You are on page 1of 6

Ann Nutr Metab 2014;65:205–210 Published online: November 18, 2014

DOI: 10.1159/000360229

Group-Based Trajectory Modeling:


An Overview
Daniel S. Nagin
Carnegie Mellon University, Pittsburgh, Pa., USA

Key Words an outcome over age or time. A detailed account of the


Trajectory groups · Finite mixture modeling · Group-based statistical underpinnings of the method and a full range
trajectory modeling of applications are provided in the literature [1, 2].
In this discussion, the term developmental trajectory
is used to describe the progression of any phenomenon,
Abstract whether behavioral, biological or physical. Charting and
This article provides an overview of a group-based statistical understanding developmental trajectories is among the
methodology for analyzing developmental trajectories – the most fundamental and empirically important research
evolution of an outcome over age or time. Across all applica- topics in the social and behavioral sciences and medicine.
tion domains, this group-based statistical method lends it- Longitudinal data – data with a time-based dimension –
self to the presentation of findings in the form of easily un- provide the empirical foundation for the analysis of de-
derstood graphical and tabular data summaries. In so doing, velopmental trajectories. Most standard statistical ap-
the method provides statistical researchers with a tool for proaches for analyzing developmental trajectories, in-
figuratively painting a statistical portrait of the predictors cluding hierarchical modeling [3, 4] and latent curve
and consequences of distinct trajectories of development. analysis [5, 6], are designed to account for individual vari-
Data summaries of this form have the great advantage of be- ability about a mean population trend. However, many of
ing accessible to nontechnical audiences and quickly com- the most interesting and challenging problems in longi-
prehensible to audiences that are technically sophisticated. tudinal analysis have a qualitative dimension that allows
Examples of the application of the method are provided. A for the possibility that there are meaningful subgroups
detailed account of the statistical underpinnings of the within a population that follow distinctive developmental
method and a full range of applications are provided by the trajectories that are not identifiable ex ante based on some
author in a previous study. © 2014 S. Karger AG, Basel measured set of individual characteristics (e.g. gender or
socioeconomic status). In psychology and medicine, for
example, there is a long tradition of taxonomic theorizing
about distinctive developmental progressions of these
Introduction subcategories. For research problems with a taxonomic
dimension, the aim is to chart out the distinctive trajec-
This paper provides an overview of group-based tra- tories, to understand what factors account for their dis-
jectory modeling (GBTM), a statistical methodology for tinctiveness and to test whether individuals following the
analyzing developmental trajectories – the evolution of different trajectories also respond differently to a treat-

© 2014 S. Karger AG, Basel Daniel S. Nagin


0250–6807/14/0653–0205$39.50/0 Carnegie Mellon University
5000 Forbes Avenue
E-Mail karger@karger.com
Pittsburgh, PA 15213 (USA)
www.karger.com/anm
E-Mail dn03 @ andrew.cmu.edu
ment, such as a medical intervention, or a major life event,
such as the birth of a child. GBTM, which was first ad- Low (14.4%) High declining (27.6%)
vanced by Nagin and Land [7] in 1993, provides the ca- 4.0 Moderate declining (53.7%) Chronic (4.3%)
pacity for conducting group-based analysis with time- 3.5
and age-based data.
3.0
Muthén and colleagues [8–10] have since developed

Physical aggression
an alternative group-based approach called growth mix- 2.5
ture modeling [for comparative discussion of these alter- 2.0
native approaches, see 1, 2, 10]. However, for the purpose
of this overview, the differences between the methods are 1.5

