pattern changes. Again, for each set of mutually
exclusive and exhaustive codes, a kappa table can be constructed, tallying
agreement and disagreement for each second (or whatever time unit is
used) coded. When timing pattern changes, codes are always parceled into
74 Assessing observer agreement

Ok*
At TUt
13 A$ 22 4 26
1
583 587 8 1166 1174

15 585 600 To 50 1170 1200

ClS»t3MS8S»587) p.(3O»26Mll7O»l»7H)
600*600 rc
" 1200*1200

Figure 4.5. When time intervals are tallied, and a reasonable time interval is used,
the value of kappa is not changed by halving the time interval (which doubles the
tallies).

mutually exclusive and exhaustive sets. In other cases, constructing a set

of mutually exclusive and exhaustive codes is no problem, as the examples
presented in Figures 4.4 and 4.5 demonstrate. Adding the "none of the
above" code to a code for affective display or to codes for distress calls
makes the set mutually exclusive and exhaustive.
When the remaining two recording strategies - coding intervals and
cross-classifying events - are used, computing agreement using kappa is
straightforward. An example of what a kappa table might look like when
intervals are coded was given earlier (see Figure 4.1). When events are
cross-classified, the only difference is that there is one kappa table for each
classification scheme (or dimension) and that events not detected by both
observers cannot be entered into the table. For example, in their study of
social rules, Bakeman and Brownlee (1982) reported kappas for each of
their three dimensions: (a) prior possession, (b) resistance, and (c) success
(see section 2.13). Kappa, however, is not the only agreement statistic
there is. As we discuss in the next section, other statistics and approaches
to observer agreement have their advantages.
 Generalizability theory 75

4.7 Generalizability theory

Cronbach, Gleser, Nanda, and Rajaratnam (1972; see also Brennan, 1983)
presented what amounts to a conceptual breakthrough in thinking about
both reliability and validity. To understand their notions, let us introduce
the concept of the "work we want the measure to do." For example, we
would like to be able to use our observations of the amount of negative-
affect reciprocity to discriminate satisfied from dissatisfied marriages. Or,
we might want our measure of the amount of negative affect to predict the
husband's health in three years. This is the work our measure is to do. It
is designed to discriminate or to predict something of interest.
Cronbach and colleagues' major point is that this work is always relative
to our desire in measurement to generalize across some facet of our exper-
imental design that we consider irrelevant to this work. For example, our
test scores should generalize across items within a measurement domain. It
should not matter much if we correlate math achievement and grade point
average (GPA) using even or odd math achievement items to compute the
correlation coefficient. We are generalizing across the irrelevant facet of
odd/even items. The work the measure does is to discriminate high- from
low-GPA students.
In a similar way, if we have a measure of negative affect, we expect it to
discriminate among happily and unhappily married couples, and not to dis-
criminate among coders (the irrelevant facet). We wish to generalize across
coders. Let us briefly discuss the computations involved in this analysis.
Figure 4.6 presents the results of one possible generalizability study. For
five persons, each of two observers computed the frequency of code A for
a randomly selected segment of videotape. The setup is the same as a sim-
ple repeated-measures experiment. The within-subject factor is observer
(with two levels, i.e., data from two observers) and there is no between-
subject factor as such; subjects represent total between-subject variability.
The analysis of variance source table for the data shown in Figure 4.6,
expanded to include R2 as recommended by Bakeman (1992), is shown
in Table 4.1. Given these data and the current question, an appropriate
coefficient of generalizability, or reliability, is:
MSp-MSr
a = (4.1)
MSP + (no - \)MSr
where n0 is the number of observers (2 in this case) and MSP and MSr are
the mean squares for persons and residual (or error), respectively (the first
edition of this book omitted no — 1 in the denominator because it equaled
1, but this proved confusing). This is an intraclass correlation coefficient
based on the classical assumption that observed scores can be divided into a
true and an error component (X = T + e), so that the appropriate intraclass
76 Assessing observer agreement

ft**
\ Z 1 Q
=• LET
I zo Q a (7O
3 3O 22. = 26.O
H 3 7 = SO
5 iZO 84 P5 s I02.O
5^=25:6 h * 30.3

= 0.95

Figure 4.6. A generalizability approach to observer agreement: no is the number

of observers or 2, n p is the number of persons or 5, MSP is the mean square for
persons, MSO is the mean square for observers, and MSr is the mean square for
residual or, in this case, the P x O interaction; see also Table 4.1. The formulas
for MSP and MSO were incorrect in the first edition of this book; they are also
given incorrectly in Wiggins (1973, p. 289).

correlation is defined as

a = -J^-J. (4.2)

(Equation 4.1 is derived from 4.2 by substitution and algebraic manip-

ulation.) This statistic estimates the reliability of observations made by
a randomly selected observer, selected from the pool that contained the
two observers used here for the reliability study (Table 4.1), and further
 Generalizability theory 11

Table 4.1. Source Table for the Generalizability Study of

Observer Reliability
Source R2 AR2 SS df MS

Person 0.95 0.95 13,611.6 4 3,402.9

Observer 0.97 0.02 220.9 1 220.9
PxO 1.00 0.03 485.6 4 121.4
Total 14,318.1 9

assumes that data will be interpreted within what Suen (1988) terms a norm-
referenced (i.e., values are meaningful only relatively; rank-order statistics
like correlation coefficients are emphasized) as opposed to a criterion-
referenced framework (i.e., interpretation of values references an absolute
external standard; statistics like unstandardized regression coefficients are
emphasized). Equation 4.1 is based on recommendations made by Hart-
mann (1982) and Wiggins (1973, p. 290). For other possible intraclass
correlation coefficients (generalizability coefficients), based on other as-
sumptions, see Fleiss (1986, chapter 1) and Suen (1988), although, as a
practical matter, values may not be greatly different. For example, val-
ues for a criterion-referenced fixed-effect and random-effect model per
Fleiss (1986) were .926 and .921, respectively, compared to the .931 of
Figure 4.6. In contrast, assuming that observers were item scores and we
wished to know the reliability of total scores based on these items, Cron-
bach's internal-consistency alpha (which is MSP — MSr divided by MSP;
see Wiggins, 1973, p. 291) was .964.
This way of thinking has profound consequences. It means that reli-
ability can be high even if interobserver agreement is moderate, or even
low. How can this be? Suppose that for person #5 in Figure 4.6, Observer
1 detected code A 120 times, as shown but only 30 of these overlapped
in time with Observer 2's 84 entries. Then the interobserver agreement
would be only 30/120 = .25. Nonetheless, the generalizability coefficient
of equation 4.1 is .93. The reliability is high because either observer's data
distinguishes equally well between persons. The agreement within person
need not be high. The measure does the work it was intended to do, and
either observer's data will do this work. This is an entirely different notion
of reliability than the one we have been discussing.
Note that the generalizability or reliability coefficient estimated by equa-
tion 4.1 is a specific measure of the relative variance accounted for by an
interesting facet of the design (subjects) compared to an uninteresting one
(coders). This is an explicit and specific proportion, but it does not tell us
78 Assessing observer agreement

how large a number is acceptable, any more than does the proportion of
variance accounted for in a dependent variable by an independent variable.
The judgement must be made by the investigator. It is not automatic, just
as a judgment of an adequate size for kappa is not an automatic procedure.
Note also that what makes the reliability high in the table in Figure 4.6
is having a wide range of people in the data, ranging widely with respect to
code A. Jones, Reid, and Patterson (1975) presented the first application of
Cronbach and colleagues' (1972) theory of measurement to observational
data.
The reliability analysis just presented, although appropriate when scores
are interval-scaled (e.g., number of events coded A by an observer), is
inadequate for sequential analysis. The agreement required for sequential
analysis cannot be collapsed over time, but must match point for point, as
exemplified by the kappa tables presented in the previous section. Such
matching is much more consistent with "classical" notions of reliability,
i.e., before Cronbach et al. (1972).
Still, agreement point-for-point could be assessed in the same manner as
in Figure 4.6. The two columns in Figure 4.6 would be replaced by sums
from the confusion matrix. Specifically, the sums on the diagonal would
replace Observer l's scores, and the sums of diagonal plus off-diagonal
cells (i.e., the row marginals) would replace Observer 2's scores. If the
agreement point-for-point were perfect, all entries for code A in the confu-
sion matrix would be on the diagonal and the two column entries would be
the same. There would then be no variation across "observers," and alpha
would be high. In this case, the "persons" of Figure 4.6 become codes,
"Observer 1" becomes agreements and "Observer 2" becomes agreements
plus disagreements.
This criterion is certainly sufficient for sequential analysis. However, it
is quite stringent. Gottman (1980a) proposed the following: If independent
observers produce similar indexes of sequential connection between codes
in the generalizability sense, then reliability is established. For example,
if two observers produced the data in Figure 4.7 (designed so that they
are off by one time unit, but see the same sequences), their interobserver
agreement would be low but indexes of sequential connection would be very
similar across observers. Some investigators handle this simple problem
by having a larger time window within which to calculate the entries in the
confusion matrix. However, that is not a general solution because more
complex configurations than that of Figure 4.7 are possible, in which both
observers detect similar sequential structure in the codes but point-for-point
agreement is low. Cronbach et al.'s (1972) theory implies that all we need
to demonstrate is that observers are essentially interchangeable in doing
the work that our measures need to do.
 Unreliability as a research variable 79

A B B A B A B A B A B
A A B B A B A B A B A 2.'s

B
1 5

4 1

Figure 4.7. A confusion matrix when observers "see" the same sequence but one
observer lags the other. In such cases, a "point-for-point" agreement approach
may be too stringent.

4.8 Unreliability as a research variable

Raush (personal communication, 1974) once suggested that one source of
unreliability is to be found in the nature of the social interaction itself.
He referred to a message called a "probe" that a person might send to the
receiver. The probe is designed to be taken one of two ways, depending on
the state of the receiver. For example, in a potentially sexual interaction a
sender may send a probe with a subtle sexual invitation. If it is ignored,
the sender gains information that directs the interaction one way; if it is
responded to, the interaction may proceed in another direction. Krokoff
(1983) recently tested this notion in marital interaction. He reasoned that
such probe messages would be more common during high-conflict inter-
action because of the delicate nature of the subject matter and the great
danger that the conflict would escalate. If this were true, Krokoff rea-
soned, then reliability would be significantly reduced for those videotapes
high in negative affect. This hypothesis was strongly supported by the data.
Patterson's (1982, p. 50) book quoted Reid as noting that "observer
agreement is largely a function of the complexity of the interaction. By
selecting simple interaction segments, one may obtain very high observer
agreement." When complexity was defined as the number of different
codes entered divided by the total entries for 5 minutes, this hypothesis
80 Assessing observer agreement

was strongly supported; reliability was lower for more complex segments.
This could partly be due to the increased task demands on the coder, but it
could also be partly a property of the interaction itself, if in more complex
interactions people sent more probe messages. The point of this section is
to suggest that reliability can itself become a research variable of interest.

4.9 Summary
There are at least three major reasons for examining agreement among
observers. The first is to assure ourselves and others that our observers are
accurate and that our procedures are replicable. The second is to calibrate
multiple observers with each other or with some assumed standard. This
is important when the coding task is too large for one observer or when it
requires more than a few weeks to complete. The third reason is to provide
feedback when observers are being trained.
Depending on which reason is paramount, computation of agreement
statistics may proceed in different ways. One general guiding principle is
that agreement need be demonstrated only at the level of whatever scores
are finally analyzed. Thus if conditional probabilities are analyzed, it is
sufficient to show that data derived from two observers independently cod-
ing the same stream of behavior yielded similar conditional probabilities.
Such an approach may not be adequate, however, when training of ob-
servers is the primary consideration. Then, point-by-point agreement may
be demanded. Point-by-point agreement is also necessary when data de-
rived from different observers making multiple coding passes through a
videotape are to be merged later.
When an investigator has sequential concerns in mind, then, point-by-
point agreement is necessary for observer training and is required for at
least some uses of the data. Moreover, if point-by-point agreement is
established, it can generally be assumed that scores derived from the raw
sequential data (like conditional probabilities) will also agree. If agreement
at a lower level is demonstrated, agreement at a higher level can be assumed.
For these reasons, we have stressed Cohen's kappa in this chapter because
it is a statistic that can be used to demonstrate point-by-point agreement.
A Pascal program that computes kappa and weighted kappa is given in the
Appendix. Kappa is also computed by Bakeman and Quera's Generalized
Sequential Querier or GSEQ program (Bakeman & Quera, 1995a). At the
same time, we have also mentioned other approaches to observer reliability
(in section 4.7). This hardly exhausts what is a complex and far-ranging
topic. Interested readers may want to consult, among others, Hollenbeck
(1978) and Hartmann (1977, 1982).
 Representing observational
data

5.1 Representing versus recording

When observational data are being recorded, it is reasonable that practical
concerns dictate what is done. One uses whatever equipment is available
and convenient. One records in ways that are easy and natural for the human
observers. One tries to preserve in some form the information that seems
important. Thus recorded data can appear in literally a multitude of forms.
Some of these were described in chapter 3; however, we recognize that
what investigators actually do often combines the simple forms described
there into more complex, hybrid forms.
However, the form used for data recording - the data as collected - should
not become a straitjacket for the analysis to follow. A format that works
well for data recording may not work so well for data analysis, and a
format that works well for one analysis may not work well for another
analysis. The solution is to figure out simple ways to represent data, ways
that make different kinds of analysis simple and straightforward. There is
nothing especially sacrosanct about the form of the recorded data, after all,
and there is no merit in preserving that form when it proves awkward for
subsequent uses.
It would be very useful if just a few relatively standard forms for repre-
senting observational data could be defined. Not only would this help to
standardize terminology with respect to sequential analysis, thus facilitat-
ing communication, it would also make analysis easier and would facilitate
designing and writing general-purpose computer programs for sequential
analysis. We assume that most investigators use computers for their data
analysis tasks, but even if they do not, we think that representing the data
by use of one (or more) of the five standard forms defined in this chapter
will make both thinking about and doing data analysis a simpler and more
straightforward affair. Further, there is nothing exclusive about these five
forms. Depending on how data were recorded, investigators can, and prob-
ably often will, extract different representations from the same recorded
data for different purposes.

81
82 Representing observational data

The first four forms are defined by Bakeman and Quera's (1992, 1995a)
Sequential Data Interchange Standard (SDIS), which defines a standard
form for event, state, timed-event, and interval sequences, respectively.
Sequential data represented by any of these forms can be analyzed with
the Generalized Sequential Querier (GSEQ; Bakeman & Quera, 1995a).
The fifth form is an application of the standard cases by variables rectan-
gular matrix and is useful for analyzing cross-classified events, including
contingency table data produced by the GSEQ program.

5.2 Event sequences

The simplest way to represent sequential behavior is as event sequences. As
an example, imagine that observers used the following mnemonic codes
for the parallel play study described in chapter 1: Un = Unoccupied,
Sol = Solitary, Tg = Together, Par = Parallel, and Gr = Group. A child
is observed for just a minute or two. First she is unoccupied, then she
shifts into together play, than back to unoccupied, back to together, then
into solitary play, and back to together again. If each code represents a
behavioral state, then the event sequence data for this brief observation
would look like this:
Un Tg Un Tg Sol Tg. . .
In this case, there are no Par or Gr codes because the child was not observed
in parallel or group play during this observation session.
The data might have been recorded directly in this form or they might
have been recorded in a more complex form and only later reduced to event
sequences. The behavior to be coded could be thought of as behavioral
states, as in this example, or as discrete events. Behavior other than just the
states or events that form the event sequences might have been recorded but
ignored when forming event sequences for subsequent analysis. In general,
if the investigator can define a set of mutually exclusive and exhaustive
codes of interest, and if the sequence in which those codes occurred can be
extracted from the data as recorded, and if reasonable continuity between
successive codes can be assumed (as discussed in section 3.4), then some
(if not all) of the information present in the data can be represented as event
sequences.
This is often very desirable to do. For one thing, the form is very simple.
A single stream of codes is presented without any information concerning
time, whether onsets or offsets. Lines of an event-sequential data file consist
simply of codes for the various events, ordered as they occurred in time,
along with information identifying the participant(s) and sessions. Thus
event sequential data are appropriate when observed behavior is reduced
 State sequences 83

to a single stream of coded events (which are thus mutually exclusive and
exhaustive by definition), and when information about time (such as the
duration of events) is not of interest. Event sequences are both simple and
limited. Yet applying techniques described in chapter 7 to event sequences,
Bakeman and Brownlee were able to conclude that parallel play often
preceded group play among 3-year-old children.

5.3 State sequences

For some analyses, the duration of particular behavioral states or events
may matter, either because the investigator wants to know how long a
particular kind of event lasted on the average or because the investigator
wants to know what proportion of the observation time was devoted to a
particular kind of event. For example, it may be important to know that the
mean length for a bout of parallel play was 34 seconds and that children
spent 28% of their time engaged in parallel play, on the average. In such
cases, the form in which data are represented needs to include information
about how long each event or behavioral state lasted.
As we define matters, state sequences are identical to time sequences with
the simple addition of timing information. The terminology is somewhat
arbitrary, and in the next section we discuss timed-event sequences, but
our intent is to provide a few simple forms for representing sequential data,
some of which are simpler than others, so that investigators can choose a
form that is no more complex than required for their work. For example, if
duration were important, the SDIS state-sequential representation for the
sequence given earlier would be:
Un=12 Tg = 8 Un = 21 Tg=ll
Sol = 34 Tg = 6 • • •
Assuming a time unit of a second, this indicates that Unoccupied lasted 12
seconds, followed by 8 seconds of Together, 21 more seconds of Unoccu-
pied, etc. The same sequence can also be represented as:
Un,8:01 Tg,8:13 Un,8:21
Tg,8:42 Sol,8:53 Tg,9:27 ,9:33 • • •
which indicates an onset time for Unoccupied of 8 minutes and 1 second,
for the first Together of 8 minutes and 13 seconds, etc. The offset time for
the session is 9 minutes and 33 seconds; because the first onset time was
8:01, the entire session lasted 92 seconds.
State sequences, like simple event sequences, provide a useful way to
represent aspects of the recorded data when a single stream of mutually
84 Representing observational data

exclusive and exhaustive (ME&E) coded states (or events) captures infor-
mation of interest. It would be used instead of (or in addition to) state
sequences when information concerning proportion of time devoted to a
behavior (e.g., percentage of time spent in parallel play) or other timing
information (e.g., average bout length for group play) is desired. Addition-
ally, it is possible to define multiple streams of ME&E states; for details see
Bakeman and Quera (1995a). In sum, both simple event and state sequences
are useful for identifying sequential patterns, given a simple ME&E cod-
ing scheme. But they are not useful for identifying concurrent patterns
(unless multiple streams of states are defined) or for answering relatively
specific questions, given more complex coding schemes. In such cases, the
timed-event sequences described in the next section may be more useful.

5.4 Timed-event sequences

If codes can cooccur, and if their onset and offset times were recorded, then
the data as collected can be represented as timed-event sequences. This is a
useful and general-purpose format. Once data are represented in this form,
an investigator can determine quite easily such things as how often specific
behavioral codes cooccur (does the baby smile mainly when the mother is
looking at him or her?), or whether certain behavioral codes tend to follow
(or precede) other codes in systematic ways (does the mother respond to
her baby's smiling within 5 seconds?).
For example, imagine that two people engaged in conversation were
videotaped, and four codes were defined: Alook, meaning person A looks
at person B; Blook, meaning B looks at A; Atalk, and Btalk. A brief segment
of the coded conversation, depicted as though it had been recorded with
an old-fashioned, deflected-pen, rolling-paper event recorder, is shown in
Figure 5.1. Following SDIS conventions for timed-event sequences, this
same segment would be represented as follows:

,1 Alook, 2-4 Atalk,3-10 Blook,4-7 Alook,6

Blook,8-10 Alook, 12-16 Btalk, 12-19
Blook, 13 Alook, 17-20 AtaM8 , 21
Assuming time units are a second, this session began at second 1 and
ended at second 21; thus the session lasted 20 seconds. Person A first
began looking at second 2 and stopped at second 4; thus As first looking
bout lasted 2 seconds, etc. By convention, when offset times are omitted,
durations are assumed to be one time unit; thus Alook, 6 implies an offset
time of 7. As you can see, offset times are assumed to be exclusive, but
 Interval sequences 85

6 8 10 12 14 16 18 20
Time Interval

Figure 5.1. An example of the kind of output produced by a deflected-pen event

recorder.

SDIS conventions also allow inclusive offset times (signaled by a right

parenthesis). Thus the segment just given could also be represented as:

, 1 Alook, 2-3) Atalk, 3-9) Blook, 4 - 6 ) Alook, 6

Blook, 8-9) Alook, 12-15) Btalk, 12-18)
Blook, 13 Alook, 17-19) Atalk, 18 ,20)
which some investigators may find more convenient. Other conventions for
timed-event sequences, including inferred offset times and context codes,
are detailed in Bakeman and Quera (1995a).
The flexibility of the timed-event format is very useful. As in other SDIS
formats, codes can be defined for any number of behaviors (although, as
a practical matter, the SDIS program limits the number to 95), but unlike
event sequence data, timed-event sequential data (TSD) preserves onset and
offset times, and unlike a single stream of state sequences, TSD preserves
cooccurrences among behaviors. This allows for the level of complexity
represented in the data to approach "real life," and it certainly imposes a
minimum of constraints on the questions investigators can ask.

5.5 Interval sequences

The fourth format defined by SDIS, interval sequences, is designed to ac-
commodate interval recording in a simple and straightforward way. Codes
86 Representing observational data

are simply listed as they occur and interval boundaries are represented by
commas. For example:

Interval = 5;
, , Ivoc, Ivoc, Aofr, Rain Aofr Ivoc Ismi, Rain, . . .

indicates that intervals have a duration of 5 time units. No behavior is

coded for the first two intervals. The infant vocalizes in intervals 3 and 4
and an adult offers in interval 5. The adult continues to (or again) offers
in interval 6, at which time the infant both vocalizes and smiles and it be-
gins to rain. The rain continues in interval 7, etc. Other conventions for
interval sequences, including context codes, are detailed in Bakeman and
Quera (1995a).
If the interval width for interval sequences were 1 second, for example,
and if the time unit for timed-event sequences were likewise 1 second, then
the same observed sequence could be represented with either timed-event
sequences or interval sequences. Both allow behavioral codes to cooc-
cur. But when an investigator begins with an interval recording strategy,
it is often easier (and requires fewer key strokes) to represent such data
directly as interval sequences, which is why we included this format as
part of SDIS.
Many computer-based data collection devices and programs in common
use produce data in their own format. For example, timed-event sequences
in SDIS place the code first, followed by a comma, followed by the time the
code occurred, whereas at least a few data collection systems with which
we are familiar place the time before the code. This is hardly problematic.
In such cases, it is relatively easy to write a simple computer program
(in Pascal, for example) that reformats the data as initially collected into
SDIS format.
No matter the particular form - whether event, state, timed-event, or
interval sequences - the advantages of a few standard forms for sequen-
tial data are considerable. They make it easier, and more worthwhile,
to develop general-purpose programs for sequential data - such as GSEQ
(Bakeman & Quera, 1995a) - that compute, not just simple frequencies and
percentages for different codes, but a variety of conditional probabilities
and sequential and other statistics as well. Such general-purpose programs
can then be shared among different laboratories. In addition, relying on a
standard form for data representation should enhance the development of
special-purpose software within a given laboratory as well.
 Cross-classified events 87

5.6 Cross-classified events

In chapter 3, we distinguished between two general approaches to data
collection: intermittent versus continuous recording. Because this is a
book about sequential analysis, we stressed continuous recording strategies,
defining four particular ones: (a) coding events, (b) recording onset and
offset times, (c) timing pattern changes, and (d) coding intervals. We
also said that a fifth strategy - cross-classifying events - could result in
sequential data if the major categories used to cross-classify the event could
be arranged in a clear temporal order.
When continuous recording strategies are used, there is some choice
as to how data should be represented. When events are cross-classified,
however, there seems only one obvious way to do it: Each line represents
an event, each column a major category. For example, as described in
section 2.13, Bakeman and Brownlee coded object struggles. The first
major category was prior possession, the second was resistance, and the
third was success. Thus (if 1 = yes, 2 = no) the following

1 1 1
2 1 2

would code two events. In the first, the child attempting to take the object
had had prior possession (1), his take attempt was resisted (1), but he
succeeded in taking the object (1). In the second, the attempted taker had
not had prior possession (2), he was likewise resisted (1), and in this case,
the other child retained possession (2).
Unlike the SDIS data formats discussed in the previous several sections,
data files that contain cross-classified event data are no different from the
usual cases by variables rectangular data files analyzed by the standard sta-
tistical packages such as SPSS and SAS. Typically, cross-classified data are
next subjected to log-linear analyses. When events have been detected and
cross-classified in the first place, the data file could be passed, more or less
unaltered, to the log-linear routines within SPSS or other standard pack-
ages, or to a program like ILOG (Bakeman & Robinson, 1994) designed
specifically for log-linear analysis. Moreover, often analyses of SDIS data
result in contingency table data, which likewise can be subjected to log-
linear analysis with any of the standard log-linear programs. In fact, GSEQ
is designed to examine sequential data and produce contingency table sum-
maries in a highly flexible way; consequently it allows for the export of
such data into files that can subsequently be read by SPSS, ILOG, or other
programs.
88 Representing observational data

Table 5.1. Relationship between data recording and data

representation
This data recording strategy Allows for this data representation

Coding events, no time Events sequences

Recording onset and offset Events, state, timed-event, or interval
times, or timing pattern sequences
changes, or coding intervals
Cross-classifying events Cross-classified events

5.7 Transforming representations

Early in this chapter, we suggested that the data as collected should not
become a straitjacket for subsequent analyses; that it was important to put
the data into a form convenient for analysis. A corollary is that the data as
collected may take various forms for different analyses and that one form
may be transformed into another. There are limits on possible transmuta-
tions, of course. Silken detail cannot be extracted from sow-ear coarseness.
Still, especially when onset and offset times have been recorded, the data as
collected can be treated as a data "gold mine" from which pieces in various
forms can be extracted, tailored to particular analyses.
The relationship between data recording strategies and data representa-
tion form is presented in Table 5.1. As can be seen, when onset and offset
or pattern-change times are recorded, aspects of that collected data can
be represented as event, state, timed-event, or interval sequences. Further
data represented initially as state or timed-event sequences can be trans-
formed into event or even interval sequences; again, see Bakeman and
Quera (1995a) for details.
One example of the potential usefulness of data transformation is pro-
vided by Bakeman and Brown's (1977) study of early mother-infant inter-
action. Desiring to establish an "ethogram" of early interactive behavior in
the context of infant feeding, they defined an extensive number of detailed
behavioral codes, more than 40 for the infant and 60 for the mother. Some
of these were duration events, some momentary, the whole represented as
timed-event sequences. For one series of analyses, Bakeman and Brown
wanted to examine the usefulness of viewing interaction as a "behavioral
dialogue." To this end, they defined some of the mother codes and some of
the infant codes as representing "communicative acts," actions that seemed
potentially communicative or important to the partner.
This done, they then proceeded to extract interval sequences from the
timed-event sequential data. Each successive interval (they used a 5-second
 Transforming representations 89

interval, but any "width" interval could have been used) was categorized
as follows: (a) interval contains neither mother nor infant "communicative
act" codes, (b) interval contains some mother but no infant codes, (c)
interval contains some infant but no mother codes, and (d) interval contains
both mother and infant codes. This scheme was inspired by others who
had investigated adult talk (Jaffe & Feldstein, 1970) and infant vocalization
and gaze (Stern, 1974). With it, Bakeman and Brown were able to show, in
a subsequent study, differences between mothers interacting with preterm
and full-term infants (Brown & Bakeman, 1980). But the point for now is
to raise the possibility, and suggest the usefulness, of extracting more than
one representation from the data as originally recorded.
A second example of data transformation is provided by the Bakeman
and Brownlee (1980) study of parallel play described in chapter 1. There an
interval recording strategy was used. Each successive 15-second interval
was coded for predominant play state as follows: (a) Unoccupied, (b)
Solitary, (c) Together, (d) Parallel, or (e) Group play. Thus the data as
collected were already in interval sequential form and were analyzed in
this form in order to determine percentage of intervals assigned to the
different play states.
However, to determine if these play states were sequenced in any system-
atic way, Bakeman and Brownlee transformed the interval sequence data
into event sequences, arguing that they were concerned with which play
states followed other states, not with how long the preceding or follow-
ing states lasted. But once again, the moral is that for different questions,
different representations of the data are appropriate.
Throughout this book, the emphasis has been on nominal-scale measure-
ment or categorization. Our usual assumption is that some entity-an event
or a time interval - is assigned some code defined by the investigator's
coding scheme. But quantitative measurement can also be useful, although
such data call for analytic techniques not discussed here. (Such techniques
are discussed in Gottman, 1981.) The purpose of this final example of data
transformation is to show how categorical data can be transformed into
quantitative time-series data.
Tronick, Brazelton, and their co-workers have been interested in the
rhythmic and apparently reciprocal way in which periods of attention and
nonattention, of quiet and excitation, seem to mesh and merge with each
other in the face-to-face interaction of mothers with their young infants
(e.g., Tronick, Als, & Brazelton, 1977). They videotaped mothers and in-
fants interacting and then subjected those tapes to painstaking coding, using
an interval coding strategy. Several major categories were defined, each
containing a number of different codes. The major categories included,
for example, vocalizations, facial expressions, gaze directions, and body
90 Representing observational data

movement for both mother and infant. The tapes were viewed repeatedly,
often in slow motion. After each second of real time, the observers would
decide on the appropriate code for each of the major categories. The end
result was interval sequential data, with each interval representing 1 second
and each containing a specific code for each major category.
Next, each code within each of the major categories was assigned a num-
ber, or weight, reflecting the amount of involvement (negative or positive)
Tronick thought that code represented. In effect, the codes within each ma-
jor category were ordered and scaled. Then, the weights for each category
were summed for each second. This was done separately for mother and in-
fant codes so that the final result was two parallel strings of numbers, or two
times series, in which each number represented either the mother's or in-
fant's degree of involvement for that second. Now analyzing two time series
for mutual influence is a fairly classic problem, more so in astronomy and
economics than psychology, but transforming observational data in this way
allowed Gottman and Ringland (1981) to test directly and quantitatively
the notion that mother and infant were mutually influencing each other.

