Epidemiologia e Psichiatria Sociale

http://journals.cambridge.org/EPS

Additional services for Epidemiologia e Psichiatria Sociale:

Email alerts: Click here

Subscriptions: Click here
Commercial reprints: Click here
Terms of use : Click here

Event-based categorical sequential analyses of the medical interview: a

review

Maria Angela Mazzi, Lidia Del Piccolo and Christa Zimmermann

Epidemiologia e Psichiatria Sociale / Volume 12 / Special Issue 02 / June 2003, pp 81 - 85

DOI: 10.1017/S1121189X00006126, Published online: 11 October 2011

Link to this article: http://journals.cambridge.org/abstract_S1121189X00006126

How to cite this article:

Maria Angela Mazzi, Lidia Del Piccolo and Christa Zimmermann (2003). Event-based categorical sequential analyses of
the medical interview: a review. Epidemiologia e Psichiatria Sociale, 12, pp 81-85 doi:10.1017/S1121189X00006126

Request Permissions : Click here

Downloaded from http://journals.cambridge.org/EPS, IP address: 128.122.253.212 on 12 May 2015

 Event-based categorical sequential analyses
of the medical interview: a review
MARIA ANGELA MAZZI, LIDIA DEL PICCOLO, CHRISTA ZIMMERMANN
Department of Medicine and Public Health, Service of Medical Psychology, University of Verona, Verona, Italy

SUMMARY. When the doctor-patient interaction is viewed as a series of utterances, the temporal position of utterances
becomes a central information in understanding the nature of interaction. Important concepts are interdependence and serial depen-
dence which account for the fact that two partners influence each other in their talk and that each partner influences him/herself.
Lag sequential analysis studies the associations between doctor and patient utterances in a two-way contingency table (lag one
sequences) and is used for exploratory purposes. Log-linear modelling, based on multi-way contingency tables, is used as an exten-
sion of lag-sequential analysis to study longer sequences.
Markov chains test sequences in terms of processes with the aim to find predictive models and require a theory driven approach.
Pattern recognition aims to discover regularities in the temporal evolution of the utterance sequences. Theory driven applications
analyse manifest patterns in terms of their conditional probability distribution while empirically driven applications are used to
detect "hidden" patterns. These different approaches to sequential data can be regarded as complementary tools to describe the doc-
tor patient consultations at various levels of complexity.

Declaration of Interest: none.

KEY WORDS: doctor patient conversation, sequential dyadic interaction, lag-sequential analysis, categorical data series.

INTRODUCTION ral position becomes a central information in understan-

Sequential methods allow to study the doctor-patient
consultation in terms of interaction between two indivi-
duals.
After the translation of the doctor-patient consultation
into a series of codes, interaction may be studied from
many different points of view. The focus of research may
be on detecting specific recurring sequential patterns, on
studying the time lag in response to some given stimulus
(to evidence the relationship between what doctor and
what patient says), or on testing differences between
groups of interviews in terms of antecedent-consequent
events (to distinguish, for example, interviewing styles
before and after a training or to detect gender differences
in conducting the consultation).
Bakeman & Gottman (1986) described the relation-
ship between the behaviours of two persons as two sepa-
rate series of events in interaction. When the dyadic inte-
raction is between a series of codes/events, their tempo-
ding the relationship between occurrences of events.
The theory of the social exchange (Dumas, 1986)
claims that in a dyadic interaction process each partner's
behaviour follows from his/her previous behaviour and it
is only partially influenced by the behaviour of the other.
According to Stiles et al. (1998), instead, participants of
a conversation continually adjust their actions to achieve
desirable outcomes; the authors called this the "responsi-
veness of human interaction". Interaction in terms of
"responsiveness" seems the more convincing description
of what happens during a medical consultation.
Expressed in statistical terms this assumes that the serial
dependence between physician and patient speech is
high, but must be quantified and distinguished from the
dependence "within speakers" stressed by Dumas (1986),
when we want to understand the complex structure of the
verbal interaction.
We will therefore focus our attention on the statistical
methods which take into account these serial dependen-
cies between and within subjects (here patients and phy-
sicians).