secondary to their both being group-based approaches 1.0


involving application of finite mixture modeling.
0.5
Across all application domains, the group-based tra-
jectory statistical method, whether of the GBTM or 0
growth mixture modeling variety, lends itself to the pre- 6 10 11 12 13 14 15
Age (years)
sentation of findings in the form of easily understood
graphical and tabular data summaries. In so doing, the
method provides statistical researchers with a tool for fig-
Fig. 1. Trajectories of physical aggression.
uratively painting a statistical portrait of the predictors
and consequences of distinct trajectories of development.
Data summaries of this form have the great advantage of
being accessible to nontechnical audiences and quickly 15 they continue to display a modest level of physical ag-
comprehensible to audiences that are technically sophis- gression. Finally, there is a small group of ‘chronics’, com-
ticated. prising less than 5% of the population, who display high
levels of physical aggression throughout the observation
period.
An Illustration of GBTM Much could be said about the implications of these
trajectories for the development of physical aggression,
Figure 1 reports a well-known application of GBTM but for our purposes here, two implications are empha-
that was first reported by Nagin and Tremblay [11]. It is sized. One implication follows from the observation that
based on data assembled as part of the Montreal Longitu- all the trajectories are either stable or declining from the
dinal-Experimental Study of Boys that has tracked 1,037 initial assessment at age 6. This implies that to understand
males from school entry through young adulthood. As- the developmental origins of physical aggression it is im-
sessments were made on a wide range of factors. Among portant to begin studying physical aggression at an even
these were teacher reports of each boy’s physical aggres- earlier age. A second and related observation is that the
sion at the age of 6 years and again annually from the age onset of physical aggression is not in adolescence as sug-
of 10 to 15 years. The scale was based on items such as gested by many theories of delinquent behavior (see
frequency of fighting and physically bullying. Tremblay and Nagin [12] for a full development of these
The best model was found to involve four groups. A two observations).
group called ‘lows’ is comprised of individuals who dis- These two points are highlighted because they illus-
play little or no physically aggressive behavior. This group trate the value of conducting longitudinal analysis in
is estimated to comprise about 15% of the sample popula- terms of groups. Nagin [1] and Nagin and Odgers [2] em-
tion. A second group, comprising about 50% of the popu- phasize that the groups should not be interpreted as lit-
lation, is best labeled ‘moderate declining’. At age 6, boys eral entities. Instead, they should be thought of as latent
in this group displayed a modest level of physical aggres- longitudinal strata in the data that are composed of indi-
sion, but by 10 years they had largely desisted. A third viduals following approximately the same development
group, comprising about 30% of the population, is labeled course on the outcome of interest. These strata identify
‘high declining’. This group starts off scoring high on distinctive longitudinal features of the data. In this appli-
physical aggression at age 6 but scores far lower by the age cation, the fact that all of the trajectories are stable or de-
of 15 years. Notwithstanding this marked decline, at age clining is a feature of the data that is of great substantive

206 Ann Nutr Metab 2014;65:205–210 Nagin


DOI: 10.1159/000360229
Color version available online
Table 1. Physical aggression group profiles
Low High declining
Variable Group 0.7 Moderate declining Chronic
low moderate high chronic 0.6
declining declining
0.5

Probability
Years of school – mother 11.1 10.8 9.8 8.4 0.4
Years of school – father 11.5 10.7 9.8 9.1 0.3
Low IQ, % 21.6 26.8 44.5 46.4
Completed 8th grade on 0.2
time, % 80.3 64.6 64.6 6.5 0.1
Juvenile record, % 0.0 2.0 6.0 13.3
Sexual partners at age 17 0

IQ

om

om

s
isk
ris

m
(past year), % 1.2 1.7 2.2 3.5

ho

lr
o

Lo
N

Al
d.