5.8 Summary
Five standard forms for representing observational data are presented here.
The first, event sequences, consists simply of codes for the events, or-
dered as they occurred. The second, state sequences, adds onset times
so that information such as proportions of time devoted to different codes
and average bout durations can be computed. The third, timed-event se-
quences, allows for events to cooccur and is more open-ended; momen-
tary and duration behaviors are indicated along with their onset and offset
times, as required. The fourth, interval sequences, provides a convenient
way to represent interval recorded data. And the fifth form is for cross-
classified events.
An important point to keep in mind is that data as collected can be rep-
resented in various ways, depending on the needs of a particular analysis.
Several examples of this were presented in the last section. The final exam-
ple, in fact, suggested a sixth data representation form: time series. Ways to
analyze all six forms of data are discussed in the next chapters, although the
emphasis is on the first five. (For time-series analyses, see Gottman, 1981.)
Some of the analyses can be done by hand, but most are facilitated by using
computers. An advantage of casting data into these standard forms is that
such standardization facilitates the development and sharing of computer
software to do the sorts of sequential analyses described throughout the rest
of this book. Indeed, GSEQ (Bakeman & Quera, 1995a) was developed to
analyze sequential data represented according to SDIS conventions.
 Analyzing sequential data:
First steps

6.1 Describing versus modeling

If all the steps described in previous chapters - developing coding schemes,
recording behavioral sequences reliably, representing the observational
data - are in order, then the first fruits of the research should be simple
description. Introductory textbooks never tire of telling their readers that
the basic tasks of psychology are, one, description, and two, explanation.
Similarly, almost all introductory textbooks in statistics distinguish be-
tween descriptive statistics, on the one hand, and inferential statistics, on
the other. This distinction is important and organizes not just introductory
statistics texts but this and the next four chapters as well.
Much of the material presented in this and the following four chapters,
however, assumes that readers want first to describe their data, and so
description is emphasized. Problems of inference and modeling - deter-
mining if data fit a particular model, estimating model parameters - are
touched on only slightly here. These are important statistical topics and
become especially so when one wants to move beyond mere description
to a deeper understanding of one's data. That is why so many books and
courses, indeed huge specialized literatures, are devoted to such topics.
We assume that readers will use scores derived from observing behav-
ioral sequences as input for anything from simple chi-square or analyses
of variance, to log-linear modeling, to the modeling approach embodied in
programs like LISREL.
Our task, fortunately, is not to describe all the modeling possibilities
available. Instead, we have set ourselves the more manageable task of
discussing how to derive useful descriptive scores from sequential data.
Still, we describe some simple instances of model testing and try to point
out when sequential data analysis presents particular problems for statis-
tical inference. Throughout, we attempt to maintain the distinction be-
tween description and modeling. Thus we would never talk of "doing" a
Markov analysis. Instead we would describe how to compute transitional

91
92 Analyzing sequential data: First steps

probabilities, on the one hand, and how to determine if those transitional

probabilities fit a particular Markov model, on the other (see section 8.3).
In the remainder of this chapter, we note some of the simpler descriptive
statistics that can be derived from sequential observational data.

6.2 Rates and frequencies

Some statistics are so obvious that it almost seems an insult to the reader's
intelligence to mention them at all. Still, they are so basic, and so useful,
that omitting them would seem negligent. Certainly, logical completeness
demands their inclusion. One such statistic is the rate or frequency with
which a particular event occurred.
What is recorded is the event. This could be either a momentary or
a duration event. Just the occurrence of the event might be recorded, or
its onset and offset times might be recorded as well, or the occurrence of
events might be embedded in another recording strategy such as coding
intervals or cross-classifying events.
What is tallied is the frequency - how often a particular event occurred.
The data could be represented as simple event sequences, as state sequences,
as timed-event sequences, or as cross-classified events. In all these cases,
it is possible to count how often particular events occurred. Only interval
sequences pose potential problems. Two successive intervals that contain
the same code may or may not indicate a single bout; thus frequencies
of intervals that contain a particular code may overestimate the number
of instances, and this needs to be taken into account when interpreting
frequency data for interval sequences.
In most cases, raw frequencies should be transformed to rates, fre-
quencies, of course, depend on how long observation continues, whereas
rates have the merit of being comparable across cases (individuals, days,
etc.). The total observation time needs to be recorded, of course, even for
event sequences, which otherwise do not record time information; other-
wise rates cannot be computed. For example, Adamson and Bakeman
(1985) observed infants at different ages (6, 9, 12, 15, and 18 months)
and with different partners (with mothers, with peers, and alone). A coder
recorded whenever infants engaged in an "affective display." For a variety
of reasons, the observation sessions (with each partner, at each age) varied
somewhat in length. Thus Adamson and Bakeman divided the number of
affective displays (for each observation session) by the observation time
(in their case, some fraction of an hour), yielding a rate per hour for affec-
tive displays. This statistic, just by itself, has descriptive value (there were
59 affective displays per hour, on the average), but further, it can be (and
 Probabilities and percentages 93

was) used as a score for traditional analysis of variance. Thus Adamson

and Bakeman were able to report that the rate of affective displays was
greater when infants were with mothers instead of peers, and that this rate
increased with age.

6.3 Probabilities and percentages

A second obvious and very useful descriptive statistic is the simple prob-
ability or percentage. It can be either event-based or time-based. When
event-based, the simple probability or percentage tells us what proportion
of events were coded in a particular way, relative to the total number of
events coded. Initial procedures are the same as for rates. Events are
recorded (using any of the recording strategies described in chapter 3) and
the frequencies tallied. The only assumption required is that codes be mu-
tually exclusive and exhaustive. Then the number of events coded in a
particular way is divided by the total number of events coded. This gives
the simple probability for that particular kind of event. Or, the quotient can
be multiplied by 100, which gives the percentage for that particular kind of
event. For example, in the Bakeman and Brownlee study of parallel play,
12% of the events were coded Unoccupied, 18% Solitary, 26% Together,
24% Parallel, and 21% Group. (Recall that they used an interval-coding
strategy; thus a single event was defined as any contiguous intervals coded
the same way.)
When simple probabilities or percentages are time based, the interpre-
tation is somewhat different. These widely used statistics convey "time-
budget" information, that is, they indicate how the cases observed (animals,
infants, children, dyads, etc.) spent their time. The recording strategy must
preserve time information; thus simple event coding would not work, but
recording onset and offset times of events, or even coding intervals (re-
membering the approximate nature of the statistics estimated) would be
fine. Similarly, data would need to be represented as state, timed-event,
or interval sequences, not simple event-sequences. (Cross-classified event
data would also work if duration had been recorded, but frequently when
this recording and representation approach is used, the proportion of time
devoted to the event being cross-classified may not be recorded.) One final
note: When proportion of time coded in a particular way is of interest,
codes need not be mutually exclusive and exhaustive.
The Bakeman and Brownlee study of parallel play, just used as an exam-
ple of event-based probabilities, also provides an example of time-based
probabilities, or percentages as well. Using an interval-coding strategy,
they reported that 9% of the 15-second intervals were coded Unoccupied,
94 Analyzing sequential data: First steps

25% Solitary, 21% Together, 28% Parallel, and 17% Group. Thus, for ex-
ample, although 24% of the events were coded Parallel play, Parallel play
occupied 28% of the time. (These are estimates, of course. Recording on-
set times for behavioral state changes would have resulted in more accurate
time-budget information than the interval-recording strategy actually used;
see section 3.7.) A second example could be provided by the Adamson and
Bakeman study of affective displays. They recorded not just the occurrence
of affective displays, but their onset and offset times as well. Thus we were
able to compute that affective displays occurred, on the average, 4.4% of
the time during observation sessions. Put another way, the probability that
the infant would display affect in any given moment was .044.
Event-based (e.g., proportion of events coded Solitary) and time-based
(e.g., proportion of time coded Solitary) probabilities or percentages pro-
vide different and independent information; there is no necessary correla-
tion between the two. Which then should be reported? The answer is, it
depends. Whether one or both are reported, investigators should always
defend their choice, justifying the statistics reported in terms of the research
questions posed.

6.4 Mean event durations

Whenever time information is recorded, mean event durations can be re-
ported as well as (or instead of) proportions of total time devoted to par-
ticular kinds of events. In fact, mean event durations provide no new
information not already implied by the combination of rates and time-
based percentages. After all, mean event (or bout, or episode) durations
are computed by dividing the amount of time coded for a particular kind
of event by the number of times that event was coded. But in some cases,
mean event durations may be more useful descriptively than time-based
probabilities or percentages.
Because of the clear redundancy among these three descriptive statistics
(rates or frequencies, time-based probabilities or percentages, mean event
durations), we think investigators should report, or at least analyze, only
two of them. The question then is, which two? The answer will depend
on whatever an investigator thinks most useful descriptively, given the sort
of behavior coded. However, we suspect that when behavioral states are
being coded (see section 3.2), time-based percentages and mean durations
are more useful than rates, but that when the events being coded occur just
now and then and do not exhaustively segment the stream of behavior as
behavioral states do, then rates and mean event durations may prove more
useful than time-based percentages.
 Transitional probabilities: An introduction 95

For example, although in the previous section we computed the (time-

based) probability of an affective display from the Adamson and Bakeman
data, they in fact reported just rates and mean durations for affective dis-
plays. As with rates, mean durations were used as a score for subsequent
analysis of variance. Thus Adamson and Bakeman reported, not just that
the average length of an affective display was 2.5 seconds, but that the
length became shorter as the infants became older. This was an effect
Adamson and Bakeman had predicted on the basis of a hypothesized shift
in the function of affective displays during these infant ages.

6.5 Transitional probabilities: An introduction

The statistics discussed in the three preceding sections can be extremely
useful for describing aspects of sequential observational data (and all can
be computed with the GSEQ program), but they do not themselves convey
anything uniquely sequential. Perhaps the simplest descriptive statistic that
does capture a sequential aspect of such data is the transitional probability;
however before transitional probabilities can be described, some definitions
are in order.
A simple (or unconditional) probability is just the probability with which
a particular "target" event occurred, relative to a total set of events (or inter-
vals, if time based). For example, if there were 20 days with thunderstorms
last year, we could say that the probability of a thunderstorm occurring on
a particular day was .055 (or 20 divided by 365).
A conditional probability, on the other hand, is the probability with which
a particular "target" event occurred, relative to another "given" event. Thus
if it rained 76 days last year, and if on 20 of those 76 days there were thunder-
storms, then we would say that the probability of a thunderstorm occurring,
given that it was a rainy day, was .263 (or 20 divided by 76). If T stands
for thunderstorms and R for a rainy day, then the simple probability for
thunderstorms is usually written p(7"), whereas the conditional probability
for thunderstorms, given a rainy day, is usually written p(T\R); in words,
this is "the probability of T, given R."
A transitional probability is simply one kind of conditional probability.
It is distinguished from other conditional probabilities in that the target
and given events occur at different times. Often the word "lag" is used to
indicate this displacement in time. For example, if data are represented
as event sequences, then we might want to describe the probability, given
event A, of the target event B occurring immediately after (lag 1), occur-
ring after an intervening event (lag 2), etc. These event-based transitional
probabilities can be written p(B+i \AQ), p(B+2\AQ), etc.
96 Analyzing sequential data: First steps

Similarly, if data are represented as state, timed-event, or interval se-

quences, then we might want to describe the probability, given event A, of
the target event B occurring in the next time interval (often written t + 1,
where t stands for time), in the time interval after the next (t + 2), etc.
Such time-based transitional probabilities are often written p(Bt+i\At),
p(fl, +2 |A,),etc.
For example, imagine the following data (assume that each letter stands
for an event and that the coding scheme contains three mutually exclusive
and exhaustive codes):
BCAAABBCBCAC
In this case, it turns out that each code occurred four times. That is, f(A) =
f(B) = f(C) = 4 (where "f" stands for frequency), and p(A) = p(B) =
p(C) = 4/12 = .33, because 12 events were coded in all. The transitional
frequency matrix for these data is given in Figure 6.1. Note that the labeling
of rows and columns is somewhat arbitrary. We could just as well have
labeled rows lag—1 and columns lag 0. Either way, rows refer to events
occurring earlier, columns to events occurring later in time. This is the
usual convention (probably because we are a left-to-right reading society)
and one we recommend following.
Transitional frequency matrices are easy to construct. Each cell indicates
the number of times a particular transition occurred. For example, B
was followed by C three times in the sequence given in the preceding
paragraph; thus the cell formed by the Bth row and the Cth column in the
lag-one frequency matrix contains a "3." Symbolically, f(C+i |2?o) = 3, or
fBC = 3; this if often written XBC, letting x represent a score in general or,
less mnemonically but more conventionally, X23. Note that if N successive
events or intervals are coded, then there will be N — 1 lag-one transitions,
N — 2 lag-two transitions, etc. (The reader may want to verify the other
frequencies for the transitional frequency matrices given in Figure 6.1.)
To tally transitions, we use a "moving time-window." For the lag 1
transition matrix, we slide the two-event moving time-window along as
follows:
(BC)AAABBCBCAC
Then,
B(C A)AAB B C BC AC
Then,
BC(AA)ABBCBCAC
etc. The first position adds a tally to the XBC cell of the first table in
Figure 6.1. The second position of the window adds a tally to the XQA
cell, and so forth. Note that the consequent code takes its turn next as
an antecedent. This raises questions of independence of tallies, and may
 Transitional probabilities: An introduction 97

Data: B C A A A B B C B C A C

Lagl

A B C
A 2 1 1 4

LagO B 0 1 3 4

C 2 1 0 3

11

Data: A B A B C B A C B C A C

Lagl
A B C
A 0 2 2 4

LagO B 2 0 2 4

C 1 2 0 3

11

Figure 6.1. Examples of transitional frequency matrices and state transition dia-
grams.

matter for some subsequent inferential statistical procedures; this issue

is discussed further in section 8.1, although Bakeman and Dorval (1989)
believe it is not practically consequential.
Similarly, transitional probability matrices are easy to compute. For
example, for the data sequence given earlier, the lag 1 transitional prob-
ability matrix would be as given in Table 6.1. As the reader can see, a
98 Analyzing sequential data: First steps

Table 6.1. Transitional probability matrix for

the first data sequence given in Figure 6.1
Lagl

LagO A B C

A .50 .25 .25

B .00 .25 .75
C .67 .33 .00

transitional probability is just the frequency for a particular cell divided by

the frequency for that row. (A consequence of this definition is that the
transitional probabilities in each row sum to 1.) Symbolically,

For example, the probability of code C occurring, given that code B just
occurred, is p(C+i|2?o) — tBC = *BC + XB+ = 3/4 = .75. This means
that 75% of the time, code C followed code B.
Transitional probabilities are often presented graphically, as state transi-
tion diagrams (for examples, see Bakeman & Brown, 1977; Stern, 1974).
Such diagrams have the merit of rendering quite visible just how events (or
time intervals) were sequenced in time. Circles represent the codes, and
arrows represent the transitional probabilities among them. Examples are
given in Figure 6.1, and the reader may want to verify that they were drawn
and labeled correctly.
Figure 6.1 contains a second data sequence, along with its associated
transitional frequency matrix and state transition diagram. For both, the
simple probabilities are the same, that is, for both sequences each different
code occurred four times. The point of presenting these two examples
is to show that, even when simple probabilities indicate no differences,
events may nonetheless be sequenced quite differently. And when they
are, transitional probabilities and their associated state transition diagrams
can reveal those differences in a clear and informative way.
One final point: The discussion here treats transitional probabilities as
simple descriptive statistics, and state transition diagrams as simple de-
scriptive devices. Others (e.g., Kemeny, Snell, & Thompson, 1974) dis-
cuss transitional probabilities and state transition diagrams from a formal,
mathematical point of view, as parameters of models. Interesting questions
for them are, given certain models and model parameters, what sorts of out-
comes would be generated? This is a formal, not an empirical exercise, one
in which data are generated, not collected. For the scientist, on the other
 Summary 99

hand, the usual question is, first, can I accurately describe my data, and
second, does a particular model I would like to support generate data that
fit fairly closely with the data I actually got? The material in this chapter,
as noted in its first section, is concerned mainly with the first enterprise -
accurate description.

6.6 Summary
This "first steps" chapter discusses four very simple, very basic, but very
useful, statistics for describing sequential observational data: rates (or
frequencies), simple probabilities (or percentages), mean event durations,
and transitional probabilities.
Rates indicate how often a particular event of interest occurred. They
are probably most useful for relatively momentary and relatively infrequent
events like the affective displays Adamson and Bakeman (1985) described.
Simple probabilities indicate what proportion of all events were of a par-
ticular kind (event based) or what proportion of time was devoted to a
particular kind of event (time based). Time-based simple probabilities are
especially useful when the events being coded are conceptualized as behav-
ioral states (see section 3.2). Indeed, describing how much time individuals
(dyads, etc.) spent in various activities (or behavioral states) is a frequent
goal of observational research.
With the knowledge of how often a particular kind of event occurred,
and what percentage of time was devoted to it, it is possible to get a sense
of how long episodes of the event lasted. Mean event durations can be
computed directly, of course. But because these three statistics provide
redundant information, the investigator should choose whichever two are
most informative, given the behavior under investigation.
Finally, transitional probabilities capture sequential aspects of observa-
tional data in a simple and straightforward way. They can be presented
graphically in state transition diagrams. Such diagrams have the merit of
rendering visible just how events are sequenced in time.
 7
Analyzing event sequences

7.1 Describing particular sequences: Basic methods

Throughout this chapter, we assume that the reader is interested in event-
sequence data. That is, no matter how the data may have been recorded
and represented initially, we assume that it is possible to extract event
sequences (see section 5.2) from the data, and that the investigator has
good reasons for wanting to do so. This means that the data to be analyzed
are represented as sequences or chains of coded events (or behavioral states
but without time information) and that those events are defined in a way that
makes them mutually exclusive and exhaustive. Sometimes the chains will
be unbroken, collected all during one uninterrupted observation session.
Other times, several sequences may be pooled together for an analysis,
either because there were breaks in the observation session or because
observation sessions occurred at different times. Data from more than one
subject may even be pooled for some analyses (see section 8.4). In all
cases, the data to be analyzed consist of chains or sequences of codes.
The codes for event sequences are mutually exclusive and exhaustive,
as already noted. In addition, often the logic of the situation does not
permit consecutive codes to repeat. For example, when coders are asked to
segment the stream of behavior into behavioral states, it follows naturally
that two successive states cannot be coded the same say. If they were,
they would not be two states, after all, but just one. The restriction that
the same code may not follow itself in event sequences occurs relatively
often, especially when the events being coded are thought of as states. This
restriction affects how some statistics, especially expected frequencies, are
computed, as we discuss later.
As an example, let us again assume the coding scheme used in the
Bakeman and Brownlee study of parallel play. Behavioral states were
classified as one offivekinds: Unoccupied, Solitary, Together, Parallel, and
Group. When sequences were analyzed, adjacent codes were not allowed
to be identical. Thus there were 20 (or 5 x 4 ) different kinds of two-event
sequences (not 5 2 , which would be the case if adjacent codes could be the

100
 Determining significance of particular chains 101

same); 80 (or 5 x 42) different kinds of three-event sequences (not 53); 320
(or 5 x 43) different kinds of four-event sequences (not 54); etc.
Determining how often particular two-event, three-event, etc., sequences
occurred in one's data is what we mean by "basic methods." This involves
nothing more than counting. The investigator simply defines particular
sequences, or all possible sequences of some specified length, and then
tallies how often they appear in the data. For example, Bakeman and
Brownlee were particularly interested in transitions from Parallel to Group
play. For one child in their study, 127 two-event sequences were observed,
10 of which were from Parallel to Group. Thus they could report that, for
that child, f(PG) = 10 and p(PG) = .079 (10 divided by 127). (Note
that p(PG) is not a transitional probability. It is the simple or zero-order
probability for the two-event sequence, Parallel to Group.) In sum, the
most basic thing to do with event-sequence data is to define particular
sequences, count them, and then report frequencies and/or probabilities for
those sequences.

7.2 Determining significance of particular chains

We might now ask, how should these values be evaluated? One possibility
would be to compute expected frequencies, based on some model, for
particular observed frequencies, and then compare observed and expected
with a chi-square goodness-of-fit test. For example, there are 20 different
kinds of two-event sequences possible (U to 5, U to T9 U to P, U to G, S
to [/, S to T, etc.). Thus, we might argue, the expected probability for any
one kind is .05 (1/20), and so the expected frequency in this case is 6.35
(.05 x 127). What we are doing is assuming a particular model - in this
case, a "zero order" or "equiprobable" model, so called because it assumes
that the five codes occur with equal probability - and then comparing the
expected values the model generates for a particular sequence with those
actually observed. We note that the observed value for the Parallel to
Group sequence, 10, is greater than the expected value, 6.35. If we were
only concerned with the Parallel to Group sequence, we might categorize
all sequences as either Parallel to Group (10) or not (117), and compare
observed to expected using the familiar Pearson chi-square statistic,

2 _ ^ jobs - expf
exp
(10 - 6.35)2 (117 - 120.65)2
+
_
635 120.65
102 Analyzing event sequences

which, with one degree of freedom, is not significant (we use a Roman X
to represent the computed chi-square statistic to distinguish it from a Greek
chi, which represents the theoretical distribution).
Alternatively, we might make use of what we already know about how
often the five different codes occurred. This "first-order" model assumes
that codes occurred as often as they in fact did (and were not equiprobable),
but that the way codes were ordered was determined randomly. For the
child whose data we are examining, 143 behavioral states were coded; 34
were coded Parallel and 30 Group. (Because 127 two-event sequences
were tallied, and 143 states were coded, there must have been 15 breaks
in the sequence.) Now, if codes were indeed ordered randomly, then we
would expect that the probability for the joint event of Parallel followed
by Group would be equal to the simple probability for Parallel multiplied
by the simple probability for Group (this is just basic probability theory).
Symbolically,

p(PG)exp=p(P)xp(G)
The p(P) is .238 (34/143, the frequency for Parallel divided by the total, N).
In this case, however, the p(G) is not the f(G) divided by N. Because
a Parallel state cannot follow a Parallel state, the probability of group
(following Parallel) is computed by dividing the frequency for Group, not
by the total number of states coded, but by the number that could occur
after Parallel - that is, the total number of states coded, less the number of
Parallel codes. Symbolically.

P(G)=
- f(P)
when adjacent codes must be different and when we are interested in the
expected probability for Group following Parallel. Now we can compute
the expected probability for the joint event of a Parallel to Group transition.
It is:
f(P) f(G) 34 30
p(PG)exp = — x j r = W ) = — x ^ - ^ = .0654

The expected frequency, then, is 8.31 (.0654, the expected probability for
this particular two-event sequence, times 127, the number of two-event
sequences coded).
The chi-square statistic for this modified expected frequency is

2 (10-8.31)2 ( 1 1 7 - 118.69)2
X +
8.31
 Transitional probabilities revisited 103

which likewise is not statistically significant. Neither analysis suggests

that Group is any more likely to follow Parallel than an equiprobable or
first-order independence model would suggest.
The methods presented in this section are fairly limited. First, as se-
quences become longer, the number of possible sequences increases expo-
nentially. For example, with just five codes when consecutive codes cannot
repeat, there are 20 (5 x 4) two-event sequences, 80 (5 x 4 x 4) three-event
sequences, 320 ( 5 x 4 x 4 x 4 ) four-event sequences, etc. Consequently, ex-
pected probabilities for any one sequence may become vanishingly small,
requiring staggering amounts of data before expected frequencies become
large enough to evaluate with any confidence. Second, rarely are investiga-
tors interested in just one particular sequence such as the Parallel to Group
sequence used here as an example. More general methods, described in
subsequent sections and chapters, are required.

7.3 Transitional probabilities revisited

In the last section, we presented data derived from observing one child in
the Bakeman and Brownlee study of parallel play. We noted that her event-
sequence data contained 127 two-event sequences and that 10 of them rep-
resented transitions from Parallel to Group play. Thus we were able to say
that p(PG), the probability of a Parallel to Group sequence, was .0787, or
10 divided by 127. We then discussed ways of determining whether this ob-
served probability differed significantly from expected. We also noted that
p(PG) was a simple, not a transitional probability. In other words, p(PG)
is the probability for this particular sequence; if the probabilities for all 20
possible two-event sequences were summed, they would add up to one.
Occasionally it may be useful to describe probabilities for particular se-
quences, no matter whether chains are two-event, three-event, or longer.
When longer sequences are considered (e.g., 4 or 5 events long instead of
just 2), the number of possible sequences increases exponentially. When
there are many possible sequences, probabilities for particular sequences
can become almost vanishingly small and, as a result, less useful descrip-
tively. Thus usually attention focuses on transitional probabilities involving
two events. As discussed in section 6.5, these are usually symbolized t and,
unless noted otherwise, refer to lag 1.
For example, consider again the child in the parallel play study, (fre-
quencies and simple and transitional probabilities derived from her data
are given in Tables 7.1 through 7.3.) Considering just simple probabili-
ties, we note that the probability of a Parallel to Group sequence, p(PG),
was .0787, whereas the probability of a Parallel to Unoccupied sequence,
104 Analyzing event sequences

Table 7.1. Observed frequencies for two-event sequences

Given code, Target code, lag 1

lagO Un. Sol. Tog. Par. Gr. Totals

Unoccupied — 6 5 2 2 15
Solitary 5 — 6 7 5 23
Together 5 6 12 10 33
Parallel 2 7 11 10 30
Group 2 4 11 9 26
Totals 14 23 33 30 27 127

Table 7.2. Simple probabilities for two-event sequences

Given code, Target code, lag 1

lagO Un. Sol. Tog. Par. Gr.

Unoccupied — .0472 .0394 .0157 .0157

Solitary .0394 — .0472 .0551 .0394
Together .0394 .0472 .0945 .0787
Parallel .0157 .0551 .0866 .0787
Group .0157 .0315 .0866 .0709

Note: The tabled probabilities do not sum exactly to 1 because of

rounding.

p(PU), was .0157. This certainly conveys the information that Group was
more common after Parallel than Unoccupied. But somehow it seems both
clearer and descriptively more informative to say that the probability of
Group, given a previous Parallel, p(G|P) or tpc, was .333, whereas the
probability of Unoccupied, given a previous Parallel, p(U\P) or tpjj, was
.067. Immediately we know that 33.3% of the events after Parallel were
Group, whereas only 6.7% were Unoccupied.
We just considered transitions from the same behavioral state (Parallel)
to different successor states (Group and Unoccupied), but the descriptive
value of transitional probabilities is portrayed even more dramatically when
transitions from different behavior states to the same successor state are
compared. For example, the simple probabilities for the Unoccupied to
Solitary, p(C/5), and for the Together to Solitary, p(TS), transitions are
both .0472. Yet the probability of Solitary, given a previous Unoccupied,
p(S\U) or tus, is .400, whereas the probability of Solitary, given a previous
Together, p(S\T) or tjs, is .182. The transitional probabilities "correct"
 Transitional probabilities revisited 105

Table 7.3. Transitional probabilities for two-event sequences

Given code, Target code, lag 1

lagO Un. Sol. Tog. Par. Gr.

Unoccupied .400 .333 .133 .133

Solitary .217 — .261 .304 .217
Together .152 .182 — .364 .303
Parallel .067 .233 .367 — .333
Group .077 .154 .423 .346 —

Note: Rows may not add to 1 because of rounding.

for differences in base rates for the "given" behavioral states and, there-
fore, clearly reveal that in this case, Solitary was relatively common after
Unoccupied, and considerably less so after Together, even though the Un-
occupied to Solitary and the Together to Solitary transitions appeared the
same number of times in the data.
Moreover, as already noted in section 6.5, transitional probabilities form
the basis for state transition diagrams, which, at least on the descriptive
level, are a particularly clear and graphic way to summarize sequential
information. The only problem is that, even with as few as five states, the
number of possible arrows in the diagram can produce far more confusion
than clarity. The solution is to limit the number of transitions depicted
in some way. In this case, for example, we could decide to depict only
transitional probabilities that are .3 or greater, which is what we have done
in Figure 7.1. This reduces the number of arrows in the diagram from a
possible 20, if all transitions were depicted, to a more manageable 9.
The nine transitions shown in Figure 7.1 are not necessarily the most
frequent transitions; this information is provided by the simple probabilities
for two-event sequences (see Table 7.2). Nor are the transitions necessarily
significantly different from expected; to determine this we would need to
compute and evaluate a z score for each transition (see next section). What
the state transition diagram does show are the most likely transitions, taking
the base rate for previous states into account. In other words, it shows the
most likely ways of "moving" from one state to another. For this one child,
Figure 7.1 suggests frequent movement from Unoccupied to both Solitary
and Together, from Solitary to Parallel, and reciprocal movement among
Together, Parallel, and Group.
One final point: Transitional probabilities can be used to describe re-
lationships between two nonadjacent events as well. Not only can we
compute, for example, the probability of Group in the lag 1 positions given
an immediately previous Parallel: p(G+i | PQ), but we can also compute the
106 Analyzing event sequences

Figure 7.1. A state transition diagram. Only transitional probabilities greater than
0.3 are shown. U = unoccupied, S = solitary, T = together, P — parallel, and
G = group.

probability of Group in the lag 2 position: p(G+2\Po\ the lag 3 position,

etc., with as many events intervening between the "given" and the "target"
code as desired. The ability of transitional probabilities to describe such
lagged relationships, in fact, forms the basis for the lag sequential method
described in section 7.5.
Transitional probabilities, although useful descriptively, also have lim-
itations. In particular, when several "subjects" take part in a study, and
when transitional probabilities are computed separately for each subject,
those transitional probabilities should not, under most circumstances, be
used as scores for testing for individual or group differences.
The reason is as follows: A transitional probability is valuable descrip-
tively, to be sure, and certainly reflects the "experience" of a particular
subject. For example, the experience of the child whose data are given in
Table 7.3 was that one-third of the time after Parallel, Group followed. This
may be important, but it may or may not be "significant." It depends on
how probable Group was for the child. If Group was especially probable
for one child, then a score of .33 for the transitional probability might not
be very high; it could even be less than expected. On the other hand, if
Group was unlikely for another child, then a score of .33 might be quite
high, considerably above expected.
 Transitional probabilities revisited 107

The major problem, then, with using transitional probabilities when

analyzing for individual or group differences, is that similar numeric values
may have quite different meanings for different subjects, or for the same
subject at different times, rendering the results difficult to interpret at best.
A secondary problem is that the values of transitional probabilities are
"contaminated" with the values for the simple probabilities, which means
that what appear to be analyses of transitional probabilities may in fact
reflect little more than simple probability effects.
An example may help to clarify this. Imagine that we have two kinds
of children, farm kids and city kids, and that the mean value for the simple
probability of Group computed for the city kids is significantly higher than
the mean value computed for farm kids. If we then turn around and test
the transitional probability, p(G+i | Po)> we would expect that mean values
for city and farm kids would differ significantly as well. After all, the
expected value for this transitional probability is directly related to the
simple probability for Group. The more often the Group code appears
in the data, the more often we would expect it to follow the Parallel code
as well.
Symbolically, what we are saying is that the expected value for p(T\G)
- where T stands for "target" and G for "given" - is directly related to
p(r), the probability of the target code. It is not necessary, of course, that
observed agree with their expected values. Still, when analyses of both
p(T\G) and p(T) reveal significant group differences, it seems unjustified
to us to claim that the differences with respect to p(T\G) are explained by
anything more complex than the differences detected for p(T). In sum,
for both reasons given above, analyses that use transitional probabilities as
scores rarely provide much insight into sequential aspects of the data.
Almost always, when individual or group differences are at issue, the
appropriate score to use is not the transitional probability, but some in-
dex of the strength of the effect, like Yule's Q (as discussed in section
7.7). Analyzing such scores, we would be able to determine, for example,
whether the Parallel to Group sequence was significantly more character-
istic of city kids, on the average, than of farm kids, or whether the extent
to which Group tended to follow Parallel was a stable characteristic of
children measured at two different ages.
In the previous edition of this book, we suggested that z scores (see next
section) be used as an individual index for such analyses, but the magnitude
of z scores is affected by the number of tallies. For an effect of a specific
size, the z score becomes larger as the number of tallies increases. This
is an example of the well-known fact that power increases with sample
size. But it also makes the z score an inappropriate choice for analyzing
individual differences. Unless the number of tallies is the same for all
108 Analyzing event sequences

subjects (dyads, etc.) analyzed, z scores conflate how strongly a particular

transitional probability deviates from its expected value with the number
of tallies available for the analysis.

