0% found this document useful (0 votes)
94 views37 pages

Nagin (2005)

-

Uploaded by

Adda Barco
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
94 views37 pages

Nagin (2005)

-

Uploaded by

Adda Barco
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 37

American Academy of Political and Social Science

What Has Been Learned from Group-Based Trajectory Modeling? Examples from Physical
Aggression and Other Problem Behaviors
Author(s): Daniel S. Nagin and Richard E. Tremblay
Source: Annals of the American Academy of Political and Social Science, Vol. 602,
Developmental Criminology and Its Discontents: Trajectories of Crime from Childhood to Old
Age (Nov., 2005), pp. 82-117
Published by: Sage Publications, Inc. in association with the American Academy of Political and Social
Science
Stable URL: http://www.jstor.org/stable/25046143 .
Accessed: 25/01/2011 14:09

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=sage. .

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

Sage Publications, Inc. and American Academy of Political and Social Science are collaborating with JSTOR
to digitize, preserve and extend access to Annals of the American Academy of Political and Social Science.

http://www.jstor.org
The focus of this article is model
group-based trajectory
ing. Its purpose is threefold. The first is to clarify the
statistical of a
proper interpretation trajectory group.
The second is to summarize some on the
key findings
course of aggression and other
developmental problem
behaviors that have from the of
emerged application
models and that in the authors'
group-based trajectory
judgment are important to the fields of developmental
criminology and developmental psychopathology. The
third is to lay out some guidelines on the types of prob

What Has Been lems

may
for which
be
particularly
use of
group-based
productive.
trajectory modeling

Learned from Keywords: group-based trajectory modeling; physical


aggression; trajectory group

Group-Based
Trajectory use the term
developmental
Psychologists
trajectory to describe the course of a behav
Modeling? ior or outcome over age or time. Until about a
decade ago, the two main branches of methodol
Examples ogy for analyzing developmental trajectories
were hierarchical and
from Physical modeling
Raudenbush 1987, 1992; Goldstein 1995) and
(Bryk

latent curve (McArdle and Epstein


Aggression and analysis
1987;Meredith and Tisak 1990;Muth?n 1989;
Willett and Sayer 1994).A1993 article by Nagin
Other Problem and Land laid out a third alternative?group

Behaviors
Daniel S. Nagin is Teresa and H. John Heinz HI Profes
sor of Public Policy and Statistics at the Heinz School,
Carnegie Mellon University. He is an electedfellow of the
American Society of Criminology and of the American
Society for the Advancement of Science and is a 1985
By recipient of
the Northeastern Association of Tax Admin
in Tax Administration. He
DANIEL S. NAGIN istrators Award for Excellence
is author of Development
of Group-Based Modeling
and
(Harvard University Press, 2005).
RICHARD E. TREMBLAY Richard E. Tremblay isCanada Research Chair in child
development, professor of pediatrics/psychiatry/
on Chil
psychology, and director of the Research Unit
dren's at the
Psychosocial Maladjustment University of
Montreal. He is also the Joannes Groen at
Professor
Utrecht University in the Netherlands, director of the
Centre of Excellence for Early Child Development, afel
low of the Royal Society of Canada, afellow of the Acad
emy of Experimental Criminology, and theMolson fel
low of the Canadian Institute Advanced Research.
for
DOI: 10.1177/0002716205280565

82ANNALS, AAPSS, 602, November 2005


GROUP-RASED TRAJECTORYMODELING 83

based trajectory modeling. The group-based trajectory model is a


specialized
application of finite
mixture modeling. Using mixtures of suitably defined proba
to
bility distributions, the method is designed identify distinctive clusters of devel
opmental trajectories within the population.
The introduction of "canned" software for estimating group-based models has
resulted in a growing body of research based on this method. At this time, there are
two excellent software alternatives for estimating group-based trajectories models.
One is a SAS-based called Proc Traj. It is described in Jones, Nagin, and
procedure
Roeder (2001) and Jones and Nagin (2005) and in documentation available at
www.ncovr.org. Proc is to be inserted into the SAS software package.
Traj designed
Once inserted, SAS treats it like any other standard SAS procedure. The other
is a software package called
alternative widely used structural equation modeling
M-Plus, developed by Bengt Muth?n, Linda Muth?n, and colleagues (Muth?n and
Muth?n 1998-2004). (2004) reported that more than arti
Piquero fifty published
cles use group-based trajectory modeling.
The purpose of this article is threefold. The first is to clarify the proper statistical
interpretation of a trajectory group. The second is to summarize some key findings
on the course of
developmental aggression and other problem behaviors that have
from the application of group-based trajectory models, which in our
emerged
are to the fields of and develop
judgment important developmental criminology
mental psychopathology. The third is to lay out some guidelines on the types of

problems for which use of group-based trajectory modeling may be particularly


productive.

Group-Based Trajectory Modeling


As previously stated, hierarchical modeling and latent curve analysis are two
alternative to the for modeling
important approaches group-based methodology
processes. Like the these two alternatives
developmental group-based approach,
are to a statistical tool for measuring and explaining differences
designed provide
across members in their course. Because all three
population developmental
share the common of individual-level in
approaches goal modeling heterogeneity
trajectories, each must make technical assumptions about the dis
developmental
tribution of trajectories in the It is these assumptions that distinguish
population.
the three approaches.
While the assumptions underlying hierarchical modeling and latent curve anal
ysis differ in important respects, they also have important commonalities
(MacCallum et al. 1997; Willett and Sayer 1994; Raudenbush 2001). For our pur
poses, one commonality is crucial: both model the population distribution of tra
on continuous distribution functions. Unconditional models esti
jectories based
mate two key features of the population distribution of trajectory parameters?
their mean and covariance structure. The former defines average growth within
the population, and the latter calibrates the variance of growth throughout the
Conditional models are to explain this variability by relating
population. designed
84 THE ANNALS OF THE AMERICAN ACADEMY

to one or more
trajectory parameters explanatory variables. Modeling individual
level differences requires that assumptions be made about the distribution of tra
jectory parameters in the population. Both hierarchical modeling and latent curve
assume that the parameters are distributed the
analysis generally throughout
to the multivariate normal distribution.
population according
Group-based trajectory modeling takes a qualitatively different approach
to

modeling individual differences. Rather than assuming that the population distri
bution of trajectories varies continuously across individuals and in a fashion that
can a
ultimately be explained by particular multivariate distribution of population
it assumes that there may be clusters or
parameters (usually normal), groupings of
distinctive developmental trajectories that themselves may reflect distinctive etiol
In some the groups may be literal entities. For the
ogies. applications, example,
of some on the users' In many other
efficacy drugs depends genetic makeup. appli
cation domains, however, the groups should not be as
thought of literally distinct
entities. They serve rather as a statistical to a more
approximation complex
underlying reality.
One use of finite mixture models is to approximate a continuous distribution
function (Everitt and Hand 1981; Heckman and Singer 1984; Titterington, Smith,
and Makov 1985). For example, Everitt and Hand (1981) described the use of a
mixture of univariate normal distributions to approximate any unspecified
univariate distribution function. McLachlan and Peel (2000,8) described such use
of finite mixture as a "niche between and nonparametric
modeling parametric
to statistical estimation. . . . [M]ixture model-based are
approaches approaches
in are
parametric that parametric forms specified for the component density func
tions, but that they can also be regarded as nonparametric by allowing the number
of components to Pickles and (2003, 541) noted
[groups] grow." Similarly, Angold
that "both theoretical and empirical work in statistics has shown that the

nonparametric estimator of the underlying distribution, essentially the best fitting


distribution, is just such a set of discrete classes of this kind, even when the underly
ing distribution is continuous." For this reason, the group-based trajectory method
is often described as a semiparametric method (e.g., Nagin 1999, 2005; Nagin and
Tremblay 1999, 2001a).
The idea of using a finite number of groups to approximate a continuous distri
bution is easily illustrated with an example. Suppose that panel A
in Figure 1
the distribution of some behavior z. In B, this same distri
depicts population panel
bution is replicated and overlaid with a histogram that approximates its
shape.
Panel B illustrates that any continuous distribution with finite end points can be
a discrete distribution (i.e., a histogram) or a
approximated by alternatively by
finite number of "points of support" (i.e., the dark shaded "pillars"). A higher num
ber of support points yields a discrete distribution that more closely approximates
the true continuous distribution. However, simulation evidence in
reported
Brame, Nagin, and Wasserman (forthcoming) and Nagin (2005) suggests that rela
are to approximate even
tively few points of support required reasonably complex
continuous distributions of trajectories.
GROUP-RASED TRAJECTORYMODELING 85

FIGURE 1
USING GROUPS TO APPROXIMATE AN UNKNOWN DISTRIBUTION

Panel A

0.10

M
0.05 ?\

O.OO ?\

Panel B

0.10 -H

.M
0.05

O.OO ?\

use groups to approximate a continuous distribution?


Why population
Heckman and Singer (1984) built upon the approximating mix
capability of finite
ture models to construct a maximum likelihood estimator for the
nonparametric
distribution of unobservables in duration models. The motivation for this seminal
innovation was their observation that social science theory rarely provides theoreti
cal guidance on the distribution of unobserved individual differences, yet statisti
cal models of duration data were often sensitive to the assumed form of the distri
bution of such differences. Their proposed estimator finessed the problem of
86 THE ANNALS OF THE AMERICAN ACADEMY

to a distribution of unobserved individual differences


having specify by approxi
the distribution with a finite mixture model. This same motivation under
mating
lies the use of groups in group-based trajectory modeling?theory rarely provides
on the tra
guidance specific form of the population distribution of developmental
jectories. As McLachlan and Peel (2000, xix) observed, "Because of their flexibility,
[finite] mixture models are as a convenient,
being increasingly exploited
in which to model unknown distributional
semiparametric way shapes."
This brings us back to the key distinction between standard growth curve mod
eling and the group-based modeling method. Both approaches model individual
with a that links age to behavior. The
trajectories polynomial relationship
differ in their strategy for incorporating population hetero
approaches modeling
in the curve parameters (i.e., ?0, ?x, ?2, and ?3). In conventional
geneity growth
curve the parameters describing individual-level are
growth modeling, trajectories
assumed to be distributed to a the multivariate
according specific function, usually
normal distribution. In the semiparametric group-based trajectory model, the dis
tribution is a finite number of trajectory groups, aka points of
approximated by
support.
While users of conventional curve have demonstrated
growth modeling great
in the basic model to accommodate data that is
ingenuity adapting longitudinal
not distributed data or skewed data), these
clearly normally (e.g., binary highly
adaptations of the basic model do not resolve the more fundamental problem
described by Raudenbush (2001): the standard growth curve modeling structure is
not well mixtures of developmental
adapted for modeling complex trajectories
within a population in which members are not a common
population following
of or decline. He offered as an
developmental process growth depression example.
He observed, "Itmakes no sense to assume that everyone is increasing (or decreas
. . .
ing)
in
depression many persons will never be high in depression, others will

always be high, while others will become increasingly depressed" (p. 509). Instead,
he recommended amultinomial statistical framework such as that provided by the
return to this point in a later section, where we
group-based trajectory model. We
out on domains in which trajectory
lay guidelines application group-based
modeling may be particularly useful.