en
en

-e

Te
ok

w
Br

Lo
significance. Further, the absence of a feature, namely a
Fig. 2. Impact of risk factors on group membership probabilities.
trajectory reflecting the adolescent onset of physical ag-
ed. = Educated.
gression, also has important substantive significance.
The group-based methodology is intended to be re-
sponsive to calls for the development of ‘person-based’
approaches to analyzing development [13, 14]. Such ap- low aggression group are least likely to suffer from these
peals are motivated by a desire for methods that can pro- risk factors. Further, 90% of the chronic aggression group
vide a statistical snapshot of the distinguishing charac- fail to reach the eighth grade on schedule and 13% have a
teristics and behaviors of individuals following distinc- juvenile record by age 18. By comparison, only 19% of the
tive developmental pathways. The group-based method low aggression group had fallen behind grade level by the
lends itself to creating such profiles. Table 1 reports pro- eighth grade and none have a juvenile record. In between
files of the characteristics of individuals following the are the moderate- and high-declining groups.
four physical aggression trajectories shown in figure 1. Table 1 demonstrates that trajectory group member-
As developed in chapter 5 of Nagin [1], the parameter ship varies systematically with the individual’s psycho-
estimates of the model can be used to calculate the prob- social characteristics. An important generalization of
ability of an individual’s belonging to each of the trajec- the base model that is laid out in chapter 6 of Nagin [1]
tory groups. These probabilities are called the posterior allows for joint estimation of both the shapes of the tra-
probability of group membership. To create the profiles jectory groups and the impact of psychosocial charac-
reported in table 1, individuals were assigned to the tra- teristics on the probability of trajectory group mem-
jectory group to which they most likely belonged based bership. For example, such an analysis shows that the
on their measured history of physical aggression. The probability of physical aggression trajectory group
summary statistics reported in table 1 are simply the membership is significantly predicted by low IQ, low
product of a cross-tabulation of group membership with paternal education, family breakup prior to age 6 and
the various individual characteristics and outcomes. being born to a mother who began childbearing as a
(The posterior probabilities can also be used to compute teenager [15]. Figure 2 reports calculations of trajectory
weights that account for uncertainty in individual-level group membership for various combinations of these
trajectory group membership. However, use of these risk factors for physical aggression based on estimated
weights usually does not materially alter the profiles in model coefficients.
well-fitting models.) Trajectories are not immutable. Life events or inter-
The profiles conform to long-standing findings on the ventions may alter trajectories for the better or worse. Na-
predictors and consequences of problem behaviors such gin et al. [16] explored the effect of grade retention from
as physical aggression. Individuals in the chronic aggres- the age of 6 to 15 years on the trajectories of physical
sion group tend to have the least educated parents and aggression shown in figure 1. They found that grade re-
most frequently score in the lowest quartile of the IQ dis- tention seems to exacerbate physical aggression in the
tribution of the sample. By contrast, individuals in the low- and high-declining trajectory groups but has no ap-