7.4 Computing z scores and testing significance

Once lagged events have been tallied in a table like Table 7.1, where rows
represent lag 0 and columns lag 1, and after we have examined transitional
probabilities like those in Table 7.3, next we usually want to identify which
transitional probabilities deviate significantly from their expected values.
Commonly, a z score of some sort has been used for this purpose. Assuming
that a computed score is distributed normally, then z scores larger than 1.96
absolute are often regarded as statistically significant at the .05 level. But
there are a variety of ways to compute z scores, and simply calling one a z
score hardly guarantees that it will be normally distributed.
Assume that the behavioral event of interest is called the "target" event,
7\ and that we want to relate that to another event, called the "given" event,
G. In other words, we are interested in p(7\ \G$) or tor, the probability of
the target even occurring after the given event.
A z score compares observed to expected, so the first task is to compute
the expected value for XGT •> the observed value for the transition from given
to target behavior. When consecutive codes may repeat (and so the upper-
left to lower-right diagonal tallies would not all be zero as in Table 7.1),
expected values are computed using the familiar formula,
X+T XG+X+T
mGT = x
*G+ = x
(7.1)

where mcr is an estimate of the expected frequency (m because often

expected values are means), XG+ is the sum of the observed frequencies in
the Gth or given row, X+T is the sum of the observed frequencies in the Tth
or target column, and x++ is the total number of tallies in the table (also
symbolized as N2 or the number of two-event chains tallied, as compared
to N\9 the number of single events coded). This formula yields expected
values assuming independence, that is, no association between the rows
and columns of the table.
However, when consecutive codes cannot repeat, resulting in what are
called structural zeros on the diagonal (structural because logical definition
and not data collection resulted in their being zero), expected frequencies
cannot be computed with a simple formula but require an iterative proce-
dure, best performed with a computer. Two of the most widely used are iter-
ative proportional fitting (IPF, also called the Deming-Stephan algorithm)
 Computing z scores and testing significance 109

Table 7.4. Expected frequencies for two-event sequences

Given code, Target code, lag 1

lagO Un. Sol. Tog. Par. Gr.

Unoccupied 2.82 4.69 4.04 3.45

Solitary 2.65 — 7.83 6.75 5.76
Together 4.40 7.83 — 11.21 9.56
Parallel 3.80 6.75 11.21 — 8.24
Group 3.14 5.59 9.27 8.00 —

and the Newton-Raphson algorithm; for descriptions see Bishop, Fienberg,

& Holland (1975) and Fienberg (1980). Both methods yields identical val-
ues. Expected frequencies for the data shown in Table 7.1 are given in
Table 7.4.
When consecutive codes may repeat, z scores are computed as follows:
XGT - mGT
ZGT = / n V1 = (7.2)
(l )(1 - P+T) n
where pc+ is XQ+ -^*++ and p+r is *+r -i~*++- I the log-linear literature,
this is called an adjusted residual (Haberman, 1978, p. 111). When consec-
utive codes cannot repeat, matters are more complex. The formula (actually
several formulas, many of which are used by the Newton-Raphson algo-
rithm) is given in Haberman (1979, p. 454; but see footnote 1 in Bakeman
& Quera, 1995b); a number of relatively complex matrix operations are
involved. The most practical way for you to compute these values, given
structural zeros, is to use a computer program like GSEQ or a general-
purpose log-linear program. Adjusted residuals, computed with the SPSS
for Windows General Log-Linear routine, are given in Table 7.5.
If you are new to sequential analysis, you may want to skip the fol-
lowing paragraphs, which are included largely for historical purposes and
for readers who wish to reconcile the preceding paragraphs with the first
edition of this book and with earlier literature. Early on, Sackett (1979)
suggested that z be computed as follows:
XGT - mGT
ZGT = , n == (7.3)
V ^ ( l - pT)
where
mGT = p ( D x f(G) = ^-xG = ^ (7.4)

which is almost but not quite the same as Equation 7.1 because it is based
110 Analyzing event sequences

Table 7.5. Adjusted residuals for two-event sequences

Given code, Target code, lag 1

lagO Un. Sol. Tog. Par. Gr.

Unoccupied 2.25 0.19 -1.29 -0.96

Solitary 1.71 — -0.94 0.13 -0.42
Together 0.37 -0.94 — 0.37 0.22
Parallel -1.17 0.13 -0.10 — 0.88
Group -0.79 -0.88 0.85 0.51 —

Note: Row and column totals may not add exactly to those shown in
Table 7.1 because of rounding.

on simple frequencies and probabilities for given and target behaviors, not
on values from two-dimensional tables as is Equation 7.1. Equation 7.3 is
based on the normal approximation for the binomial test,

(7.5)

where N is f(G) or JCG, P is p(T) or xj -T- N\ (also symbolized as

and Q is 1 — p?. Almost immediately, however, Allison and Liker (1982)
objected, noting that Equation 7.5 would only be appropriate if pj were
derived theoretically instead of from the data at hand. They wrote that
Equation 7.3 should be
XGT - mGT
ZGT = , n V1 , (7.6)
> c r ( l - PG)(1 - PT)
instead, which is almost but not quite the same as Equation 7.2 because, like
Equation 7.3, it is based on single occurrences and not two-event chains.
We prefer Equation 7.2 to 7.6 because it seems more grounded in a well-
developed statistical literature, that dealing with log-linear models (e.g.,
see Bishop, Fienberg, & Holland, 1975; Fienberg, 1980; Wickens, 1989),
and because it is based on two-event chains, which seems more faithful
to the situation at hand. True, if a sequence of N\ consecutive events is
tallied using overlapped sampling (i.e., tallying first the e\e2 chain, then
^2^3, ^3^4, and so forth, where e stands for an event), so that N\ — l chains
are tallied, then XQ and JCG+> for example, will differ by at most 1. But
overlapped sampling, while common, is not always used; moreover, often
breaks occur in sequences and then the number of two-event chains tallied
is N\ — 5, where S is the number of separate segments coded. In such
cases, xG and *G+ could differ by quite a bit.
 Classic lag sequential methods 111

Moreover, the log-linear tradition, from which Equation 7.2 is derived,

offers a statistically based solution when consecutive codes cannot repeat
(see Bakeman & Quera, 1995b). Sackett (1979), recognizing that Equa-
tion 7.1 would not compute expected frequencies correctly when structural
zeros occupied the diagonal, suggested
XT *G*T (n -.
— xG = — . (7.7)
iVi —XG N\ — xG
He reasoned that when consecutive events cannot repeat, the expected
probability for the target code at lag 1 (assuming independence) is the
frequency for that code divided by the number of events that may occur at
lag 1, which is N\ minus the frequency for the given code. Thus expected
frequencies on the diagonal are set to zero and off-diagonal ones are the
frequency for the given code times this probability, as indicated by Equation
7.7. However, expected frequencies, when summed across rows and down
columns, should equal the observed row and column totals, which expected
frequencies computed per Equation 7.7 do not (Bakeman & Quera, 1995b),
whereas expected frequencies computed with an iterative procedure do (see
Table 7.4). Thus, as mentioned earlier, when consecutive codes cannot
repeat, we would compute adjusted residuals using log-linear methods
(and an appropriate computer program), not Equations 7.6 and 7.7.

7.5 Classic lag sequential methods

So far in this chapter, much of our discussion and most of our examples
have been confined to two-event sequences. We have mentioned how
longer sequences can be described and tested for significance (mainly in
sections 7.1 and 7.2), but we have also noted that such tests may require
prohibitive amounts of data. The approaches already discussed are "ab-
solute" in the sense that they define particular sequences and then tally
how often each occurred. As we attempt to investigate longer and longer
sequences, the expected frequencies for particular sequences become van-
ishingly small, the number of possible sequences increases at a staggering
rate, and it becomes almost impossible to make sense out of the wealth
of information produced about so many different sequences. Clearly, a
less absolute, more probabilistic and more flexible approach to the inves-
tigation of sequences comprising more than two events would be useful.
One such approach is usually called the "lag sequential method." It was
first developed by Sackett (1974, 1979, 1980) and later described by oth-
ers (Bakeman & Dabbs, 1976; Bakeman, 1978; Gottman & Bakeman,
1979).
112 Analyzing event sequences

The reader of this book will already be familiar with the basic elements
of the lag sequential method. As an example, assume that our code catalog
defines several events, five of which are:
1. Infant Active
2. Mother Touch
3. Mother Nurse
4. Mother Groom
5. Infant Explore
(These codes are suggested by Sackett's work with macaque monkeys. The
example here is based on one given in Sackett, 1974.) Assume further that
successive events have been coded so that, as throughout this chapter, we
are analyzing event-sequence data. Finally, assume that we are particularly
interested in what happens after times when the infant is active, that is, we
want to know whether there is anything systematic about the sequencing
of events beginning with Infant Active episodes.
To begin with, the investigator selects one code to serve as the "criterion"
or "given" event. In this case, that code would be Infant Active. Next, an-
other code is selected as the "target." For example, we might select Mother
Touch as our first target code. Then, a series of transitional probabilities are
computed: for the target immediately after the criterion (lag 1), after one
intervening event (lag 2), after two intervening events (lag 3), etc. Symbol-
ically, we would write these lagged transitional probabilities as p(T\ |Go),
p(?21 Go), p(?31 Go), etc. (remember, if we just write p(T | G), target at lag
1 and given at lag 0 are assumed). The result is a series of transitional
probabilities, each of which can then be tested for significance. For exam-
ple, given Infant Active at lag "position" 0, if we had computed transitional
probabilities for Mother Touch at lags 1 through 6, but only the lag 1 tran-
sitional probability significantly exceeded its expected value, we would
conclude that Mother Touch was likely to occur just after Infant Active,
but was not especially likely in the other lag "positions" investigated.
If we stopped now, we would have examined transitional probabilities for
one particular target code, at different lags after a particular criterion code.
This is not likely to tell us much about multievent sequences. The next step
is to compute other series of transitional probabilities (and determine their
significance), using the same criterion code but selecting different target
codes. For example, given a criterion of Infant Active at lag 0, we could
compute lag 1 through 6 transitional probabilities for Mother Nurse, Mother
Groom, and Infant Explore. Imagine that the transitional probabilities for
Mother Nurse at lag 2, for Mother Groom at lag 3, and for Infant Explore
at lags 4 and 5 were significant. Such a pattern of results could suggest the
four-event sequence: Infant Active, Mother Touch, Mother Nurse, Mother
 Classic lag sequential methods 113

Table 7.6. Results required to confirm the Active, Touch, Nurse,

Groom sequence
Lag

Criterion Target 1 2 3 4 5

Infant Active Mother Touch p* P P P P

Mother Nurse P p* P P P
Mother Groom P P P* P P
Mother Touch Mother Nurse P* P P P P
Mother Groom P P* P P P
Mother Nurse Mother Groom P* P P P P

Note: Asterisks indicate transitional probabilities whose values significantly ex-

ceed expected. Numerical values for transitional probabilities have not been
given for this hypothetical example.

Groom (we shall return to Infant Explore in a moment), even though the
lagged transitional probabilities examined only two codes at a time.
As stated before, the lag sequential is a probabilistic, not an absolute
approach. To confirm the putative Active, Touch, Nurse, Groom sequence,
we should do the following. First, compute lagged transitional probabilities
with Mother Touch as the criterion and Mother Nurse and Mother Groom
as targets, then with Mother Nurse as the criterion and Mother Groom as
the target. If the transitional probabilities for Mother Nurse at lag 1 and
Mother Groom at lag 2, with Mother Touch as the lag 0 criterion, and
for Mother Groom at lag 1 with Mother Nurse as the lag 0 criterion, are
all significant, then we would certainly be justified in claiming that the
Active, Touch, Nurse, Groom sequence was especially characteristic of the
monkeys observed (see Table 7.6).
Recall, however, that Infant Explore was significant at lags 4 and 5 after
Infant Active. Does this mean that we are dealing with a six-event instead
of a four-event sequence? The answer is, not necessarily. For example,
if the transitional probabilities for Infant Explore at lags 3 and 4 given
Mother Touch as the criterion, at lags 2 and 3 given Mother Nurse, and at
lags 1 and 2 given Mother Groom were not significant, then there would
be no reasons to claim that Infant Explore followed the Active, Touch,
Nurse, Groom sequence already identified. Instead, if the results were as
suggested here, we would conclude that after a time when the infant was
active, next we would likely see either the Touch, Nurse, Groom sequence
or else three more or less random events followed by Infant Explore in the
fourth or fifth position.
114 Analyzing event sequences

If we let X stand for a "random" event, then in effect we have detected

the following sequence: Active, X, X, X, Explore, Explore. More accu-
rately, because in this example adjacent codes must be different, we have
detected the following two sequences: Active, X, X, X, Explore, X and
Active, X, X, X, X, Explore. Given other data, we might have detected
a sequence like Active, X, Nurse, Groom, which we would interpret as
follows: Whatever happens after times when the infant is active is not
systematic - it could be almost any code, randomly chosen. After a ran-
dom event in lag 1, however, the Nurse, Groom sequence is likely (in lag
positions 2 and 3). Such a sequence would not be easily detected with
"absolute" methods. One advantage, then, of the lag sequential approach
is the ease with which sequences containing random elements can still be
detected. The main advantage of this approach, however, remains its abil-
ity to detect sequences involving more than two events without requiring
as much data as absolute methods would. When interpreting lag sequential
results, a number of cautions apply. First, it is important to keep in mind
whether adjacent codes can be the same or not, because this affects how ex-
pected frequencies or probabilities are computed. In the previous section,
we noted that when consecutive codes cannot repeat, expected frequencies
for lag 1 are best computed using an iterative procedure, although Equation
7.7 from the lag sequential literature provides an approximation. A similar
approximation is suggested by Sackett (1979) for lags greater than 1. It is

(7.8)
-xG
where mcr represents the expected frequency for the target behavior at
lag L when preceded by the given behavior at lag 0, and XGT the observed
frequency for the target at lag L — 1 preceded by the given behavior at lag 0.
Sackett (1979) reasoned that when adjacent codes cannot repeat, the
expected probability for a particular target code at lag L (assuming a par-
ticular given code at lag 0) is the frequency for that target code diminished
by the number of times it appears in the lag L — 1 position (because then
it could not appear in the L position, after itself) divided by the number of
events that may occur at lag L (which is the sum of the lag L minus the lag
L — 1 frequencies summed across all K target codes). Simply put, this sum
is the number of all events less the number of events assigned the given
code. As with Equation 7.7, Equation 7.8 assumes overlapped sampling;
and again like Equation 7.7, marginals for expected frequencies based on
Equation 7.8 do not match the observed marginals.
Nonetheless, when consecutive codes cannot repeat, traditional lag
sequential analysis (Sackett, 1979; the first edition of this book) has
 Classic lag sequential methods 115

estimated expected frequencies at lag 1 with Equation 7.7 and at longer

lags with Equation 7.8, and then has determined statistical significance
based on the z computed per Allison and Likers's (1982) Equation 7.6. As
already noted, at lag 1 we recommend the log-linear methods described in
the previous section, and in the next section we develop log-linear methods
that apply at longer lags. But note, when consecutive codes may repeat,
and when overlapped sampling is used, at lag 1 traditional lag sequential
(Equations 7.4 and 7.6) and log-liner (Equations 7.1 and 7.2) analyses pro-
duce almost identical results. The same is essentially true at longer lags,
although then log-linear analyses offers certain additional advantages, as
described in the next section.
Two additional cautions should be mentioned. As is always true, no
matter the statistic, before significance is assigned to any z score, the in-
vestigator should determine that there are sufficient data to justify this (see
section 8.5). Finally, as always, the investigator should keep in mind the
type I error problem (see section 8.6).
This is probably not a serious problem if sequences beginning with just
one, or at most a few, criterion codes are investigated in the context of a
confirmatory study. However, if many codes are defined, and if all serve
exhaustively as criteria and targets, with many lags, then interpretation of
such exploratory results should be guided by the almost certain knowledge
that some chance findings are contained therein.
We end this section with a second example of the lag sequential method.
For a study of marital communication Gottman, Markman, and Notarius
(1977) coded the sequential utterances of a number of nondistressed and
distressed couples observed discussing marital problems. Among other
questions, these investigators wanted to know what happened after the
husband complained about a marital problem. In other words, given Hus-
band Complaint as the criterion code, they computed lagged transitional
probabilities for a number of target codes, including Wife Agreement, Wife
Complaint, Husband Agreement, and Husband Complaint. (The same code
may serve as both criterion and target; however, when adjacent codes must
be different, both its observed and expected frequencies at lag 1 will be
0. This is an example of a "structural" zero.) Although 24 codes were
defined, the four just listed always included the highest z scores.
An interesting difference was noted between nondistressed and dis-
tressed couples. For nondistressed couples, significant z scores occurred
only when Wife Agreement (at lags 1,3, and 5) and Husband Complaint (at
lags 2 and 4) served as targets. The process of cycling between Husband
Complaint and Wife Agreement, Gottman et al. called "validation." For
distressed couples, on the other hand, significant z scores occurred only
when Wife Complaint (at lags 1 and 3) and Husband Complaint (at lags 2,
116 Analyzing event sequences

Figure 7.2. A lagged probability profile for nondistressed couples. Triangles

represent transitional probabilities for Husband Complaint at the lag specified,
given Husband Complaint at lag 0. Circles represent transitional probabilities for
Wife Agreement at the lag specified, given Husband Complaint at lag 0. Asterisks
(*) indicate that the corresponding z score is significant.

4, and 6) served as targets, a process that Gottman et al. termed "cross-

complaining." Lagged probability profiles for these results are presented
in Figures 7.2 and 7.3.

7.6 Log-linear approaches to lag-sequential analysis

Since lag-sequential analysis was first developed, log-linear analyses have
become more widely understood and used by social scientists (Bakeman
& Robinson, 1994; Wickens, 1989). They offer a number of advantages
over traditional lag-sequential methods. As already noted, structural zeros
are handled routinely and do not require the ad hoc formulas described in
the previous section. But primarily, use of log-linear methods allows inte-
gration of sequential analysis into an established and well-supported sta-
 Log-linear approaches to lag-sequential analysis 117

3 4
Lag Position

Figure 7.3. A lagged probability profile for distressed couples. As in Figure 7.2,
triangles represent transitional probabilities for Husband Complaint at the lag
specified, given Husband Complaint at lag 0. Squares, however, represent transi-
tional probabilities for Wife Complaint, given Husband Complaint at lag 0. Again
asterisks (*) indicate that the corresponding z score is significant.

tistical tradition (Bakeman & Quera, 1995b). As Castellan (1979) pointed

out some time ago, almost always sequential questions can be phrased in
terms of multidimensional contingency tables, which log-linear analysis
was developed to analyze.
In this section, we describe three advantages of a log-linear view to
sequential questions. First, log-linear analysis promotes a whole-table
view, whereas often traditional lag-sequential analysis focused, almost
piecemeal, on individual transitions in a table, which invites type I er-
ror. Moreover, log-linear analysis provides ways of disentangling the web
of connected results in a table, as we demonstrate shortly. Finally, log-
linear analysis, using well-established techniques, provides an integrated
method for determining whether there are effects at various lags, no matter
whether consecutive codes may or cannot repeat and no matter whether or
not overlapped sampling was employed.
118 Analyzing event sequences

Omnibus tests
Used in an exploratory way, traditional lag-sequential analysis invites type
I error (although among statistical techniques it is hardly unique in this
respect). When 10 codes are defined, for example, the lag 0 by lag 1 table
contains 100 cells when consecutive codes may repeat and 90 when not
(K2 and K[K — 1] generally, where K is the number of codes defined).
Assuming the usual .05 value for alpha, if there were no association be-
tween lag 0 and lag 1 behavior, approximately 5 of 100 transitions would
be identified, on average and incorrectly, as statistically significant. One
solution is to take an omnibus or whole-table view. Absent specific pre-
dictions that one or just a few transitions will be significant, individual cell
statistics should be examined for significance only when a tablewise statis-
tic, such as the Pearson or likelihood-ratio chi-square (symbolized as X2
and G 2 , respectively), is large, just as post hoc tests are pursued in analysis
of variance only when the omnibus F ratio is significant (Bakeman, 1992;
Bakeman & Quera, 1995b).
Applying this test, we would not have examined the data presented in
Table 7.1 further. For these data,

X\U, N = 127) = f : y ( X G r m G r )
= H.O (7.9)
m
V T
and
K K
G2(ll,N = 127) = 2 y V x G r l o g - ^ - = 10.3 (7.10)

where log represents the natural logarithm (i.e., the logarithm to the base
e)\ X2 and G 2 both estimate chi-square, although usually G 2 is used in
log-linear analyses (for a discussion of the differences between them, see
Wickens, 1989). These estimates fall short of the .05 critical value of
19.7. Moreover, only 1 of 20 adjusted residuals exceeded 1.96 absolute
(Unoccupied to Solitary; see Table 7.5), which suggests it was simply a
chance finding, unlikely to replicate.
In log-linear terms, we ask whether expected frequencies generated by
the model of independence (i.e., Equation 7.1), which is symbolized [0][l]
and indicates the independence of the lag 0 and lag 1 dimension, are similar
to the observed frequencies. If they are, then the chi-square statistic will
not exceed its .05 critical value, as here (i.e., observed frequencies fit those
expected tolerably well). However, if the computed chi-square statistic is
large, exceeding its .05 critical value, then we reject the model of indepen-
dence and conclude that the dimensions of the table are in fact related and
not independent.
 Log-linear approaches to lag-sequential analysis 119

Table 7.7. Observed frequencies and adjusted residuals for 100

two-event sequences
Observed frequencies: Adjusted residuals:
Given code, Target code, lag 1 Target code, lag 1

lagO A B C Totals A B c
A 23 5 15 43 2.02 -1.82 -.056
B 11 1 7 19 1.56 -1.78 -0.12
C 8 14 16 38 -3.32 3.30 0.66
Totals 42 20 38 100

Note: This example was also used in Bakeman and Quera (1995b).

Winnowing results
As a second example, consider the data given in Table 7.7 for which K is
3; the codes are labeled A, B, and C; and consecutive codes may repeat.
For these data, X 2 (4, N = 100) is 15.7 and G2(4, N = 100) is 16.4. Both
exceed the .05 critical value of 9.49, which suggests that lag 0 and lag
1 are associated. Moreover, three of nine adjusted residuals exceed 1.96
absolute (again, see Table 7.7). But now we confront a different dilemma.
Adjusted residuals in a table form an interrelated web. If some are large,
others necessarily must be small, and so, rather than attempting to interpret
each one (thereby courting type I error), now we need to determine which
one or ones should be emphasized.
The initial set of all statistically significant transitions in a table can be
winnowed using methods for incomplete tables (i.e., tables with structural
zeros). Assume the C-B chain, whose adjusted residual is 3.30, is of
primary theoretical interest. In order to test its importance, we declare the
C-B cell structurally zero, use an iterative procedure to compute expected
frequencies (e.g., using Bakeman & Robinson's, 1994, ILOG program),
and note that now the [0][l] model fits the remaining data (G 2 [3, N =
86] = 5.79; .05 critical value for 3 df= 7.81; df= 3 because one is lost to the
structurally zero cell). We conclude that interpretation should emphasize
the C-B chain as the other two effects (decreased occurrences for C-A,
increased occurrences for A-A) disappear when the C-B chain is removed.
Had the model of independence not fit the reduced data, we would have
declared another structural zero and tested the data now reduced by two
cells. Proceeding step wise (but letting theoretical considerations not raw
empiricism determine the next chain to delete, else one risks capitalizing on
chance as with backward elimination in multiple regression and compro-
120 Analyzing event sequences

mising type I error control), we would identify those chains that prevent
the [0][l] model from fitting. (A logically similar suggestion, not in a
log-linear context, is made by Rechten & Fernald, 1978; see also Wick-
ens, 1989, pp. 251-253.) The ability to winnow results in this way is one
advantage of the log-linear view over traditional lag-sequential analysis.

Sequential log-linear models

Perhaps the major advantages of a log-linear approach to sequential prob-
lems are the statistical foundation and generality provided. Assuming an
interest in lags no greater than L, we begin by assembling (L + l)-event
chains. For example, if we were interested in lags no greater than 2, we
would collect three-event chains in which the first event was associated
with lag 0, the second with lag 1, and the third with lag 2.
The three-event chains might be derived from just one or a few longer
sequences using overlapped sampling, selecting first e\e2e?>, then £ 2 ^ 4 ,
then e^e^es, etc., where e\ represents an event in the longer sequence.
In such cases, if S segments were coded comprising N events in all, the
number of three-event chains derived would be N — SL (assuming none
of the segments consisted of fewer than L + 1 events). For example, in
the simplest instance, if one segment consisting of N events were coded,
N — 2 three-event chains would be tallied. Alternatively, the three-event
chains might be derived from a few longer sequences using nonoverlapped
sampling (selecting first ^^2^3, then £41516> etc.), in which case the number
of three-event chains derived would be N -=- 3 (assuming all segments are
multiples of three). Or the three-event sequences might be sampled directly
from a population of three-event sequences.
Overlapped sampling is often used, and usually assumed, in traditional
lag-sequential analysis. But nonoverlapped sampling is both quite common
and useful. For example, imagine that we are only interested in tallying
speaker 1 (e.g., husband) to speaker 2 (e.g., wife) lag 1 transitions. Then,
assuming two-event chains were derived from a single segment of N coded
events, the number of speaker 1 to 2 transitions would be N -7- 2 if the
speakers always alternated turns (tallyingfirste\ei, then £3^4, etc.), or some
smaller number if a speaker's turn may contain more than one thought unit
(which could result in tallying ^ 3 , e^ee, eje%, e\\e\i, etc.; see Bakeman
& Casey, 1995).
No matter the sampling strategy, the (L + l)-event chains are tallied
in a KL+l table; thus each chain adds a tally to one, and only one, cell.
To demonstrate the log-linear approach, let us begin with an example rep-
resenting the simplest of circumstances, assuming codes of A, B, and C
 Log-linear approaches to lag-sequential analysis 121

1: A B C
0:A 21 25 49
B 23 26 21
C 50 19 15

Figure 7.4. Observed frequencies for 249 two-event chains derived from a se-
quence of 250 events.

(thus K = 3) that may repeat and an initial interest in lag 1. Then occur-
rences of each of the nine possible two-event chains (AA, AB, etc.) would
be tallied in one of the cells of a 3 2 table. For example, we (Bakeman &
Quera, 1995b) generated a sequence of 250 coded events and tallied the 249
overlapped two-event chains; the results are shown in Figure 7.4. For this
two-dimensional table, the [0][l] model (implying independence of rows
and columns) fails to fit the data (G2[4, N=249]=35.2) and so we would
conclude that events at lag 0 and lag 1 are associated and not independent.
This much seems easy and, apart from the preliminary omnibus test, not
much different from traditional lag-sequential methods.
Next, assume that our interest expands from lag 1 to lag 2. Still assuming
K = 3 and consecutive codes that may repeat, then each of the 27 possible
3-event chains (AAA, AAB, etc.) would be tallied in one of the cells of
a 3 3 table. Tallies for the 248 overlapped three-event chains derived from
the same sequence used earlier are shown in Figure 7.5.
Cells for the 3 3 table shown in Figure 7.5 are symbolized *//*, where i9j,
and k represent the lag 0, lag 1, and lag 2 dimensions, respectively. Tradi-
tional lag-sequential analysis would test for lag 2 effects in the collapsed
02 table, that is, the table whose elements are JC;+&, where

This table is shown in Figure 7.6. For this two-dimensional table, the [0][2]
model (implying independence of rows and columns) fails to fit the data
(G 2 [4, TV = 248] = 10.70) and so, traditionally, we would conclude that
events at lag 0 and lag 2 are associated and not independent. But this fails
to take into account events at lag 1.
A hierarchic log-linear analysis of the 3 3 table shown in Figure 7.5
provides more information, and in this case leads to a different conclusion,
than a traditional lag-sequential analysis of the collapsed table shown in
Figure 7.6. The complete or saturated model for a three-dimensional table
122 Analyzing event sequences

2: A B
0:A A 7 6 8
B 10 9 6
C 31 8 10

B A 3 6 14
B 6 11 9
C 12 4 4

A 11 13 26
B 7 6 6
C 7 7 1
Figure 7.5. Observed frequencies for 248 three-event chains derived from the
same sequence of 250 events used for Figure 7.4.

2: A B C
0:A 48 23 24
B 21 21 27
C 25 26 33
Figure 7.6. Observed frequencies for the collapsed lag Ox lag 2 table derived from
the observed frequencies for three-event chains shown in Figure 7.5.

is represented as [012] and includes seven terms: 072, 01, 12, 02, 0, 1, and
2. The saturated model is not symbolized as [012][01][12][02][0][l][2]
because the three two-way and three one-way terms are implied by (we
could say, nested hierarchically within) the three-way term, and so it is
neither necessary nor conventional to write them explicitly.
Typically, a hierarchic log-linear analysis proceeds by deleting terms,
seeking the simplest model that nonetheless fits the data tolerably well
(Bakeman & Robinson, 1994). Results for the data shown in Figure 7.4
are given in Table 7.8. The best-fitting model is [01][12]; the term that
represents lag 0-lag 2 association (i.e., the 02 term) is not required. Thus
 Log-linear approaches to lag-sequential analysis 123

Table 7.8. Hierarchic log-linear analysis of the data shown in

Figure 7.4
G2 Term AG2
2 2
Model G df Deleted AG df
[012] 0.0 0
[01][12][02] 7.67 8 012 7.67 8
[01][12] 11.52 12 02 3.85 4
[01][2] 45.88* 16 12 34.35* 4
[0][l][2] 81.34* 20 01 35.46* 4

**p < .01

the log-linear analysis reveals that, when events at lag 1 are taken into
account, events at lag 0 and lag 2 are not associated, as suggested by the
analysis of the 02 table, but are in fact independent. Such conditional
independence - that is, the independence of lag 0 and lag 2 conditional
on lag 1 - is symbolized 0_U_2|l by Wickens (1989; see also Bakeman &
Quera, 1995b), and the ability to detect such circumstances represents an
advantage of log-linear over traditional lag-sequential methods. Readers
who wish to pursue the matter of conditional independence further should
read Wickens (1989, especially chapter 3).
As just described, when consecutive codes may repeat log-linear but not
traditional lag-sequential methods detect conditional independence. Ad-
ditional advantages accrue when consecutive codes cannot repeat because
log-linear methods handle structural zeros routinely and do not require ad
hoc and problematic formulas such as Equations 7.7 and 7.8. As an ex-
ample, we (Bakeman & Quera, 1995b) generated a sequence of 122 coded
events and tallied the 120 overlapped three-event chains. Tallies for the 12
permitted sequences are given in Figure 7.7. Cells containing structural
zeros are also indicated; when consecutive codes cannot repeat, the 012
table will always contain the pattern of structural zeros shown.
A summary of the log-liner analysis for the data given in Figure 7.7 is
shown in Table 7.9. When K is 3, and only when K is 3, the [01][12][02]
model is completely determined; its degrees of freedom are 0 and expected
frequencies duplicate the observed ones (as in the [012] model, when con-
secutive codes may repeat). Unlike in the previous analysis, for these data
the model of conditional independence - [01] [12] - fails to fit the data
(G 2 [3, N = 120] = 10.83, p < .05). Thus we accept the [01] [12] [02]
model and conclude that events at lag 0 and lag 2 are associated (and both
are associated with lag 1).
124 Analyzing event sequences

2: A B C
0:A —
B 15 6
C 10 9 -

A - 12 8
B - - -
C 9 12 -

A - 8 11
B 5 - 15
C - - -

Figure 7.7. Observed frequencies for the 12 possible three-event chains derived
from a sequence of 122 events for which consecutive codes cannot repeat. Struc-
tural zeros are indicated with a dash.