Address for correspondence: Dr. M.A. Mazzi, Department of

Medicine and Public Health, Section of Psychiatry, Service of Medical
Psychology, University of Verona, Policlinico G.B. Rossi, Piazzale
L.A. Scuro 10, 37134 Verona (Italy).
Fax: +39-045-585871
E-mail: mariangela.mazzi@univr.it
DEFINITIONS AND CONCEPTS
Generally sequential analyses pursue two aims
(Bakeman & Gottman, 1986). The first one is to discover

Epidemiologia e Psichiatria Sociale, 12, 2, 2003

81
 M.A. Mazzi et al.
probabilistic patterns in the stream of code events; in
other words, the interest is centred on the order and the
prevailing sequences that characterise the data. The
second aim is to assess the effect of contextual or expla-
natory variables on the sequential structure of interac-
tions (here the doctor-patient dialogue).
There is no unique definition of sequential analysis. In
statistics the term refers to the experimental theory, in
which probability tests are posed in sequence, as is done
when applying hierarchical, or nested, methods. These
techniques are of no pertinence to conversational analy-
ses; at least as discussed here. In the biological and medi-
cal context "sequential analysis" is a proxy term of the
linkage analysis. It is used in the field of human genetics
which study the patterns hidden in the DNA "alphabet".
The application of some of these sophisticated statistical
techniques might be considered also for the exploration
of patterns in conversational analyses.
In psychology sequence analysis is mainly intended in
terms of lag sequential analysis by which sequentially
recorded behavioural data are analysed in observational
studies of social systems (Sackett, 1979; Gottman & Roy,
1990; Bakeman & Quera, 1995b). This approach investi-
gates the position of events of interest in the temporary
distribution of their occurrences and seems appropriate
when the intent is exploratory.
Serial dependence and interdependence of event codes
The literature on behavioural processes distinguishes
two kinds of dependencies in the dyadic conversation
(Dumas, 1986; Van Beek et al, 1992): the serial depen-
dence and the interdependence.
Serial dependence is also called autodependence or
autocontingency (Altham, 1979; Tavare & Altham,
1983), or "dependence within subject" and accounts for
the fact that a subject may influence him/herself. When
the patient (or doctor) says something, it may be influen-
ced by what s/he has said before. The dependence within
subject is an important element in the basic statistical
assumptions underlying sequential analysis, because
autodependence is a non-random aspect of a behavioural
sequence.
Interdependence refers to the "dependence between
subjects" which occurs when two subjects influence each
other. It is also called cross-contingency or joint depen-
dence, which derives from the interaction of two distinct
sources (either patient or doctor).
The choice of the statistical technique has to take into
account this complexity of the units of analysis (code
categories of speech).
Epidemiologia e Psichiatria Sociale, 12, 2, 2003
82
LAG-SEQUENTIAL ANALYSIS
This approach was introduced by Sackett (1979) in
order to analyse complex interactions between mother
and infant. This approach was adopted by Gottman &
Bakeman (1979), Bakeman & Gottman (1986; 1997) and
Bakeman & Quera (1992; 1995a, b), who applied the lag-
sequential analysis to study the association among cate-
gories in a two-way contingency table. The aim was to
use relatively simple summary statistics on rather com-
plex data. This approach did not require strong a-priori
hypotheses and the applications were essentially explora-
tory.
For an easier comprehension of the temporal position
of events, they adopted the notation of transitional matri-
ces from the Markov chain theory (Lehoczky, 1998). The
Markov chain theory actually implies an inferential con-
firmatory approach to the sequential processes and requi-
res strong assumptions in terms of distribution, temporal
homogeneity and dependency structure (Faraone &
Dorfman, 1987). In contrast, the exploratory characteri-
stic of lag sequential analysis permits to ignore these
restrictions, but this advantage reduces the power of the
statistical tests.
The information related to the sequence of codes is
structured by using a transitional contingency table which
is the basis for any sequence analysis. This table usually
represents coded events in terms of their displacement in
time (lag). Rows represent the first or given event (for
example, physician utterance) and columns the following
or target event (what the patient says subsequently, that is
at lag one or greater than one). The lag is the distance, in
terms of subsequent code units or events, between the
given and the target event considered. Each cell repre-
sents the frequency of chains that begin with the given
event and end with the target event. The same chain may
be expressed in terms of transition probabilities, dividing
the frequency of the cell by the frequency of the row.
The advantage of transition probabilities is that "they
"correct" for differences in base rates for the "given"
behavioural states ... [and] easily show the most likely
ways of "moving" from one state to another" (Bakeman
& Gottman, 1997, p. 104).
However, the disadvantages, according to the authors,
are the great variability among subjects in transition pro-
babilities and their questionable use for comparing grou-