Group-Based Trajectory Modeling Ann Nutr Metab 2014;65:205–210 207


DOI: 10.1159/000360229
parent effect on the physical aggression of the extreme munity, block face or an organization, or it can measure a
groups – the lows and the chronics. The model extension quantity such as a poverty rate or a mean salary level.
allowing for this sort of analysis is developed in chapter 7 Let P(Yi) denote the probability of Yi. As developed in
of Nagin [1] (see also Haviland et al. [17, 18] for a discus- chapter 2 of Nagin [1], for count data P(Yi) is specified as
sion of the use of propensity score matching in combina- the zero-inflated Poisson distribution, for censored data
tion with GBTM in making causal inferences about the it is specified as the censored normal distribution and for
effect of life events and interventions on developmental binary data it is specified as the binary logit distribution.
trajectories). Whatever the probability distribution, the ultimate objec-
Other important extensions of the basic model include tive is to estimate a set of parameters, Ω, that maximizes
joint/multitrajectory modeling and accounting for non- the probability of Yi. The particular form of this param-
random subject dropout. Joint and multitrajectory mod- eter set is distribution specific. However, across all distri-
eling are designed to link trajectories of behaviors or out- butions, these parameters perform the basic function of
comes that are thought to be theoretically related, such as defining the shapes of the trajectories and the probability
trajectories of the level of physical activity in childhood of group membership. As in standard growth curve mod-
and of body mass index in childhood and beyond, for ex- eling, the shapes of the trajectories are described by a
ample. Joint trajectory modeling probabilistically links polynomial function of age or time.
trajectories of two different outcomes. Multitrajectory If the parameters of this polynomial function were
modeling is designed to link trajectories for two or more constant across population members, the expected trajec-
outcomes by defining a trajectory in terms of trajectories tory of all population members would be identical. Nei-
for all of the outcomes of interest. This modeling exten- ther standard growth curve methods nor the group-based
sion is described in chapter 8 of Nagin [1]. The other ex- method assume such homogeneity. Indeed the assump-
tension generalizes the basic model laid out in the next tion of homogeneity is antithetical to the objective of ei-
section to account for nonrandom subject dropout. In ther approach because both aim to analyze the reason for
this model extension, each trajectory group is described individual differences in development. Standard growth
by a trajectory and the probability of trajectory group curve modeling assumes that the parameters defining the
membership at baseline and also by the probability of polynomial describe only a population mean and that the
dropout for each period after baseline. This extension is trajectories of individual population members vary con-
described in Haviland et al. [19]. tinuously about this mean, usually according to the mul-
tivariate normal distribution. The group-based method
assumes that individual differences in trajectories can be
Likelihood Function summarized by a finite set of different polynomial func-
tions of age or time. Each such set corresponds to a trajec-
Group-based trajectory models are a specialized appli- tory group which is hereafter indexed by j. Let Pj(Yi) de-
cation of finite mixture models. While the conceptual aim note the probability of Yi given membership in group j,
of the analysis is to identify clusters of individuals with and πj denote the probability of a randomly chosen popu-
similar trajectories, the estimated parameters of the mod- lation member belonging to group j.
el are not the result of a cluster analysis. Rather they are If it were possible to observe group membership, the
the product of maximum likelihood estimation. As such, sampled individuals could be sorted by group member-
they share the many desirable characteristics of maxi- ship and their trajectory parameters estimated with read-
mum likelihood parameter estimates – they are consis- ily available Poisson, censored normal (Tobit) and logit
tent and asymptotically normally distributed. regression software packages. However, group member-
The specific form of the likelihood function to be max- ship is not observed. Indeed, the proportion of the popu-
imized depends on the type of data being analyzed, but all lation comprising each group j, πj, is an important param-
are a special form of the following underlying likelihood eter of interest in its own right. Thus, construction of the
function: let Yi = yi1, yi2, ...., yit denote a longitudinal se- likelihood function requires the aggregation of the J con-
quence of measurements on individual i over T periods. ditional likelihood functions, Pj(Yi), to form the uncondi-
For expositional convenience, yit will generally be de- tional probability of the data, Yi:
scribed as the behavior of an individual. However, the out- J
come of interest does not have to pertain to an individual P Yi  œ ␲ j P j Yi , (1)
j
or a behavior – yit can reference an entity such as a com-