Table 7.9. Hierarchic log-linear analysis of the data shown in

Figure 7.7
G2 Term AG 2
2 2
Model G df Deleted AG df
[01][12][02] 0.0 0
[01][12] 10.83* 3 02 10.83* 3
[01][12]-CBC 1.64 2 cell XCBC 9.19* 1

*p < .05
**p < .01

Moreover, we can winnow these results, exactly as described for the

earlier example that permitted consecutive codes to repeat. For theoretic
reasons, assume that the CBC chain is of particular interest. An exam-
ination of the residuals for the [01] [12] model (i.e., the differences be-
tween observed frequencies and expected frequencies generated by the
[01] [12] model) showed that 4 of the 12 chains were associated with quite
large absolute residuals (the ABA, ABC, CBA, and CBC chains), which
made us think that the observed frequencies for these chains, in particular,
 Log-linear approaches to lag-sequential analysis 125

might be responsible for the failure of the [01] [12] model to fit the ob-
served data. Because the CBC chain is of primary interest, we replaced
the XCBC cell (which contained a tally of 15) with a structural zero. As
shown in Table 7.9, the model of conditional independence now fit the data
(G 2 [2, AT = 105] = 1.64, NS\ and so we conclude that the CBC chain can
account for the lag 2 effect detected by the omnibus analysis described in the
previous paragraph. This can be tested directly with a hierarchic test, as in-
dicated in Table 7.9. The difference between two hierarchically related G2s
is distributed approximately as chi-square with degrees of freedom equal to
the difference between the degrees of freedom for the two G 2 s; in this case,
AG 2 (1) = 9.19, p < .01. (Replacing a different chain with a structural
zero might also result in a fitting model, which is why it is so important that
selection of the chain to consider first be guided by theoretic concerns.)

Minimizing data demands

Quantitative data analysis always requires sufficient data, and log-linear
and traditional lag-sequential approaches are no exception. Still, the data
required for the multidimensional tables of log-linear analysis can be quite
intimidating. Several rules of thumb for log-linear analysis are available,
usually stated in terms of expected frequencies or even degrees of free-
dom for hierarchic tests like the one for cell XCBC shown in Table 7.9,
but one suggested requirement (a necessary minimum, but not necessarily
sufficient) is that the total sample be at least 4 or 5 times the number of
cells not structurally zero (Wickens, 1989, p. 30). This number is KL+l
when consecutive codes may repeat and K(K — 1)L when they cannot (e.g.,
when K = 3 and L = 2, the number of cells is 27 and 12 when codes
may and cannot repeat, respectively). Thus the number of cells, and so
the total sample desired, increases exponentially with increases in K and
L (although the increase is somewhat less pronounced when consecutive
codes cannot repeat).
Especially for larger values of L, unless the number of events observed
is almost astronomically large, the average number of events per cell may
be distressingly small. Further, expected frequencies for far too many of
the cells may be near zero, which is problematic for log-linear analysis.
To minimize data demands, Bakeman and Quera (1995b) have suggested
a sequential search strategy for explicating lagged effects. Although the
details are somewhat different, the general strategy is the same when con-
secutive codes may and cannot repeat, which once again demonstrates the
generality of the log-linear approach.
Consider first the strategy when consecutive codes may repeat. Lag 1 ef-
fects, which require only the two-dimensional 01 table, are unproblematic,
126 Analyzing event sequences

although of course we would test for lag 1 effects. Thus the search begins
by looking for complex lag 2 effects in the 012 table. At each lag (L > 1),
the complex effects we seek first implicate lags 0 and L with lag L — 1.
If present, collapsing over the L — 1 dimension (which would reduce the
number of cell and so data demands) is unwarranted. For example, if L is
2, then complex effects are present if the simplest model that fits the 012
table includes any of the following:
1. [012] because then lag 0 and lag 2 are associated and interact with
lag 1 (three-way associations), or
2. [01] [12] [02] because then lag 0 and lag 2 are associated with each
other and lag 1 but do not interact with lag 1 (homogeneous asso-
ciations), or
3. [01] [12] because then lag 0 and lag 2 are independent conditional
on lag 1.
If complex effects are found, we would explicate them, as demonstrated
in the previous section. However, if simpler models fit (e.g., [01][2] or
any others not in the list just given), which means no complex effects were
found, then collapsing over the L — 1 dimension is justified (Wickens,
1989, pp 79-81, pp. 142-143), resulting in the 0L table of traditional
lag-sequential analysis (e.g., the 02 table when L = 2).
Assuming no complex effects are found in the 012 table, after collapsing
we would first test whether 0X2 (unconditional independence) in the 02
table and, if not, examine residuals in order to explicate the lag 0-lag 2
effect just identified (exactly as we would have done for the 01 table).
Next we would create a new three-dimensional table by adding the lag 3
dimension, tally sequences in this 023 table, and then look for lag 3 effects
in the 023 table exactly as described for the 012 table. This procedure is
repeated for successive lags. In general terms, beginning with lag L, we test
whether the three-dimensional 0(L — 1)L table can be collapsed over the
L — 1 dimension. If so, we collapse to the 0L table, add the L + 1 dimension
thereby creating a new three-dimensional table, increment L, and repeat the
procedure, continuing until we find a table that does not permit collapsing
over the L — 1 dimension. Once such a table is found, we explicate the lag
L effects in this three-dimensional table. If data are sufficient, we might
next analyze the four-dimensional 0(L — 1)L(L +1) table, and so forth, but
further collapsing is unwarranted because of the lag L effects just found.
Nonetheless, this strategy may let us examine lags longer than 2 without
requiring tables larger then K3 when consecutive codes may repeat.
The sequential search strategy described in the previous paragraph ap-
plies when consecutive codes cannot repeat with one modification. When
consecutive codes may repeat, and no complex lag L effects are found (i.e.,
each table examined sequentially permits collapsing) then the test series
 Computing Yule's Q or phi 127

becomes 0X2, then 0J_3 and so forth (i.e., 0J_L is tested in the OL table),
as just described. When consecutive codes cannot repeat, the unconditional
test makes no sense because it fails to reflect the constraints imposed when
consecutive codes cannot repeat. Then, when no complex lag L effects
are found, the analogous series becomes 0_ll_2| 1, 0_JL3|2, and so forth [i.e.,
0_ll_L \L — 1 is tested in the 0(L — \)L table]. Models associated with these
tests include the (L — \)L term. The corresponding marginal table has
structural zeros on the diagonal, which reflect the cannot-repeat constraint.
This strategy may let us examine lags longer than 2 without requiring tables
larger than K2(K — 1) when consecutive codes cannot repeat (K[K — I] 2
when L = 2). These matters are discussed further in Bakeman and
Quera (1995b).

7.7 Computing Yule's Q or phi and testing for individual

differences
Often more than a single individual, dyad, family, or whatever, is observed;
these units are embedded in a design (e.g., a two-group design might include
clinic and nonclinic couples), and investigators want to ask questions about
the importance of their research factors (e.g., is a particular sequential
pattern more characteristic of clinic than nonclinic couples). The previous
edition of this book suggested than z scores might serve as scores for
subsequent analyses (e.g., analyses of variance, multiple regression, etc.),
but that was not sound advice. The z score is affected by the number
of tallies (if the number of tallies doubled but the association remained
the same, the z score would increase), and so is not comparable across
experimental units (subjects, dyads, families, etc.) unless the total number
of tallies remains the same for each. Some measure that is unaffected by the
number of tallies, such as a strength of association or effect size measure,
should be used instead (Wampold, 1992).
Strength of association or effect size measures are especially well devel-
oped for 2 x 2 tables (to give just two examples from an extensive literature,
see Conger & Ward, 1984, and Reynolds, 1984; much of the material in
this and subsequent paragraphs is summarized from Bakeman, Me Arthur,
& Quera, 1996). This is fortunate, because when interest centers on one
cell in a larger two-dimensional table, the larger table can be collapsed into
a 2 x 2, and statistics developed for 2 x 2 , tables can be used (as Morley,
1987, noted with respect to phi). Assume, for example, that we want to
know whether event B is particularly likely after event A. In this case, we
would label rows A and ~ A and columns B and ~ B (where rows repre-
sent lag 0, columns lag 1, and ~ represents not). Then the collapsed 2 x 2
128 Analyzing event sequences

table can be represented as

B -B
a b
c d

where individual cells are labeled a,b,c, and d as shown and represent
cell frequencies.
One of the most common statistics for 2 x 2 tables (perhaps more so
in epidemiology and sociology than psychology) is the odds ratio. As its
name implies, it is estimated by the ratio of a to b divided by the ratio
of c to d,
est. odds ratio = ">" (7.1.)
c/d
(where a, &, c, and d refer to observed frequencies for the cells of a 2 x 2
table as noted earlier; notation varies, but for definitions in terms of pop-
ulation parameters, see Bishop Fienberg, & Holland, 1975; and Wickens,
1993). Multiplying numerator and divisor by d/c, this can also be ex-
pressed as
ad
est. odds ratio = — . (7.12)
be
Equation 7.12 is more common, although Equation 7.11 reflects the name
and renders the concept more faithfully. Consider the following example:

A 10 10 20
A 20 60 80
30 70 100

The odds for B after A are 1:1, where as the odds for B after any other
(non-A) event are 1:3; thus the odds ratio is 3. In other words, the odds for
B occurring after A are three times the odds for B occurring after anything
else. When the odds ratio is greater than 1 (and it can always be made > 1
by swapping rows), it has the merit, lacking in many indices, of a simple
and concrete interpretation.
The odds ratio varies from 0 to infinity and equals 1 when the odds are
the same for both rows (indicating no effect of the row classification). The
natural logarithm (In) of the odds ratio, which is estimated as
(ad\
est. log odds ratio = In — (7.13)
\bcj
 Computing Yule's Q or phi 129

extends from minus to plus infinity, equals 0 when there is no effect, and
is more useful for inference (Wickens, 1993). However Equation 7.13
estimates are biased. An estimate with less bias, which is also well defined
when one of the cells is zero (recall that the log of zero is undefined), is
obtained by adding 1/2 to each count,
(a + l/2)(d + 1/2)
= <714)
^ "V+l/2X*+l/2)
(Gait & Zweifel, 1967; cited in Wickens, 1993, Equation 8). As Wickens
(1993) notes when recommending that the log odds ratio computed per
Equation 7.14 be analyzed with a parametric t test, this procedure not
only provides protection for a variety of hypotheses against the effects of
intersubject variability when categorical observations are collected from
each member of a group (or groups), it is also easy to describe, calculate,
and present.

Yule's Q
Yule's Q is a related index. It is a transformation of the odds ratio designed
to vary, not from zero to infinity with 1 indicating no effect, but from — 1
to +1 with zero indicating no effect, just like the familiar Pearson product
- moment correlation. For that reason many investigators find it more
descriptively useful than the odds ratio. First, c/d is subtracted from the
numerator so that Yule's Q is zero when a/b equals c/d. Then, a/b is
added to the denominator so that Yule's Q is +1 when b and/or c is zero
and —1 when a and/or d is zero, as follows:
a c ad — be
Yule's Q = | — \ = , bf , = ab~bc (7.15)
^ £ + f* bc + ad ad + bc
d b bd
Yule's Q can be expressed as a monotonically increasing function of both
the odds and log odds ratio; thus these three indices are equivalent in the
sense of rank ordering subjects the same way (Bakeman, McArthur, &
Quera, 1996).

Phi
Another extremely common index for 2 x 2 tables is the phi coefficient.
This is simply the familiar Pearson product-moment correlation coefficient
computed using binary coded data (Cohen & Cohen, 1983; Hays, 1963).
130 Analyzing event sequences

One definition for phi is

where z is computed for the 2 x 2 table and hence equals \[~y}. Thus phi
can be viewed as a z score corrected for sample size. Like Yule's Q, it
varies from —1 to +1 with zero indicating no association. In terms of the
four cells, phi is defined as

(7.17)
V(a + b)(c + d)(a + c)(b + d)
Multiplying and rearranging terms this becomes
ad — be
(7.18)
bd + ad + bc)(ab + cd + ad + be)
If we now rewrite the expression of Yule's Q, first squaring the denominator
of Equation 7.15 and then taking its square root

ad
Yule's Q = ~bC (7.19)
y/(ad + bc)(ad + be)
the value of Yule's Q is not changed but similarities and differences between
phi and Yule's Q (Equations 7.18 and 7.19) are clarified.
Does it matter which index is used, Yule's Q or phi? The multiplier and
multiplicand in the denominator for Yule's Q (Equation 7.19) consist only
of the sum of ad and be, whereas multiplier and multiplicand in the phi
denominator (Equation 7.18) add more terms. Consequently, values for
phi are always less than values for Yule's Q (unless b and c, or a and d,
are both zero, in which case both Yule's Q and phi would be +1 and —1,
respectively). Yule's Q and phi differ in another way as well. Yule's Q is
+1 when either b or c is zero and — 1 when either a or d is zero (this is called
weak perfect association, Reynolds, 1984), whereas phi is +1 only when
both b and c are zero and — 1 only when both a and d are zero (this is called
strict perfect association). Thus phi achieves its maximum value (absolute)
only when row and column marginals are equal (Reynolds, 1984). Some
investigators may regard this as advantageous, some as disadvantageous,
but in most cases it probably matters little which of these two indices is
used (or whether the odds ratio or log odds ratio is used instead). In fact,
after running a number of computer simulations, Bakeman, McArthur, and
Quera (1996) concluded that, when testing for group differences, it does
not matter much whether Yule's Q or phi is used since both rank-order cases
essentially the same. Transformed kappa, a statistic proposed by Wampold
 Computing Yule's Q or phi 131

(1989, 1992), however, did not perform as well. For details see Bakeman,
Me Arthur, and Quera (1996).

Type I error considerations

In an ideal confirmatory world, investigators would pluck the one transition
from a larger table needed to answer their most important research question;
compute a single Yule's Q or phi based on a collapsed A, ~ A|B, ~ Btable,
such as the one shown earlier; and proceed to test for group differences
(or other questions as their design permits). But much of the world is
rankly exploratory. Indeed, it is tempting to compute some index for each
of the K2 cells of a table (K[K — 1] cells when consecutive codes cannot
repeat), one set for each subject, and then subject all K2 scores to standard
parametric tests (t test, analyses of variance, etc.). This courts type I
error in a fairly major way. At the very least, no more indices should
be derived than the degrees of freedom associated with the table, which is
(K — l)(K — 1) when consecutive codes may repeat and (K — l)(K — \) — K
when not (assuming a table with structural zeros on the diagonal). This is
somewhat analogous to decomposing an omnibus analysis of variance into
single-degree-of-freedom planned comparisons or contrasts.
One systematic way to derive indices from a larger table requires that
one code be regarded as something of a baseline, or base for comparison,
such as unengaged or no activity. For example, imagine that codes are
labeled A, 5, and C, and that code C represents some sort of baseline. Then
following Reynolds's (1984) suggestion for decomposing the odds ratio in
tables larger than 2 x 2 , and labeling the cells in the 3 2 table as follows:
A B C
P
A a b c
B d e f
C 8 h i

four 2 x 2 tables would be formed for each subject, as follows:

AC B C
a c b c
8 i h i

B d f B e f
C g i C h i
132 Analyzing event sequences

and a Yule's Q or phi computed for each. These statistics could then be sub-
jected to whatever subsequent analyses the investigator deems appropriate.
In this section we have suggested that sequential associations between
two particular events (e.g., an A to B transition) be assessed with an index
like Yule's Q or phi. These statistics gauge the magnitude of the effect
and, unlike the z score, are unaffected by the number of tallies. Thus
they are reasonable candidates for subsequent analyses such as the familiar
parametric tests routinely used by social scientists to assess individual
differences and effects of various research factors (e.g., t tests, analyses of
variance, and multiple regression). But the events under consideration may
be many in number, leading to many tests and thereby courting type I error.
It goes without saying (which may be why it is so necessary to restate)
that guiding ideas provide the best protection against type I error. Given
K codes and an interest in lag 1 effects, a totally unguided and completely
exploratory investigator might examine occurrences of all possible K2 two-
event chains (or K[K — 1] two-event chains when consecutive codes cannot
repeat). In this section, we have suggested that a more justifiable approach
would limit the number of transitions examined to the (K — I) 2 degrees of
freedom associated with the table (or [K — I] 2 — K degrees of freedom
when consecutive codes cannot repeat) and have demonstrated one way
that this number of 2 x 2 subtables could be extracted from a larger table.
Presumably a Yule's Q or some other statistic would be computed for each
subtable. Positive values would indicate that the pair of events associated
with the upper-left-hand cell is associated more than expected, given the oc-
currences observed for the baseline events associated with the second row
and second column of the 2 x 2 table. The summary statistic for the 2 x 2 ta-
bles, however many are formed, could then be subjected to further analysis.
Investigators are quite free - in fact, encouraged - to investigate a smaller
number of associations (i.e., form a smaller number of 2 x 2 tables). For ex-
ample, a larger table might be collapsed into a smaller one, combining some
codes that seem functionally similar, or only those associations required to
address the investigator's hypotheses might be subjected to analysis in the
first place. Other transitions might be examined later, and those analyses
labeled exploratory instead of confirmatory. For further discussion of this
"less is more" and "least is last" strategy for controlling type I error, see
Cohen and Cohen (1983, pp. 169-172).

7.8 Summary
Investigators often represent their data as sequences of coded events. Some-
times, data are recorded as event sequences in the first place; other times,
 Summary 133

in order to answer particular questions, event sequences are extracted from

data initially recorded and represented in some other way. The purpose
of this chapter has been to describe some ways of analyzing such event
sequences, although much of what has been presented here can apply to
the analysis of time sequences as well.
Sometimes consecutive events cannot be assigned the same code in event
sequences. For example, when coders are asked to segment the stream of
behavior into mutually exclusive and exhaustive behavioral states, often
adjacent states cannot be coded the same way, by definition. It they were
the same, they would be just one state. However, we can imagine other
ways of defining event boundaries that would allow adjacent codes to be
the same. (The codes used to categorize events would still be mutually
exclusive and exhaustive, but that is a different matter.) For example,
if utterances were being coded, two successive utterances might both be
coded the same. Whether adjacent codes can be the same or not is an
important matter because it affects the way expected frequencies, expected
probabilities, and hence adjusted residuals (i.e., z scores) are computed.
One approach to sequence detection we have called "absolute." Investi-
gators define particular sequences of some specified length, categorize and
tally all sequences of that length, and report the frequencies and probabil-
ities for particular sequences. A z score can be used to gauge the extent
to which an observed frequency (or probability) for a particular sequence
exceeds its expected value. However, if the z score is to be tested for sig-
nificance, its computation should be based on sufficient data to justify the
normal approximation to the binomial distribution.
In theory, absolute methods apply to sequences of any length. In prac-
tice, certain limitations may prevail. In particular, the number of possible
sequences increases dramatically as longer and longer sequences are con-
sidered. Unless truly mammoth amounts of data are available, expected
frequencies for a particular sequence may be too small to justify assign-
ing significance. Moreover, the number of occurrences for a particular
sequence may be so few that the investigator has little confidence in the
accuracy of the observed frequency, even descriptively. Another exacer-
bating circumstance is the number of codes defined. In general, when there
are more codes, the expected frequencies for particular sequences are likely
to be smaller, and hence more data will be required.
Even when z-score computations are based on sufficient data, the type
I error problem remains. This is usually not a problem for confirmatory
studies, assuming, of course, that just a few theoretically relevant tests of
significance are made. But when the number of tests is large, as it typically
is for exploratory studies, then some thought should be given to ways to
control the type I error rate. As discussed in this chapter, the number
134 Analyzing event sequences

of initially significant transitions can be winnowed using structural zeros

and log-linear methods, and the number of transitions examined in the
first place can be limited to those of clear theoretic interest. Nonetheless,
interpretation of results may need to take into account that some of the
apparently significant findings are due simply to chance.
One way to describe two-event sequences is to report their simple prob-
abilities. Another way is to report (lag 1) transitional probabilities instead.
Of the two, transitional probabilities (which "control" for differences in the
base rate of the first or "given" code) often seem more informative descrip-
tively. The values of both, however, are affected by the values for the base
rates of the two codes involved. This does not affect their descriptive value,
but in cases in which transitional probabilities have been computed sepa-
rately for different subjects, it does make such scores poor candidates for
subsequent analyses of individual or group differences. First, depending
on base rates, similar numerical values for the same transitional probabil-
ity may have quite different meanings for different subjects. And second,
what appear to be analyses of transitional probabilities may in fact reflect
little more than simple probability effects. The z scores are not hampered
by these problems, but have additional problems of their own. Their values
are affected by the number of tallies, and so larger values may reflect, not
a larger effect, but simply more data. Whenever individual or group dif-
ferences or effects of research factors generally are of interest, magnitude
of effect statistics, and not z scores, should be used. Examples include the
odds ratio, the log odds ratio, Yule's Q, and phi.
A second approach to sequence detection, the lag-sequential method,
we have characterized as "probabilistic" instead of "absolute." Like that of
absolute methods, its purpose is to detect commonly occurring sequences,
but because it examines codes pairwise only, it can detect sequences longer
than two events without invoking the same restrictions and limitations
involved with absolute methods. Moreover, sequences containing random
elements can be detected as well. The method is based on an examination
of z scores associated with transitional probabilities computed for various
lags; thus any limitations and cautions (including the type I error problem)
that apply to two-event transitional probabilities also apply to the lag-
sequential approach as well.
Perhaps the most adequate approach to sequence detection is log-linear.
Log-linear analysis promotes a whole-table view, whereas often traditional
lag-sequential analysis focused, almost piecemeal, on individual transi-
tions in a table. This is not necessarily problematic when, in the context
of a confirmatory study, only a few transitions are of interest, but a narrow
focus on repeated tests tied to all cells in the context of an exploratory
study invites type I error. Additionally, log-linear analysis provides ways
 Summary 135

of disentangling the web of connected results in a table, and makes rou-

tine the analysis of sequences in which, for logical reasons, consecutive
codes cannot repeat. Finally, log-linear analysis, using well-established
statistical techniques, provides an integrated method of broad generality
for determining whether there are effects at various lags, no matter whether
consecutive codes may or cannot repeat and no matter whether or not over-
lapped sampling was employed. Whenever possible, it seems the analytic
approach of choice for the analysis of coded sequences.
 8
Issues in sequential analysis

8.1 Independence
In classical parametric statistics, we assume that our observations are in-
dependent, and this assumption forms part of the basis of our distribution
statistics. In the sequential analysis of observational data, on the other
hand, we want to detect dependence in the observations. To do this we
compare observed frequencies with those we would expect if the observa-
tions were independent. Thus, dependence in the data is not a "problem."
It is what we are trying to study.
The statistical problem of an appropriate test is not difficult to solve. It
was solved in a classic paper in 1957 by Anderson and Goodman (see also
Goodman, 1983, for an update). Their solution is based on the likelihood-
ratio chi-square test.
The likelihood-ratio test applies to the comparison of any two statisti-
cal models if one (the "little" model) is a subcase of the other (the "big"
model). The null-hypothesis model is usually the little model. In our case,
this model is often the assumption that the data are independent (or quasi
independent); i.e., that there is no sequential structure. Compared to this is
the big, interesting model that posits a dependent sequential structure. As
discussed in section 7.6, the difference between the G 2 for the big model
(e.g., [01]) and the G 2 for the little model (e.g., [0][l]) is distributed asymp-
totically as chi-square, with degrees of freedom equal to the difference in
the degrees of freedom for the big and little models. "Asymptotic" means
that it becomes increasingly true for large N, where N is the number of
observations.
When the data have "structural zeros," e.g., if a code cannot follow itself
(meaning that the frequency for that sequence is necessarily zero), the
number of degrees of freedom must be reduced (by the number of cells that
are structural zeros). These cells are not used to compute chi-square (see
Goodman, 1983).
We shall now discuss the conditions required to reach asymptote. In
particular, we shall discuss assigning probability values to z scores. We

136
 Independence 137

should note that most observational data are stochastically dependent. They
are called in statistics "ra-dependent" processes, which means that the
dependencies are short lived. One implication of this is that there is poor
predictability from one time point to another, as the lag between time
points increases. In time-series analysis, forecasts are notoriously poor
if they exceed one step ahead (see Box & Jenkins, 1970, for a graph of
the confidence intervals around forecasts). It also means that clumping m
observations gives near independence. For most data, m will be quite small
(probably less than 4), and its size relative to n will determine the speed at
which the asymptote is approached.
We conclude that assigning probability values to pairwise z scores (or
tablewise chi-squares) is appropriate when we are asking if the observed
frequency for a particular sequence is significantly different from expected
(or whether lag 0 and L are related and not independent). We admit,
however, that more cautious interpretations are possible, and would quote
a paragraph we wrote earlier (Gottman & Bakeman, 1979, p. 190):
As N increases beyond 25, the binomial distribution approximates a normal dis-
tribution and this approximation is rapidly asymptotic if P is close to 1/2 and
slowly asymptotic when P is near 0 or 1. When P is near 0 or 1, Siegel (1956)
suggested the rule of thumb that NP(l — P) must be at least 9 to use the normal
approximation. Within these constraints the z-statistic above is approximately
normally distributed with zero mean and unit variance, and hence we may cau-
tiously conclude that if z exceeds ± 1.96 the difference between observed and
expected probabilities has reached the .05 level of significance (see also Sackett,
1978). However, because dyadic states in successive time intervals (or simply
successive dyadic states in the case of event-sequence data) are likely not inde-
pendent in the purest sense, it seems most conservative to treat the resulting z
simply as an index or score and not to assign /?-values to it.
As the reader will note, in the chapter just quoted we were concerned
with two issues: the assumption of independence and the number of tallies
required to justify use of the binomial distribution. On reflection, we find
the argument that the categorizations of successive n-event sequences in
event-sequence data are not "independent" less compelling than we did
previously, and so we are no longer quite so hesitant to assign probability
values on this score.
Our lack of hesitancy rests in part on a simulation study Bakeman and
Dorval (1989) performed. No matter the statistic, for the usual sorts of
parametric tests, p values are only accurate when assumptions are met. To
those encountering sequential analysis for the first time, the common (but
not necessarily required) practice of overlapped sampling (tallying first the
e\e2 chain, then ^ 3 , £3^4, etc.) may seem like a violation of independence.
The two-event chain is constrained to begin with the code that ended the
138 Issues in sequential analysis

previous chain (i.e., if a two-event chain ends in B, adding a tally to the 2nd
column, the next must add a tally to the 2nd row), and this violates sampling
independence. A Pearson or likelihood-ratio chi-square could be computed
and would be an index of the extent to which observed frequencies in this
table tend to deviate from their expected ones. But we would probably
have even less confidence than usual that the test statistic is distributed as
X2 and so, quite properly, would be reluctant to apply a p value.
Nonoverlapped sampling (tallying first the e\e2 chain, then ^3^4, e^e^,
etc.) does not pose the same threat to sampling independence, although it
requires sequences almost twice as long in order to extract the same num-
ber of two-event chains produced by overlapped sampling. However, the
consequences of overlapped sampling may not be as severe as they at first
seem. Bakeman and Dorval (1989) found that when sequences were gen-
erated randomly, distributions of a test statistic assumed their theoretically
expected form equally for the overlapped and nonoverlapped procedures
and concluded that the apparent violation of sampling independence asso-
ciated with overlapped sampling was not consequential.

8.2 Stationarity
The term "stationarity" means that the sequential structure of the data is
the same independent of where in the sequence we begin. This means that,
for example, we will get approximately the same antecedent/consequent
table for the first half of the data as we get for the second half of the data.
1st Half 2nd Half
HNice HNasty HNice HNasty
WNice
WNasty

We then compute the pooled estimates over the whole interaction:

HNice HNasty
WNice
WNasty
To test for stationarity of the data, we compare the actual data to the
expected values under the null hypothesis that the data are stationary. Let's
assume that we are interested in only the lag-1 antecedent/consequent table
for s codes. Let N(IJ, t) be the joint frequency for cell / / in segment r, and
 Describing general orderliness 139

P(IJ, t) be the transition probability for that cell. Let P(IJ) be the pooled
transition probability. Then G 2 , computed as

is distributed as chi-square. If there are s codes, and T segments, then G2

has degrees of freedom (T — l)(s)(s — 1). A more general formula for the
rth-order transition table is given by Gottman & Roy (1990, pp. 62-63),
where r is the order of the chain. The sum is across segments of the data.
For the example data, the value computed was 6.44, which is compared to
df = 2; this quantity is not statistically significant. Hence, the example
data are stationary.
This test can be used to see if the data have a different sequential structure
for different parts of the interaction. This appears to be the case for conflict
resolution in married couples; the first third is called the "agenda building
segment," the second third is called the "disagreement segment," and the
final third is called the "negotiation segment" (Gottman, 1979a). However,
a problem with this test is that as the number of observations increases, the
power we have to detect violations of absolute stationarity increases, and
yet, for all intents and purposes the data may be stationary enough.
In this case we can do a log-linear analysis of the data and evaluate the
Q2 statistic, as recommended by Bakeman and Robinson (1994, p. 102
ff.). For example, let C = consequent (husband nice/husband nasty),
A = antecedent (wife nice/wife nasty), and T = segment (first half/second
half). If the CAT term is required for a fitting model, this suggests that the
association between antecedent and consequent varies as a function of time
(i.e., is not stationary). For the present data, the loss in fit when this term is
removed is significant at the .10 but not the .05 level (G 2 [l] = 3.11). The
G 2 for the base model (i.e., [C][A][T]) is 227.4, so the 3.11 represents less
than 1.4% of the total. Even when the loss in fit is significant at the .05
or a more stringent level, Knoke and Burke (1980) recommend ignoring
terms that account for less than 10% of the total. This can be useful when
even small effects are statistically significant, as often happens when the
number of tallies is large.