ps. An example of this application can be seen in Eide et
al. (2003).
The sequential techniques presented below are distin-
guished into those which restrict the analyses to imme-
diately following (lag 1) behaviours and to those which
 Event-based categorical sequential analyses of the medical interview: a review
refer to a multivariate contest, considering more than lag
1 sequences. The assumption underlying lag 1 sequence
analysis, rarely declared, is the so-called "fist-order
Markov chain". It implicitly assumes that there is only a
short-term memory effect in the sequences, namely that
each behaviour is influenced by the immediately prece-
ding behaviour. It would be therefore more correct to call
lag 1 sequential analysis "contiguity analysis".
Lag 1 sequential analyses
The basic conditions for contingency tables require
that observations are probabilistically independent, show
the same distribution and that their number is large
(Wickens, 1989). To satisfy the last condition the total
sample should be at least four or five times the number of
cells (or more if the distribution of the marginal category
frequencies is skewed), with the exception of a few cells
with low frequencies (in a large table no more than 20%
of total cells may have frequencies less than 5).
The classical methods of contingency analysis (for
example, chi-square test, phi index, kappa index) should
not be applied to transitional tables that describe sequen-
ces of behaviour, because these last requirements are
unlikely to be met by these type of data.
The scientific debate about the appropriateness of the
statistical procedures regarding sequence analysis was
already lively twenty years ago (Allison & Liker, 1982;
Tavare & Altham, 1983; Dumas, 1986; Faraone &
Dorfman, 1987). In particular it was pointed out that the
simple chi-square index is inadequate to test the presence
of interdependence between two "interactants", and that
the autodependence needs to be controlled and quanti-
fied. Tavare & Altham (1983) proposed the gamma cor-
rection factor to multiply with the chi-square in order to
infer cross contingency (interdependency) between event
categories of two partners. Gamma index is a function of
the two autocontingencies, estimated by the Phi coeffi-
cients of each separate within-partner sequence (summa-
rised in a distinct transition contingency table for either
patient or doctor). Because the gamma value is comprised
between 0 and 1, it decreases the value of the Pearson
chi-square by removing the component due to autocon-
tingencies within partners.
Several solutions were proposed also to resolve the
question of the distribution requirements of statistical
tests. Faraone & Dorfmann (1987) used two robust tech-
niques to estimate the correct variance of a cross-depen-
dence index and its confidence intervals, based on jack-
knife and data-split approaches. Bakeman et al. (1996a)
introduced the permutation test approach on some asso-
ciation or transition indices (Yule's Q, odds ratio,
Epidemiologia e Psichiatria Sociale, 12, 2, 2003
83
Wampold's K, z-score) to improve the estimation criteria
when the parametric assumptions are not confirmed or
when the sample size is small. They then compared the
transition indices between groups, to test different
sequence characteristics.
LOG-LINEAR MODELLING
An extension beyond lag 1 needs multivariate statisti-
cal techniques, which take into account more than one
subsequent dependency. This approach allows to analyse
for how long (at lag 2, 3, 4, etc.) the reciprocal influence
of physician and patient speech is operating (sequence
memory). Log-linear models are often applied to lag-
analysis, because they test the association among catego-
ries in a multi-way contingency table (Bakeman & Quera,
1995; Van-Beek et al, 1992).
The log-linear model is a regression function for fre-
quency data, in which the logarithm of the expected value
of a count variable is modelled as a linear function of
parameters (explanatory variables). The parameters can
evidence the main effects (the common part shared by all
interviews), the marginal effects (the information provi-
ded from marginal rows and columns - they represent the
specific style of each partner) and the interaction effects
(the association among variables indicates the autodepen-
dence within a subject, when even lag distances are con-
sidered, and the interdependence between subjects, when
odd distances are explored). The best model (the one that
fits the observed data in the simplest way) is usually
selected from hierarchical models by a stepwise criterion
which extracts the relevant interaction parameters. By
this way it is possible to quantify the length of the
memory effect.
Once obtained the best model, the analysis of residuals
(Agresti, 1996) is used to compare observed and fitted
counts and to indicate those that contain sequences occur-
ring with a significantly higher or lower frequency than
expected. The residual frequency is thus a measure of the
lack of fit (distance of observed from expected counts)
where the expected frequencies reflect the hypothesis of
independence.
Another interesting technique to study multi-way con-
tingency tables is logit-linear modelling (Budescu, 1984),
by which a more traditional distinction between the
response variables and the explicative variables factors,
as in ANOVA, is maintained. It permits to investigate