208 Ann Nutr Metab 2014;65:205–210 Nagin


DOI: 10.1159/000360229
where P(Yi) is the unconditional probability of observing model assumes that the sequential realizations of yit are
individual i’s longitudinal sequence of behavioral mea- independent, conditional upon the individual’s random
surements, Yi. It equals the sum across the J groups of the effect. Thus, in the group-based model, the conditional
probability of Yi given i’s membership in group j weight- independence assumption is made at the level of the
ed by the probability of membership in group j. Equation group, whereas in the random effect model it is invoked
1 describes what is called a ‘finite mixture model’ because at the level of the individual. In this sense, the condition-
it sums across a finite number of discrete groups that al independence assumption is stronger in the group-
comprise the population. The term ‘mixture’ is included based model than in the standard random effect model.
in the label because the statistical model specifies that the Balanced against this disadvantage is the advantage that
population is composed of a mixture of unobserved the group-based model does not make the very strong as-
groups. sumption that the random effect is independently and
For given j, conditional independence is assumed for identically distributed according to the normal distribu-
the sequential realizations of the elements of Yi, yit, over tion.
the T periods of measurement. Thus, The likelihood for the entire sample of N individuals is
T
simply the product of the individual likelihood functions
P j Yi   p j yit , (2) of the N individuals comprising the sample:
N
where pj(yit) is the probability distribution function of yit L   P Yi .
given membership in group j.
The rationale for the conditional independence as- Intuitively, the estimation procedure for all data types
sumption deserves elaboration. This assumption implies identifies distinctive trajectory groups as follows. Sup-
that for each individual within a given trajectory group j, pose a population is composed of two distinct groups: (i)
the distribution of yit for period T is independent of youth offenders (comprising 50% of the population) who
the realized level of the outcome in prior periods, yit – 1, up to age 18 have an expected offending rate, λ, of 5 and
yit – 2, …. Thus, pj(yit) does not include prior values of yit who after age 18 have a λ of 1 and (ii) adult offenders
in its specification. This assumption greatly reduces the (comprising the other 50% of the population) whose of-
complexity of an already complex model. Due to this re- fending trajectory is the reverse of that of the youth of-
duction in complexity, most applications of finite mix- fenders – through age 18 their λ = 1 and after age 18 their
ture modeling with longitudinal data assume conditional λ increases to 5. Longitudinal data on the recorded of-
independence for the sake of tractability. fenses of a sample of individuals from this population
On its face, the conditional independence assumption would reveal two distinct groups: a clustering of about
may seem implausible because it would seem to imply 50% of the sample who have had many offenses prior to
that current behavioral outcomes are uncorrelated with 18 years and relatively few offenses after age 18, and an-
past outcomes. At the level of the group which is not ob- other 50% clustering with the reverse pattern.
served, this is indeed the case. For individuals within a If these data were analyzed under the assumption
given group j, behavioral outcomes over time are as- that the relationship between age and λ was identical
sumed not to be serially correlated in the sense that indi- across all individuals, the estimated value of λ would be
vidual level deviations from the group trend are uncor- a ‘compromise’ estimate of about 3 for all ages. From
related. However, even with the assumption of condition- this, one might mistakenly conclude that the rate of of-
al independence at the level of the latent group, there will fending is invariant with age in this population. If the
still be serial dependence over time at the level of the pop- data were instead analyzed using the group-based ap-
ulation. Specifically, past outcomes will be correlated proach, which specifies the likelihood function as a mix-
with current outcomes. Such serial dependence results ing distribution, no such mathematical ‘compromise’
from the group-specific specification of pj(yit). Differenc- would be necessary. The parameters of one component
es in this specification across groups allow for persistent of the mixture would effectively be used to accommo-
differences in the outcome variable across population date (i.e. match) the youth offending portion of the data
members. whose offending declines with age and another compo-
The conditional independence assumption is also in- nent of the mixing distribution would be available to
voked in the standard random effect model that underlies accommodate the adult offender data whose offending
conventional growth curve models. The random effect increases with age.

Group-Based Trajectory Modeling Ann Nutr Metab 2014;65:205–210 209


DOI: 10.1159/000360229
Concluding Remarks vide researcher’s with a valuable tool for identifying, sum-
marizing and communicating complex patterns in longi-
A hallmark of modern longitudinal studies is the vari- tudinal data.
ety and richness of measurements that are made about the Summarizing data necessarily requires reduction. Re-
study subjects and their circumstances. Less often ac- duction requires approximation. In the case of group-
knowledged is that this abundance of information is ac- based models, the approximation involves the grouping
companied by a difficult companion – complexity. Com- of individuals who are not entirely homogenous. Bal-
monly, researchers are confronted with the dilemma of anced against this reduction error is a greatly expanded
how best to explore and communicate the rich set of mea- capability for creating dense, yet comprehensible, de-
surements at their disposal without becoming so bogged scriptions of groups of people through time.
down in complexity that the lessons to be learned from
the data are lost on them and their audience.
Acknowledgment
An important motivation for my commitment to de-
veloping and promoting the group-based trajectory This research has been supported by the National Science
method is the belief that alternative methods for analyz- Foundation (NSF: SES-99113700 and SES-0647576) and the Na-
ing development in longitudinal data sets too often leave tional Institute of Mental Health (RO1 MH65611-01A2).
the researcher with a Hobson’s choice of balancing com-
prehensibility against an adequate exploration of com- Disclosure Statement
plexity. GBTM does not solve the problem of balancing
comprehensibility and complexity. However, it does pro- The author has nothing to disclose.