Useful Findings from Applications


of Group-Based Trajectory Modeling
next summarize some that derive from group
We findings and conclusions
based that in our showcase the strengths of the
trajectory modeling judgment
method. We organize these findings under three headings: (1) late onset physical
aggression is the exception not the rule, (2) clarifying developmental taxonomies,
and (3) clarifying the predictors and consequence of developmental trajectories.
All of these involve important themes in developmental criminology and

psychopathology.
GROUP-BASED TRAJECTORYMODELING 87

At the outset, we that we are not suggesting that these results were
emphasize
not obtainable some alternative method but
using only that the group-based
approach greatly facilitated their discovery and communication. In the next sec
tion, we discuss guidelines for the types of analyses for which group-based model
ing is particularly well suited.
Throughout this review, we will refer to trajectory groups as if they are literal
entities. It should be understood that this is literary convenience for clear commu
nication. As discussed in the prior section, trajectory groups, like all statistical mod
els, are not literal depictions of reality. Rather, they are only meant as a convenient
use of the term use of taxo
approximation. Our trajectory group is similar to the
nomic distinctions in Strickland, a and
biology. Hugh leading zoologist geologist
who was trying to change traditions in zoological nomenclature, observed,

Of course I mean a conventional


you will understand that by type-species only distinction,
to words, not to and like human titles, used as amatter of conve
referring only things; only
nience. Nature knows no more or than she does of Dukes
type-species "typical groups"
and Marquesses, (cited in Burkhardt and Smith 1988, 216)

We reiterate this point because the pervasive use of group-based thinking


in
the world around us and in our economic and social institu
explaining structuring
tions is testimony to the foundational role of groups in human cognition. Cognitive
use the term sch?mas as a label for the cognitive role of
psychologists describing
groups. Anderson (1980, 128) defined sch?mas as units of knowl
"large complex
of what we about of objects,
edge that organize much know general categories
classes of events, and types of people [italics added]." Anderson went on to observe,
"Schematic is a to process sets of information.
thought powerful way complex
However, schematic is to biases and distortion" In the
thought subject (p. 129).
context of the present discussion, one such cognitive distortion is the reification of
groups as nonexistent realities.
The risk of reification is particularly great when the groups are identified using a
statistical method. The group's reality is reinforced by the patina of scientific objec
tivity that accompanies statistical analysis and by the very language that is used to
describe the statistical findings, for example, "group 1,which comprises x percent
of the population, is best labeled ..."
The tendency to reify groups has important risks. One is in the conduct of public
a is small and its behavior is socially undesirable, such as committing
policy If group
crimes, the reification of the group as a distinct entity?rather than as an extreme on
a to the behavior,
continuum?may provoke draconian responses by creating the
of a line of between "them" and "us." Human
impression bright separation history
is instances in which a fictional group-based separation is the
replete with tragic
first step in the dehumanization of the "them." Two other risks are not inherently
insidious but still important. One is that reification creates the impression of the
a
immutability of the groups. The reification of groups may also trigger quixotic
quest for assessment instruments to assign individuals to their true trajec
designed
tory group. See Nagin and Tremblay (2005) for a discussion of these two risks.
88 THE ANNALS OF THE AMERICAN ACADEMY

Late onset is the exception, not the rale


physical aggression
Two important scientific committees reports in the 1990s on the
published
causes and of violence. The U.S. National Research Council panel con
prevention
cluded that "modern psychological perspectives emphasize that aggressive and
violent behaviors are learned responses to frustration, that can also be learned
they
as instruments for occurs
achieving goals, and that the learning by observing mod
els of such behaviors. Such models may be observed in the
family, among peers,
elsewhere in the mass media, or in violent
neighborhood, through the pornogra
phy, for example" (Reiss and Roth 1994, 7). Four years later, the report of the
Human Capital Initiative Committee on
Coordinating Reducing Violence (1996,
12) concluded, "In short, watching violent movies and television shows year after

[T]heories of desistance need tofocus


on multiple domains of behavior and
the linkages between them.

year and listening to brutal lyrics set to throbbing music can change one s attitudes
about antisocial, aggressive behavior. In children, it can lead to more aggressive
behavior and also can evoke unwarranted fears and defensive actions. Whatever
the violent content, movies and television exert powerful influences through visual
imagery and dramatic characterizations; video games may have similar effects."
Over the past decade, the authors have collaborated on a series of studies that
the held view, articulated in these two blue-ribbon
challenge widely panel reports,
that physical aggression is a learned behavior. Instead, we have
argued that physi
cal aggression is an innate behavioral not need to be learned.
tendency that does
Rather, it is a behavior that we must learn to control. a few studies have traced
Only
the developmental origins of physical aggression; the available evidence suggests
that, on average, it at a very young age?perhaps as young as two years old.
peaks
Cairns, Cairns, and colleagues (Cairns and Cairns 1994; Cairns et al. 1989) found
that physical aggression, on average, decreased rather steadily from ages ten to
et al. (1999) a in in
eighteen. Tremblay reported steady decline physical aggression
a of Montreal males from ages six to fifteen. Similar results are found in a
sample
Pittsburgh, Pennsylvania-based longitudinal study (Loeber and Hay 1997). Stud
ies of to age six are even rarer, but the little evidence that
physical aggression prior
is available suggests that humans begin physically aggressing others as soon as they
have the physical capacity to do so (Tremblay et al. 1999, 2004).
GROUP-BASED TRAJECTORYMODELING 89

Still, average tendencies may conceal the late emergence of physical aggression
among subpopulations. It is on this that has
point group-based trajectory modeling
a crucial role in the case the view that the
played methodological building against
late emergence of physical aggression is the norm rather than the
exception. Space
does not permit a full summary of this evidence (but see Tremblay and Nagin
[2005] for a full elaboration). Instead, we summarize evidence from several key
studies that highlight the important methodological role group-based trajectory
has in this inquiry into the
modeling played developmental origins of physical
aggression.
The first published study using group-based trajectory modeling that chal
the idea of late onset was and (1999).
lenged physical aggression Nagin Tremblay
This study was based on a prospective of about one thousand
longitudinal study
white, French-speaking males from low-socioeconomic-status in
neighborhoods
Montreal. Among many other variables, the data include teacher ratings of
physi
cal aggression at age six and again at ages ten to fifteen. Teachers were asked to rate
the frequency with which each
boy kicks, bites, and/or hits other children; fights
with other children; and bullies or intimidates other children. Figure 2 reports the
trajectory model in that
preferred four-group identified analysis. A group called
"lows" is composed of individuals who display little or no physically aggressive
behavior. This group is estimated to comprise about 15 percent of the
sample pop
ulation. A second group, comprising about 50 percent of the population, is best
labeled "moderate declining." At age six, boys in this group displayed a modest
level of physical aggression, but by age ten
they had largely desisted. A third group,
about 30 of the is labeled
comprising percent population, "high declining." This
group starts off on at age six but scores far lower
scoring high physical aggression by
age fifteen. this marked decline, at age fifteen continue to
Notwithstanding they
a modest a small group of
display level of physical aggression. Finally, there is
"chronics," making up less than 5 percent of the population, who display high levels
of physical aggression throughout the observation period.
These trajectories are notable both for what is present and what is not present.
As for what is present, all are stable or from age six on. Thus, over the
declining
from age six to fifteen, there is no evidence of rising physical aggression
period
even among a small in these data. As for what is not present, we see
subpopulation
no evidence of late-onset-like
trajectories of physical aggression, namely, a trajec
tory that rises from a zero or level at some point between six and fifteen.
negligible
Because the trajectories are at their highest at age six, this suggests that to under
stand the developmental origins of physical aggression in these we need to
boys,
look back in time prior to age six rather than forward in time into their adolescence.
Our failure to find late-onset-type trajectories of physical aggression is not
unique to these Montreal males. A of five additional
follow-up analysis prospective
studies?one more from Canada, two from New Zealand, and two
longitudinal
from the United States?again found no evidence of the onset of physical aggres
sion after age six et al. 2003).
(Broidy
It is also important to emphasize that the absence of late onset trajectories in
these analyses of six different data sets from around the world is not an artifact of
90 THE ANNALS OF THE AMERICAN ACADEMY

FIGURE 2
TRAJECTORIES OF PHYSICAL AGGRESSION

6 10 11 14 15

Low (14.4%) ? ? Moderate Declining (53.7%; ?Age ?*?Chronic


?High Declining (27.6%) (4.3%)

SOURCE: Nagin and Tremblay (1999).

the methodology itself. Figure 3 shows trajectories of indirect aggression reported


in C?t?, Vaillancourt, Barker, et al. (2004) from ages four to eight. The trajectories
are the an a
product of analysis of representative sample of about twelve hundred
Canadian children. Indirect aggression was measured with five items (Lagerspetz,
Bjorkqvist, and Peltonen 1988): "becomes friends with another as revenge," "says
bad things behind the other s back," "when mad at someone, gets others to dislike
"
him/her," "says to others: lets not be with him/her,' and "tells the other one s
secrets to a third
person." Observe that for a sizable minority of children, about 30
of the indirect aggression is rising over these ages.
percent sampled population,
This is not terribly surprising because, unlike physical aggression, indirect aggres
sion requires verbal facility and social awareness, two personal capacities that must
be learned. By comparison, the three trajectories of physical aggression from ages
two to that were also identified in this were all
eight analysis declining.
Another of evidence on the course of
important piece developmental physical
is in another analysis by C?t? and colleagues (C?t?,
aggression reported
et same data set used
Vaillancourt, LeBlanc, al, 2004). This analysis also used the
in C?t?, Vaillancourt, Barker, et al. (2004) but examined the developmental course
of physical aggression from ages two to eleven in ten cohorts of approximately
=
1,000 children (N 10,658). Group-based trajectory modeling again identified
three trajectory groups, which are reported in Figure 4. One-third of children
(31.1 percent) followed a low reflected in use of
desisting trajectory, infrequent
in toddlerhood and virtually no The
aggression aggression by preadolescence.
of children (52.2 percent) followed a moderate trajectory,
majority desisting
reflected in occasional use of
aggression in toddlerhood and infrequent use by
GROUP-BASED TRAJECTORYMODELING 91