non only the presence of interdependence, but also
whether a partner dependence prevails over the other one
(for example, patient behaviour is more dependent on
 M.A. Mazzi et al.
physician behaviour than vice versa). This is called "lag-
ged dominance".
Markov Chain models
The Markov chains are defined as a collection of ran-
dom variables with a specific dependency structure, cal-
led Markov property. This theory assumes that to make
predictions about "future" behaviours it is sufficient to
consider only its present state and not its past history.
Faraone & Dorfman (1987) discuss the applicability of
this technique to the study of behavioural sequences.
They discuss the underlying assumptions and their vali-
dity. The assumptions are stationarity, ergodicity and
order of dependency. The first postulates that the proba-
bilities of the occurrences of events and their transition
probabilities do not change over time. Stationarity is also
called time homogeneity. In terms of doctor-patient con-
sultations this means, for example, that the probability of
a specific sequence during the dialogue remains inva-
riant. A chain is ergodic when all components (sequence
of two code events) have a likelihood of occurrence (tran-
sition probability from one code to the other) greater than
zero. The order of dependency is the duration of the
memory effect under study (at what lag a code event
stops to affect subsequent code events) and has to be defi-
ned a priori. Van-Beek et al. (1992) applied Markov
chains to mother-infant interactions, but they limited
their analysis at the second order dependence (lag 2).
The analysis of the complex structure of the doctor
patient consultation would require a higher order Markov
technique or the more sophisticated Hidden Markov
Chain (MacDonald & Zucchini, 1997; Pentland & Liu,
1999) or interconnected Markov Chain (Avery &
Henderson, 1999; Avery, 2002), which so far have found
application only in the field of genetics.
Recognition of the occurrence of defined sequences
(pattern recognition): Conditional probability theory
applied to specific sequences of events
When the aim is to discover regularities in the tempo-
ral evolution of the conversation, the appropriate metho-
dology to investigate the behavioural interaction is the
pattern recognition. The pattern research can be both
hypothesis driven (an earlier part of the pattern may be
seen as a likely cause of the later part of the same pattern)
and empirically driven (the search for associations
between parts of sequences without specific hypotheses).
Since the structure of the human behaviour is very
complex, it is convenient to distinguish between the
manifest beliavioural patterns (visible and recurrent pat-
terns in behavioural streams) and the hidden beliavioural
Epidemiologia e Psichiatria Sociale, 12, 2, 2003
84
patterns (particular relations between groups or pairs of
events in a time series). In the last case the aim is to find
some nested relations among a complex sequence of
occurrences.
In the first approach the statistical techniques are easy
and are based on the conditional probability theory, but
they require to test theory driven hypotheses. Prior know-
ledge, theory based or based on previous studies, plays an
important role in determining the choice of the sequence
pattern under study. The aim is to test whether the expec-
ted pattern occurs among the observed patterns more
often than by chance in terms of its conditional probabi-
lity distribution. The conditional probability of a sequen-
ce is estimated by dividing the number of times it occur-
red by the numbers of times it could have occurred in the
observational records (the numbers of occurrences plus
the number of non-occurrences).
An example of this application can be seen in
Zimmermann et al. (2003).
This approach to describe behavioural sequence,
however, is not exhaustive, because patterns easily beco-
me invisible to the naked eye when other behaviours
occur in between. The "hidden pattern" approach helps to
distinguish which pattern is relevant (signal) and which is
irrelevant (noise), by reducing the extraneous sources of
variability (e.g. by increasing signal-to-noise ratio). The
tools for the study of hidden patterns refer to a specific
branch of statistics (called also classification theory)
which explore structure and patterns in large data sets,
without recourse to the classical assumptions of the con-
firmatory approach, often too rigid for practical use
(Lange, 1998). Magnusson (2000) defined hidden pat-
terns as "patterns of patterns of patterns" (T-patterns). He
proposed for the study of human interaction the "T-pat-
tern detection algorithm" to discover hierarchical pattern
(patterns of simpler sequence patterns). Magnusson
(2002) has suggested this approach to study DNA
sequences. This technique seems promising also for the
study of doctor-patient interactions at a high level of
complexity.
CONCLUSION
The different approaches illustrated above can be
regarded as complementary tools to treat the complex
structure of doctor patient consultations. None of these
techniques seems better than the other, but each of them
can show specific characteristics of the involved sequen-
ces. As Moran et al. (1992) suggest, to better understand
the relationships among behavioural events "efforts must
 Event-based categorical sequential analyses of the medical interview: a review