8.3 Describing general orderliness

The material already presented in the last chapter assumes that investigators
want to know how particular behavioral codes are ordered. For example,
they may want to confirm that a particular, theoretically important sequence
(like Touch, Nurse, Groom) really occurs more often than expected, given
140 Issues in sequential analysis

base rates for each of these codes. Or, in a more exploratory vein, they may
want to identify whichever sequences, if any, occur at greater than chance
rates. However, there is another quite different kind of question investiga-
tors can ask, one that concerns not how particular codes in the stream of
behavior are ordered, but how orderly the stream of behavior is overall.
In this section, we shall not describe analyses of overall order in any
detailed way. Instead, we shall suggest some references for the interested
reader, and shall try to give a general sense of what such analyses reveal and
how they proceed. Primarily, we want readers to be aware that it is possible
to ask questions quite different from those discussed earlier in this chapter.
One traditional approach to the analysis of general order is provided
by what is usually called "information theory." A brief explication of
this approach, along with appropriate references and examples, is given by
Gottman and Bakeman (1979). Although the classical reference is Shannon
and Weaver (1949), more useful for psychologists and animal behaviorists
are Attneave (1959) and Miller and Frick (1949). A well-known exam-
ple of information theory applied to the study of social communication
among rhesus monkeys is provided by S. Altmann (1965). A closely re-
lated approach is called Markovian analysis (e.g., Chatfield, 1973). More
recently, problems of gauging general orderliness are increasingly viewed
within a log-linear or contingency-table framework (Bakeman & Quera,
1995b; Bakeman & Robinson, 1994; Bishop, Fienberg, & Holand, 1975;
Castellan, 1979; Upton, 1978).
No matter the technical details of these particular approaches, their goals
are the same: to determine the level of sequential constraint. For example,
Miller and Frick (1949), reanalyzing Hamilton's (1916) data concerning
trial-and-error behavior in rats and 7-year-old girls, found that rats were
affected just by their previous choice whereas girls were affected by their
previous two choices. In other words, if we want to predict a rat's current
choice, our predictions can be improved by taking the previous choice into
account but are not further improved by knowing the choice before the
previous one. With girls, however, we do improve predictions concerning
their current choice if we know not just the previous choice but the one
before that, too.
If data like these had been analyzed with a log-linear (or Markovian)
approach, the analysis might have proceeded as follows: First we would
define a zero-order or null model, one that assumed that all codes occurred
with equal probability and were not in any way sequentially constrained.
Most likely, the data generated by this model would fail to fit the observed
data. Next we would define a model that assumed the observed probabilities
for the codes but no sequential constraints. Again, we would test whether
the data generated by this model fit the observed. If this model failed to
 Individual vs. pooled data 141

fit, we would next define a model that assumed that codes are constrained
just by the immediately previous code (this is called a first-order Markov
process). In terms of the example given above, this model should generate
data that fit those observed for rats but not for girls. Presumably, a model
that assumes that codes are constrained by the previous two codes should
generate data that pass the "fitness test" for the girl's data.
In any case, the logic of this approach should be clear. A series of models
are defined. Each imposes an additional constraint, for example, that the
data generated by the model need to take into account the previous code,
the previous two codes, etc. The process stops when a particular model
generates data similar to what was actually observed, as determined by a
goodness-of-fit test. The result is knowledge about the level of sequential
constraint, or connectedness, or orderliness of the data, considered as a
whole. (For a worked example, analyzing mother-infant interaction, see
Cohn & Tronick, 1987.)

8.4 Individual versus pooled data

In the previous chapter, we discussed two different uses of sequential statis-
tics such as z scores (i.e., adjusted residuals and Yule's Q's). First, assuming
that successive codings of events are independent of previous codings, and
assuming that enough data points are available, we have suggested that
z scores can be tested for significance (see section 7.4). Second, when
data are collected from several different "units" (e.g., different partici-
pants, dyads, or families), we have suggested that scores such as Yule's Q
can be used in subsequent analyses of individual or group differences (see
section 7.7). Because the familiar parametric techniques (e.g., analyses
of variance) are both powerful and widely understood, such a course has
much to recommend it.
Not all studies include readily discernible "units" however. For ex-
ample, just one individual or couple might be observed, or the animals
observed might not be easily identifiable as individuals. In such cases,
the issue of pooling data across units does not arise; there are not multi-
ple units. In other cases, for example, when several different children are
observed, so few data might be collected for each child that pooling data
across all children could seem desirable, if for no other reason than to in-
crease the reliability of the summary statistics reported. At the same time,
assuming enough data, it might then be possible to test z scores for signif-
icance on the basis of the pooled data. Properly speaking, however, any
conclusions from such analyses should be generalized just to other behavior
of the group observed, not to other individuals in the population sampled.
142 Issues in sequential analysis

Thus, even though investigators who pool data over several subjects usu-
ally do so for practical reasons, it has some implications for how results
are interpreted.
How seriously this last limitation is taken seems to vary somewhat by
field. In general, psychologists studying humans seem reluctant to pool
data over subjects, often worrying that some individuals will contribute
more than others, thereby distorting the data. Animal behaviorists, on the
other hand, seem to worry considerably less about pooling data, perhaps
because they regard their subjects more as exemplars for their species and
focus less on individuality. Thus students of animal behavior often seem
comfortable generalizing results from pooled data to other members of the
species studied.
As we see it, there are three options: First, when observations do no de-
rive from different subjects (using "subject" in the general sense of "case"
or "unit"), the investigator is limited to describing frequencies and proba-
bilities for selected sequences. Assuming enough data, these can be tested
for significance. Second, even when observations do derive from different
subjects, but when there are few data per subject, the investigator may opt
to pool data across subjects. As in the first case, sequences derived from
pooled data can be tested for significance, but investigators should keep in
mind the limits on interpretation recognized by their field.
Third, and again when observations derive from different subjects, in-
vestigators may prefer to treat statistics (e.g., Yule's Q's) associated with
different sequences just as scores to be analyzed using standard techniques
like t test or the analysis of variance (see Wickens, 1993). In such cases,
statistics for the sequences under consideration would be computed sepa-
rately for each subject. However, analyses of these statistics tell us only
whether they are systematically affected by some research factor. They do
not tell us whether the statistics analyzed are themselves significant. In
order to determine that, we could test individual z scores for significance,
assuming enough data, and report for how many participants z scores were
significant, or else compute a single z score from pooled data, assuming
that pooling over units seems justified.
For example, Bakeman and Adamson (1984), for their study of infants'
attention to people and objects, observed infants playing both with their
mothers and with same-age peers. Coders segmented the stream of behavior
into a number of mutually exclusive and exhaustive behavioral states: Two
of those states were "Supported Joint" attention (infant and the other person
were both paying attention to the same object) and "Object" attention (the
infant alone was paying attention to some object). The Supported Joint
state was not especially common when infants played with peers. For
 Individual vs. pooled data 143

that reason, observations were pooled across infants, but separately for the
"with mother" and "with peer" observations.
The z scores computed from the pooled data for the Supported Joint
to Object transition were large and significant, both when infants were
observed with mother and when with peers. This indicates that, considering
these observations as a whole, the Supported Joint to Object sequence
occurred significantly more often than expected, no matter the partner. In
addition, an analysis of variance of individual scores indicated a significant
partner effect, favoring the mother. Thus, not only was this sequence
significantly more likely than expected with both mothers and peers, the
extent to which it exceeded the expected was significantly higher with
mothers, compared to peers.
In general, we suspect that most of our colleagues (and journal editors)
are uneasy when data are pooled over human participants. Thus it may be
worthwhile to consider how data such as those just described might be an-
alyzed, not only avoiding pooling, but actually emphasizing individuality.
The usual parametric tests analyze group means and so lose individuality.
Better, it might be argued, to report how many subjects actually reflected
a particular pattern, and then determine whether that pattern was observed
in more subjects than one might expect by chance.
For such analyses, the simple sign test suffices. For example, we might
report the number of subjects for whom the Yule's Q associated with the
Supported Joint to Object transition was positive when observed with moth-
ers, and again when observed with peers. If 28 infants were observed, by
chance alone we would expect to see the pattern in 14 (50%) of them, but
if the number of positive Yule's Q's was 19 or greater (p < .05, one-tailed
sign test), we would conclude that the Supported Joint to Object transition
was evidenced by significantly more infants than expected. And if the
Yule's Q when infants were observed with mothers was greater than the
Yule's Q when infants were observed with peers for 19 or more infants, we
would conclude that the association was stronger when with mothers, com-
pared to peers. The advantage of such a sign-test approach is that we learn,
not just what the average pattern was, but exactly how many participants
evidenced that pattern.
The approach presented earlier in section 8.2 can also be applied to
the issue of pooling over subjects. Again, the question we ask is one of
homogeneity, that is, whether the sequential structure is the same across
subjects, or groups of subjects (instead of across time as for stationarity).
The formula to test this possibility is similar to the formula for stationarity
(see Gottman & Roy, 1990, pp. 67 ff.). In the following formula the sum
is across s = 1, 2 , . . . S subjects, and P(IJ) represents the pooled joint
144 Issues in sequential analysis

probability across the s subjects:

The degrees of freedom are (s - 1) (NCODES)r (NCODES-1), where r is

the order of the chain (in our case, with lag— 1, r = 1), where NCODES =
the number of codes. Note that this approach is quite general. For example,
we can test whether a particular married couple's sequential data is best
classified with one group of couples or with another group. Although
tedious, this could be a strategy for grouping subjects.

8.5 How many data points are enough?

The issue of the number of tallies - of determining how many data points
are enough - remains an important matter. The question the investigator
needs to ask is this: How many events need to be coded in order to justify
assigning significance to a computed z score associated with a particular
cell or a chi-square statistic associated with a table?
At issue is not just the total number of events, but how they are distributed
among the possible codes. It is worth reflecting for a moment why - and
when - we need be concerned with sufficient tallies for the various codes in
the first place. The bedrock reason is stability. If a summary statistic like a
chi-square, a z score, or Yule's Q is based on few tallies, then quite rightly
we place little confidence in the value computed, worrying that another
time another observation might result in quite a different value.
For example, if one of the row sums of a 2 x 2 table is a very small number
(such as 1 or 2), then shifting just one observation from one column to the
other can result in a big change in the summary statistic. As a specific
example, imagine that 50 observations were classified A | not-A and B|not-
B, as follows:
B ~B
A 2 0 2
A 24 24 48
26 24 50

The Pearson chi-square for this table is 1.92 (so the z score is its square
root, 1.39) and its Yule's Q is + 1 . After all, every A (all two of them) was
followed by a B. However, if only one of the As were followed by a B,
 How many data points are enough? 145

B ~B
A 1 1 2
A 24 24 48
25 25 50
then all statistics (Pearson chi-square, z score, and Yule's Q) would be zero.
This example demonstrates summary statistics' instability when only a few
instances of a critical code are observed.
To protect ourselves against this source of instability, we compute sum-
mary statistics only when all marginals sums are 5 or greater, and regard
the value of the summary statistics as missing (too few instances to regard
the computed values as accurate) in other cases. This is only an arbi-
trary rule, of course, and hardly confers absolute protection. Investigators
should always be alert to the scanty data problem, and interpret results
cautiously when summary statistics (e.g., z scores, Yule's Q's) are based
on few instances of critical codes.
If stability is the bedrock reason to be concerned about the adequacy
of the data collected, correct inference is a secondary but perhaps more
often mentioned concern. This matters only when investigators wish to
assign a p value to a summary statistic (e.g., a X2 or a z score) based on
assumptions that the statistic follows a known distribution (e.g., the chi-
square or normal). Guidelines for amount of adequate data for inference
have long been addressed in the chi-square and log-liner literature, so it
makes sense to adapt those guidelines - many of which are stated in terms
of expected frequencies - for lag-sequential analysis.
Several considerations play a role, so absolute guidelines are as difficult
to define as they are desired. Summarizing current advice, Wickens (1989,
p. 30) noted that (a) expected frequencies for two-dimensional tables with
1 degree of freedom should exceed 2 or 3 but that with more degrees of
freedom some expected frequencies may be near 1 and with large tables up
to 20% may be less than 1, (b) the total sample should be at least four or
five times the number of cells (more if marginal categories are not equally
likely), and (c) similar rules apply when testing whether a model fits a
three- or larger-dimensional table.
As noted earlier, when lag 1 effects are studied, the number of cells is K2
when consecutive codes may repeat and K(K — 1) when they cannot. Thus,
at a minimum, the total number of tallies should exceed K2 or K(K — 1), as
appropriate, times 4 or 5. Additionally, marginals and expected frequencies
should be checked to see whether they also meet the guidelines. When lag
L effects are studied, the number of cells is KL+l when consecutive codes
may repeat and K(K — 1)L when they cannot. As L increases, the product
146 Issues in sequential analysis

of these values multiplied by 4 or 5 can become discouragingly large.

However, if attention is limited to three-dimensional 0(L — 1)L tables, as
suggested in section 7.6, then the number of cells is no greater than K3
when consecutive codes may repeat and K2(K — 1) when they cannot. But
remember, these values multiplied by 4 or 5 represent a minimum number
of tallies. Marginals and expected frequencies still need to be examined.
Further, these products provide a guideline for the minimum number of
events that should be coded.
Without question, considerable data may be required for a sequential or
log-linear analysis. Following the strategy that limits attention to three-
dimensional tables (section 7.6), K is the determining factor. For example,
the numbers of cells when K is 3, 5, 7, and 9 and consecutive codes may
repeat are
LA 2
K:3 9 27
5 25 125
7 49 343
9 81 729
(values for higher lags are the same as when L = 2; the general formula
is K2 when L = 1 and K3 when L = 2 or higher) and when consecutive
codes cannot repeat, numbers of cells (excluding structural zeros) are
LA 2 3
K:3 6 12 18
5 20 80 100
7 42 252 294
9 72 576 648
(values for higher lags are the same as when L = 3; the general formula
is K(K - 1) when L = 1, K(K - I) 2 when L = 2, and K2(K - 1) when
L = 3 or higher). Taking into account only the times-the-number-of-cells
rule, in order to compute the N needed these values should be multiplied
by 4, 5, or whatever factor seems justified (Bakeman & Quera, 1995b).
Moreover, as Bakeman and Robinson (1994) note, such guidelines should
not be regarded as minimal goals to be satisfied, but as troubled frontiers
from which as much distance as possible is desired.
If the numbers still seem onerous, reconsider expected frequencies and
note that Haberman (1977; cited in Wickens, 1989, p. 30) suggested that
 The type I error problem 147

requirements for tests based on the difference between two chi-squares -

such as the tests of 0 _1L L\L — 1 described in section 7.6 - may depend
more on frequency per degree of freedom than on the minimum expected
cell size. When doubts arise, it may be best to consult local experts. But for
many analyses, especially when the number of codes analyzed (K) is large,
expect that the number of events coded should number in the thousands,
not hundreds.
The guidelines described in the preceding paragraphs assumed that the
computed test statistics would be distributed exactly as some theoretical
distribution (e.g., the chi-square or normal), thus permitting p values to
be based on the appropriate theoretic distribution. Such tests are often
called asymptotic because as the amount of data on which the test statis-
tic is based increases, its distribution approximates the theoretical one
ever more closely. Asymptotic tests may be either parametric like those
based on z or nonparametric like those based on chi-square. Permutation
tests (Edgington, 1987; Good, 1994), although less well known, provide
an alternative. Such tests construct sampling distributions from the data
observed. No reference is made to another, theoretic distribution, so no
minimum-data assumption to justify the reference is required.
Especially when data are few, investigators should consider permutation
tests as an alternative to the usual asymptotic ones. They yield an exact,
instead of an asymptotic, p value, and render minimum-data requirements
unnecessary. If data are few, statistical significance will still be unlikely,
but that is at it should be. Interested readers are urged to consult Bakeman,
Robinson, and Quera (1996) for more details concerning permutation tests
in a sequential context. But no matter whether asymptotic or permutation
tests are used, you still should expect that the number of events coded will
often need to number in the thousands.

8.6 The type I error problem

Even when enough data are collected to justify significance testing for the
various scores computed, the problem of type I error - of claiming that se-
quences are "significant" when in fact they are not - remains. The reason
type I error is such a problem with sequential analyses such as those de-
scribed earlier is that typically investigators have many codes. These many
codes generate many more possible sequences, especially when anything
longer than two-event sequences is considered. The number of sequences
tested for significance can rapidly become astronomical, in which case the
probability of type I error (the alpha level) approaches certainty. When tests
are independent, and the alpha level for each is .05, the "investigationwise"
148 Issues in sequential analysis

or "study wise" alpha level, when k tests are performed, is 1 — .95* (Cohen &
Cohen, 1983). Thus if 20 independent tests are performed, the probability
of type I error is really .64, not .05.
What each study needs is a coherent plan for controlling type I error.
The best way, of course, is to limit drastically the number of tests made.
And even then, it makes sense to apply some technique that will assure a
desired study wise alpha level. For example, if k tests are performed and a
studywise alpha level of .05 is desired, then, using Bonferroni's correction,
the alpha level applied to each test should not be alpha, but rather alpha
divided by k (see Miller, 1966). Thus if 20 tests are performed, the alpha
level for each one should be .0025 (.05/20).
When studies are confirmatory, type I error usually should not be a
major problem. Presumably in such cases the investigator is interested in
(and will test for significance) just a few theoretically relevant sequences.
Exploratory studies are more problematic. Consider the parallel play study
discussed earlier. Only five codes were used, which is not a great number
at all. Yet these generate 20 possible two-event and 80 possible three-
event sequences. This makes us think that unless very few codes are used
(three or four, say) and unless there are compelling reasons to do so, most
exploratory investigations should limit themselves to examining just two-
event sequences, no longer - even if the amount of data is no problem.
Even when attention is confined to two-event sequences, the number of
codes should likewise be limited. For two-event sequences, the number
of possible sequences, and hence the number of tests, increase roughly as
the square of the number of codes. For this reason, we think that coding
schemes with more than 10 codes border on the unwieldy, at least when
the aims of a study are essentially exploratory.
Two ways to control type I error were described in section 7.6 when dis-
cussing log-linear approaches to lag-sequential analysis. First, exploratory
studies should not fish for effects at lag L in the absence of significant lag
L omnibus tests. And second, the set of seemingly significant sequences
at lag L should be winnowed into a smaller subset that can be viewed as
responsible for the model of independence's failure to fit. Still, as empha-
sized in section 7.7 when discussing Yule's Q, guiding ideas provide the
best protection against type 1 error. Investigators should always be alert
for ways to limit the number of statistical tests in the first place.

8.7 Summary
Several issues important if not necessarily unique to sequential analysis
have been discussed in this chapter. Investigators should always worry
 Summary 149

whether summary indices (means, Yule's Q's, etc.) are based on sufficient
data. If not, confidence in computed values and their descriptive value is
seriously compromised. Further, when inferential statistics are used, data
sufficient to support their assumptions are required. Guidelines based on
log-linear analyses were presented here, but the possibility of permutation
tests, which require drastically fewer assumptions, was also mentioned.
Again, investigators should always limit the number of statistical tests in
any study, else they court type I error. Of help here is the discipline provided
by guiding ideas and theories, clearly stated. In contrast, issues of pooling
may arise more in sequential than other sorts of analyses because of data
demands. Pooling data over units such as individuals, dyads, families, etc.,
is rarely recommended, no matter how necessary it seems. When data per
unit are few, a jackknife technique (computing several values for a sum-
mary statistic, each with data for a different unit removed, then examining
the distribution for coherence) is probably better than pooling. Finally,
common to almost all statistical tests is the demand for independence (or
exchangeability; see Good, 1994). When two-event chains are sampled
in an overlapping manner from longer sequences, this requirement might
seem violated, but simulation studies indicate that the apparent violation
in this particular case does not seem consequential.
 Analyzing time sequences

9.1 The tyranny of time

Events unfold in time. What distinguishes users of sequential analysis
from many other researchers is that they attempt to move beyond this banal
observation and try to capture aspects of this unfolding in quantifiable ways.
For such purposes, it is often sufficient to know just the order of events, and
to use the techniques for analyzing event sequences discussed in chapter 7.
Often, however, we want to know more than just the order of events.
We want to know in addition how people (dyads, animals, etc.) spent their
time. For that reason, it is common for investigators to record not just
the order of events, but their times as well. In chapter 3 we described
three ways such time information could be recorded (timing onsets and
offsets, timing pattern changes, interval recording), and in chapter 5 we
described three ways data could be represented preserving time (state se-
quences, timed-event sequences, interval sequences). Once recorded and
represented, however, sometimes time information exerts a tyrannical hold
on investigators who then seem reluctant to omit time from their analyses,
even when this would be appropriate.
For many studies, especially when behavioral states are coded, we
think time is worth recording primarily because "time-budget" informa-
tion (amounts or percentages of time devoted to different kinds of events or
behavioral states) has such descriptive value. For example, Bakeman and
Adamson (1984), in their study of infants' attention to objects and people,
recorded onset times for behavioral states, represented these data as state
sequences, and computed and reported percentages of time devoted to the
various behavioral states when with different partners (mothers or peers)
at different ages.
However, when examining how behavioral states were sequenced, Bake-
man and Adamson ignored time, in effect reducing their state-sequential
data to event sequences. This approach has merit whenever investigators
want to describe how a simple stream of events or states unfolded in time.
For example, if state sequences are analyzed instead of event sequences,

150
 Taking time into account 151

using a time unit instead of the event as the basic unit of analysis, values for
transitional probabilities are affected by how long particular events lasted
- which is undesirable if all the investigator wants to do is describe the
typical sequencing of events.
As an example, recall the study of parallel play introduced earlier, which
used an interval recording strategy (intervals were 15 seconds). Let U =
unoccupied, S = solitary, and P = Person Play. Then:
Interval = 15; U, U, U, S, P, P . . .
Interval = 15; U, S, S, S, P, P . . .
Intervals 15; U, U, S, S, P, P . . .
represent three slightly different ways an observational session might have
begun. All three interval sequences clearly represent one event sequence:
Unoccupied to Solitary to Parallel. Yet values for transitional probabilities
and their associated z scores would be quite different for these three interval
sequences. For example, the p(Pt+\\St) would vary from 1.00 to 0.33 to
0.50 for the three sequences given above. Worse, the z score associated with
0.33 would be negative, whereas the other two z scores would be positive.
No one would actually compute values for sequences of just six intervals, of
course, but if we had analyzed longer interval sequences like these with the
techniques described in chapter 7, it is not clear that the USP pattern would
have been revealed. Very likely, especially if each interval had represented 1
second instead of 5 seconds, we might have discovered only that transitions
from one code to itself were likely, whereas all other transitions were
unlikely. In short, we urge investigators to resist the tyranny of time. Even
when time information has been recorded, it should be "squeezed out"
of the data whenever describing the typical sequencing of events is the
primary concern. In fact, the GSEQ program (Bakeman & Quera, 1995a)
includes a command that removes time information, thereby transforming
state, timed-event, or interval sequences into event sequences when such
is desired.

9.2 Taking time into account

The simple example of a USP sequence presented in the previous section
hints at the underlying unity of the four methods of representing data de-
scribed in chapter 5. This unity, in turn, suggests a welcome economy.
If we have already learned a number of techniques for analyzing event
sequences (in chapter 7), and if state, timed-event, and interval sequences
are logically the same as event sequences, then analyzing time sequences
(sequential data that take time into account, i.e., state, timed-event, and
152 Analyzing time sequences

interval sequences) does not require learning new techniques, only the
application of old ones.
In fact, one underlying format suffices for event, state, timed-event, and
interval sequences. These four forms are treated separately both here and
in the Sequential Data Interchange Standard (SDIS; see Bakeman & Quera,
1992) because this connects most easily with what investigators actually
do, and have done historically. Thus the four forms facilitate human use
and learning. A general-purpose computer program like GSEQ, however,
is better served by a common underlying format because this allows for
greater generality and hence less specific-purpose computer code. Indeed,
the SDIS program converts SDS files (files containing data that follow
SDIS conventions) into a common format (called MDS or modified SDS
files) that is easily read by GSEQ (Bakeman & Quera, 1995a).
The technical details need not concern users of these computer programs,
but understanding the conceptual unity of the four forms can be useful.
Common to all four is an underlying metric. For event sequences, the
underlying metric is the discrete event itself. For state and timed-event
sequences, the underlying metric is a unit of time, often a second. And for
interval sequences, the underlying metric is a discrete interval, usually (but
not necessarily) defined in terms of time.
The metric can be imagined as cross marks on a time line, where the
space between cross marks is thought of as bins to which codes may be
assigned, each representing the appropriate unit. For event sequences, one
code and one code only is placed in each bin. Sometimes adjacent bins may
be assigned the same code (consecutive codes may repeat), sometimes not
(for logical reasons, consecutive codes cannot repeat). For state sequences,
one (single stream) or more codes (multiple streams) may be placed in each
bin. Depending on the time unit used and the typical duration of a state,
often a stretch of successive bins will contain the same code. For timed-
event sequences, one or more codes or no codes at all may be placed in each
bin. And for interval sequences, again one or more codes or no codes at all
may be placed in each bin. As you can see, the underlying structure of all
forms is alike. Successive bins represent successive units and, depending
on the form, may contain one or more or no codes at all.
Interval sequences, in particular, can be quite useful, even when data
were not interval recorded in the first place. For example, imagine that
a series of interactive episodes are observed for particular children and
that attributes of each episode are recorded (e.g., the partner involved, the
antecedent circumstance, the type of the interaction, the outcome). Here
the event (or episode) is multidimensional, so the event sequential form
is not adequate. But interval sequences, which permit several codes per
bin, work well, and permit both concurrent (e.g., are certain antecedents
 Micro to macro 153

often linked with particular types of interaction) and sequential (e.g., are
consequences of successive episodes linked in any way) analyses. Used
in this way, each episode defines an interval (instead of some period of
elapsed time); in such cases, interval sequences might better be called
multidimensional events. Further examples of the creative and flexible use
of the four forms for representing sequential data are given in Bakeman
and Quera (1995a, especially chapter 10).
Because the underlying form is the same for these four ways of represent-
ing sequential data, computational and analytic techniques are essentially
the same (primarily those described in chapter 7). New techniques need not
be introduced when time is taken into account. Only interpretation varies,
depending on the unit, whether an event or a time unit. For event sequences,
coders make a decision (i.e., decide which code to assign) for each event.
For interval sequences, coders make decisions (i.e., decide which codes
occurred) for each interval. For state and timed-event sequences, there is
no simple one-to-one correspondence between decisions and units. Coders
decide and record when events or states began and ended. They may note
the onset of a particular event, the moment it occurs. But just as Charles
Babbage questioned the literal accuracy of Alfred Lord Tennyson's cou-
plet "Every minute dies a man, / Every minute one is born" (Morrison &
Morrison, 1961), so too we should question the literal accuracy of a claim
that observers record onsets discretely second by second and recognize the
fairly arbitrary nature of time units. For example, we can double tallies by
changing units from 1 to 1/2 second. The connection, or lack of connec-
tion, between coders' decisions and representational units is important to
keep in mind and emphasize when interpreting sequential results.

9.3 Micro to macro

In this section we would like to share some wisdom based on our expe-
rience doing programmatic observational research with the same kinds of
data for over a decade. Sequential analysis is interesting because so much
theoretical clarity about interacting people is provided by the study of tem-
poral patterns. Often when we have begun working in an area, we start
with fairly small units and a large catalog of precise codes. A microana-
lytic description often is the product of these initial efforts (e.g., Brown,
Bakeman, Snyder, Fredrickson, Morgan, & Hepler, 1975).
Sequential analysis of the micro codes then can be used to identify in-
dexes of more complex social processes (e.g., Bakeman & Brown, 1977).
For example, Gottman (1983) found that a child's clarification of a message
after a request for clarification (Hand me the truck/ Which truck?/ The red
154 Analyzing time sequences

truck) was an index of how connected and dialogic the children's conver-
sations were. This sequence thus indexed a more macro social process.
Gottman could have proceeded to analyze longer sequences in a statisti-
cal fashion, but with 40 codes the four-element chain matrix will contain
2,560,000 cells! Most of these cells would have been empty, of course, but
the task of even looking at this matrix is overwhelming. Instead, Gottman
designed a macro coding system whose task it was to code for sequences,
to code larger social processes. The macro system used a larger interaction
unit, and it gave fewer data for each conversation (i.e., fewer units of ob-
servation). However, the macro system used a larger interaction unit, and
it gave fewer data for each conversation (i.e., fewer units of observation).
However, the macro system was extremely useful. First, it was far more
rapid to use than the micro system. Second, because a larger unit was now
being used, new kinds of sequences were being discovered. This revealed
an organization of the conversations that Gottman did not notice, even with
the sequential analysis of the micro data (see Gottman & Parker, 1985 in
press).
To summarize, one strategy we recommend for sequential analysis is not
looking for everything that is patterned by employing statistical analysis
of one data set. It is possible to return to the data, armed with a knowledge
of patterns, and to reexamine the data for larger organizational units.

9.4 Time-series analysis

Time-series analysis offers a wide range of analytic options (see Gottman,
1981, for a comprehensive introduction), and, furthermore, it is possible to
create time-series data from a categorial stream of codes, as we mentioned
in section 5.7. In this section we shall review a few of the advantages
provided by time-series analysis, and discuss further how to create time-
series data from a stream of categorical data.
What are the advantages of creating time-series data from a stream of
categorial data? First, one can obtain an overall visual picture of the inter-
action. This can be useful in two ways: One use is to create taxonomies of
interactions. For example, Gottman (1979b) created a time series from data
obtained from marital interaction (see Figure 9.1). The variable was the
total positive minus the total negative interaction up to that point in time.
In some couples, both the husband's and wife's graphs were quite negative.
These couples tended to be high in reciprocating negative affect. In some
couples, one partner's graph was negative and the other's was positive.
These couples tended to have one partner who gave in most of the time in
response to the partner's complaints. Couples whose graphs were flat at
 Time-series analysis 155

4
COUPLE NO. 27 IMPROVISATION NO. 3 CLINIC
0
- HUSBAND
- WIFE -4

-8

-12

-16

-20

-24

-28

-32

-36 -36
40 80 120
FLOOR SWITCHES

COUPLE NO. 29 IMPROVISATION NO. 2 CLINIC

12

., HUSBAND
-4 t WIFE -4

40 80 120
FLOOR SWITCHES

Figure 9.1. Time series for two clinic couples. From Gottman (1979b, p. 215).
156 Analyzing time sequences

the start of an interaction tended to have social skill deficits in listening,

whereas couples whose graphs were flat at the end of an interaction tended
to have social skill deficits in negotiation.
A second use of a time-series graph is that it makes it possible to discover
the limitations of a coding system. One can use the graphs to scan for
shifts in slope or level. These may be rare critical events that the coding
system itself does not know about. An example of this comes from a
videotape that Ted Jacob had which was coded using the Marital Interaction
Coding System (MICS). The interaction began in a very negative way but
changed dramatically in the middle and became quite positive. The event
that triggered the change seemed to be the husband's summarizing what
he thought was the wife's complaint and then accepting responsibility for
the problem. The MICS has no code for summarizing the other (which is
a very rare event, but quite powerful); it does have a code for accepting
responsibility, but this was miscoded on this tape. Despite the fact that
the critical event was missed by the coding system, time-series analysis of
the data detected the shift in the overall positivity of the interaction and
pinpointed the time of the switch. Gottman refers to this use of time series
as "Godeling" because, like Kurt Godel's work, it is concerned with using
a system to view itself and discover its own limitations.

Creating time series from categorical data

In addition to these reasons for creating time-series data from categorical
data, time-series analysis has some powerful analytic options, which we
shall discuss briefly in a moment. First we would like to mention three
options for creating time-series data from categorical data.
One option, used by psychophysiologists, is the interevent interval. This
involves graphing the time between events of a certain type over time.
Cardiovascular psychophysiologists, for example, plot the average time
between heart beats within a time block of sufficient size. Interevent in-
tervals can be computed with quite simple data. For example, imagine we
had asked a smoker to keep a diary, noting each time he or she smoked a
cigarette, and that a small portion of the data looks like that portrayed in
Table 9.1. (For convenience, all times are given to the nearest 6 minutes or
0.1 hour.) Such interevent intervals are often averaged in some way. For
example, if we used 2-hour time blocks, the averaged interevent intervals
would be as given in Table 9.1. These scores are then available for graphing
or subsequent time-series analysis.
A second option is the moving probability-window. Here we compute
the proportion of times an event was observed within a particular block of
 Time-series analysis 157

Table 9.1. Computing interevent interval time series

Interevent Block
Time interval Block average

7:54 a.m.
2:12(2.2)
10:06
0:42(0.7) 10-12 1.45
10:48
1:30(1.5)
12:18
0:24 (0.4) 12-2 0.95
12:42
1:48(1.8)
2:30
1:18(1.8) 2-4 1.55
3:48

Note: Interevent intervals are given both in hours : minutes and in decimal hours.

observations, and then we slide the window forward in time. This option
smooths the data as we use a larger and larger time unit, a useful procedure
for graphical display, but not necessary for many time-series procedures
(particularly those in the frequency domain).
A third option is the univariate scaling of codes. For two different ap-
proaches to this, see Brazelton, Koslowski, and Main (1974), and Gottman
(1979b).
Each option produces a set of time series for each variable created, for
each person in the interacting unit. Analysis proceeds within each in-
teracting unit ("subject"), and statistics of sequential connection are then
extracted for standard analysis of variance or regression. A detailed ex-
ample of the analysis of this kind of data obtained from mother-infant
interaction appears in Gottman, Rose, and Mettetal (1982). A review of
time-series techniques is available in Gottman's (1981) book, together with
10 computer programs (Williams & Gottman, 1981).