FIGURE 3
OF INDIRECTAGGRESSION
TRAJECTORIES

c
0 1.5
"3
CO
0>
o>
CD
2
c
c 0.5
8
2

Age inyears

-Low 69.5%;

-High rising:30.5%

SOURCE: C?t?, Vaillancourt, Barker, et al. (2004).

preadolescence. One-sixth of children (16.6 percent) followed a high stable trajec


we see no evidence of
tory of physical aggression. Here again late-onset-type physi
cal aggression trajectories even a
in sample that includes measurement on children
who were tracked from as early as two years of age.
Even if late onset of childhood physical aggression is
highly unusual, this does
not rule out the that a sizable fraction of adolescents who were not
possibility physi
as young children become violent during adolescence. Adoles
cally aggressive
cence is a critical in Parental control and influence
period personal development.
generally diminishes and children take important steps toward independence.
to
Unstructured time with peers increases, opportunities organize life's activities
without adult supervision grow, and experience with the successes and failures that
determine life chances become more prominent (Osgood
et al. 1996;
Thornberry
et al. 2003; Warr 2002). In various ways, each of these changes in life circumstances
has been implicated in the onset of delinquency.
Brame, Nagin, andTremblay (2001) andNagin andTremblay (2002)built from
the findings of Nagin andTremblay (1999) and Broidy et al. (2003)by exploring the
in childhood and trajectories of
linkage between trajectories of physical aggression
violent delinquency in adolescence. Like Nagin and Tremblay (1999), both of
these studies are also based on the Montreal longitudinal study. The key question
that was addressed in each was whether there is any substantial evidence of late
92 THE ANNALS OF THE AMERICAN ACADEMY

FIGURE 4
TRAJECTORIES OF PHYSICAL AGGRESSION FROM AGES THREE TO ELEVEN

Age (in years)

-^^? Low Desisters 31.1% -Low Desisters predicted


?^k? Moderate Desisters 52.2% -Moderate Desisters predicted
High Stable 16.6% -High Stable predicted

SOURCE: C?t?, Vaillancourt, Barker, et al. (2004).

no
onset of physical aggression in adolescents among individuals with history of
childhood physical aggression, and if there is, whether that violence is
physical
chronic and sustained rather than episodic.
Both analyses are based on dual trajectories models, which are designed to
two distinct but related measurement
explore the linkage between conceptually
series (Nagin and Tremblay 2001a; Nagin 2005). In both analyses, the two measure
ment series were six to thirteen based on the teacher ratings
physical aggression from
described above and self-reported violent delinquency from thirteen to seventeen.
The self-reported violent delinquency scale sums the boy s report of the frequency
with which he engaged in a
fistfighting, gang fighting, carrying/using deadly
or someone, and an at someone.
weapon, threatening attacking throwing object
Brame, Nagin, and Tremblay (2001) used what Nagin (2005) called a con
strained dual trajectory model. In the constrained model, each trajectory for child
hood physical aggression is uniquely combined with a trajectory of adolescent vio
shows two of the seven combined were
lent delinquency. Figure 5 trajectories that
in that Both a level of
reported analysis. groups display consistently high physical
aggression from age six to thirteen. Despite the similarity of their childhood

aggression trajectories, the two groups display markedly different trajectories of


violent delinquency in adolescence. At age thirteen, the self-reported violent
of one group is than that of any other group. Their violence
delinquency higher
rises steadily to a peak at age fifteen and thereafter declines. This group was called
the high childhood aggression/high adolescent violence group and was estimated
to account for about 3 percent of the population. In contrast, the second group is
made up of individuals who report a negligible level of violence in their adoles
GROUP-BASED TRAJECTORYMODELING 93

FIGURE 5
TWO TRAJECTORIES OF CHILDHOOD PHYSICAL
AGGRESSION AND ADOLESCENT VIOLENT DELINQUENCY
BASED ON THE CONSTRAINED DUAL TRAJECTORY MODEL

ChildhoodPhysicalAggression AdolescentViolentDelinquency

-High Child-Low
Adol(3%)
-High-Child-High
Adol(10%)
13 14 15 16 17
10 11 12 13

Age

SOURCE: Brame, Nagin, and Tremblay (2001).

cence, despite their high aggression in childhood. This group, which was called the

high childhood aggression/low adolescent violence group, was estimated to


comprise 10 percent of the population.
The remaining five combined dual trajectories groups were distinguished by
their levels of childhood physical aggression. A cluster of three groups were char
acterized by low and steadily declining physical aggression in childhood. The
larg
est group, which was estimated to account for 33 percent of the self
population,
no violent in adolescence. A second group had a tra
reported virtually delinquency
jectory of adolescent violent delinquency that started off low at age thirteen but
thereafter increased to a moderate level by ages sixteen to seventeen. We called
this category the low childhood aggression/increasing adolescent aggression class.
contrast, the third latent class started off with amoderate level of self-reported
By
at age thirteen, but in the ensuing years their violence decreased
aggression gradu
to a near-zero level. The latter two groups were of nearly equal size, about 11
ally
percent to 13 percent of the population, respectively.
The final two groups were characterized by medium but declining levels of
childhood physical aggression. The larger of the two groups, estimated to comprise
21 percent of the no violent
population, reported virtually delinquency during ado
lescence. The other class, estimated to account for about 10 percent of the
popula
tion, reported a relatively high but declining level of physical aggression during
adolescence.

Thus, Brame, Nagin, and Tremblay (2001) found only one group that might be
characterized as a late onset violence group. The group called the low childhood
adolescent was estimated to make up
aggression/increasing aggression group only
10 percent of the
sampled population. The remaining 90 percent followed dual tra
of either or from ages six to
jectories steady declining physical aggression
seventeen.
94 THE ANNALS OF THE AMERICAN ACADEMY

The Nagin andTremblay (2002) analysiswas based onwhat Nagin (2005) called
the general dual trajectory model. Three key outputs of the are
general dual model
(1) the shape of the trajectory of each group for both measurement series, (2) the
in each such trajectory group, and (3) the
probability of membership joint proba
of in trajectory groups across behaviors. This final
bility membership output is key.
It provides the capacity for
specifying the linkage between teacher-rated childhood
physical aggression from ages six to thirteen with self-reported violent delinquency
in adolescence from ages thirteen to seventeen. Thus, the general model relaxes
the constraint of the model used in Brame, Nagin, and
Tremblay (2001 )of a unique
identification between the trajectories of childhood physical aggression and of
adolescent violence. The Nagin and Tremblay (2002) analysis examined whether
this constraint might have been more substantial evidence of late onset
concealing
violence.

The trajectories of physical aggression from ages six to thirteen and violent
from thirteen to seventeen are in
delinquency ages reported Figures 6 and 7,
The and sizes of in are vir
respectively. shapes the childhood trajectories Figure 6
identical to those in Nagin and (1999) for the age period
tually reported Tremblay
six to fifteen. The one exception concerns the size of the chronic group, which is
three times than in the earlier analysis, 11 percent versus 4
nearly larger percent.
We attribute the increased size of the chronic group to the two-year truncation in
the observation period in this analysis.
The five trajectories in Figure 7 for the adolescent not been
period, which had
previously reported, deserve greater elaboration. The two
largest trajectory
groups, labeled the low 1 and 2 groups, combined to account for 65.8 percent of the

population. These individuals average one or less on the seven-item inventory of


self-reported violent delinquency described earlier. This amounts to
responding
with "1 or 2 times" in the past year to one item and "never" to the remaining six
items. This a low level of an occa
finding suggests that physical aggression (e.g.,
sional fistfight) is normal among adolescent males.
on this
Another perspective point is provided by consideration of the cumula
tive distribution of the sum of the individual-level self-reports of violent delin
quency from ages thirteen to seventeen. To score 0 on this sum, the individual must
report "never" to all seven items for all five years. Only about 10 of the sam
percent
so The median score of about 5 on average, to a
ple reported. corresponds,
response of "1 or 2 times per year" on one item in the scale in each
only yearly
assessment. occasional in adolescence is
Clearly, physical aggression
commonplace.
The one-third of the boys who engaged inmore than occasional physical aggres
sion are a
split among three groups: high chronic group (6.6 percent) following the
classic hump-shaped a
trajectory of delinquency, declining group (13.3 percent)
that starts at a rate that is nearly as high as the chronics at age thirteen but declines
thereafter, and a rising group (15.3 percent) who are low at age thirteen but rise
thereafter to a level that is about half that of the chronics. Note that because the
declining group is twice the size of the chronic group, among the who start
boys
at age thirteen, two in three are "desisters." This is consistent with our conten
high
GROUP-BASED TRAJECTORYMODELING 95

FIGURE 6
PHYSICAL AGGRESSION TRAJECTORIES FROM AGES SIX TO
THIRTEEN BASED ON THE GENERALIZED DUAL TRAJECTORY MODEL

10 11 12 13
Age
-
Never (14.5%) Moderate Declining (44.4%)
High Declining (30.1%) -Chronic (11.0%)

SOURCE: Nagin and Tremblay (2002).

tion that overall physical aggression tends to decline with age. Still, there is a very
of While are a small
interesting group "rising" boys. they minority of the population
(16.5 percent), their size is nontrivial. Indeed, they comprise nearly 50 percent of
the 35.3 percent in the three trajectory groups that engage in more than very
occasional physical aggression.
How do the adolescent trajectory groups link up with the
physical aggression
childhood physical aggression trajectories? Table I reports the conditional proba
bility of "transitioning" from each of the age six to thirteen physical aggression tra
jectory groups to the various age thirteen to seventeen violent delinquency trajec
tory groups. For example, the estimated probability of an individual in the
moderate childhood group transitioning to either the low 1 or 2 adoles
declining
cent trajectory group is .73. to the
By contrast, the probability of his transitioning
chronic group is less than .02.
The transition probabilities conform to long-standing results on the continuity
of problem behaviors. Across the childhood physical aggression groups, the low
childhood group is most likely to transit to the low 1 or 2 adolescent trajectories
(.90), whereas the chronic group is least likely (.30) to transit to this group. In
between are the moderate and high declining childhood physical aggression
groups with probabilities of .73 and .53, respectively. Conversely, the low child
hood group is least likely to join any of the three high adolescent trajectory groups
96 THE ANNALS OF THE AMERICAN ACADEMY

FIGURE 7
TRAJECTORIES OF VIOLENT DELINQUENCY FROM AGES THIRTEEN
TO SEVENTEEN BASED ON THE GENERALIZED DUAL TRAJECTORY MODEL

14 13 15 17 16
Age

- -
-*- Low 1 (29.5%) Low 2 (36.3%)
-T*- Rising (16.5%) -*- Declining (10.8%)
-*- Chronic (6.8%)

SOURCE: Nagin and Tremblay (2002).