References
1 Nagin DS: Group-Based Modeling of Devel- 9 Muthén B, Brown CH, Masyn K, Jo B, Khoo 15 Nagin D S, Tremblay R E: Parental and early
opment. Cambridge, Harvard University ST, Yang CC, et al: General growth mixture childhood predictors of persistent physical
Press, 2005. modeling for randomized preventive inter- aggression in boys from kindergarten to high
2 Nagin DS, Odgers CL: Group-based trajecto- ventions. Biostatistics 2002;3:459–475. school. Arch Gen Psychiatry 2001; 58: 389–
ry modeling in clinical research. Annu Rev 10 Muthén B: Latent variable analysis: growth 394.
Clin Psychol 2010;6:109–138. mixture modeling and related techniques for 16 Nagin D, Pagani L, Tremblay R, Vitaro F: Life
3 Bryk AS, Raudenbush SW: Application of hi- longitudinal data; in Kaplan D (ed): Hand- course turning points: a case study of the ef-
erarchical linear models to assessing change. book of Quantitative Methodology for the fect of school failure on interpersonal vio-
Psychol Bull 1987;101:147–158. Social Sciences. Newbury Park, Sage, 2004, pp lence. Dev Psychopathol 2003;15:343–361.
4 Goldstein H: Multilevel Statistical Models. 345–368. 17 Haviland A, Nagin DS, Rosenbaum PR: Com-
London, Arnold, 1995. 11 Nagin DS, Tremblay RE: Trajectories of boys’ bining propensity score matching and group-
5 McArdle JJ, Epstein D: Latent growth curves physical aggression, opposition, and hyperac- based trajectory modeling in an observational
within developmental structural equation tivity on the path to physically violent and study. Psychol Methods 2007;12:247–267.
models. Child Dev 1987;58:110–113. nonviolent juvenile delinquency. Child Dev 18 Haviland A, Nagin DS, Rosenbaum PR,
6 Willett JB, Sayer AG: Using covariance struc- 1999;70:1181–1196. Tremblay RE: Combining group-based tra-
ture analysis to detect correlates and predic- 12 Tremblay RE, Nagin DS: Aggression in hu- jectory modeling and propensity score match-
tors of individual change over time. Psychol mans; in Tremblay RE, Hartup WW, Archer ing for causal inferences in nonexperimental
Bull 1994;116:363–381. J (eds): Developmental Origins of Aggression. longitudinal data. Dev Psychol 2008; 44: 422–
7 Nagin DS, Land KC: Age, criminal careers, New York, Guilford, 2005, pp 83–106. 436.
and population heterogeneity – specification 13 Bergman LR: A pattern-oriented approach to 19 Haviland A, Jones B, Nagin DS: Group-based
and estimation of a nonparametric, mixed studying individual development: snapshots trajectory modeling extended to account for
Poisson model. Criminology 1993; 31: 327– and processes; in Cairns RB, Bergman LR, Ka- non-random subject attrition. Sociol Meth-
362. gan J (eds): Methods and Models for Studying ods Res 2011;41:367–390.
8 Muthén B, Shedden K: Finite mixture model- the Individual. Thousand Oaks, Sage, 1998,
ing with mixture outcomes using the EM al- pp 83–122.
gorithm. Biometrics 1999;55:463–469. 14 Magnusson D: The logic and implications of
a person-oriented approach; in Cairns RB,
Bergman LR, Kagan J (eds): Methods and
Models for Studying the Individual. Thou-
sand Oaks, Sage, 1998, pp 33–64.

210 Ann Nutr Metab 2014;65:205–210 Nagin


DOI: 10.1159/000360229

You might also like