Brief overview of time-series analysis

Brazelton, Koslowski, and Main (1974) wrote about the interactive cyclicity
and rhythmicity that they believe characterizes the face-to-face play of
mother and infant. They described a cycle of attention followed by the
withdrawal of attention, with each partner waiting for a response from the
158 Analyzing time sequences

other. They described the interactional energy building up, then ebbing and
cycling in synchrony. Part of this analysis had to do with their confidence
in the validity of their time-series variable. They wrote:
In other words, the strength of the dyadic interaction dominates the meaning of
each member's behavior. The behavior of any one member becomes a part of
a cluster of behaviors which interact with a cluster of behaviors from the other
member of the dyad. No single behavior can be separated from the cluster for
analysis without losing its meaning in the sequence. The effect of clustering and
of sequencing takes over in assessing the value of particular behaviors, and in
the same way the dyadic nature of interaction supercedes the importance of an
individual member's clusters and sequences, (p. 56)

Time-series analysis is ideally suited to quantitative assessments of such

descriptions. It is a branch of mathematics that began to be developed
in the mid-1700s, on the basis of a suggestion by Daniel Bernoulli, and
later developed into a theorem by Jean Baptiste Joseph Fourier in 1822.
The idea was that any continuous function could be best approximated in a
least-squares sense by a set of sine and cosine functions. At first this idea
seemed counterintuitive to mathematicians. However, it is true, and it is
true even if the function itself is not periodic. Sines and cosines have an
advantage over polynomials, because polynomials tend to wander off to
plus or minus infinity after the approximation period, which is very poor
for extrapolation if one believes that the data have any repetitive pattern.
This is usually the case in our scientific work. For example, we tend to
believe that if we observe the brightness fluctuations of a variable star, it
really doesn't matter very much if we begin observing on a Monday or a
Tuesday; we believe that the same process is responsible for generating the
data, and that it has a continuity or stability (which, in time-series language
is referred to as "stationarity").
Fourier proved his famous theorem incorrectly, and proving it correctly
took the best mathematical minds over a century; furthermore, it has led
to the development of much of a branch of modern mathematics called
"analysis." Time-series analysis underwent a major conceptual revolution
in the 20th century because of the thinking of an American, Wiener, and a
Russian, Khintchine.
Without time-series analysis, attempts to describe cyclicity and syn-
chronicity end in a hopeless muddle of poetic metaphor about interactive
patterns. For example, Condon and Ogston (1967) tried to summarize 15
minutes of the dinner time interaction of one family. They wrote:
We are dealing with ordered patterns of change during change, which exhibit
rhythmic and varying patterns in the temporal sequencing of such changes.
Metaphorically, there are waves within waves within waves, with complex yet
 Time-series analysis 159

determinable relationships between the peaks and troughs of the levels of waves,
which serve to express organized processes with continually changing relation-
ships, (p. 224)

Time-series analysis offers us a well-elaborated set of procedures for

passing beyond metaphor to a precise analysis of cyclicity, synchronicity,
and the analysis of the relationship between two time series (called "cross
correlation"), controlling for regularity within each time series (called "au-
tocorrelation"), among other options.

The notion of cycles

A good way to begin thinking of cycles is to imagine a pendulum moving
back and forth. The amplitude of the oscillation of the pendulum is related
to the energy with which it is shoved. In fact, the variance of the pendulum's
motion is proportional to this energy, and this is proportional to the square
of the amplitude. The period of oscillation of the pendulum is the time it
takes for the pendulum to return to the same spot; it is usually measured
as the time from peak to peak of oscillations. Now imagine that we attach
a second pendulum to the first and permit them to be able to oscillate
independently. We can generate very complex oscillations just with the
oscillations of two pendula. A new variable enters into the picture when
we imagine two pendula, the relative phase of oscillation of the two. They
can be moving in synchrony, or exactly opposite (180 degrees out of phase),
or somewhere in between. So now we have three parameters: the energy
of each pendulum (proportional to the amplitude squared), the frequency
of oscillation of the pendulum (which is the reciprocal of the period of
oscillation), and the relative phases of the pendula. These are the basic
dimensions of what is called "frequency domain" time-series analysis.

Intuitive motions of the spectrum

What is the spectrum? Imagine Isaac Newton holding a prism through
which white light passes on one side and the rainbow emerges from the
other. The rainbow is called the spectrum; in general, it is the resolution
of some incoming wave into its basic components. In the case of white
light, all colors (or frequencies) are present in all brightnesses (energies,
variances). For different kinds of oscillations, some of the colors would
be missing entirely, some would be weaker, some would be stronger. Fur-
thermore, the phase relationships of the different colors could vary. The
160 Analyzing time sequences

Table 9.2. Guessing that the period t = 5

0.00 0.95 0.59 -0.59 -0.95

0.00 0.95 0.59 -0.59 -0.95
0.00 0.95 0.59 -0.59 -0.95
0.00 0.95 0.59 -0.59 -0.95
0.00 0.95 0.59 -0.59 -0.95
0.00 0.95 0.59 -0.59 -0.95

Means 0.00 0.95 0.59 -0.59 -0.95

Table 9.3. Guessing incorrectly that the period t — A

1 2 3 4

0.00 0.95 0.59 -0.59

0.95 0.00 0.95 0.59
0.59 -0.95 0.00 0.95
0.59 -0.59 -0.95 0.00
0.95 0.59 -0.59 -0.99

Means 0.00 0.00 0.00 0.00

spectrum, or the "spectral density function," tells us only which frequen-

cies are present and to which degree; that is, we learn how much variance
each frequency accounts for in the given time series.
We shall illustrate how this spectral density function is computed by
using an old 19th-century method, called the periodogram (a technique
no longer used). Suppose we generate a very simple time series, xt =
sin( 1.257f), and let t go from 0 to 10. The values of the time series are
0.00, 0.95, 0.59, -0.59, -0.95, 0.00, 0.95, 0.59, -0.59, -0.95, 0.00.
Suppose we did not know that these data repeated every five time intervals.
Suppose we guess at the period and guess correctly that the period is 5, and
we arrange the data as shown (Table 9.2). In this table, we can see that the
variance of the series is equal to the variance of the means. The ratio of
these two variances is always one when we have guessed the right period.
Suppose we had guessed the wrong period, say t = 4. This situation is
illustrated in Table 9.3.
In the case of Table 9.3, the variance of the means is zero, so that the
ratio of the variance of the means to the variance of the series is zero. If we
 Time-series analysis 161

were to plot the value of this ratio for every frequency we guess (remember
frequency equals l/t)9 this graph is an estimate of the spectral density
function. Of course, this is the ideal case. In practice, there would be a lot
of noise in the data, so that the zero values would be nonzero, and also the
peak of the spectral density function would not be such a sharp spike. This
latter modification in thinking, in which the amplitudes of the cycles are
themselves random variables, is a major conceptual revolution in thinking
about data over time; it is the contribution of the 20th century to this area.
(For more discussion, see Gottman, 1981.)

Rare events
One of the uses of univariate time-series analysis is in evaluating the effects
of rare events. It is nearly impossible to assess the effect of a rare but
theoretically important event without pooling data across subjects in a
study by the use of sequential analysis of categorical data. However, if we
create a time-series variable that can serve as an index of the interaction,
the problem can be solved by the use of the interrupted, time-series quasi-
experiment.
What we mean by an "index" variable is one that is a meaningful the-
oretical index of how the interaction is going. Brazelton, Koslowski, and
Main (1974) suggested an index time series that measured a dimension of
engagement and involvement to disengagement. The dimension assessed
the amount of interactive energy and involvement that a mother expressed
toward her baby and that the baby expressed toward the mother. This is an
example of such an index variable. Gottman (1979) created a time-series
variable that was the cumulative positive-minus-negative affect in a mari-
tal interaction for husband and wife. The interactive unit was the two-turn
unit, called a "floor switch." Figure 9.1 illustrates the Gottman index time
series for two couples.
Now suppose that the data in Figure 9.2 represented precisely such a
point graph of a wife's negative affect, and that a rare but interesting event
occurred at time 30, when her husband referred to a previous relationship
he had had. We want to know whether this event had any impact on
the interaction. To answer this question, we can use an interrupted time-
series analysis. There are many ways to do this analysis (see Gottman,
1981). We analyzed these data with the Gottman-Williams program ITSE
(see Williams & Gottman, 1981) and found that there was no significant
change in the slope of the series |>(32) = —0.3], but that there was a
significant effect in the change in level of the series [t@2) = 5.4]. This is
one important use of time-series analysis.
162 Analyzing time sequences

Time

Figure 9.2. Interrupted time-series experiment.

Cyclicity
Univariate time-series analysis can also answer questions about the cyclicity
of the data. Recall that to assess the cyclicity of the data, a function called
the "spectral density function" is computed. This function will have a
significant peak for cycles that account for major amounts of variance
in the series, relative to what we might have expected if there were no
significant cycles in the data (i.e., the data were noise). We used the data
in Figure 9.2 for this analysis, and used the Gottman-Williams program
SPEC. The output of SPEC is relatively easy to interpret. The solid line
in Figure 9.3 is the program's estimate of the spectral density function,
and the dotted line above and below the solid line is the 0.95 confidence
interval. If the entire confidence interval is above the horizontal dashed
line, the cycle is statistically significant. The x axis is a little unfamiliar to
most readers, because it refers to "frequency," which in time-series analysis
means cycles per time unit. It is the reciprocal of the period of oscillation.
The peak cycle is at a frequency of 0.102, which corresponds to a period
of 9.804 time periods. The data are cyclic indeed. It is important to realize
that cyclicity in modern time-series analysis is a statistical concept. What
we mean by this is that the period of oscillation is itself a random variable,
with a distribution. Thus, the data in Figure 9.3 are almost periodic, not
precisely periodic. Most phenomena in nature are actually of this sort.
 Time-series analysis 163

.3960 r

.3448

.2928

.t? .2407

I
CO

& .1*87

.1367

.0847

.0327

Frequency

Figure 9.3. Plot of density estimates. The density is indicated by a solid line, and
the 0.95 confidence interval by a dotted line. The white-noise spectrum is shown
by a dashed line.

Multivariate time-series analysis

There are quite a few options for the multivariate analysis of time-series
data. We shall discuss only one option here, a bivariate time-series analysis
that controls for autocorrelation (predictability within each time series) in
making inferences about cross correlation between two time series. This
option is discussed in a paper by Gottman and Ringland (1981).
164 Analyzing time sequences

• Turns
i n 1/111111fti111111111111111 n 11 ii ii 1111

<
-10-

-15-

u
-20
Figure 9.4. Couple mp48 conflict discussion. The solid line indicates the hus-
band's affect; the broken line, the wife's affect.

The data in Figure 9.4 are a summary of the interactions of a happily

married couple discussing a marital conflict. The y axis is a cumulative
sum of positive minus negative affect in the discussion. This graph is fairly
typical of happily married couples' discussions of an area of continued
disagreement. The central third of the conversation shows a downward drift
because most disagreements occur in the middle phase of the discussion;
the final third shows the graph coming back up as the couple moves closer
to resolving the issue. In making an inference of influence from one partner
to another, what this procedure does is to attempt to predict as well as it
can from the parts of each series and then see if any additional information
is provided by the other series.
Table 9.4 summarizes the results of this analysis, which was accom-
plished with the program BIVAR from the Gottman-Williams package
(see Williams & Gottman, 1981). Row 1 of the table shows that the initial
starting values for the models is 8; this means that we shall go back 8 units
into the past; this number is arbitrary, but should probably not exceed 10
(it is limited by the total number of observations). The second row shows
results of the program's successive iterations to find the smallest model that
loses no information; this is the model with one autoregressive term for the
husband and four cross-regressive terms for the wife. The test of model 1
 Time-series analysis 165

Table 9.4. Bivariate-time series analysis of couple mp48's data

Model Husband Wife SSE T LN(SSE/T)«

1 8 8 133.581 44.499
2 1 4 150.232 52.957
3 1 0 162.940 58.803

1 vs 2: Q = 8.458 df = 11 z = -.542
2vs3: 2 = 5.846 df = 4 z = .653

Model Husband Wife SSE T LN(SSE/T)

207.735 76.291
224.135 81.762
298.577 102.410

Ivs2: 2 = 5.471 df = 9 z = -.832

2vs3: 2 = 20.648 *//= 2 z = 9.324

a
Weighted error variance; see Gottman and Ringland (1981), p. 411.

versus model 2 should be nonsignificant, which is the case. The third row
of the table shows the model with all of the wife's cross-regressive terms
dropped. If this model is not significantly different from model 2, then we
cannot conclude that the wife influences the husband; we can see that the
z score (0.653) is not significant. Similar analyses appear in the second
half of the table for determining if the husband influences the wife; the
comparison of models 2 and 3 shows a highly significant z score (this is the
normal approximation to the chi-square, not to be confused with the z score
for sequential connection that we have been discussing). Hence we can
conclude that the husband influences the wife, and this would be classified
as a husband-dominant interaction according to Gottman's (1979) defini-
tion. We note in passing that although these time series are not stationary,
it is still sensible to employ BIVAR.

Interevent Interval
In the study of the heart, in analyzing the electrocardiogram (ECG), there
is a concept called "interbeat interval," or the "IBI." It is the time between
the large R-spike of the ECG that signals the contraction of the ventricles
of the heart. It is usually measured in milliseconds in humans. So, for
example, if we recorded the ECG of a husband and wife talking to each
166 Analyzing time sequences

other about a major issue in their marriage, and we took the average of the
IBIs for each second (this can be weighted or prorated by how much of
the second each IBI took up), we would get a time series that was a set
of consecutive IBIs (in milliseconds) that looked like this: 650, 750, 635,
700, 600, 625, 704, and so on.
We wish to generalize this idea of IBI and discuss a concept we call
the "interevent interval." This is like an old concept in psychology, the
intertrial interval. When we are recording time, which we get for free with
almost all computer-assisted coding systems, we also can compute the time
between salient events, and these times can become a time series. When
we do this, we are interested in how these interevent times change with
time within an interaction.

Other ways of transforming observational data to a time series

When the observational data consist of ratings, the data are already in the
form of a time series. For example, if we coded the amount of positive
affect in the voice, face, body, and speech of a mother during parent - child
interaction, we can add the ratings on each channel and obtain an overall
time series.

When the data are categorical codes

There are other ways to obtain a time series from a stream of categorical
codes. One method is the idea of a "moving probability window," which
would be a window of a particular sized time block (say 30 seconds) within
which we compute the frequencies of our codes (estimating their probabili-
ties); then the window moves forward one time unit, and we compute these
probabilities again. Another approach we have used (Gottman, 1979b;
Gottman & Levenson, 1992), as has Tronick, Als, and Brazelton (1977;
1980) is to weight the codes along some composite dimension. In the Tron-
ick case the dimension was a combination of engagement/disengagement
and positive/negative, so that positive scores meant engaged and/or with
positive affect, and negative scores meant disengaged and/or with negative
affect. In the Gottman cases, positive and negative codes were given pos-
itive or negative integer weights, and the number of total positive minus
negative points was computed for each turn at speech (a turn lasts about 6
seconds on marital interaction). We have found it very fruitful for visual
inspection of a whole interaction to cumulate these time series as the inter-
action proceeds. Figure 9.5 (a) and (b) shows these time series cumulated
 Autocorrelation and time-series analysis 167

to illustrate two very different marital interactions. Losada, Sanchez and

Noble (1990), in their research on six-person executive work groups, code
power and affect on 3- point scales and then turn these into numerical scores
within time blocks. They then compute directional cross-correlations be-
tween people, and use these numbers to create an animated graphic of a
"moving sociometric" of affect and power over time. This moving so-
ciometric is time-locked so that the researchers can provide instant replay
video feedback to the group so that the group can see a video illustration
of their moving sociometric.

Why transform the data to a time series?

Reducing the observational coding to these summary time-series graphs
is very profitable. In marital interaction having these time series made it
possible for us (see Cook et al., 1995) to construct a mathematical model of
the data that led us to a new theoretical language for describing our divorce
prediction data and led us to a new discovery about what predicts divorce.

9.5 Autocorrelation and time-series analysis

Autocorrelation function
We will begin by examining the nature of the dependent time-structure
of time-series data. To accomplish this we start by examining what is
called the "autocorrelational structure" of the time series. This gives us
information about how predictable the present data are from its past. To
explain the concept of "lag-1" autocorrelation, we draw a lag-1 scatterplot
of the data, where the jc-axis is the data at time t, and the j-axis is the data at
time t +1. This means that we plot the data in pairs. The first point we plot
is (x 1, JC2); the second point is (JC2, JC3); the third point is (x3, x4); and so
on. This gives a scatterplot similar to the ones we plot when we compute a
regression line between two variables. The correlation in this case is called
the "lag-1 autocorrelation coefficient," r\. In a similar fashion we can pair
points separated by two time points. The pairs of points would then be
(JC 1, JC3), (JC2, JC4), (JC3, JC5), ...; the two axes are x(t) and x(t + 2), and the
autocorrelation coefficient would be the "lag-2 autocorrelation coefficient,"
7*2. We can continue in this fashion, plotting r*, the lag-/: autocorrelation
coefficient against the lag, k. The autocorrelation function is useful in
identifying the time-series model.
168 Analyzing time sequences

A regulated couple

20 40 60 80 100 120 140

Turns at speech

Husband Wife

A nonregulated couple
20
0 -
c -20 -
1
C/)
o -40 -
Q_
-60 -
ted

-80 -

1
o
o
-100 -
-120 -
-140 -
-160
20 40 60 80 100 120 140 160
Turns at speech

Husband Wife

Figure 9.5. (a) Cumulative point graphs for a regulated couple, for which pos-
itive codes generally exceed negative codes, (b) Cumulative point graphs for a
nonregulated couple, for which negative codes generally exceed positive codes.
 Autocorrelation and time-series analysis 169

Model identification
Using the autocorrelation function, an autoregressive model for the time
series can be identified exactly if the series is stationary (this means that
the series has the same correlation structure throughout and no local or
global trends) using a computer program (see the Williams & Gottman,
1981, computer programs). A wide variety of patterns can be fit using the
autoregressive models, including time series with one or many cycles.

Spectral time-series analysis

In 1822 Jean Baptiste Fourier, a French mathematician, discovered a very
important theorem (one that took mathematicians over a hundred years to
prove correctly and led to the development of a branch of modern math-
ematics called mathematical analysis). He discovered that any piecewise
continuous function could be fit best (in the least-squares sense of distance)
by a series of specially selected sine and cosine functions. These functions
are, of course, cyclic, but the representation is still the best one ever if the
function being fit is not itself cyclic. This was quite an amazing theorem.
In time-series analysis we can actually use the data to compute the cycles
present in the data, even if these cycles are obscured with noise. This
analysis is called a Fourier analysis, or a "spectral time-series analysis."
In a spectral time-series analysis, the end result is that we generate a plot,
called the "spectral density function," of the amount of variance accounted
for by each of a set of cycles, from slow to fast. We do not use all cycles,
but only a particular set, called "the overtone series." If a particular time
series were actually composed of two cycles, a slow one and a fast one, the
spectral density function would have peaks at these two cycle frequencies.
Usually, however, real data do not look as spikelike as this figure, but
instead the spectral density function is statistically significant across a
band of frequencies.
We can then actually use these frequencies to fit a function to the data
and use it to model the time series. In most cases this type of modeling
is not feasible, because the models we obtain this way tend to be poor fits
of the data. For this reason we usually use models like the autoregressive
model.

Interrupted time-series experiments

Once we have a time series, and we have established that the time series
itself has some validity (e.g., Gottman and Levenson's time series predicted
170 Analyzing time sequences

divorce versus marital stability), we can model the series, and then we can
scan the time series for statistically significant changes in overall level
or slope. This is called an "interrupted time-series experiment," or ITSE
(see Figure 9.6). An ITSE consists of a series of data points before and
after an event generally called the "experiment." The "experiment" can
be some naturally occurring event, in which case it is actually a quasi-
experiment. We then represent the data before the intervention as one
function b\ +m \t+ Autoregressive term, and the data after the intervention
as &2 + rri2t+ Autoregressive term. We need only supply the data and the
order of the autoregressive terms we select, and the computer program tests
for statistically significant changes in intercept (the 6's) and slope (the m's).
The experiment can last 1 second, or just for one time unit, or it can last
longer. One useful procedure is to use the occurrence of individual codes
as the event for the interrupted time-series experiment. Thus, we may ask
questions such as "Does the wife's validation of her husband's upset change
how he rates her?" Then we can do an interrupted time-series experiment
for every occurrence of Wife Validation. For the data in Figure 9.6, there
were 48 points before and 45 points after the validation. The order of the
autoregressive term selected was about one tenth of the preintervention
data, or 5. The t for change in intercept was t(19) = —2.58, p < .01, and
the t for change in level was f(79) = -1.70, p < .05.
For this example, we used the Williams and Gottman (1981) computer
program ITSE to test the statistical significance of changes in intercept and
slope before and after the experiment; an autoregressive model of any order
can be fit to the data. Recently, Crosbie (1995) developed a powerful new
method for analyzing short time-series experiments. In these analyses only
the first-order autoregressive parameter is used, and the preexperiment data
are fit with one straight line (intercept and slope) and the postexperiment
data are fit with a different straight line (intercept and slope). An omnibus
F test and t tests for changes in level and slope are then computed. This
method can be used to determine which codes in the observational system
have potentially powerful impact on the overall quality of the interaction.

Phase-space plots
Another useful way to display time-series data is by using a "phase-space"
plot. In a phase-space plot, which has an x-axis and a j-axis, we plot the
data as a set of pairs of points: (x 1, JC2), (JC2, x3), (x3, x4), The x-axis
is x{t), and the y-axis is x(t + 1), where t is time, t = 1, 2, 3, 4, and so
on. Alternatively, if we are studying marital interaction, we can plot the
interevent intervals for both husband and wife separately, so we have both
 Autocorrelation and time-series analysis 171

Husband rates wife

Time

Figure 9.6. Plot of rating dial in which the husband rated his wife's affect during
their conversation.

a husband and a wife time series. A real example may help clarify how we
might use this idea (Gottman, 1990).

An example from an actual couple who subsequently divorced

For marital interaction many possible behavioral candidates exist for "the
event" selected to be used for computing interevent intervals. As we have
suggested, one promising event to select is negative affect, in part because
it has a reasonably high base rate during conflict discussions, and because it
tends to be high in dissatisfied marriages. Our interest here in computing the
phase-space plot is not whether negative affect is high among an unhappily
married couple, but whether this system is homeostatically regulated and
stable, or whether it is chaotic. In this interaction of a married couple
we computed the time between negative affect, independent of who it was
(husband or wife) who displayed the negative affect. The times between
negative affects were thus selected for analysis. These were interevent
intervals for either partner in a marital interaction (Figure 9.7).
In phase-space plot the data look a lot like the scatterplot for the first-
order autocorrelation coefficient, except for one thing. In the phase-space
172 Analyzing time sequences

Time
Figure 9.7. Interevent integrals for negative affect in a marital conversation.

plot we connect the successive dots with straight lines (see Figures 9.8 and
9.9). This gives us an idea of the "flow" of the data over time.
What does this figure mean in terms of the actual behavior of the marital
system? It means that, insofar as we can ascertain from these data, we have
a system whose energy balance is not stable, but dissipative; that is, it runs
down. Like the pendulum winding down, this system tends toward what is
called an attractor; in this case the attractor represents an interevent interval
of zero. However, for the consequences of energy balance, this movement
toward an attractor of zero interevent interval between negative affect may
be disastrous. Specifically, this system tends, over time, toward shorter
response times for negative affect. Think of what that means. As the
couple talk, the times between negativity become shorter and shorter. This
interaction is like a tennis match where that ball (negative affect) is getting
hit back and returned faster and faster as the game proceeds. Eventually
the system is tending toward uninterrupted negativity.
We can verify this description of this marital interaction using more
standard analytic tools, in this case by performing the mathematical pro-
cedure we discussed called spectral time-series analysis of these IEIs (see
Gottman, 1981). Recall that a spectral time-series analysis tells us whether
there are specific cyclicities in the data, and, if there are, how much variance
each cycle accounts for. See Figure 9.10.
 Autocorrelation and time-series analysis 173

20 -r

20

Data
Time Code IEI
5:14 Husb Anger
5:27 Wife Sad 13 Sees
5:40 Husb Disgust 13 Sees
5:44 Husb Anger 04 Sees
(X 1 ,X 2 ) = (13,13)
(X 2 ,X 3 ) =

Figure 9.8. A scatterplot of interevent interval times with consecutive points

connected.

Note that the overall spectral analysis of all the data reveals very little.
There seem to be multiple peaks in the data, some representing slower and
some faster cycles. However, if we divide the interaction into parts, we
can see that there is actually a systematic shift in the cyclicities. The cycle
length is 17.5 seconds at first, and then moves to 13.2 seconds, and then
to 11.8 seconds. This means that the time for the system to cycle between
negative affects is getting shorter as the interaction proceeds. This is exactly
what we observed in the state space diagram in which all the points were
connected. Hence, in two separate analyses of these data we have been led
to the conclusion that this system is not regulated, but is moving toward
more and more rapid response times between negative affects. From the
data we have available, this interaction seems very negative, relentless, and
unabated. Of course, there may be a more macro-level regulation that we
174 Analyzing time sequences
X
i+2

Figure 9.9. System is being drawn toward faster and faster response times in the
IEI of negative affect.

do not see that will move the system out toward the base of the cone once
it has moved in, and it may oscillate in this fashion. But we cannot know
this. At the moment it seems fair to conclude that this unhappy marriage
represents a runaway system.
There are lots of other possibilities for what phase-space flow diagrams
might look like. One common example is that the data seem to hover quite
close to one or two of what are called "steady states." This means that the
data really are quite stable, except for minor and fairly random variations.
Another common example is that the data seem to move in something
like a circle or ellipse around a steady state. The circle pattern suggests
one cyclical oscillation. More complex patterns are possible, including
chaos (see Gottman, 1990, for a discussion of chaos theory applied to
families); we should caution the reader that despite the strange romantic
appeal that the chaos theory has enjoyed, chaotic patterns are actually
almost never observed. Gottman (1990) suggested that the cyclical phase-
space plot was like a steadily oscillating pendulum. If a pendulum is
steadily oscillating, like the pendulum of a grandfather clock, energy is
constantly being supplied to drive the pendulum, or it would run down (in
phase space, it would spiral in toward a fixed point).
 Summary 175

Overall
Spectrum

o
c
CO

cz
co

I
Frequency

.057 .076 .085

Figure 9.10. Spectrum for three segments showing higher frequencies and shorter
periods.

9.6 Summary
When successive events have been coded (or when a record of successive
events is extracted from more detailed data), event-sequential data result.
When successive intervals have been coded, interval-sequential data result.
And when event or state times have been recorded, the result is timed-event
or state data. At one level, the representation of these data used by the GSEQ
program, these four kinds of data are identical: All consist of successive
bins, where bins are defined by the appropriate unit (event, time, or interval)
and may contain one, more, or no codes.
In general, all the analytic techniques that apply to event sequences can
be applied to state, timed-event, and interval sequences as well, but there
are some cautions.
Primarily, the connection, or lack of connection, between coders' deci-
sions and representational units is important to keep in mind and emphasize
when interpreting sequential results because in some cases units represent
decisions and in other cases arbitrary time units. It is also important to keep
in mind how different forms of data representation suit different questions.
176 Analyzing time sequences

When time is an issue - whether percent of time devoted to certain activities

or average bout lengths for those activities - time must be recorded and
represented, probably using either state or timed-event sequences. When
the simple sequencing of events is an issue, event sequences may suffice,
or an event-sequential version of state or timed-event sequences created
(this can be done with the GSEQ program). Finally, when several dimen-
sions of an event are important, data can be represented using the interval
sequential form but, in this case, might better be termed multidimensional
event sequential data.
 10
Analyzing cross-classified
events

10.1 Describing cross-classified events

Earlier, in chapter 3, we distinguished between continuous and intermit-
tent recording strategies, noting that because this is a book about sequen-
tial analysis, we would emphasize continuous approaches. Four of these
strategies (coding events, recording onset and offset times, timing pattern
changes, coding intervals) are continuous both in the sense that observers
are continuously alert, ready to record whenever required, and in the sense
that the data recorded reflect a continuous record of some aspect of the
passing stream of behavior. Such data are typically represented as event,
state, timed-event, or interval sequences (see chapter 5). We also men-
tioned a fifth strategy (cross-classifying events), which is continuous only
in the sense that observers are continuously alert, ready to record infor-
mation about certain events whenever they occur. This strategy results in
sequential data only if the information coded represents logically sequential
aspects of the event.
This is an appealing strategy for a number of reasons. When an inves-
tigator knows beforehand that a particular kind of event is of interest, this
approach provides focused information about that event with little extrane-
ous detail. Moreover, the data can be represented and analyzed in simple,
well-known, and quite straightforward ways.
As an example, let us again use the Bakeman and Brownlee (1982) study
of object struggles, referred to earlier. The event of interest, of course,
was an object conflict or, to put it in more neutral terms, a "possession
episode." As they defined it, a possession episode required that one child
(the taker) touch and attempt to take away an object currently being held
by another child (the holder). Observations took place in both a toddler
(ages ranged from 12 to 24 months) and a preschool (ages ranged from 40
to 48 months) classroom. In each, a target or focal child strategy was used,
that is, observers would focus first on one child for a period of time (in this
case, 5 minutes), then on another, switching their attention from child to
child according to a predetermined random order. Each time the current

177
178 Analyzing cross-classified events

Table 10.1. Resistance cross-classified by prior possession i2nd

dominance for toddlers and preschoolers
Resistance
Age Prior
group Dominance possession Yes No

Toddlers Yes Yes 19 7

No 42 30
No Yes 16 4
No 61 13
Preschoolers Yes Yes 6 5
No 18 5
No Yes 9 6
No 27 4

: Counts for the toddlers are based on 20.7 hours of observations; for the
preschoolers, on 16.6 hours; see Bakeman and Brownlee (1982).

target child was involved in a possession episode, observers recorded (a)

the name of the taker and the name of the holder; (b) whether the taker
had had "prior possession," that is, had played with the contested object
within the previous minute; (c) whether the holder resisted the attempted
take; and (d) whether the taker was successful in gaining the object.
In addition, a linear dominance hierarchy was determined for the chil-
dren in each class and so, for each episode, it was possible to determine
whether the taker was dominant to the holder. Thus each episode was
cross-classified as follows: taker dominant (yes/no), taker had prior posses-
sion (yes/no), taker resisted (yes/no), and taker successful (yes/no). These
were viewed as sequential, in the order given, although, properly speaking,
dominance is a "background variable" and not a sequential aspect of the
possession episode.
Bakeman and Brownlee chose to regard both resistance and success as
outcomes, as "response variables" to be accounted for by the "explanatory
variables" of dominance and prior possession. As a result, they presented
their data as a series of 2 x 2 x 2 tables (dominance x prior possession
x success or dominance x prior possession x resistance) instead of in-
cluding resistance and success as dimensions in the same table. Because
our purpose here is to discuss, in a general way, how to describe (and an-
alyze) cross-classified event data, we shall use an an example just data on
dominance, prior possession, and resistance, ignoring success.
 Log-linear models: A simple example 179

Describing cross-classified event data is almost embarrassingly simple.