TABLE 1
PROBABILITY OF ADOLESCENT VIOLENT
DELINQUENCY TRAJECTORY GROUP MEMBERSHIP CONDITIONAL
ON CHILDHOOD PHYSICAL AGGRESSION TRAJECTORY GROUP

Adolescent Group
Childhood Group Low 1& 2 Rising Declining Chronic

.90
Low .09 .00 .01

Moderate .73 .02


.12 .13
declining
.53 .11 .17
.19
High declining
Chronic .30 .27 .21 .22

= .09 + .01 +
(.10 .00) and the childhood chronics are most likely to transit to these
= .27 + .21 +
trajectories (.70 .22).
The transition probabilities are also consistent with prior research that finds
with age. For all groups, including the
generally declining physical aggression
childhood chronics, the modal transition is to the low 1 or 2 adolescent trajectory
GROUP-BASED TRAJECTORYMODELING 97

groups. Furthermore, with the exception of the childhood chronic group, this
exceeds .5. the of in the two lowest child
probability Conversely, probability boys
hood groups transiting to the chronic adolescent group is negligible. Still, a sizable
minority of individuals in the two lowest childhood trajectory groups transit into
one of the three
higher adolescent trajectory groups?10 percent and 23 percent,
respectively, for the low and moderate declining childhood groups.
The results of this application of group-based a
trajectory modeling suggest
more nuanced version of the "no late onset of
physical aggression" hypothesis?
physical aggression ismostly declining from childhood through adolescence. How
ever, the finding that a small minority of boys appear to have a late onset physical
aggression should not obscure the central finding that has emerged from these
applications of group-based trajectory modeling. The dominant trend among all
is either or stable Thus, to under
trajectory groups declining physical aggression.
stand the developmental of it is that studies
origins physical aggression, imperative
at birth or
begin (ideally) prenatally.
Longitudinal studies of physical aggression with large samples that were initi
ated before the 1990s focused on schoolchildren and adolescents. A few studies
initiated in the past decade attempted to chart the
developmental origins of physi
cal aggression in the years. Results indicate that
preschool physical aggression
appears during the first year after birth (Tremblay 2004a, 2004b). All applications
of group-based trajectory modeling show that the frequency of aggression
increases rapidly during the second year after birth, reaches a peak between

twenty-four and forty-two months after birth, and then decreases steadily (C?t?,
Vaillancourt, LeBlanc, et al. 2004; National Institute of Child Health and Develop
ment [NICHD] 2004; Tremblay et al. 2004). These data map on very well to the
data from studies on school-age children described above. Thus, the story that is
appearing from the patchwork of longitudinal studies from birth to adulthood is
that the peak in frequency of physical aggression for humans is during early child
hood, not during kindergarten, or
adolescence, early adulthood. This suggests that
rather than learning to physically aggress, children are learning not to physically
aggress. From this perspective, itwould be exists a sig
extremely surprising if there
nificant group of individuals who never used physical aggression during the pre
school years and initiated this behavior during adolescence. Based on the available
data, we hypothesize that children with relatively low levels of physical aggression
during the elementary school years who show an increase in physical aggression
were on a tra
during adolescence high trajectory during the preschool years. Only
with data from to late adolescence will enable
jectory analyses longitudinal infancy
us to test the "late onset"
hypothesis.

taxonomies
Clarifying developmental
There is a long tradition in developmental psychology of group-based theorizing
about both normal and pathological development. Examples include theories of
use (Kandel 1975),
personality development (Caspi 1998), drug learning (Holyoak
and Spellman 1993), language and conceptual development (Markman 1989),
98 THE ANNALS OF THE AMERICAN ACADEMY

depression (K?sen et al. 2001), eating disorders (Tyrka, Graber, and Brooks-Gunn
2000), alcoholism (Cloninger 1987), anxiety (Cloninger 1986), and the develop
ment of behaviors such as conscience (Kochanska 1997) and of antisocial
prosocial
behaviors such as delinquency (Loeber 1991; Moffitt 1993; Patterson, DeBaryshe,
and Ramsey 1989).

Human history is replete with

tragic instances inwhich afictional


group-based separation is thefirst step
in the dehumanization of the "them."

When testing these taxonomic theories, developmental researchers have com


to rules based on crite
monly resorted using assignment subjective categorization
ria to construct categories of developmental For in their
trajectories. example,
research on the developmental origins of violence, Haapasalo and Tremblay (1994)
a taxonomy
proposed composed of five groups?stable high fighters, desisting high
onset These
fighters, late high fighters, variable high fighters, and nonfighters.
were created from the annual teacher of a child's
groups ratings physical aggression
in the described Montreal-based longitudinal study. Haapasalo and
previously
labeled who scored on this scale in any as
Tremblay boys high given year "high
into their
fighters" for that year. They then defined rules for assigning individuals
were based on the
five-group taxonomy. These rules frequency and trend of each
classification as a For were
boys high fighter. example, "desisting high fighters"
who were in but who were classified as
boys high fighters kindergarten high
in no more than one of the ensuing assessment periods.
fighters
Moffitt s (1993) well-known more
taxonomy is parsimonious. She posited only
two distinct developmental trajectories of behavior. One group follows
problem
what she calls a life course persistent (LCP) trajectory of antisocial behavior, and
the other group is posited to follow an adolescent limited AL)( trajectory. In empir
ical tests of her theory, such as Moffitt et al. (1996), she has used classification rules
similar to those used and (1994). LCPs are
conceptually by Haapasalo Tremblay
defined as individuals who score one or more standard deviations above the mean
in three of four assessments of a conduct disorder index between the ages five and
eleven and who also score at least one standard deviation above the mean in self
at least once at either age fifteen or eighteen. The ALs are
reported delinquency
defined as individuals who do not meet the LCP criteria for childhood conduct
problems but who do achieve the LCP threshold for adolescent delinquency.1
GROUP-BASED TRAJECTORYMODELING 99

While such assignment rules are generally reasonable, limitations and pitfalls
are attendant on their use. One is that the existence of distinct tra
developmental
must be assumed a Thus, the cannot test for their pres
jectories priori. analysis
ence, a fundamental shortcoming. A second and related pitfall is the risk of simul
taneously "overfitting and underfitting" the data by creating trajectory groups that
reflect only random variation and failing to identify unusual but still real develop
mental patterns. Third, ex ante specified rules provide no basis for calibrating the
precision of individual classifications to the various groups that comprise the taxon
omy. Thus, the about an individuals group membership cannot be
uncertainty
in the form of
quantified probabilities.
To illustrate these limitations, consider the Haapasalo and Tremblay (1994)
study. While Haapasalo and Tremblay proposed five taxonomic groups, the appli
cation of group-based trajectory analysis described in
Nagin and Tremblay (1999)
found that the four-group model shown in Figure 2 best fitted the data. The Nagin
and Tremblay analysis provided formal statistical support for the presence of three
of the groups hypothesized in the and Tremblay taxonomy: the stable
Haapasalo
to the chronic trajectory group), the desisting high
high fighters (who correspond
(who correspond to the trajectory and the
fighters high declining group),
nonfighters (who correspond to the low trajectory group). However, there was no
a
evidence of trajectory corresponding to and Tremblay s late onset high
Haapasalo
or of a variable These are
fighter group high fighter group. examples of classifica
tions that likely resulted from overfitting the data, where random variation is
confounded with real structural differences.
Two other
examples of the utility of the formal group-based trajectory method
to ad hoc classification are studies and
compared procedures by Nagin, Farrington,
Moffitt (1995) and Lacourse et al. (2003). The former study was intended to test
several predictions of Moffitt s (1993) two-group taxonomic theory, including test
ing for the very presence of the trajectories predicted by her taxonomy. Based on
an of a classic data set assembled and West (1990), which
analysis by Farrington
included data on convictions from age ten to thirty-two in a sample of more than
four hundred males from a poor neighborhood in London, three offending trajec
tories were found. These trajectories are shown in Figure 8. One trajectory, which
in late adolescence, the adolescent limited group
peaks sharply closely matches
predicted by Moffitts theory. The high-hump-shaped trajectory, labeled high
chronic, is similar in some to Moffitt s second group, the LCP group. This
respects
group is in at age ten. However, their fre
already actively engaged delinquency
of antisocial at least as measured is
quency behavior, by conviction, very age
s rises until about
dependent?a pattern that is not anticipated by Moffitt theory. It
age eighteen and then begins a steady decline. By age thirty, it has dropped below
its starting point at age ten and is about equal to the rate of a third group called low
chronic offenders. This third group was not included inMoffitt s taxonomy. Thus,
the application of the group-based method provides basic confirmation of the
presence of the AL and LCP trajectories that Moffitt predicted in her taxonomy
but also suggests that the LCP trajectory may be more age-dependent than antici
100 THE ANNALS OF THE AMERICAN ACADEMY

FIGURE 8
TRAJECTORIES OF CONVICTIONS

o) 2.5
c
.2 2
S
> 1.5
c
? 1

Age
--
Never (69.5%) Adolescent Limited(12.4%)

-?r- Low Chronic (12.2%) -^?HighChronic(5.9%)


SOURCE: Nagin (2005).

it identified the low chronic trajectory, which was


pated by the theory. Moreover,
not the taxonomy.
predicted by
In Lacourse et al. (2003), group-based was used to
trajectory modeling analyze
the developmental course of gang membership. This analysis was based on the
same Montreal-based in
sample that forms the basis for the trajectories depicted
Figures 2 and 4 through 6. The trajectories of gang membership identified in this
were based an annual from age eleven to seventeen of delin
analysis self-reports
quent group involvement in the past year. The resulting trajectories of probability
of gang membership at each age are shown in Figure 9. One trajectory, called the
never group, is estimated to make up 74.4 percent of the population. This groups
of was very small over all ages. The second group,
probability gang membership
onset group, at age eleven with a
called the childhood began high probability of
gang membership that modestly rises till age fourteen and declines thereafter. The
third group, called the adolescent onset group, had a near-zero probability of gang
at age eleven, but thereafter the rose to a rate that
membership probability actually
exceeded that of the childhood onset group. The latter two groups are each
estimated to constitute 12.8 percent of the sampled population.
The identified in the Lacourse et al. (2003) analysis illustrate two
trajectories
valuable properties of the group-based modeling to the use of
approach compared
classification rules. One is the capacity to identify qualitatively distinct develop
mental progressions that are not readily identifiable using ad hoc, ex ante classifica
GROUP-BASED TRAJECTORYMODELING 101