be focused to paint different pictures of the same scene, Eide H., Quera V. & Finset A. (2003). Exploring rare patient behaviour
with sequential analysis: an illustration. Epidemiologia e Psichiatria
each accurate in its own way" (p. 89). However, we Sociale 12, 109-114.
should not forget Wickens' (1993) warnings which help Faraone S.V. & Dorfman D.D. (1987). Lag sequential analysis: Robust
to avoid the pitfalls connected with the wrong use of sta- statistical methods. Psychological Bulletin 101 , 312-323.
Gottman J.M. & Backeman R. (1979). The sequential analysis of obser-
tistical techniques. He admonishes not to ignore the basic vational data. In Social interaction analysis: Metliodological issues
assumptions of each technique which have to be satisfied (ed. ME. Lamb, S.J. Soumi and G.R. Stephenson). University of
in order to have unbiased results. Wiscounsin Press: Madison.
Gottman J.M. & Roy A.K. (1990). Sequential Analysis: a Guide for
Behavioral Researchers Cambridge University Press: Cambridge.
Lange N. (1998). Pattern Recognition. In Encyclopedia of Biostatistics
REFERENCES (ed. P. Armitage and T. Colton), pp. 3298-3304. John Wiley &
Sons: West Sussex.
Lehoczky J. (1998). Markov Chains. In Encyclopedia of Biostatistics
Agresti A. (1996). An Introduction to Categorical Data Analysis. John
(ed. P. Armitage and T. Colton), pp. 2423-2428. John Wiley &
Wiley & Sons: New York.
Sons: West Sussex.
Allison P.D. & Liker J.K. (1982). Analyzing sequential categorical data
MacDonald I.L. & Zucchini W. (1997). Hidden Markov and Other
on dyadic interaction: A comment on Gottman. Psychological
Models for Discrete-valued Time Series. Chapman & Hall: London.
Bulletin 91, 393-403.
Magnusson M.S. (2000). Discovering hidden time patterns in behavior:
Altham P.M. (1979). Detecting relationships between categorical varia-
T-patterns and their detection. Behavioral Research. Methods,
bles observed over time: a problem of deflating a Chi-squared stati-
Instruments and Computers 32, 93-110.
stic. Applied Statistics 28, 115-125.
Magnusson M.S. (2002). T-patterns in behavior and DNA: detection
Avery P.J. (2002). Fitting interconnected Markov chain models: DNA
and analysis with Theme and GeneTheme. In Measuring Beliavior
sequences and test cricket matches. Statistician 51, 267-278.
2002, 4* International Conference on Methods and Techniques in
Avery P.J. & Henderson D.A. (1999). Fitting Markov chain models to
Behavioral Research, 27-30 August 2002, Amsterdam, The
discrete state series such as DNA sequences. Applied Statistics 48,
Netherlands.
153-161.
Moran G., Dumas J.E. & Symons D.K. (1992). Approaches to sequen-
Bakeman R. & Gottman J.M. (1986). Observing Interaction: an