According to time-honored tradition, such data are presented as contin-
gency tables (see Table 10.1). Data can be presented as raw counts or
tallies as here, or various percentages can be computed. For example, 72%
of the possession episodes met with resistance among toddlers; the cor-
responding percentage for preschoolers was 75%, almost the same. We
could then go on and report percentages of possession episodes encoun-
tering resistance when the taker had had prior possession, was dominant,
etc., but a clearer way to present such data is as conditional probabilities.
The simple probability that a take attempt would encounter resistance
was .72 among toddlers. The conditional probability for resistance, given
that the toddler taker had had prior possession, p(/?/P), was actually some-
what higher, .76, but the conditional probability for resistance, given that
the taker was dominant, p(R/D), dropped to .62. This suggests that, at
least among toddlers, the likelihood that the taker will encounter resistance
is affected by dominance but not by prior possession.
Among preschoolers, the situation appears reversed. For them, the sim-
ple probability that a take attempt would meet with resistance was .75. The
conditional probability for resistance, given that the preschooler taker had
had prior possession, was only .58, whereas the conditional probability for
resistance, given that the taker was dominant was .71, close to the simple or
unconditional probability for resistance. Thus preschool takers were less
likely to encounter resistance when attempting to "take back" a contested
object - as though prior possession conferred some current right. However,
this presumed "right" was evident only among the preschoolers, not the
toddlers. (The reader may want to verify that the conditional probability
values were correctly computed from the data given in Table 10.1.)
This is only descriptive data, however. The question now is, are the
differences between simple and conditional probabilities just described
larger than one would expect, based on a no-effects or chance model?

10.2 Log-linear models: A simple example

Cross-classified categorical data like those in Table 10.1 can be analyzed
relatively easily with what is usually called a log-linear modeling approach
(e.g., see Bakeman & Robinson, 1994; Wickens, 1989; for applications
see Bakeman, Adamson, & Strisik, 1989, 1995; Bakeman, Forthman, &
Perkins, 1992). In principle, this approach is fairly simple and flexible.
In addition, results can be expressed in familiar analysis-of-variance-like
terminology. Finally, few assumptions need to be made about the data. For
all these reasons, this approach, or some variety of it, can prove useful.
180 Analyzing cross-classified events

Although there are a variety of different kinds of models (e.g., log-

linear, logistic, etc.), the general approach is the same. The investigator
defines a set of hierarchical modes ("hierarchical" in the sense that simpler
models are proper subsets of more complex ones). The simplest model
- the null or equiprobale model - contains no terms at all and generates
the same expected value for each cell in the contingency table. The most
complex model - the saturated model - contains sufficient terms to gen-
erate expected values for each cell that are identical to the values actually
observed. The idea is to find the least complex model that nonetheless
generates expected values not too discrepant from the observed ones, as
determined by a goodness-of-fit test. Sometimes the investigator begins
with the null or some other simple model. If the data generated by it fail to
fit the actual data, then more complex models are tried. If all else fails, the
saturated model will alwaysfitthe data because it generates values identical
to the observed ones. Alternatively, and more typically, the investigator
begins with the saturated model and deletes terms until the simplest fitting
model is found.
The simplest example of this logic is provided by the familiar chi-square
test of independence, although introductory textbooks rarely present it in
these terms. The model typically tested first - the no-interaction model
- consists of two terms, one for the row variable and one for the column
variable. In other words, the row and the column totals for the data gen-
erated by the model are forced to agree with the row and column totals
actually observed. In introductory statistics, students almost always learn
how to compute these expected frequencies, although they are rarely taught
to think of them as generated by a particular model.
For example, if we look just at the 2 x 2 table detailing prior possession
and resistance for dominant toddlers (shown in the upper right-hand corner
of Table 10.1), most readers would have no trouble computing the expected
frequencies. The total tally for row 1 is 26 (19 + 7), for row 2,72; similarly,
the total for column 1 is 61 (19 + 42), for column 2, 37; the grand total is
98. Thus the expected frequencies are 16.2 (61 times 26 divided by 98) and
9.8 for row 1 and 44.8 and 27.2 for row 2. In this case, the generated data fit
the observed quite well, and the chi-square would be small. In the context
of a chi-square analysis just of this 2 x 2 table, the investigator would
conclude that, among dominant toddlers, prior possession and resistance
are not related.
However, if chi-square were big, meaning that data did not fit this partic-
ular model, the investigator would need to invoke a more complex model.
In the case of a 2 x 2 table, the only model more complex than the no-
interaction model is the saturated model, which includes (in addition to the
row and column effects) an interaction term. To recapitulate, the expected
 Log-linear models: A simple example 181

frequencies computed for the common chi-square test of independence are

in fact generated by a no-interaction model. If the chi-square is significantly
large, then this model fails to fit the data. An interaction term is required,
which means that the row and column variables are not independent of each
other, but are in fact associated. Often this lack of independence is exactly
what the investigator hopes to demonstrate.
Following usual notational conventions (e.g., Fienberg, 1980), the no-
interaction model mentioned in the previous paragraph would be indicated
as the [R] [C] model, meaning that cell values are constrained by the model
to reflect just the row and column totals. The saturated model would be
represented as [R] [C] [RC], meaning that in addition to the row and column
constraints, the cell totals must also reflect the row x column [RC] cross-
classification totals. In the case of a two-dimensional table, this means that
the model is saturated and in fact generates data identical with the observed.
Because meeting the [RC] constraints means that necessarily the [R] and
[C] constraints are met, the saturated model is usually indicated simply
by [RC]. In general, to simplify the notation, "lower"-level terms implied
by higher-order terms are usually not included when particular models are
described. They are simply assumed.
At this point, we hope the reader has a clear idea of the logic of this
approach, at least with respect to two-dimensional contingency tables. The
situation with respect to tables with three or more dimensions is somewhat
more complex. For one thing, computing expected frequencies for various
models is not the simple matter it is with two-dimensional tables; usually,
computer programs are used. For another, the notation can become a little
confusing. Our plan here is to demonstrate the approach with the three-
dimensional dominance x prior possession x resistance tables for toddlers
and preschoolers, as given in Table 10.1. This hardly exhausts the topic,
but we hope it will given interested readers a start, and that they will
then move on to more detailed and specialized literature (e.g., Bakeman &
Robinson, 1994).
For the 2 x 2 x 2 tables for the toddlers and preschoolers, it seems
clear that the null or equiprobable model (all cells have the same value)
would not fit these data. In fact, we would probably not test the null
model first, but, as with the chi-square test of independence, would begin
with a more complex one. Earlier we said that resistance (R) was regarded
as a response or outcome variable, and that dominance (D) and prior pos-
session (P) were regarded as explanatory variables. That is, substantively
we want to find out if dominance and/or prior possession affected resis-
tance. Thus we would begin testing with the [R] [DP] model. This model
simply constrains the generated data to reflect the resistance rates and the
dominance cross-classified by prior possession rates actually observed. In
182 Analyzing cross-classified events

particular, it does not contain any terms that suggest that either dominance,
or prior possession, or their interaction, is related to the amount of resis-
tance encountered.
In analysis-of-variance terms, the [/?][DP] model is the "no effects"
model. In a sense, the [DP] term just states the design, whereas the fact
that the response variable, [/?], is not combined with any of the explanatory
variables indicates that none affect it. If the [R][DP] model failed to fit
the data, but the [RD][DP] model did, we would conclude that there was a
main effect for dominance - that unless dominance is taken into account,
we fail to make very good predictions for how often resistance will occur.
Similarly, if the [RP][DP] model fit the data, we would conclude that there
was a main effect for prior possession. If the [RD][RP][DP] model fit,
main effects for both dominance and prior possession would be indicated.
Finally, if only the [RDP] modelfitthe data (the saturated model), we would
conclude that, in order to account for resistance, the interaction between
dominance and prior possession must be taken into account.
In the present case, the no-effects model failed to fit the observed data.
For both toddlers and preschoolers, the chi-square comparing generated to
observed was large and significant (values were 11.3 and 7.2, df= 3, for tod-
dlers and preschoolers, respectively; these are likelihood-ratio-chi-squares,
computed by Bakeman & Robinson's [1994] ILOG program). However,
for toddlers the [RD] [DP] model, and for preschoolers the [RP] [DP] model
generated data quite similar to the observed (chi-squares were L9 and 0.8,
df = 2, for toddlers and preschoolers, respectively; these chi-squares are
both insignificant, although in both cases the difference between them and
the no-effects model is significant). This is analogous to a main effect for
dominance among toddlers and a main effect for prior possession among
preschoolers. In other words, the dominance of the taker affected whether
his or her take attempt would be resisted among toddlers, but whether
the taker had had prior possession of the contested object or not affected
whether he or she would meet with resistance among preschoolers. Thus
the effects noted descriptively in the previous section are indeed statistically
significant. Bakeman and Brownlee (1982) interpreted this as evidence for
shared possession rules, rules that emerge somewhere between 2 and 3
years of age.

10.3 Summary
Sometimes investigators who collect observational data and who are inter-
ested in sequential elements of the behavior observed seem compelled both
to obtain a continuous record of selected aspects of the passing stream of
 Summary 183

behavior and to analyze exhaustively that continuous record. An alterna-

tive is to record just selected aspects of certain kinds of events. The kind of
event is defined beforehand (for example, possession struggles) as well as
the aspects of interest. Each aspect corresponds to a codable dimension. In
the case of the possession struggles described in the previous sections, the
dimensions were dominance, prior possession, and resistance. The codes
for each dimension were the same (either yes or no), but in other cases the
mutually exclusive and exhaustive codes for the various dimensions could
well be different. In all cases, cross-classified event data result.
A major advantage of such data is, first, cross-classified categorical data
are hardly exotic or new, and second, ways of analyzing such data are rela-
tively well understood (e.g., Bakeman & Robinson, 1994; Kennedy, 1983,
1992; Upton, 1978; Wickens, 1989). In this chapter, a simple example
was given, showing how log-linear results can be expressed in analysis-of-
variance-like terms. The technique, however, is not confined just to contin-
gency tables, but can also be applied to event-sequence and time-sequence
data; this was mentioned earlier in section 7.6. The interested reader is
advised to consult Bakeman and Quera (1995b) and Castellan (1979).
 11
Epilogue

11.1 Kepler and Brahe

When Johannes Kepler went to work for the astronomer Tycho Brahe, he
found himself in the midst of a strange set of circumstances. Brahe lived
in a castle on an island, and there he conducted his painstakingly careful
observations. Brahe had a dwarf who scrambled for food in the dining hall.
Brahe also had a silver nose that replaced the natural nose he had lost in a
duel. Kepler was given the problem of computing the orbit of Mars, and
he wrote in one of his letters that Brahe fed him data in little bits, just as
the dwarf received crumbs under the table.
Kepler had come to Brahe because of the observations. He was hungry
for the observations. Without them, the principles of planetary motion -
the patterns he discovered - would never have been discovered. Without
these patterns, Newton's explanation - the universal theory of gravitation
- would never have emerged.
We need to observe, and we need to do it very well, with imagination,
with boldness, and with dedication. If we do not observe, we shall never
see what is there. If we never see what is there, we shall never see the
patterns in what is there. Without the patterns, there will never be the kind
of theory that we can build with.
Observing and discovering pattern is what this book is about. We do this
kind of thing for a living, and we have chosen to do it because it is what we
think science is about. Obviously we think it is not really that hard to do.
But we are lonely. We want company in this enterprise. Only about 8% of
all psychological research is based on any kind of observation. A fraction
of that is programmatic research. And, a fraction of that is sequential in its
thinking.
This will not do. Those of us who are applying these new methods of
observational research are having great success. We are beginning to find
consistent results in areas previously recalcitrant to quantitative analysis:
how babies learn to interact with their parents and organize their behav-
ior; how young children make friends or are rejected by their peers; how

184
 Soskin and John on marital interaction 185

marriages succeed or fail by how spouses interact; how families weather

stress or create pathology.
In many senses, in developmental, social, and clinical psychology we
are returning to earlier efforts with a new conceptual technology, and it
is paying off. We are able to return to the spirit of observation that char-
acterized developmental psychology in the 1930s and ask questions about
social and emotional development more clearly. We are now able to return
to initial clinical theorizing about families in the 1950s and bring a great
deal of precision to the task.
What is new is not a new statistical technique. It is a whole new way
of thinking about interaction in terms of its temporal form, about pattern
that occurs and recurs in time. We believe that great conceptual clarity
can be obtained by thinking about temporal patterning, and we believe that
anyone who has collected observational data over time and ignores time is
missing an opportunity.
Kepler needed the observations. The observations needed Kepler.

11.2 Soskin and John on marital interaction

Over 30 years ago, Roger Barker edited a classic book titled The Stream
of Behavior, which was seminal in the history of naturalistic observational
research. The book also focused on social interaction. We believe that it
would be useful to review a chapter in that book by Soskin and John that
tackled the description of marital interaction. We shall compare their anal-
ysis with what we currently know about marital interaction. Our objective
is to make the point that we can now confront old problems with new tools.
Soskin and John (1963) spent a year pilot-testing a 3-pound radio trans-
mitter to be worn in a kind of backpack arrangement by two young husband-
wife pairs. The subjects received an expense-free vacation in a resort com-
munity; they lived in a small cottage at the edge of a large lake, though they
mingled freely with other guests. The transmitters could be disconnected
by the subjects if they felt the need for privacy, but neither couple did so
during the day. Transmitters were turned off after midnight. Soskin and
John presented an extensive fragment of one couple's conversation while
they were rowing. The episode contains good natural humor:
Jock: Yo-ho, heave ho. You do the rowing.
Roz: Nuts to that idea. You're a big strong man. Mmmm!
Jock: Yeah, but I have a handicap.
Roz: Yeah, you have a handicap in your head.
The episode also contains come conflict:
Roz: You're dipping your oars too deep, dear.
Jock: I can't feather them, either.
186 Epilogue

Roz: You should be able to . . .

Jock: Change places with me.
Roz: Let me get this off first.
Jock: No, change places first. Hold it up and... you can stand up. It's perfectly
all right.
Roz: That's the first thing I learned about water safety, love. Don't stand up
in a boat...
Jock: Well I wanted you to stay into the waves, whatever you did.
Roz: Well, why didn't you tell me so!
Jock: Go up that way.
Roz: (slightly irritated) Which way do you want to go?
Jock: This way.
Soskin and John analyzed their corpus of data in three ways: (a) a "struc-
tural analysis"; (b) a "functional analysis"; and, (c) a "dynamic analysis."
The structural analysis primarily computed talk times. The functional anal-
ysis began by distinguishing between "informational" messages ("It's four
miles from here") and "relational" messages ("Then get going"). They
wrote that relational talk "encompasses the range of verbal acts by which
a speaker manages his interpersonal relations" (p. 253).
In their functional analysis they eventually identified six categories of
messages: (a) expressive statements ("Ouch!" "Wow!" "Darn!"); (b) "ex-
cogitative" statements, which are "most commonly described as 'thinking
aloud' "(p. 255) ("Humm... what have I done wrong here?" "Oh, I see!");
(c) "signomes," which are "messages that report the speaker's present phys-
ical or psychological state" (p. 255) ("I'm cold!" "Oh, Jock, I like that!");
(d) "metrones," which are evaluative statements arising out of the subject's
belief system ("What a fool I've been!" "You shouldn't do that!"); (e) "re-
gones," which control or channel the behavior of the listener ("Why don't
you do it right now?"); and (f) "structones," which include information
statements ("I weigh 181 pounds").
The dynamic analysis distinguished three variables: state, locus-direc-
tion, and bond. State involved affective information: (a) joy, glee, high
pleasure; (b) satisfaction, contentment, liking; (c) ambivalence; (d) mild
apprehension, dislike, frustration, disappointment; (e) pain, anger, fear,
grief; and (f) neutrality. The locus-direction variable was indicated with
arrows up or down "for the direction and locus of state change it would
produce from the point of view of a neutral observer" (p. 258). These
categories were (a) wants, wishes, self-praise; (b) mutually complimen-
tary statements; (c) derogation, reproof, rebuffs, which imply the speaker's
superiority; (d) self-criticism; (e) apology, praise; (f) compliments, per-
mission; (g) mutually unfavorable statements; (h) accusations, reproof;
and (i) no inferable change. "Bonds" referred to "the degree of intimacy
the speaker was willing to tolerate in the relationship" (p. 258).
 Soskin and John on marital interaction 187

Soskin and John reported several findings based on their analysis of their
tapes. The structural analysis showed that, in most situations, Jock talked
a lot and could be described as highly gregarious (he talked about 36%
of the total talk time in a four-person group). His longer utterances were
predominantly structones (factual-information exchanges).
The functional analysis of 1850 messages of Roz and Jock's talk showed
that Roz was significantly more expressive (8.6% vs. Jock's 3.8%), less con-
trolling (fewer regones: 11.0% vs. Jock's 13.9%), and less informational
(fewer structones: 24.5% vs. Jock's 31.3%). They concluded:
Roz produced a high percentage of expressive messages whenever the two were
out of the public view and became noticeably more controlled in the presence of
others. Jock's output, on the other hand, was relatively low throughout, (p. 267)

This is not quite consistent with the earlier analysis of Jock as gregarious
and a high-output talker. They then turned to the results of their dynamic
analysis, which they began describing as follows:
The very dimensions by which it was hoped to identify inter- and intrapersonal
changes in the sequential development of an episode proved most difficult to
isolate, (p 267)

Unfortunately, they could find no consistent patterns in the way Roz and
Jock tried to influence and control one another. They wrote:
The very subtle shifts and variations in the way in which these two people at-
tempted to modify each other's states sequentially throughout this episode obliged
us to question whether summaries of very long segments of a record reflect the
actual sequential dynamics of the behavior in a given episode, (p. 268)

Equally disappointing was their analysis of affect; they wrote:

As with locus-direction shift, the assessment of affective state changes met with
only marginal success, (p. 270)

However, they did conclude that:

the coders saw Roz as producing a higher percent of mildly negative statements
than her husband in all five of the episodes, in three of which the difference
was statistically significant. By contrast, in all five episodes Jock was seen as
producing a higher percent of neutral statements, and in four of the five episodes
the difference between them was significant, (p. 272)

In these results, we can see a struggle with coding systems that are unwieldy,
that are hard to fit together, and that lack a clear purpose or focus. The
research questions are missing.
188 Epilogue

11.3 Marital interaction research since 1963

In the early 1970s, psychologists began systematically applying observa-
tional methods to the study of marital interaction. The first important such
study was Raush, Barry, Hertel, and Swain's (1974) book Communication,
Conflict, and Marriage. This was a longitudinal study that followed cou-
ples from a newly wed stage through the pregnancy and birth of the couple's
first child. There were major research questions: (1) Was there consistency
over time in couples' interactional style in resolving conflict? (2) How did
the pregnancy period affect the marriage? (3) What sex differences ex-
ist? (4) How are happily and unhappily married couples different in the
way they resolve conflict? Raush et al. employed one coding system, a
series of improvised conflict situations, and they used sequential analysis
(employing multivariate information theory).
Raush et al. found that couples had an interactive style that was con-
sistent over time. In a later study, Gottman (1980b) found that there was
greater cross-situational consistency within couples (from high- to low-
conflict interactions) when sequential z scores rather than unconditional
probabilities were used.
Contrary to what Soskin and John reported about Jock and Roz, Raush
et al. found no evidence to support the contention that in marriages men
were less expressive and more instrumental than women. However, con-
sistent with Sokin and John's affect results, they did report that women
were more coercive than men, whereas men were more reconciling. These
sex differences were enhanced during the pregnancy phase. However, in
general, "discordant" couples were far more coercive than "harmonious"
couples.
Subsequent sequential analytic research on the question of differences
between satisfied and dissatisfied couples has produced a clear picture in
conflict-resolution strategies. There is now a body of literature that can be
called upon. We shall not review this literature here because our purposes
are methodological. For a recent review, see Noller (1984). One series
of studies with one coding system was reported in a research monograph
by Gottman (1979a). We shall summarize the consistent results of these
studies by giving examples of conversation sequences.
Historically, three methodological innovations were necessary before
this work could proceed: appropriate observational methods, sequential
analysis, and the systematic study of affect displayed through nonverbal
behavior. Consider the following dialogue of a dissatisfied couple dis-
cussing how their day went (H = husband, W = wife; extracts are from
unpublished transcripts):

H: You'll never guess who I saw today, Frank Dugan.

W: So, big deal, you saw Frank Dugan.
 Marital interaction research since 1963 189

H: Don't you remember I had that argument with him last week?
W: I forgot.
H: Yeah.
W: So I'm sorry I forgot, all right?
H: So it is a big deal to him.
W: So what do you want me to do, jump up and down?
H: Well, how was your day, honey?
W: Oh brother, here we go again.
H: (pause) You don't have to look at me that way.
W: So what d'ya want me to do, put a paper bag over my head?

Using the Couples Interaction Scoring System (CISS), which codes both
verbal and nonverbal behaviors of both speaker and listener, Gottman and
his associates coded the interaction of couples who were satisfied or dis-
satisfied with their marriages. Among other tasks, couples were studied
attempting to resolve a major area of disagreement in their marriages.

Gottman's results
The major question in this research was "what were the differences between
satisfied and dissatisfied couples in the way they resolve conflict?"
Basically, these differences can be described by using the analogy of
a chess game. A chess game has three phases: the beginning game, the
middle game, and the end game. Each phase has characteristic good and
bad maneuvers and objectives. The objectives can, in fact, be derived
inductively from the maneuvers. The goal of the beginning phase is control
of the center of the chessboard and development of position. The goal of
the middle game is the favorable exchange of pieces. The goal of the end
game is checkmate. Similarly, there are three phases in the discussion of a
marital issue. The first phase is "agenda-building," the objective of which
is to get the issues out as they are viewed by each partner. The second
phase is the "arguing phase," the goal of which is for partners to argue
energetically for their points of view and for each partner to understand the
areas of disagreement between them. The third phase is the "negotiation,"
the goal of which is compromise.
It is possible to discriminate the interaction of satisfied and dissatisfied
couples in each phase. In the agenda-building phase, cross-complaining
sequences characterize dissatisfied couples. A cross-complaining sequence
is one in which a complaint by one person is followed by a countercomplaint
by the other. For example:
W: I'm tired of spending all my time on the housework. You're not doing your
share.
H: If you used your time efficiently you wouldn't be tired.
190 Epilogue

A validation sequence recognizes the potential validity of the other per-

son's viewpoint before making a counterstatement. Usually the validation
sequence differs from the cross-complaining sequence by the use of "as-
sent codes" such as "Yeah," "Oh," "Mmmhmmm," and so on. For example,
below is a cross-complaining sequence followed by the same exchange as
a validation sequence.
Cross-complaining:
W: I've been home alone all day, cooped up with the kids.
H: I come home tired and just want to relax.

Validation:
W: I've been home alone all day.
H: Uh-huh.
W: Cooped up with the kids.
H: Yeah, I come home tired.
W: Mmm.
H: And just want to relax.
W: Yeah.

In the negotiation phase, counterproposal sequences characterize the

interaction of dissatisfied couples, whereas contracting sequences charac-
terize the interaction of satisfied couples. In a counterproposal sequence, a
proposal by one partner is met immediately by a proposal by the other part-
ner, whereas in a contracting sequence there is first some acceptance of the
partner's proposal. The agreement codes that discriminate counterproposal
and contracting sequences are very different from those that discriminate
cross-complaining and validation sequences. Instead of simple agreement
or assent, contracting sequences include direct modification of one's own
view:
Counterproposal:
W: We spent all of Christmas at your mother's last year. This time let's spend
Christmas at my mother's.
H: Let's spend it again at my mother's this year. It's too late to change it. We
can discuss our plans for next year now.

Contracting:
W: We spent all of Christmas at your mother's last year. This time let's spend
Christmas at my mother's.
H: Yeah you're right, that's not fair. How about 50-50 this year?

At some points the conversation of the two groups of couples would be

indistinguishable without the use of nonverbal codes.
 Marital interaction research since 1963 191

The anatomy of negative affect

The deescalation of negative affect, not the reciprocation of positive affect
(known in the literature on marital therapy as the quid pro quo hypothesis),
discriminated happy from unhappy marriages in these studies. Another
finding concerned the role of statements about the process of communica-
tion (metacommunication), such as "You're interrupting me." There were
no differences in the amount of metacommunication between satisfied and
dissatisfied couples, but the sequences in the two groups differed markedly.
Metacommunication tends to be what is called, in Markov model theory,
an "absorbing state" for unhappily married couples, i.e., it becomes nearly
impossible to exit once entered. For satisfied couples, metacommunica-
tive chains are brief and contain agreements that lead rapidly to other
codes. For example, a metacommunicative chain in a satisfied marriage
might be:
H: You're interrupting me.
W: Sorry, what were you saying?
H: I was saying we should take separate vacations this year.

For a dissatisfied couple the chain might be:

H: You're interrupting me.
W: I wouldn't have to if I could get a word in edgewise.
H: Oh, now I talk too much. Maybe you'd like me never to say anything.
W: Be nice for a change.
H: Then you'd never have to listen to me, which you never do anyway.
W: If you'd say something instead of jabbering all the time maybe I would
listen.
It is not the amount of metacommunication, but how it is delivered that
determines the sequence that follows and whether its role facilitates com-
munication. This fact could not have been discovered without a sequential
analysis of the data. Note that what makes the metacommunication effec-
tive is that it changes the affective nature of the interaction. If the affect
simply transfers to the metacommunication, it cannot function as a repair
mechanism.
Another pattern common to both satisfied and dissatisfied couples is
called "mind reading" - making attributions of emotions, opinions, states
of mind, etc., to a spouse. The effect of mind reading depends entirely
on the affect with which it is delivered. If mind reading is delivered with
neutral or positive affect, it is responded to as if it were a question about
feelings; it is agreed with and elaborated upon, usually with neutral affect:
H: You always get tense at my mother's house.
W: Yes, I do. I think she does a lot to make me tense.
192 Epilogue

If mind reading is delivered with negative affect, it is responded to as if

it were a criticism; it is disagreed with and elaborated upon, usually with
negative affect:
H: You always get tense at my mother's house.
W: I do not. I start off relaxed until she starts criticizing me and you take her
side.

Satisfied couples continually intersperse various subcodes of agreement

into their sequences. In the agenda-building phase, this is primarily a
simple "assent" form of agreement, as in "oh yeah," "uh huh," "I see," and
so on, whereas in the negotiation phase, this is primarily actual acceptance
of the other's point of view and modification of one's own point of view.
These listener responses or "backchanneling" (Duncan & Fiske, 1977)
are clear communications to the speaker that the listener is "tracking" the
speaker. But they do more than regulate turns, especially in the beginning
phases of marital conflict resolution. They communicate agreement not
with the speaker's point of view or content, but with the speaker's affect.
By indicating that it might make sense to see things as the other does, these
responses grease the wheels for affective expression.
In the negotiation phase of the discussion, the agreement codes are very
different. They are not "assent," but direct agreement with the other's point
of view ("Yes, you're right," or "I agree with that"). They may even involve
accepting some modification of one's own point of view in order to reach
a solution to the problem. This creates a climate of agreement that has
profound consequences for the quality of the interaction.
We see in all these results that in dissatisfied marriages couples are far
more defensive and less receptive to their partners. To investigate the nature
of this defensiveness, Robert Levenson and John Gottman began collect-
ing autonomic nervous system (ANS) data during marital interaction. They
discovered that ANS arousal during conflict resolution is highly predictive
(simple correlations in the 90s) of changes in marital satisfaction over a
3-year longitudinal period, controlling initial levels of marital satisfaction.
Here we have the possibility of a theoretical basis for the observational re-
sults. What are the setting conditions of ANS arousal? Will all discussions
of disagreements produce arousal? What are the consequences of ANS
arousal? Are there sex differences in ANS arousal in response to intense
negative affect? These questions are currently being pursued in Gottman's
and Levenson's laboratories.
To summarize, in this section we have suggested that we are now in a
position to tackle old and venerable questions with new tools. Observa-
tional techniques have been successful in the study of marital interaction
(and in other areas) in identifying stable phenomena. This is the first step
 Marital interaction research since 1963 193

toward the construction of theory, which seeks elegant, far-reaching, and

parsimonious explanations for the observed phenomena.
Just as Kepler needed Brahe's observations, the observations needed
Kepler. Kepler found the patterns. Newton explained them with the theory
of gravitation. We need our Brahes, our Keplers, and our Newtons. This
book is an advertisement for all of the "instrument makers" who have
developed observational methodology to its current handy state.
 Appendix: A Pascal program
to compute kappa and
weighted kappa
This appendix contains the source code for a Pascal program that computes
kappa and weighted kappa. The program was compiled using Borland's
Pascal 6.0 and should run on IBM-compatible microcomputers in a Win-
dows or DOS environment. The program contains essentially no error
checking for the numbers the user enters, so these must be correct. All
values entered (the number of rows, weights if any, the tallies themselves)
are integers, so entering a letter or even a decimal point instead of a digit,
for example, will cause the program to fail (unless error checking code is
added).
Program Kappa; { A simple Pascal no-bells, no-whistles,}
{ no-errors-permitted-in-input program }
uses Crt; { to compute kappa and weighted kappa. }
{ (c) Roger Bakeman, September 1995. }
const Max = 20;
Enter = chr(13);
N : word = 0;
var i, j : word;
Ch : char;
X, W : array [1..Max,1..Max] of word;
WeightsDefined : boolean;

Procedure ComputeKappa;
var i, j : word;
M : array[1..Max+1,1..Max+1] of real;
Kappa, Numer, Denom : real;

begin
for i := 1 to N do { Set row & col sums to zero. }
begin
M[i,N+l] := 0 . 0 ;

194
 Appendix: A Pascal program to compute kappa 195

M[N+l,i] := 0 . 0 ;
end;
M[N+1,N+1] := 0 . 0 ;
{ T a l l y row & col t o t a l s . }
f o r i := 1 t o N do for j := 1 t o N do
begin

M[N+1,N+1] := M[N+1,N+1] + X [ i , j ] ;
end;
{ Compute exp. f r e q u e n c i e s . }
f o r i := 1 t o N do f o r j := 1 t o N do
*

Numer := 0.0; { Compute kappa.