FIGURE 9
TRAJECTORIES OF GANG MEMBERSHIP

1
CO
0.9
I
I 0.8
| 0.7
a 0.6

I 0.5
O
? 0.4
O
= 0.3

0.2
I
?0.1
I I I I I I I

11 12 13 14 15 16 17
Age

~+~
Childhood 12.8% Adolescence 12.8% Never 74.4%
SOURCE: Lacourse et al. (2003).

tion rules. In the childhood onset and adolescent onset groups shown in
principle,
are identifiable ex ante, but given the specific developmental course of
Figure 9
each, itwould be very difficult to identify them without a formal statistical method
A second, closely related advantage also stems from the use of a formal statis
ology.
tical structure. It is because of this structure that the methodology has the capacity
for chance variation across individuals from real differences and for
distinguishing
whether individual is real or random variation in behavior.
calibrating change only
onset trajectories are the a for
Because the childhood and adolescent product of
mal statistical model, there is a firmer basis for their reality than if they had been
constructed based on subjective classification rules.
Another important example of the clarifying role of the group-based trajectory
model involves the taxonomic distinction in the criminal career model between
an active offender and a desistor. Bushway, Thornberry, and Krohn (2003)
being
a "static" versus a
explored the distinction between what they call "dynamic" defini
tion of desistence in a of about nine hundred individuals from Rochester,
sample
New York, for whom have data on from
they self-reported general delinquency
ages thirteen to The static definition of desistance was in a
twenty-three. specified
form that conformed with much research in the tradition of the criminal career
102 THE ANNALS OF THE AMERICAN ACADEMY

as articulated
in the seminal work of Blumstein et al. (1986).
paradigm Specifically,
the static an individual was defined as a desistor
definition, if he or she self
by
at least one act prior to age none
reported committing delinquent eighteen but
thereafter. The dynamic definition was based on a comparison of trajectories of

offending.
of the group-based trajectory modeling to the Rochester data
Application yields
the seven-group model shown in Figure 10. Two of these seven trajectories, called
the very-low- and low-level offenders, involve individuals with very limited num
bers of self-reported delinquent offenses during adolescence and young adult
hood. They are basically nonoffenders. these individuals were esti
Collectively,
mated to comprise 55 percent of the sample. They also accounted for 79 percent of
so-called desistors by the static definition. Thus, if one were to compare desistors
the static definition with no offenders, the resulting analyses would amount to
by
little more than a of very infrequent offenders in adolescence with
comparison
their counterparts who remained completely inactive in adolescence. By contrast,
if the trajectory groups are used to define desistors, desistance is defined a
by
a a
group that goes through a period of substantial offending followed by decline to
diminished rate. By this definition, the bell-shaped desistors and possibly the inter
mittent offenders should be defined as desistors. Collectively, these two groups are
estimated to comprise about 17 percent of the sampled population, amuch smaller

group than that defined the static definition. Furthermore, there is little
by overlap
in the of desistors these two definitions of desistance. In our
population by judg
ment, the trajectory group-based definition of a desistor identifies a substantively
far more interesting and distinctive group of individuals than the static definition.

the predictors and consequences of developmental trajectories


Clarifying
as latent strata in
Trajectory groups can be thought of longitudinal data
(Haviland and Nagin forthcoming) that clusters of individuals follow
distinguish
ing distinctive developmental paths. Theories of development provide much guid
ance on the and of favorable and unfavorable devel
predictors likely consequences
However, very little on the
opmental trajectories. theory provides guidance
functional form that relates risk and protective factors to alternative developmen
tal trajectories and in turn relates the effects of the trajectories themselves to still
other developmental consequences.
11 illustrates the importance of this functional form issue in the context
Figure
a a
of identifying the dose-response relationship between single predictor, x, and
outcome variable, y. Suppose y is a "bad" outcome we
single (e.g., crime) and that
have a strong theoretical basis for stipulating that x aggravates the likelihood of a
higher y. Mathematically, this prediction can be faithfully represented by any func
tional relationship inwhich y never declines with an increase in x. One possibility is
that the relationship between y and x iswell approximated by the linear equation A.
there are other possibilities that are not well approximated a lin
However, by such
ear
relationship. For example, the dose response relationship depicted by B is
increases after the dose of x exceeds x*. Curve C
highly nonlinear?y only steeply
GROUP-BASED TRAJECTORYMODELING 103

FIGURE 10
TRAJECTORY OF SELF-REPORTED DELINQUENCY

13 13.5 14 14.5 15 15.5 16 16.5 17 17.5 18 18.5 19 19.5 20 20.5 21 21.5 22 22.5

Age

SOURCE: and Krohn (2003).


Bushway, Thornberry,

an even more extreme version of this threshold effect inwhich the positive
depicts
association between x and y follows a step function inwhich y jumps to a new level
once x exceeds x*. Each of these alternative response functions is consistent
y*
with the theoretical prediction that more x tends to aggravate y.
Figure 11 depicts the statistical complexities that may attend capturing the rela
a a sta
tionship between single predictor variable and single outcome variable. The
tistical complexities that attend capturing the relationship between multiple pre
dictors and developmental are even more One
trajectories daunting.
to
straightforward approach begin exploring these complexities without assuming
functional forms is by stratifying the outcome variable and then examining the rela
tionship between the strata and other variables. One way of stratifying longitudinal
data on development is by grouping individuals into the trajectory groups identi
104 THE ANNALS OF THE AMERICAN ACADEMY

FIGURE 11
THREE HYPOTHETICAL DOSE RESPONSE RELATIONSHIPS

0 -1-L

0 1 I

fied by group-based trajectory modeling. It is in this sense that the trajectory


groups can be of as latent strata of individuals distinctive
thought following
developmental paths.
The remainder of the section summarizes results from three studies that we
believe illustrate the utility of this form of data stratification.

Life-course trajectories of different types of offenders


(Nagin, Farrington, and Moffitt 1995)
This article by Nagin, Farrington, and Moffitt (1995) explored the predictors
and consequences of the trajectories depicted in Figure 8. the most
Perhaps
was
important finding of this study that the seeming reformation of the individuals
the AL trajectory may have been more apparent than real. At ages four
following
teen and
eighteen, the offending and other misconduct by the ALs and the high
level chronics (HLCs) was generally statistically indistinguishable. Both groups
behaviors such as
self-reported similar levels of burglary, stealing, fighting, and
use that in turn were than the self-reported rates of indi
drug significantly higher
viduals following the nonconviction (NC) trajectory.2 Thus, during adolescence the
behavioral based on conviction with behavioral
trajectories closely corresponded
on themselves.
patterns based self-report of the study participants
GROUP-BASED TRAJECTORYMODELING 105

By age thirty-two, however, this close correspondence between convictions


record and self-reported behavior had evaporated. At this age, the HLCs and the
low level chronics (LLCs) were still active offenders as measured
by their
nonnegligible conviction rate, whereas the ALs had been inactive
seemingly
offenders for a decade or more. However, the self-reports of offending a
suggested
far more complex reality. The self-reports on were consistent with the
burglary
conviction data. The HLCs and LLCs reported comparable levels of burglary,
whereas the rates for the ALs and NCs were negligible. The ALs were also
signifi
more to be married and have a
cantly likely job than the LLCs and HCLs. How
ever, the ALs' self-reported levels of theft, fighting outside the home, and illicit
and alcohol use were from that of the HCLs and LLCs.
drug indistinguishable
Thus, this investigation into the and consequences of offending tra
predictors
as measured conviction raised a number of issues of mea
jectories by important
surement and
theory. As for measurement, the lack of correspondence of results
between self- and official reports at age thirty-two is reminiscent of amuch earlier
controversy about the measurement of offender participation: incompatibility
between results of official versus self-report studies of juvenile delinquency was a
pivotal methodological issue for in the 1970s. Hirschi,
criminologists Hindelang,
and Weiss that the was
(1979) argued discrepancy illusory because juvenile self
reports probably tap a more trivial domain of offending than do official data. It is
not clear that the identified in the Nagin, Farrington, and Moffitt
discrepancy
(1995) study can be with as The of the subjects in the
dispensed easily. self-reports
London sample at age thirty-two tapped nontrivial offenses: drunk driving,
fighting
outside the home, drug use, burglary, and theft. To be sure, some convictions for
certain of these offenses, such as drunk driving and brawling, are not included in
the conviction counts. Thus, the between versus official
discrepancy self-report
record findings may in part be a measurement artifact. This observation cannot

fully resolve the however, because the convictions for theft and bur
discrepancy,
were included in the conviction data. Thus, this raised a new method
glary study
ological issue?a discrepancy between official records and the self-reports of
adults who reported nontrivial crimes. Resolution of this issue remains important
to this careers on
day because almost all published findings about criminal rely
official data.
As for theory, this investigation raised important questions about the definition
and explanation of desistance. The ALs had clearly desisted in some but not all
domains of antisocial behavior. Their domain-specific reformation from crime
raises important questions of what is meant desistance. Nagin,
conceptual by
Farrington, and Moffitt (1995) went on to point out the domains of reformation
were not random. The men to be
seemingly appeared avoiding forms of antisocial
behavior that would most jeopardize their stakes in conformity. One theoretical
implication of this speculation is that theories of desistance need to focus on multi
ple domains of behavior and the linkages between them.
106 THE ANNALS OF THE AMERICAN ACADEMY

Parental and early childhood predictors of persistent physical


aggression in boys from kindergarten to high school
(Nagin and Tremblay 2001b)
This article by Nagin and (2001b) reported another example inwhich
Tremblay
stratification into trajectory groups valuable
provided insight into subtle but
differences that behavioral
important may distinguish trajectories. The analysis
involved an investigation into the factors that discriminated
among the four trajec
tories of in 2. A host of risk factors were identi
physical aggression reported Figure
fied that the two low
distinguished physical aggression trajectories from the two
high physical aggression trajectories. Concerning these risk factors, two findings
stand out. First, the most of in a
powerful predictors membership high aggression
trajectory group were levels of and assessed in kin
high hyperactivity opposition
dergarten. Individually, these risk factors increased the odds of in a
membership
high physical aggression more than a factor of three. In combination,
trajectory by
the increase ismore than ninefold. are
Individually, these risks comparable
inmag
nitude to the impact of serum cholesterol levels on risk of
high coronary heart
disease (Truett, Jerome, and William 1967). In combination, they far exceed it.

The group-based trajectory model iswell


suited for identifying and testing whether
the response to a or
turning point event
treatment is upon the individual's
contingent
developmental trajectory.