tial analysis and the description of contingency in behavioral inte-
Introduction to Sequential Analysis. Cambridge University Press:
raction. Behavioral Assessment 14, 65-92.
Cambridge.
Pentland A. & Liu A. (1999). Modelling and prediction of human beha-
Bakeman R. & Gottman J.M. (1997). Observing Interaction: an
vior. Neural Computation 11, 229-242.
Introduction to Sequential Analysis, 2nd ed. Cambridge University
Sackett, G.P. (1979). The lag sequential analysis of contingency and
Press: Cambridge.
cyclicity in behavioral interaction research. In Handbook of Infant
Bakeman R. & Quera V. (1992). SDIS: A sequential data interchange
Development (ed. J.D. Osofsky). Wiley: New York.
standard. Behavioral Research. Methods, Instruments and
Computers 24, 554-559. Stiles W.B., Honos-Webb L. & Surko M. (1998). Responsiveness in
psychotherapy. Clinical Psychology: Science and Practice 5, 439-
Bakeman R. & Quera V. (1995a). Log-linear approaches to lag-sequen-
458.
tial analysis when consecutive codes may and cannot repeat.
Tavare S. & Altham P.M. (1983). Serial dependence of observations
Psychological Bulletin 118, 272-284.
leading to contingency tables, and corrections to chi-squared stati-
Bakeman R. & Quera V. (1995b). Analysing Interaction. Sequential
stics. Biometrika 70, 139-144.
Analysis with SDIS and GSEQ. Cambridge University Press:
Van-Beek Y., de-Roos B., Hoeksma J.B. & Hopkins B. (1992).
Cambridge.
Sequential analysis of nominal data in mother-infant communica-
Bakeman R., McArthur D. & Quera V. (1996a). Detecting group diffe-
tion: Quantifying dominance and bidirectionality. Beliaviour 122,
rences in sequential association using sampled permutations: Log
306-328.
odds, kappa, and phi compared. Behavioral Research Metliods,
Instruments and Computers 28, 446-457. Wickens T.D. (1989). Multiway Contingency Tables Analysis for tlie
Social Science. Lawrence Erlbaum Associates Inc.: Hillsdale.
Bakeman R., Robinson B.F. & Quera V. (1996b). Testing sequential
Wickens T.D. (1993). Analysis of contingency tables with between-
association: Estimating exact p values using sampled permutations.
subjects variability. Psychological Bulletin 113, 191-204.
Psychological Metliods 1, 14-15.
Zimmermann C , Del Piccolo L. & Mazzi M.A. (2003). Patient cues and
Budescu D.V. (1984). Tests of lagged dominance in sequential dyadic
medical interviewing in general practice: examples of the applica-
interaction. Psychological Bulletin 96, 402-414.
tion of sequential analysis. Epidemiologia e Psichiatria Sociale 12,
Dumas J.E. (1986). Controlling for autocorrelation in social interaction
115-123.
analysis. Psychological Bulletin 100, 125-127.

Epidemiologia e Psichiatria Sociale, 12, 2, 2003

85