Denom := 0.0;
for i := 1 to N do for j := 1 to N do
begin
Numer := Numer + W[i,j] * X[i,j];
Denom := Denom + W[i,j] * M[i,j];
end;
Kappa := 1.0 - Numer/Denom;
writeln (' Kappa =;,Kappa:7:4);
writeln;
end;

Procedure DefaultWeights; { Provide standard weights, }

var i, j : word; { 0 on diagonal, 1 otherwise.}

begin
for i := 1 to N do for j := 1 to N do
if (i = j) then W[i,j] := 0 else W[i,j] := 1;
end;

Procedure AskForNumbers (Kind : word);

const Message: { Kind=l Kind=2 }
array [1..2] of string[7] = ('weights', 'tallies');
var i : word;

begin
196 Appendix: A Pascal program to compute kappa

writeln (' Enter >,N:3,' \ Message[Kind],

1
(with spaces between for) . . . ; ) ;
for i :=1 to N do
begin {Read numbers; }
write (' row',i:3,c> ' ) ; {weights if Kind=l,}
for j := 1 to N do read (X[i,j]);{tallies if Kind=2.}
readln; {Must be integers. }
end;
if (Kind = 1) then WeightsDefined := true;
end;

Function WantsSame (Kind : word) : boolean;

const Message : array [1..5] of string[12] =
('Same # rows ',
'Same labels ; , { Ask Y|N questions. Return }
c
Weighted k > , { TRUE if Y, y, or Enter. }
'Same weights ',
c
More kappas ' ) ;
var Ch : char;

begin
write (( \Message[Kind],c (Y|N)? ' ) ;
Ch := ReadKey; writeln (Ch);
WantsSame := (UpCase(Ch) = C Y J ) or (Ch = Enter);
end;

Procedure AskForOrder; { Ask for order of matrix. }

begin { Must be 1 through 20. }
repeat
write (' Number of rows (1-20)? >);
readln (N)
until (N > 0) and (N <= 20);
weightsDefined := false;
end;

BEGIN
TextColor (Black) ;
TextBackground (LightGray) ;
ClrScr;
 Appendix: A Pascal program to compute kappa 197

Writeln
('Compute kappa or wt kappa; (c) Roger Bakeman, GSU J );

repeat
if N = 0 then AskForOrder
else if not WantsSame(l) then AskForOrder;

if not WeightsDefined
then begin if WantsSame(3) then AskForNumbers(1) end
else if not WantsSame(4) then AskForNumbers(1);

if not WeightsDefined then DefaultWeights;

AskForNumbers(2) ;
ComputeKappa;
until not WantsSame(5);
END.
 References

Adamson, L. B., & Bakeman, R. (1985). Affect and attention: Infants observed with mothers
and peers. Child Development, 56, 582-593.
Ainsworth, M. D. S., Blehar, M. C , Waters, E., & Wall, S. (1978). Patterns of attachment.
Hillsdale, NJ: Lawrence Erlbaum.
Allison, P. D., & Liker, J. K. (1982). Analyzing sequential categorical data on dyadic
interaction: A comment on Gottman. Psychological Bulletin, 91, 393—403.
Anderson, T. W, & Goodman, L. A. (1957). Statistical inference about Markov chains.
Annals of Mathematical Statistics, 28, 89-110.
Altmann, J. (1974). Observational study of behaviour: Sampling methods. Behaviour, 49,
227-267.
Altmann, S. A. (1965). Sociobiology of rhesus monkeys. II. Stochastics of social commu-
nication. Journal of Theoretical Biology, 8, 490-522.
Attneave, F. (1959). Applications of information theory to psychology. New York: Henry
Holt.
Bakeman, R. (1978). Untangling streams of behavior: Sequential analysis of observation
data. In G. P. Sackett (Ed.), Observing behavior (Vol. 2): Data collection and analysis
methods (pp. 63-78). Baltimore: University Park Press.
Bakeman, R. (1983). Computing lag sequential statistics: The ELAG program. Behavior
Research Methods & Instrumentation, 15, 530-535.
Bakeman, R. (1992). Understanding social science statistics: A spreadsheet approach.
Hillsdale, NJ: Erlbaum.
Bakeman, R., & Adamson, L. B. (1984). Coordinating attention to people and objects in
mother-infant and peer-infant interaction. Child Development, 55, 1278-1289.
Bakeman, R., Adamson, L. B., & Strisik, P. (1989). Lags and logs: Statistical approaches to
interaction. In M. H. Bornstein & J. Bruner (Eds.), Interaction in human development
(pp. 241-260). Hillsdale, NJ: Erlbaum.
Bakeman, R., Adamson, L. B., & Strisik, P. (1995). Lags and logs: Statistical approaches to
interaction (SPSS Version). In J. M. Gottman (Ed.), The analysis of change (pp. 279-
308). Hillsdale, NJ: Erlbaum.
Bakeman, R., & Brown, J. V. (1977). Behavioral dialogues: An approach to the assessment
of mother-infant interaction. Child Development, 49, 195-203.
Bakeman, R., & Brownlee, J. R. (1980). The strategic use of parallel play: A sequential
analysis. Child Development, 51, 873-878.
Bakeman, R., & Brownlee, J. R. (1982). Social rules governing object conflicts in toddlers
and preschoolers. In K. H. Rubin & H. S. Ross (Eds.), Peer relationships and social

198
 References 199

skills in childhood (pp. 99-111). New York: Springer-Verlag.

Bakeman, R., & Casey, R. L. (1995). Analyzing family interaction: Taking time into
account. Journal of Family Psychology, 9, 131-143.
Bakeman, R., & Dabbs, J. M., Jr. (1976). Social interaction observed: Some approaches
to the analysis of behavior streams. Personality and Social Psychology Bulletin, 2,
335-345.
Bakeman, R., & Dorval, B. (1989). The distinction between sampling independence and
empirical independence in sequential analysis. Behavioral Assessment, 11, 31-37.
Bakeman, R., Forthman, D. L., & Perkins, L. A. (1992). Time-budget data: Log-linear and
analysis of variance compared. Zoo Biology, 11, 271-284.
Bakeman, R., McArthur, D., & Quera, V. (1996). Detecting group differences in sequen-
tial association using sampled permutations: Log odds, kappa, and phi compared.
Behavior Research Methods, Instruments, and Computers, 28, 446-457.
Bakeman, R., & Quera, V. (1992). SDIS: A sequential data interchange standard.
Behavior Research Methods, Instruments, and Computers, 24, 554-559.
Bakeman, R., & Quera, V. (1995a). Analyzing interaction: Sequential analysis with SDIS
and GSEQ. New York: Cambridge University Press.
Bakeman, R., & Quera, V. (1995b). Log-linear approaches to lag-sequential analysis when
consecutive codes may and cannot repeat. Psychological Bulletin, 118, 272-284.
Bakeman, R., & Robinson, B. R. (1994). Understanding log-linear analysis with ILOG:
An interactive approach. Hillsdale, NJ: Erlbaum.
Bakeman, R., Robinson, B. R, & Quera, V. (1996). Testing sequential association: Esti-
mating exact p values using sampled permutations. Psychological Methods. 1, 4-15.
Bishop, Y M. M., Fienberg, S. E., & Holland, P. W. (1975). Discrete multivariate analysis.
Cambridge, MA: MIT Press.
Box, G. E. P., & Jenkins, G. M. (1970). Time-series analysis: Forecasting and control. San
Francisco: Holden-Day.
Brazelton, T. B., Koslowski, B., & Main, M. (1974). The origins of reciprocity: The early
mother-infant interaction. In M. Lewis & L. A. Rosenblum (Eds.), The effect of the
infant on its caregiver (pp. 49-76). New York: Wiley.
Brennan, R. L. (1983). Elements of generalizability theory. Iowa City: ACT Publications.
Brown, J. V., & Bakeman, R. (1980). Relationships of human mothers with their infants
during the first year of life: Effects of prematurity. In R. W. Bell & J. P. Smoth-
erman (Eds.), Maternal influences and early behavior (pp. 353-373). New York:
Spectrum.
Brown, J. V., Bakeman, R., Snyder, P. A. Fredrickson, W. T, Morgan, S. T, & Helper,
R. (1975). Interactions of black inner-city mothers with their newborn infants, Child
Development, 46, 677-686.
Brownlee, J. R., & Bakeman, R. (1981). Hitting in toddler-peer interaction. Child Devel-
opment, 52, 1076-1079.
Cairns, R. B., & Green, J. A. (1979). How to assess personality and social patterns: Obser-
vations or ratings? In R. B. Cairns (Ed.), The analysis of social interactions: Methods,
issues, and illustrations (pp. 209-226). Hillsdale, NJ: Lawrence Erlbaum.
Castellan, N. J. (1979). The analysis of behavior sequences. In R. B. Cairns (Ed.), The anal-
ysis of social interactions: Methods, issues, and illustrations (pp. 81-116). Hillsdale,
NJ: Lawrence Erlbaum.
200 References

Chatfield, C. (1973). Statistical inference regarding Markov chain models. Applied Statis-
tics, 22, 7-20.
Cohen, J. A. (1960). Coefficient of agreement for nominal scales. Educational and Psycho-
logical Measurement, 20, ?>l-46.
Cohen, J. (1968). Weighted kappa: Nominal scale agreement with provision for scaled
disagreement or partial credit. Psychological Bulletin, 70, 213-220.
Cohen, J. (1977). Statistical power analysis for the behavioral sciences. New York: Aca-
demic Press.
Cohen, J., & Cohen, P. (1983). Applied multiple regression/correlation analysis for the
behavioral sciences. Hillsdale, NJ: Lawrence Erlbaum.
Cohn, J. F., & Tronick, E. Z. (1987). Mother-infant face-to-face interaction: The sequence
of dyadic states at 3, 6, and 9 months. Developmental Psychology, 23, 68-77.
Condon, W. S., & Ogston, W. D. (1967). A segmentation of behavior. Journal of Psychiatric
Research, 5, 221-235.
C o n g e r , A . J., & W a r d , D . G . ( 1 9 8 4 ) . A g r e e m e n t a m o n g 2 x 2 a g r e e m e n t i n d i c e s . Educa-
tional and Psychological Measurement, 44, 301-314.
Cook, J., Tyson, R., White, J., Rushe, R., Gottman, J., & Murray, J. (1995). The mathematics
of marital conflict: Qualitative dynamic modeling of marital interaction. Journal of
Family Psychology, 9, 110-130.
Cronbach, L. J., Gleser, G. C , Nanda, H., & Rajaratnam, N. (1972). The dependability
of behavioral measurements: Theory of generalizability for scores and profiles. New
York: Wiley.
Crosbie, J. (1995). Interrupted time-series analysis with short series: Why it is problematic;
How it can be improved. In J. Gottman (Ed.), The analysis of change (pp. 361-398).
Hillsdale, NJ: Erlbaum.
Dabbs, J. M., Jr., & Swiedler, T. C. (1983). Group AVTA: A microcomputer system for
group voice chronography. Behavior Research Methods & Instrumentation, 15, 79-84.
Duncan, S., Jr., & Fiske, D. W. (1977). Face to face interaction. Hillsdale, NJ: Lawrence
Erlbaum.
Edgington, E. S. (1987). Randomization tests (2nd Ed.). New York: Marcel Dekker.
Ekman, P. W., & Friesen, W. (1978). Manual for the facial action coding system. Palo Alto,
CA: Consulting Psychologist Press.
Fienberg, S. E. (1980). The analysis of cross-classified categorical data. Cambridge, MA:
MIT Press.
Fleiss, J. L. (1981). Statistical methods for rates and proportions. New York: Wiley.
Fleiss, J. L. (1986). The design and analysis of clinical experiments. New York: John Wiley.
Fleiss, J. L., Cohen, J., & Everitt, B. S. (1969). Large sample standard errors of kappa and
weighted kappa. Psychological Bulletin, 72, 323-327.
Gardner, W. (1995). On the reliability of sequential data: Measurement, meaning, and
correction. In J. M. Gottman (Ed.), The analysis of change (pp. 339-359). Hillsdale,
NJ: Erlbaum.
Gart, J. J., & Zweifel, J. R. (1967). On the bias of various estimators of the logit and its
variance with application to quantile bioassay. Biometrika, 54, 181-187.
Garvey, C. (1974). Some properties of social play. Merrill-Palmer Quarterly, 20, 163-180.
Garvey, C , & Berndt, R. (1977). The organization of pretend play. In JSAS Catalog of
Selected Documents in Psychology, 7, 107. (Ms. No. 1589).
 References 201

Good, P. (1994). Permutation tests: A practical guide to resampling methods for testing
hypotheses. New York: Springer-Verlag.
Goodenough, F. L. (1928). Measuring behavior traits by means of repeated short samples.
Journal of Juvenile Research, 12, 230-235.
Goodman, L. A. (1983). A note on a supposed criticism of an Anderson-Goodman test in
Markov chain analysis. In S. Karlin, T. Amemiya, & L. A. Goodman (Eds.), Stud-
ies in econometrics, time series, and multivariate statistics (pp. 85-92). New York:
Academic Press.
Gottman, J. M. (1979a). Marital interaction: Experimental investigations: New York:
Academic Press.
Gottman, J. M. (1979b). Time-series analysis of continuous data in dyads. In M. E. Lamb,
S. J. Sumoi, & G. R. Stephenson (Eds.), Social interaction analysis: Methodological
issues (pp. 207-229). Madison: University of Wisconsin Press.
Gottman, J. M. (1980a). Analyzing for sequential connection and assessing interobserver
reliability for the sequential analysis of observational data. Behavioral Assessment, 2,
361-368.
Gottman, J. M. (1980b). The consistency of nonverbal affect and affect reciprocity in marital
interaction. Journal of Consulting and Clinical Psychology, 48, 711-717.
Gottman, J. M. (1981). Time-series analysis: A comprehensive introduction for social
scientists. New York: Cambridge University Press.
Gottman, J. M. (1983). How children become friends. Monographs of the Society for Re-
search in Child Development, 48(3, Serial No. 201).
Gottman, J. M. (1990). Chaos and regulated change in family interaction. In P. Cowan and
E. M. Hetherington (Eds.), New directions in family research: transition and change.
Hillsdale, NJ: Erlbaum.
Gottman, J. M., & Bakeman, R. (1979). The sequential analysis of observational data. In
M. E. Lamb, S. J. Suomi, & G. R. Stephenson (Eds.), Social interaction analysis:
Methodological issues (pp. 185-206). Madison: University of Wisconsin Press.
Gottman, J. M., & Levenson, R. W. (1992). Marital processes predictive of later dissolution:
Behavior, physiology, and health. Journal of Personality and Social Psychology, 63,
221-233.
Gottman, J. M., Markman, H., & Notarius, C. (1977). The topography of marital conflict:
A sequential analysis of verbal and nonverbal behavior. Journal of Marriage and the
Family, 39, 461^77.
Gottman, J. M., & Parker, J. (Eds.) (1985). Conversations of friends: Speculations on
affective development. New York: Cambridge University Press.
Gottman, J. M., & Ringland, J. T. (1981). The analysis of dominance and bidirectionality
in social development. Child Development, 52, 393^-12.
Gottman, J., Rose, E, & Mettetal, G. (1982). Time-series analysis of social interaction data.
In T. Field & A. Fogel (Eds.), Emotion and interactions (pp. 261-289). Hillsdale, NJ:
Lawrence Erlbaum.
Gottman, J. M., & Roy, A. K. (1990). Sequential analysis: A guide for behavioral research.
New York: Cambridge University Press.
Haberman, S. J. (1977). Log-linear models and frequency tables with small expected cell
counts. Annals of Statistics, 5, 1148-1169.
Haberman, S. J. (1978). Analysis of qualitative data (Vol. 1). New York: Academic Press.
202 References

Haberman, S. J. (1979). Analysis of qualitative data (Vol. 2). New York: Academic Press.
Hamilton, G. V. (1916). A study of perserverence reactions in primates and rodents. Behavior
Monographs, 3 (No. 2).
Hartmann, D. P. (1977). Considerations in the choice of interobserver reliability estimates.
Journal of Applied Behavior Analysis, 10, 103-116.
Hartmann, D. P. (1982). Assessing the dependability of observational data. In D. P. Hartmann
(Ed.)., Using observers to study behavior: New directions for methodology of social
and behavioral science (No. 14, pp. 51-65). San Francisco: Jossey-Bass.
Hartup, W. W. (1979). Levels of analysis in the study of social interaction: An historical
perspective. In M. E. Lamb, S. J. Suomi, & G. R. Stephenson (Eds.), Social interaction
analysis: Methodological issues (pp. 11-32). Madison: University of Wisconsin Press.
Hays, W. L. (1963). Statistics (1st ed.). New York: Holt, Rinehart, & Winston.
Hollenbeck, A. R. (1978). Problems of reliability in observational research. In G. P. Sackett
(Ed.), Observing behavior (Vol. 2): Data collection and analysis methods (pp. 79-98).
Baltimore: University Park Press.
Hubert, L. (1977). Kappa revisited. Psychological Bulletin, 84, 289-297.
Jaffe, J., & Feldstein, S. (1970). Rhythms of dialogue. New York: Academic Press.
Johnson, S. M., & Bolstad, O. D. (1973). Methodological issues in naturalistic observation:
Some problems and solutions for field research. In L. A. Hamerlynch, L. C. Handy,
& E. J. Mash (Eds.), Behavior change: Methodology, concepts, and practice (pp. 7 -
67). Champaign, IL: Research Press.
Jones, R. R., Reid, J. B., & Patterson, G. R. (1975). Naturalistic observations in clinical
assessment. In P. McReynolds (Ed.), Advances in psychological assessment (Vol. 3,
pp. 42-95). San Francisco: Jossey-Bass.
Kemeny, J. G., Snell, J. L., & Thompson, G. L. (1974). Introduction to finite mathematics.
Englewood Cliffs, NJ: Prentice-Hall.
Kennedy, J. J. (1983). Analyzing qualitative data: Introductory log-linear analysis for
behavioral research. New York: Praeger.
Kennedy, J. J. (1992). Analyzing qualitative data: Log-linear analysis for behavioral re-
search (2nd ed.). New York: Praeger.
Knoke, D., & Burke, P. J. (1980). Log-linear models. Newbury Park, CA: Sage.
Krokoff, L. (1983). The anatomy of negative affect in blue collar marriages. Unpublished
doctoral dissertation, University of Illinois at Urbana-Champaign.
Landesman-Dwyer, S. (1975). The baby behavior code (BBC): Scoring procedures and
definitions. Unpublished manuscript.
Losada, M., Sanchez, P., & Noble, E. E. (1990). Collaborative technology and group process
feedback: Their impact on interactive sequences in meetings. CSCW Proceedings,
53-64.
Miller, G. A., & Frick, F. C. (1949). Statistical behavioristics and sequences of responses.
Psychological Review, 56, 311- 324.
Miller, R. G., Jr. (1966). Simultaneous statistical inference. New York: McGraw-Hill.
Morley, D. D. (1987). Revised lag sequential analysis. In M. L. McLaughlin (Ed.), Com-
munication year book (Vol. 10, pp. 172-182). Beverly Hills, CA: Sage.
Morrison, P., & Morrison, E. (1961). Charles Babbage and his calculating engines. New
York: Dover.
Noller, P. (1984). Nonverbal communication and marital interaction. Oxford: Pergamon
Press.
 References 203

Oud, J. H., & Sattler, J. M. (1984). Generalized kappa coefficient: A Microsoft BASIC
program. Behavior Research Methods, Instruments, and Computers, 76, 481.
Overall, J. E. (1980). Continuity correction for Fisher's exact probability test. Journal of
Educational Statistics, 5, 177- 190.
Parten, M. B. (1932). Social participation among preschool children. Journal of Abnormal
and Social Psychology, 27, 243-269.
Patterson, G. R. (1982). Coersive family process. Eugene, OR: Castalia Press.
Patterson, G. R., & Moore, D. (1979). Interactive patterns as units of behavior. In
M. E. Lamb, S. J. Sumoi, & G. R. Stephenson (Eds.), Social interaction analysis:
Methodological issues (pp. 77-96). Madison: University of Wisconsin Press.
Pedhazur, E. J., & Schmelkin, L. P. (1991). Measurement, design, and analysis: An inte-
grated approach. Hillsdale, NJ: Erlbaum.
Rechten, C , & Fernald, R. D. (1978). A sampled randomization test for examining single
cells of behavioural transition matrices. Behaviour, 69, 217-227.
Reynolds, H. T. (1984). Analysis of nominal data. Beverly Hills, CA: Sage.
Rosenblum, L. (1978). The creation of a behavioral taxonomy. In G. P. Sackett (Ed.), Observ-
ing behavior (Vol. 2): Data collection and analysis methods (pp. 15-24). Baltimore:
University Park Press.
Raush, H. L., Barry, W. A., Hertel, R. K., & Swain, M. A. (1974). Communication, conflict,
and marriage. San Francisco: Jossey-Bass.
Sackett, G. P. (1974). A nonparametric lag sequential analysis for studying dependency
among responses in observational scoring systems. Unpublished manuscript.
Sackett, G. P. (1978). Measurement in observational research. In G. P. Sackett (Ed.), Observ-
ing behavior (Vol. 2): Data collection and analysis methods (pp. 25-43). Baltimore:
University Park Press.
Sackett, G. P. (1979). The lag sequential analysis of contingency and cyclicity in behavioral
interaction research. In J. D. Osofsky (Ed.), Handbook of infant development (pp. 623-
649). New York: Wiley.
Sackett, G. P. (1980). Lag sequential analysis as a data reduction technique in social inter-
action research. In D. B. Sawin, R. C. Hawkins, L. O. Walker, & J. H. Penticuff (Eds.),
Exceptional infant (Vol. 4): Psychosocial risks in infant-environment transactions.
New York: Brunner/Mazel.
Shannon, C. E., & Weaver, W. (1949). The mathematical theory of communication. Urbana:
University of Illinois Press.
Shotter, J. (1978). The cultural context of communication studies: Theoretical and method-
ological issues. In A. Lock (Ed.), Action, gesture, and symbol: The emergence of
language (pp. 43-78). London: Academic Press.
Siegel, S. (1956). Nonparametric statistics for the behavioral sciences. New York:
McGraw-Hill.
Smith, P. K. (1978). A longitudinal study of social participation in preschool children:
Solitary and parallel play reexamined. Developmental Psychology, 14, 517-523.
Smith, P. K., & Connolly, K. J. (1972). Patterns of play and social interaction in preschool
children. In N. Blurton Jones (Ed.), Ethological studies of child behavior (pp. 65-95).
Cambridge: Cambridge University Press.
Soskin, W. F., & John, V. P. (1963). The study of spontaneous talk. In R. G. Barker (Ed.),
The stream of behavior: Explorations of its structure and content (pp. 228-287). New
York: Appleton-Century-Crofts.
204 References

Sroufe, L. A., & Waters, E. (1977). Attachment as an organizational construct. Child De-
velopment, 48, 1184-1199.
Stern, D.N. (1974). Mother and infant at play: The dyadic interaction involving facial,
vocal, and gaze behaviors. In M. Lewis & L. A. Rosenblum (Eds.), The effect of the
infant on its caregiver (pp. 187-213). New York: Wiley.
Suen, H. K. (1988). Agreement, reliability, accuracy, and validity: Toward a clarification.
Behavioral Assessment, 10, 343-366.
Suomi, S. J. (1979). Levels of analysis for interactive data collected on monkeys living in
complex social groups. In M. E. Lamb, S. J. Suomi, G. R. Stephenson (Eds.), Social
interaction analysis: Methodological issues (pp. 119-135). Madison: University of
Wisconsin Press.
Suomi, S. J., Mineka, S., & DeLizio, R. D. (1983). Short- and long-term effects of repetitive
mother-infant separations on social development in rhesus monkeys. Developmental
Psychology, 19, 770-786.
Taplin, P. S., & Reid, J. B. (1973). Effects of instructional set and experimenter influence
on observer reliability. Child Development, 44, 547-554.
Tapp, J., & Walden, T. (1993). PROCODER: A professional tape control coding and anal-
ysis system for behavioral research using videotape. Behavior Research Methods,
Instruments, & Computers, 25, 53-56.
Tronick, E. D., Als, H., & Brazelton, T. B. (1977). Mutuality in mother-infant interaction.
Journal of Communication, 27, 74-79.
Tronick, E., Als, H., & Brazelton, T. B. (1980). Monadic phases: A structural descriptive
analysis of infant-mother face to face interaction. Merrill-Palmer Quarterly, 26, 3-24.
Tuculescu, R. A., & Griswold, J. G. (1983). Prehatching interactions in domestic chickens.
Animal Behavior, 31, 1- 10.
Uebersax, J. S. (1982). A generalized kappa coefficient. Educational and Psychological
Measurement, 42, 181-183.
Upton, G. J. G. (1978). The analysis of cross-tabulated data. New York: Wiley.
Wallen, D., & Sykes, R. E. (1974). Police IV: A code for the study of police-civilian in-
teraction. (Available from Minnesota Systems Research, 2412 University Ave., Min-
neapolis, MN 55414.)
Wampold, B. E. (1989). Kappa as a measure of pattern in sequential data. Quality and
Quantity, 23, 171-187.
Wampold, B. E. (1992). The intensive examination of social interaction. In T. R. Kratochwill
& J. R. Levin (Eds.), Single-case research design and analysis: New directions for
psychology and education (pp. 93-131). Hillsdale, NJ: Erlbaum.
Wickens, T. D. (1989). Multiway contingency tables analysis for the social sciences. Hills-
dale, NJ: Erlbaum.
Wickens, T. D. (1993). Analysis of contingency tables with between-subjects variability.
Psychological Bulletin, 113, 191-204.
Wiggins, J. S. (1973). Personality and prediction. Reading, MA: Addison-Wesley.
Williams, E., & Gottman, J. (1981). A user's guide to the Gottman-Williams time-series
programs. New York: Cambridge University Press.
Wolff, P. (1966). The causes, controls, and organization of the neonate. Psychological
Issues, 5 (whole No. 17).
 Index

adjusted residuals, see z scores Couples Interaction Scoring System (CISS), 189
agreement matrix, 61-62, 62/, 64/, 73/ Cronbach's generalizability alpha, 75-76, 76/
agreement, observer, see observer agreement Cross-classified events
Allison-Liker z score, see z score, agreement for, 74
Allison-Liker's formula analysis of, 179-182
alpha, Cronbach's generalizability, see data format for, 87, 88f, 90
Cronbach's generalizability alpha describing, 177-179
autocorrelation and time series, 163 recording of, 49-50, 54t
cyclicity in time series, 162
Barker, Roger, 185
Bernoulli, Daniel, 158 discrete events, see momentary events
Bonferroni's correction, 148 duration behaviors, see duration events
Brahe, Tycho, 184 duration events, 38-39

chi-square statistic electronic recording devices, 52-55

likelihood-ratio, 118 equiprobable model, significance test for, 101
Pearson, 101, 118 event recording
coding schemes agreement for, 71-73
Bakeman and Borwnlee's children's object time not recorded, 40-43, 54t
struggle, 34-35 time recorded, 43-45
Bakeman and Brownlee's play state, 9t, 10 event sequential data (ESD), 82-83, 88r, 90
development of, 15-17 expected frequency, formula for, 108
Ekman's Facial Action (FACS), 18, 19/ exploratory (hypothesis-generating) research,
Gottman's children's conversation, 30-31 12-13
Landesman-Dwyer's baby behavior, 33-34
Parten's play state, 4-5, 9t Family Interaction Coding System (FICS), 27
physically based, 17-21, 22 first-order model, significance test for, 102
Smith's play state, 8, 9t Fourier, Jean Baptiste Joseph, 158
socially based, 17-21,22 frequencies, analysis of, 92, 99
Suomi's monkey activity, 35-36 frequency behaviors, see momentary events
Tuculescu and Griswold's chicken behavior,
28-29 generalizability theory, 75
Cohen's kappa Generalized Sequential Querier (GSEQ), 82, 152
advantages for observer training, 67-68 Gottman-Williams
computation of, 62-64 BIVAR program, 164-165
Pascal program for, 194-197 ITSE program, 161
standard error of, 65-66 SPEC program, 162
weighted, 66-67
conditional probability, see transitional hierarchical analysis, 121-125, 123f, \24t
probability hierarchical codes, 29
confusion matrix, see agreement matrix
continuous recording, 10, 39-40 independence assumption, 136-138

205
206 Index

information theory, 140 physically based coding schemes, see coding

interevent interval, 156, 165-166 schemes, physically based
intermittent recording, 39-40 pooling data, 141-144
interrupted time series, 169-170 probabilities, analysis of, 93
interval length, 48 probability, see simple probability; transitional
interval recording, 10, 46-49, 5At probability
agreement for, 74
interval sequential data (ESD), 85-86, 88f, 90 rare events, assessing effects of, 161
rates, analysis of, 92, 99
kappa, Cohen's see Cohen's kappa rating scales, 3
Kepler, Johannes, 184 recording onset and offset times, 43-45
recording pattern changes, 45-46
lag-sequential analysis, 111-116, 113f, 116/, reliability decay, 58-59
117/
likelihood ratio chi-square statistic, see sequential analysis
chi-square statistic, likelihood-ratio amount of data needed for, 144-147
log-linear analysis, 116-117, 179-182 basic or absolute methods, 100-103, 133
log-linear models, sequential, see sequential minimizing amount of data needed for,
log-linear models 125-127
lumping codes, 24-25 log-linear methods, 116-117, 134-135
probabilistic methods, 111-116, 134
Marital Interaction Coding System (MICS), 156 sequential approaches defined 8, 14
Markovian analysis, 140 Sequential Data Interchange Standard (SDIS),
mean event durations, analysis of, 94, 99 82,152-153
molar or macro coding schemes, 24-25, 154 sequential log-linear models, 120-125
molecular or micro coding schemes, 24-25, 153 simple probability, 95, 99
momentary events, 38-39 socially based coding schemes, see coding
moving probability window, 156-157 schemes, socially based
mutually exclusive and exhaustive (ME & E) spectral density function, 160
codes defined, 26-27 splitting codes, 24-25
state sequential data (SSD), 83-84, 88f, 90
narrative reports, 2-3 state transition diagram, 97/ 98, 106/
Newton, Isaac, 159, 193 stationarity, 138-139
structural zeros, 108
observer agreement systematic observation defined, 3, 13-14
about codes, 71-74
rationale for assessing, 57 thought units, 30
about unitizing, 68-71 time budget information, 7, 150
observer calibration, 57 time sampling, 50-51, 5At
observer reliability, 59-60 time series analysis
observers advantages, 154-156
as cultural informants, 22-23 brief overview, 157-159
as detectors, 23 multivariate, 163-165
odds ratio, 128 time-stamped videotapes, 44
omnibus tests, 118 Timed Event Sequential Data (TSD), 84-85,
orderliness of sequences, 139-141 88r, 90
transforming SDIS data types, 88-90, 151
paper and pencil recording, 52 transitional frequency matrices, 96, 97/
Pearson chi-square statistic, see chi-square transitional probability, 95, 99, 105f, 134
statistic, Pearson descriptive value, 103-105
percentage of agreement, 60 limitations of, 106-107
percentages, analysis of, 93, 99 nonadjacent events, 105-106, 133
periodogram, 160 transitional probability matrices, 97-98, 98f
permutation tests, 147 type I error, 131-132, 147-148
phase-space plots, 170-171
phi coefficient, 129-130 unconditional probability, see simple probability
 Index 207

units Yule's Q, 129

for agreement, 68
for recording, 38 z scores, 108
univariate scaling of codes, 157 Allison-Liker's formula, 110
unreliability as research variable, 79-80 effect of identical adjacent events on,
108-109, 111
winnowing significant adjusted residuals, Sackett's formula, 109
119-120 standard formula, 109