A second
prominent finding concerned the predictive power of
parental charac
characteristics of the mother have teen
teristics?only predictive power. Whereas
onset of and low educational attainment for mothers were
parenthood significant
predictors of their male child's high aggression, these characteristics of the father
had no predictive power.
The prominent predictive power of mother characteristics
again revealed itself
in the of factors that the two
analysis distinguished high physical aggression
groups?the high declining trajectory and the chronic trajectory. Here we were
concerned with
identifying characteristics that distinguished the modestly large
fraction of boys who start off levels of
displaying high physical aggression but sub
sequently desist (28 percent) from the small but prominent group of chronics (4
GROUP-BASED TRAJECTORYMODELING 107

percent) who continue their physical aggression unabated. Only two such charac
teristics were identified?the mothers low educational attainment and teenage
onset of The odds of male offspring of poorly educated
childbearing. teenage
mothers not
desisting from a high level of physical aggression at age six are 9.3
times greater than those of their counterparts without such mothers.
While this was the first analysis to document that these two mother characteris
tics distinguish persistence from desistance of chronic physical aggression, a large
of evidence links teen onset of to a
body childbearing litany of unfavorable behav
iors and outcomes for the offspring. These include conduct disorder and other
problem behaviors in childhood et al. 2000), and school
(Wakschlag delinquency
out in adolescence Brooks-Gunn, and 1987; Morash
drop (Furstenberg, Morgan
and Rucker 1989), and criminality as an adult (Grogger 1997; Pogarsky, Lizotte,
and Thornberry 2003).
This study does not explain why teen onset of motherhood and low maternal
education are risk factors for chronic physical aggression in her
offspring. These
two maternal factors are
likely markers of maternal problem behaviors and circum
stances that give rise to bad outcomes for her offspring rather than the causes per
se. For more women are more to become teen
example, aggressive young likely
mothers and to drop out of school and are also more likely to be unresponsive par
ents. There is also evidence that women who begin childbearing early
are more
to use harsh and erratic These mothers may also be more prone to
likely discipline.
birth complications, which have been shown to be related to subsequent conduct
in the child (Arsenault et al. 2002; Raine et al. 1996). These
problems findings sug
gest that the mothers themselves may be the agents of the intergenerational trans
fer of chronic physical aggression. The results are not nearly definitive about the
in terms of the separate and interactive roles of
specifics of the transfer mechanism
biology, parenting practice (including the father), and the larger social environ
ment.
Notwithstanding, they do suggest that the intergenerational transfer mecha
nism may have for the child and at
profound consequences society large.

Good marriages and trajectories of change in criminal


offending
(Laub,Nagin, and Sampson 1998)
The final example we summarize?Laub, Nagin, and Sampson (1998)?is as
notable for what it did not find as what it did find. This article further explored the
linkage between marriage and desistance that has been a central focus of the long
time collaboration between John Laub and Robert Sampson (cf. Sampson and
Laub 1993; Laub and Sampson 2003). Figure 12 reports the trajectories that
from the of trajectory modeling to the arrest
emerged application group-based
records from age seven to thirty-two of 480 men in Glueck and Gluecks (1950,
1968) delinquent sample. Observe that all four trajectories follow the classic
hump-shaped age crime curve, but only groups 3 and 4 peak prior to age twenty, as
is normally the case in the population average curve, whereas the peaks for groups
1 and 2 occur well past age twenty. Even more important, group 2 and
especially
108 THE ANNALS OF THE AMERICAN ACADEMY

FIGURE 12
PREDICTED OFFENDING TRAJECTORIES

3t
Predicted Group 1
Offending
Rate 2.5 +

-Group 1

-Group 2

-Group 3

-Group 4

1.5 +

0.5 +

CVI CO

Age
SOURCE: Laub, Nagin, and Sampson (1998).

group 1 have substantial arrest rates at the close of the measurement


predicted
period, age thirty-two.
Table 2 reproduces a (1998). It
key table from Laub, Nagin, and Sampson
means on a of individual characteristics, behaviors, and life
reports group variety
course outcomes. There are two in this table. One is the clear
prominent patterns
association between a
following trajectory of desistance, namely trajectories 3 and
4, versus trajectories of persistence, namely, trajectories
1 and 2, and adult social
bonds. Compared to the persisters, the desisters are more to
significantly likely
have quality marriages and to have stable jobs and are significantly less likely to be
divorced or there is virtually no association between
separated. By contrast,
d?sister trajectory group status and individual and family differences in childhood.
The absence of such associations makes it hard to argue that the correlation
between adult social bonds and desistance is simply a reflection of long-standing
individual differences that predict both desistance and quality adult social bonds.
GROUP-BASED TRAJECTORYMODELING 109

TABLE 2
COMPARISON OF GROUP MEANS FOR SELECTED
CHARACTERISTICS, BY GROUP MEMBERSHIP

Characteristic Groups 1 and 2 Group 3 Group 4


Individual differences in childhood
Verbal IQ 5.63 5.49 5.44
extroverted 67.0 53.1 55.8
Percentage
adventurous* 66.0 52.7 50.6
Percentage
12.3 15.9 9.7
Percentage egocentric
14.2 15.9 15.6
Percentage aggressive
tantrums 44.3 41.4 34.4
Percentage
Percentage difficult 55.8 60.9 57.9

Family differences in childhood


0.10 0.08 0.00
Poverty
size* 6.00 5.34 5.23
Family
Parental crime/alcohol abuse 2.06 1.94 2.04
Erratic discipline 0.13 -0.15 -0.08
1.39 1.47 1.42
Supervision
Attachment to family 3.07 3.13 3.10

Adolescent behavior
Arrest 0.45 0.45 0.37
frequency*
Unofficial delinquency* 15.6 14.0 13.6
attached to school* 21.6 35.9 39.7
Percentage
onset 15.4 14.1 9.6
Percentage early
Adult social bonds
38.5 18.9 10.3
Percentage divorce/separation by age
thirty-two*
40.8 34.4 18.4
Percentage "shotgun" marriage*
married 30.8 50.7 80.3
Percentage by age thirty-two*
of marriage at age -2.01 -0.47 0.69
Quality thirty-two*
Job stability at age thirty-two* -3.16 -1.37 0.50

Number of cases 106 220 154

SOURCE: Laub, Nagin, and Sampson (1998).


*
Differences are at p < .05 level.
jointly significant

This argument is far from sufficient for making the case that the bonds are the
cause of the desistance.
Notwithstanding, being able
to
plausibly rule out one
class of is an forward, a that was
important explanations important step step greatly
facilitated by application of group-based modeling.

Guidelines
for Choosing between Group-Based
Trajectory Modeling and Standard Growth Curve Modeling
is the more
For what types of problems group-based approach appropriate than
standard curve and for what types of is the standard
growth modeling problems
110 THE ANNALS OF THE AMERICAN ACADEMY

a better fit? This is a question without a clear answer. Still, some


approach guide
lines are possible. One guideline relates to the adjective growth that modifies
"curve modeling." The curve
prototypical application of standard growth modeling
involves a process inwhich population members follow a common developmental
or decline. Raudenbush
pattern of either increase (2001, 509) offered language
as a a process. Another
acquisition quintessential example of such good example is
time spent with peers from childhood through adolescence (Warr 2002). Standard
curve methods are well suited for such
growth analyzing developmental phenom
ena because it is reasonable to assume that most individuals experience a common
process of growth or decline, albeit at different rates. However, there are large
classes of developmental phenomena for which the conception of a common
not fit. described the
growth process does naturally Raudenbush population differ
ences for this class of as "multinomial," and for such recom
problems problems he
mended a as As noted earlier,
group-based approach particularly appropriate.
Raudenbush used depression as an
example.
The basis for Raudenbush s (2001) a distinction between the develop
making
mental processes and is fundamental
underlying language acquisition depression
and cannot be overstressed. The former are appropriately conven
analyzed by
tional analysis of variation; the latter are not. Because the vocabularies of all young
children from normal populations increase with age, it is sensible to ask questions
such as,What is the average growth curve of children s vocabulary over a specified
age range? How large is the variation across children in their individual-level lan
guage acquisition growth curves? How do such "between-person" variations relate
to factors such as the child's cognitive functioning and parental education? How
are in acquisition related to in interactions with
"within-person" changes changes
for to conflict?
primary caregivers due, example, parental
These questions are framed in the language of analysis of variance as reflected in
the use of terms such as within-person change and "between-person change." This
is only natural because standard growth curve analysis has its roots in analysis of
variance. Like analysis of variance, growth curve analysis is designed to sort out fac
tors accounting for variation about a population mean.
To meaningfully frame an analysis in the conceptual apparatus of analysis of
variance requires that it be sensible to characterize population differences in terms
of variation about the population mean. For processes such as
language acquisi
tion, the mean trend is, in fact, a sensible statistical anchor for describing individual
However, for many processes over time or age, it is not. For
variability. evolving
itmakes no sense to frame a statistical analysis of population differences
example,
in the
developmental progression of attention deficit disorder (ADD) in terms of
variation about the mean trajectory of ADD, because ADD is the exception, not
the norm, within the general population. Other examples of evolving behavioral
that are not properly described in terms of variation about a popula
phenomena
tion mean are most forms of psychopathology and abuse of both licit and illicit
More generally, a group-based to is
drugs. approach analyzing longitudinal data
GROUP-BASED TRAJECTORYMODELING 111

to inwhich there may be


usefully applied phenomena qualitatively different trajec
tories of
change
over age or time across
subpopulations that are not identifiable ex
ante based on measured characteristics such as gender or race.
The assumption that all individuals follow a process that increases or decreases
regularly within the population may also be violated because there may not be a
single explanation for the differences in the
developmental trajectories of a
As described, and (2001b) found that a
subpopulation. previously Nagin Tremblay
host of predictors involving the individuals psychological makeup and family cir
cumstances
distinguished individuals following low versus high trajectories of
in childhood. However, a of two distinct
physical aggression comparison
of high childhood a
subpopulations trajectories?those following trajectory of
chronic aggression versus those who started childhood with high aggression but
later declined?revealed that only two maternal characteristics distinguished
these groups. Using standard growth curve modeling methods, itwould have been
very difficult to identify this important difference in variables that
distinguished
among trajectories of childhood physical aggression. Identification of such differ
ences is far easier with a that clusters individuals with similar
methodology
developmental trajectories.

[Tjrajectory groups, like all statistical models,


are not literal depictions of reality. Rather, they
are meant as a convenient
only approximation.

A second
guideline
concerns the motivation for the analysis. One common aim
of analyses of longitudinal data is to uncover distinctive developmental trends in
the outcome variable of interest. For example, do sizable numbers of youths follow
a onset conduct disorder? The group-based is
trajectory of adolescent approach
suited for whether such distinctive are present in the data.
ideally testing patterns
common aim of studies is to test whether some
By contrast, another developmental
identifiable characteristic or set of characteristics are associated with individual
differences in trajectories of An example iswhether trajectories of
development.
conduct disorder differ across sexes. For this type of problem, either approach may
be most appropriate depending upon the specific objectives of the analysis. For

example, if one were


examining whether trajectories of number of sexual partners
differed across sexes, standard growth curve a natural
modeling provides starting
mean
point for framing the statistical analysis?a comparison of the trajectories for
112 THE ANNALS OF THE AMERICAN ACADEMY

were to there are


boys and girls. By contrast, if the objective distinguish whether
distinctive predictors of distinctive trajectories, the
developmental group-based
approach is particularly well suited for this type of problem. Thus, the group-based
approach lends itself to analyzing questions that are framed in terms of the shape of
the developmental course of the outcome of interest, whereas either method may
be most appropriate for analyzing questions framed in terms of predictors of the
outcome s course.3
developmental
A third concerns the in the response
guideline possibility of path dependencies
to events such as or to treatments such as for
turning point marriage hospitalization
a disorder. Path occur when the response to a
psychiatric dependencies turning
or treatment is
point event contingent upon the individual's developmental history.
For example, Nagin et al. (2003) found that the seeming
impact of grade retention
on the child's
physical aggression depended upon trajectory of physical aggression.
The subsequent physical aggression of children who had been
following trajecto
ries of little or of chronic
physical aggression physical aggression appeared to be
event of
unaffected by the being held back in school. By contrast, the physical
aggression of individuals who had been following trajectories of
declining physical
aggression seemed to be exacerbated. Such path dependencies are
commonplace
in the literature on human (Elder 1985). Indeed, the possibility of
development
is a rationale for studies. The tra
path dependencies key longitudinal group-based
jectory model iswell suited for the response to a
identifying and testing whether
or treatment is
turning point event contingent upon the individual's developmental
trajectory.
Laying out guidelines for the use of alternative statistical methods is a precari
ous exercise. Users desire line distinctions. Yet line distinc
naturally bright bright
tions are not The first that
generally possible. guideline implies developmental
can be those involving regular growth or
processes cleanly divided between
decline and those that do not. The reality is that for many developmental pro
cesses, it is not possible to confidently make this distinction. The second guideline
an can be classified as either
implies that the objective of analysis identifying dis
tinctive or of
developmental trajectories testing predictors developmental trajec
tories. The
reality
is that most
analyses have both objectives. Still a further compli
cation is that standard curve can be used to
growth modeling identify distinctive
for races or
developmental trajectories predefined groups (e.g., genders). The
third guideline might be interpreted as implying that it is not possible to identify
with conventional curve models. This is not the case.
path dependencies growth
Stated differently, both methods are to over time. The
designed analyze change
on identification of different trajectory shapes and on
group-based method focuses
examining how the prevalence of the shape and shape itself relate to predictors. By
contrast, standard growth curve modeling focuses on the population mean trajec
tory and how individual variation about that mean relates to predictors. Thus, the
alternative approaches are best as not competing.
thought of complementary,
GROUP-BASED TRAJECTORYMODELING 113

Concluding Remarks
Since its introduction less than twenty years ago, group-based trajectory model
has had a and we believe salutary impact on the fields of develop
ing significant
mental criminology and In this article, we have
developmental psychopathology.
to summarize some useful contributions that were facilitated by the use
attempted
of the methodology. These include illuminating important empirical regularities in
the developmental course and
origins of physical aggression, clarifying and testing
taxonomic theories of the development of antisocial behavior, and clarifying the
predictors and consequences of developmental trajectories.
The rapid growth in applications of group-based trajectory modeling stems at
least in part from its usefulness in summarizing
complexity. A hallmark of modern
longitudinal studies is the variety and richness of measurements that are made
about the study s subjects and their circumstances. Less often acknowledged is that
this abundance of information is a difficult
accompanied by companion?com
researchers are confronted with the dilemma of how best to
plexity. Commonly,
explore and communicate the rich set of measurements at their
disposal without
so down in that the lessons to be learned from the
becoming bogged complexity
data are lost on them and their audience. Group-based trajectory modeling does
not solve the of and it
problem balancing comprehensibility complexity. However,
does improve a researcher's ability
to
identify, summarize, and communicate
in
complex patterns longitudinal data.

Notes
1.We also note that another important finding of the applications of trajectory modeling that included
measurement into adolescence is that in all cases, a small group of chronics was identified. In a review by
Moffitt (forthcoming) of evidence concerning her life-course persistent/adolescent-limited developmental
taxonomy, she observed, "After ten years of research, what can be stated with some certainty is that the

hypothesized life-course persistent antisocial individual exists, at least during the first three decades of life.
Consensus about this group has emerged from all studies that have applied trajectory-detection to a
analyses
cohort measures of antisocial behavior"
representative sample having longitudinal repeated
2. However, the offending behavior of the adolescent limited AL)
( and high level chronic HLC)
( individu
als was not always significantly different than that of the low level chronics (LLCs) at these ages.
3.We thank Steven Durlauf and Wayne for pointing out this important distinction.
Osgood

References
and its implications. San Francisco: Freeman and Co.
Anderson, John R.1980. Cognitive psychology
Arseneault, Louise, Richard E. Tremblay, Bernard Boulerice, and Jean-Francois Saucier. 2002. Obstetrical

complications and violent delinquency: Testing two developmental pathways. Child Development 73 (2):
496-508.
Blumstein, Alfred, Jacqueline Cohen, Jeffrey A. Roth, and Christy A. Visher, eds. 1986. Criminal careers and
"career criminals." Washington, DC: National Academy Press.
Brame, Richard, Daniel S. Nagin, and Richard E. Tremblay. 2001. Developmental trajectories of physical
aggression from school entry to late adolescence./owrn?/ of Child Psychology and Psychiatry 42:503-12.
114 THE ANNALS OF THE AMERICAN ACADEMY

Brame, Richard, Daniel S. Nagin, and Larry Wasserman. some character


Forthcoming. Exploring analytical
istics of finite mixture models.Journal of Quantitative Criminology.
S. Nagin, Richard E. Tremblay, E. Bates, Robert Brame, Kenneth A. Dodge,
Broidy, Lisa M., Daniel John
David M. Fergusson, John L. Horwood, Rolf Loeber, Robert Laird, Donald Lynam, Terne E. Moffitt,
Gregory S. Pettit, and Frank Vitaro. 2003. Developmental trajectories of childhood disruptive behaviors
and adolescent A six site, cross national study Developmental 39:222-45.
delinquency: Psychology
S., and Stephen W. Raudenbush. 1987. Application of hierarchical linear models to assessing
Bryk, Anthony
change. Psychology Bulletin 101:147-58.
-. 1992. Hierarchical linear models social and behavioral research: Application and data analysis
for
methods. Newbury Park, CA: Sage.
Burkhardt, Fredrick, and Sydney Smith. 1988. The correspondence of Charles Darwin. Vol. 4., 1847-1850.

University Press.
Cambridge: Cambridge
Terence Thornberry, and Marvin Krohn. 2003. Desistance as a process: A
Bushway, Shawn, developmental
of static and dynamic 19:129-53.
comparison approaches. Journal of Quantitative Criminology
our time. New
Cairns, Robert B., and Beverley D. Cairns. 1994. Life lines and risks: Pathways of youth in
York: Cambridge University Press.
Cairns, Robert B., Beverley D. Cairns, Holly J. Neckerman, Lynda L. Ferguson, and Jean-Louis Gariepy.
1989. Growth and aggression: 1. Childhood to early adolescence. Developmental Psychology 25 (2): 320
30.
across the life course. InHandbook 5th
Caspi, Avshalom. 1998. Personality development of child psychology,
ed., ed. N. and W. Daom. New York: Wiley.
Eisenberg
C. 1986. A unified biosocial and its role in the development of anxiety
Cloninger, Robert theory of personality
states. Psychiatric Developments 3:167-225.
-. 1987. A systematic method for clinical description and classification of personality variants. Archives

of General Psychiatry 44:573-88.


C?t?, Sylvana, Tracy Vaillancourt, Ted Barker, Daniel S. Nagin, and Richard E. Tremblay. 2004. Continuity
and change in the joint development of physical and indirect aggression during early childhood. Working

Paper, University of Montreal, Canada.


E. Tremblay. 2004. The devel
C?t?, Sylvana, Tracy Vaillancourt, John LeBlanc, Daniel S. Nagin, and Richard
of physical from toddlerhood to A nation wide
opment aggression pre-adolescence: longitudinal study.
Paper, University of Montreal, Canada.
Working
Elder, Glen H., Jr. 1985. Perspectives on the life course. In Life course ed. Glen H. Elder Jr. Ithaca,
dynamics,
NY: Cornell University Press.
Everitt, S., and David J. Hand. 1981. Finite mixture distributions.
Brian London: Chapman & Hall.
David P., and Donald West. 1990. The in development: A pro
Farrington, J. Cambridge Study delinquent
spective longitudinal study of
411 males. In Criminality: Personality, behavior, and life history, ed. Hans
Kaiser. New York: Springer-Verlag.
J?rgen Kerner and G?nther
Frank, Jeanne Brooks-Gunn, and S. Philip Morgan. 1987. Adolescent mothers and their chil
Furstenberg,
dren in later life. Family Planning Perspectives 19:142-51.

Glueck, Sheldon, and Eleanor Glueck. 1950. Unraveling juvenile New York: Commonwealth
delinquency.
Fund.
-. 1968. Delinquents and nondelinquents in perspective. New York: Commonwealth Fund.

Goldstein, 1995. Multilevel statistical models. 2nd ed. London: Edward Arnold.
Harvey.
costs of In Kids having kids,
Grogger, Jeffrey. 1997. Estimating the incarceration-related early childbearing.
ed. Rebecca Maynard. Washington, DC: Urban Institute.
E. Tremblay. 1994. Physically aggressive boys from ages 6 to 12: Family back
Haapasalo, Jaana, and Richard
of delinquency. and Clinical Psychol
ground, parenting behavior, and prediction Journal of Consulting
ogy 62:1044-52.
Haviland, Amelia, and Daniel Nagin. Forthcoming. Causal inference with group-based trajectory models.
Psychometrika.
1984. A method the impact of distributional
James, and Burton Singer.
Heckman, for minimizing assump
tions in econometric models for duration data. Econometrica 52:271-320.
GROUP-BASED TRAJECTORYMODELING 115

Hindelang, Michael J., Travis Hirschi, and Joseph Weis. 1979. Correlates of delinquency: The illusion of dis
between and official measures. American
crepancy self-report Sociological Review 44:975-1014.
Holyoak, Keith, and Barbara Spellman. 1993. Thinking. Annual Review of Psychology 44:265-315.
Human Capital Initiative Coordinating Committee on Violence. 1996. Reducing violence: A
Reducing
research agenda. Washington, DC: Author.

Jones, Robert, and Daniel S. Nagin. 2005. Advances in


group-based trajectory modeling and a SAS procedure
for estimating them. Working Paper, Carnegie Mellon University, PA.
Pittsburgh,
Jones, Bobby L., Daniel Nagin, and Kathryn Roeder. 2001. A SAS procedure based on mixture models for

estimating developmental trajectories. Sociological Research and Methods 29:374-93.


Kandel, Denise B. 1975. Stages in adolescent involvement in use. Science 190:912-14.
drug
Kasen, Stephanie, Patricia Cohen, Andres E. Skodol, Jeffrey G. Johnson, Elizabeth Smailes, and Judith S.
Brook. 2001. Childhood depression and adult personality disorder?Alternative pathways of continuity.
Archives of General Psychiatry 58:231-36.
Kochanska, Grazyna. 1997. Multiple to conscience for children with different
pathways temperaments:
From toddlerhood to age 5. 33:228-40.
Developmental
Psychology
Lacourse, Eric, Daniel Nagin, Frank Vitaro, Michel Claes, and Richard E.
Tremblay. 2003. Developmental
trajectories of boys' delinquent group membership and facilitation of violent behaviors during adoles
cence. and 15:183-97.
Development Psychopathology
Kirsti M., Kaj Bjorkqvist, and Tarja Peltonen. 1988. Is indirect aggression
Lagerspetz, typical of females?
Gender differences in aggressiveness in 11- to old children. Aggressive Behavior 14:403-14.
12-year
Laub, John H., Daniel S. Nagin, and Robert Sampson. 1998. Good marriages and trajectories of change in
criminal offending. American Review 63:225-38.
Sociological
E. Sampson. 2003. Shared to age 70.
Laub, John H., and Robert beginnings, divergent lives: Delinquent boys
MA: Harvard University Press.
Cambridge,
Loeber, Rolf. 1991. Questions and advances in the study of developmental InModels and integra
pathways.
tions, Rochester on ed. D. Cicchetti and S. Toth. Roches
Symposium Developmental Psychopathology,
ter, NY: University of Rochester Press.
Loeber, Rolf, and Dale F Hay. 1997. Key issues in the development of aggression and violence from child
hood to early adulthood. Annual Review of Psychology 48:371-410.
MacCallum, Robert C, Cheongtag Kim, William B. Malarkey, and Janice K. Kiecolt-Glaser. 1997. Studying
multivariate multilevel models and latent curve models. Multivariate Behavioral Research
change using
32:215-53.
Ellen M. 1989. Categorization and naming in children: Problems MA:
Markman, of induction. Cambridge,
MIT Press.
McArdle, John J., and David Epstein. 1987. Latent growth curves within developmental structural equation
models. Child Development 58:110-13.
McLachlan, Geoffrey, and David Peel. 2000. Finite mixture models. New York: Wiley.
Meredith, William, and John Tisak. 1990. Latent curve analysis. Psychometrika 55:107-22.
Moffitt, Terrie E. 1993. Adolescence-limited and life-course persistent antisocial behavior: A developmental

taxonomy. Psychological Review 100:674-701.


-. Life-course versus adolescence-limited antisocial behavior. In Develop
Forthcoming. persistent
mental psychopathology, 2nd ed., ed. D. Cicchetti and D. Cohen. New York: Wiley.
Moffitt, Terrie E., Avshalom Caspi, Nigel Dickson, Phil Silva, and William Stanton. 1996. Childhood-onset
versus adolescent-onset antisocial conduct inmales: Natural history from age 3 to 18. Development and

Psychopathology 8:399-424.
Morash, Mary, and Lila Rucker. 1989. An exploratory connection of mother s age at childbearing
study of the
to her children's in four data sets. Crime 6- 35:45-93.
delinquency Delinquency
Muth?n, Bengt. 1989. Latent variable modeling in 54:557-85.
heterogeneous populations. Psychometrika
Muth?n, Linda K., and Bengt O. Muth?n. 1998-2004. Mplus user's guide. 3rd ed. Los Angeles: Muth?n &
Muth?n.

Nagin, Daniel. 1999. developmental trajectories: Semi-parametric, Psy


Analyzing group-based approach.
Methods 4:39-177.
chological
116 THE ANNALS OF THE AMERICAN ACADEMY

-. 2005. Group-based MA: Harvard University Press.


modeling of development. Cambridge,
Nagin, Daniel S., David Farrington, and Terrie Moffitt. 1995. Life-course trajectories of different types of
offenders. 33:111-39.
Criminology
Nagin, Daniel S., and Kenneth C. Land. 1993. Age, criminal careers, and population heterogeneity: Specifi
cation and estimation of a nonparametric, mixed poisson model. Criminology 31:327-62.
course case
Nagin, Daniel, Linda Pagani, Richard Tremblay, and Frank Vitaro. 2003. Life turning points: A
on
study of the effect of school failure interpersonal violence. Development and Psychopathology 15:343
61.

Nagin, Daniel S., and Richard E. Tremblay. 1999. Trajectories of boys' physical aggression, opposition, and
on the to physically violent and nonviolent Child Development
hyperactivity path juvenile delinquency.
70:1181-96.
-. 2001a. of distinct but related behaviors: A group-based
Analyzing developmental trajectories
method. Methods 6:18-34.
Psychological
-. 2001b. Parental and early childhood predictors of persistent physical aggression in boys from kinder

garten to high school. Archives of General Psychiatry 58:389-94.


-. 2002. Most fall but not all: Changes in from childhood
physical aggression through adolescence.
Paper, Carnegie Mellon University, PA.
Working Pittsburgh,
-. 2005. Trajectory groups: Fact or fiction. Working PA.
Paper, Carnegie Mellon University, Pittsburgh,
National Institute of Child Health and
Development (NICHD) Early Child Care Research Network. 2004.
of physical aggression from toddlerhood to middle childhood. Monograph of the Society for
Trajectories
Research on Child Development Serial no. 278, vol. 69, 4.
K. Wilson, G. Bachman, Patrick M. O'Malley, and Lloyd Johnson. 1996. Rou
Osgood, Wayne D., Janet Jerald
tine activities and individual deviant behavior. American Sociological Review 61:635-55.
Patterson, Gerald R., Barbara D. DeBaryshe, and Susan E. Ramsey. 1989. A developmental on
perspective
antisocial behavior. American 44:329-35.
Psychologist
Pickles, Andrew, and Adriane Angold. 2003. Natural or fundamental dimensions: On carving
categories
nature at the joints and the rearticulation of psychopathology. Development and Psychopathology 15:529
51.
we learned about the natural
Piquero, Alex. 2004. What have history of criminal offending from longitudinal
studies? Paper presented at the National Institute of Justice, Washington, DC.
P. Thornberry. of children born to young
Pogarsky, Greg, Alan J. Lizotte, and Terence 2003. The delinquency
mothers: Results from the Rochester Youth Development 41:101-38.
Study. Criminology
Raine, Adrain, Patricia Brennan, Bergitte Mednick, and Sarnoff Mednick. 1996. High rates of violence,
in males with both
crime, academic problem, and behavioral problems early neuromotor deficits and
unstable family environments. Archive General Psychiatry 53:544-49.

Raudenbush, Stephen
W. 2001. Comparing-personal trajectories and drawing causal inferences from longi
tudinal data. Annual Review of Psychology 52:501-25.
Reiss, Albert J., and Jeffrey A. Roth. 1994. Understanding and preventing violence. Vol. 4, Consequences and
control. Washington, DC: National Academy Press.

Sampson, Robert J., and John. H. Laub. 1993. Crime in the making: Pathways and turning points through
MA: Harvard University Press.
life. Cambridge,
Terence P., Alan J. Lizotte, Marvin D. Krohn, Carolyn A. Smith, and Pamela K. Porter. 2003.
Thornberry,
Causes and consequences of delinquency: Findings from the Rochester Youth Development Study. In
Taking stock of delinquency:
An overview of findings from contemporary longitudinal studies, ed. T.
Thornberry and M. D. Krohn. New York: Kluwer Academic.

Titterington, D. M., A. F. M. Smith, and U. E. Makov. 1985. Statistical analysis offinite mixture distributions.
New York: Wiley.
E. 2004a. The development
Richard of human physical aggression: How important is early child
Tremblay,
hood? In Social
and moral development: Emerging evidence on the toddler years, ed. D. M. B. Hall, A.
Leavitt, and N. A. Fox. New Brunswick, NJ: Johnson and Johnson P?diatrie Institute.
-. 2004b. Why socialization fails? The case of chronic physical aggression. In Causes of conduct disor
der and juvenile delinquency, ed. B. B. LaheyT. E. Moffitt, and A. Caspi, 182-224. New York: Guilford.
GROUP-BASED TRAJECTORYMODELING 117

Tremblay, E., Christa Japek, Daniel P?russe, Pierre McDuff,


Richard Michel Boivin, Mark Zoccolillo, and

Jacques Montplaisir. 1999. The search for the age of "onset" of aggression: Rousseau and
physical
Bandura revisited. Criminal Behaviour and Mental Health 9:8-23.

Tremblay, Richard E., and Daniel S. Nagin. 2005. Aggressionin humans. In Developmental origins of aggres
sion, ed. Richard E. Tremblay, William W Hartup, and John Archer. New York: Guilford.
S. Nagin, Jean R. S?guin, Mark Zoccolillo, D. Zelazo, Michel Boivin,
Tremblay, Richard E., Daniel Philip
Daniel Perusse, and Christa Japel. 2004. Physical aggression during early childhood: Trajectories and pre
dictors. Pediatrics 114:e43-e50.
Truett, Jeanne, Cornfield Jerome, and Kannel William. 1967. A multivariate analysis of the risk of coronary
heart disease in of Chronic Diseases 20:511-24.
Framingham./owrri?/
Tyrka, Audrey R., Julia A. Graber, and Jeanne Brooks-Gunn. 2000. The development of disordered eating. In
Handbook of developmental psychopathology, 2nd ed., ed. Arnold J. Sameroff, Michael Lewis, and
Suzanne M. Miller. New York: Kluwer Academic/Plenum.
S., Rachel A. Gordon,
Lauren B. Lahey, Rolf Loeber, M. Green, and
Wakschlag, Benjamin Stephanie
Bennett L. Leventhal. 2000. Maternal age at first birth and boys' risk for conduct disorder. Journal of
Research on Adolescence 10:417-41.
Warr, Mark. 2002. Companions in crime: The social aspects of criminal conduct. New York: Uni
Cambridge
versity Press.
structure analysis to detect correlates and
Willett, John B., and Aline G. Sayer. 1994. Using covariance predic
tors of individual over time. Bulletin 116:363-81.
change Psychological

You might also like