Cop yr i ght 1999 by the Amer i can Psychological Associ at i on, I nc. 1082-989X/99/$.VOO Analyzing Developmental Trajectories: A Semiparametric, Group-Based Approach Daniel S. Nagin Carnegie Mel l on Un i ver s i t y A devel opmental trajectory describes the course of a b ehavi or over age or time. A group-based method for i denti fyi ng distinctive groups of individual trajectories w i t h i n the popul ati on and for pr ofi l i ng the characteristics of group members is demonstrated. Such clusters mi ght i ncl ude groups of "increasers." "decreasers," and "no changers." Sui tab l y defi ned pr ob ab i l i t y di st r i b ut i ons are used to handl e 3 data typescount, b i nar y, and psychometr i c scale data. Four capab ilities are dem- onstrated: (a) the capab i l i ty to i dent i fy rather than assume di sti ncti ve groups of trajectories, (b ) the capab i l i t y to esti mate the proportion of the popul at i on fol l owi ng each such trajectory group, (c) the capab i l i t y to relate group memb ership probabil- ity to i n di vi dua l characteristics and circumstances, and (d) the capab i l i t y to use the gr oup memb er shi p prob ab i l i ti es for var i ous other purposes such as creati ng profiles of group members. Over the past decade, major advances have been made i n methodology for anal yzi ng i ndi vi dual -l evel developmental trajectories. The two mai n branches of methodol ogy are hi er a r c hi c a l model i ng ( Br yk & Raudenb ush, 1987, 1992; Goldstein, 1995) and l atent curve anal ysi s (McArdle & Epstein, 1987; Meredith &Ti sak, 1990;Muthen, 1989; Wi l l et t & Sayer, 1994). These advanced methods allow researchers to move beyond the use of ad hoc categorization procedures for constructing developmental trajectories. Al though these two classes of methodology differ in very important respects, they also have important c o mmo n a l i t i es ( Ma c Ca l l um, Ki m, Ma l a r ke y , & Ki ec o l t -Gl a s er , 1997; Mut h en & Cur r an, 1997; Raudenb us h, in press; Wi l l ett & Sayer, 1994). For the purposes of this article, one is key: Both model the This work was supported by National Science Foundation Grant SBR-9511412 and by the National Consortium for Vi ol ence Research. 1 t hank Robert Brame, Lisa Broidy, Pat- r i ck Cur r an, Kenneth Dodge, Laura Dugan, Anne Gar vi n, Don Lynam, Steve Raudenb ush, Kat hr yn Roeder, Robert Siegler, and Larry Wasserman for hel pful comments on previous versi ons of thi s article. Correspondence concer ni ng thi s article should be ad- dressed to Daniel S. Nagi n, H. John Hei nz III School of Public Policy & Management, Carnegie Mellon University, Pi ttsb ur gh, Penns yl vani a 15213. unconditional and conditional population distrib ution of growth curves based on conti nuous distrib ution funct i ons. Uncondi ti onal models estimate two key features of the popul at i on di s t r i b ut i on of growth curve parameterstheir mean and covariance struc- ture. The former defines average growth wi t hi n the population, and the latter calibrates the variances of growth throughout the popul ati on. The conditional models are designed to explain this variab ility by re- l ati ng growth parameters to one or more explanatory variables. This article demonstrates a distinct semiparametric, group-based approach for modeling developmental trajectories. The model i ng strategy is intended to pro- vide a fl exi b l e and easily applied approach for iden- t i fyi ng di sti ncti ve clusters of i ndi vi dual trajectories wi t hi n the population and for profiling the character- istics of i ndi vi dual s wi t hi n the clusters. Usi ng mix- tures of sui tab l y defined prob ab i l i ty distrib utions, the method i denti fi es di sti ncti ve groups of developmental trajectories wi t hi n the popul ati on (Land & Nagi n, 1996; Nagi n, Harrington, & Moffitt, 1995; Nagin & Land, 1993; Nagin & Tremblay, in press; Roeder, Lynch, & Nagi n, in press). Thus, whereas the hierar- chical and latent curve methodologies model popula- tion var i ab i l i t y i n growth wi t h mul t i var i at e cont i nuous distribution functions, the group-based approach uses a mul t i nomi al modeling strategy and is designed to 139 140 NAGI N i dent i fy relatively homogeneous clusters of develop- mental trajectories. The group-based model i ng strategy demonstrated in t hi s article is presaged in the work of Ri ndskopf (1990). Ri nds kopf' s methodology i s f ul l y nonpara- metric. It is designed for anal ys i s of repeated mea- sur ements of di chotomous response dat a. The re- peated mea s ur ement s may come i n the for m of responses to a series of test i t ems or, al ter nati vel y, in the form of panel data. Ri ndskopf' s method is de- signed to i denti fy di s t i nct groups of response se- quences across sampled i ndi vi dual s . The method de- scribed here expands on Ri ndskopf' s i nnovat i ve work in several i mpor tant respectsit increases the variety of response variab les to whi ch a group-based model- ing strategy can be applied, it provides a basis for l i nki ng group memb er shi p prob ab ility to r el evant in- di vi dual -l evel characteristics, and it describes a for- mal approach for deter mi ni ng the opti mal n umb er of groups. Rati onal e for a Group-Based Modeling Strategy Before di s cus s i ng the t echni cal feat ur es of the model, I b ri efl y speak to the under l yi ng rationale for a group-based modeling strategy. There is a long tra- di ti on in psychology of group-based theorizing about devel opment. Exampl es i ncl ude theories of personal- ity devel opment (Caspi, 1998), drug use (Kandel , 1975), l ear ni ng (Holyoak & Spel l man, 1993), l an- guage and conceptual development (Mar kman, 1989), devel opment of prosocial b ehavi ors such as al t r ui s m, and devel opment of anti soci al b ehaviors such as de- l i n quen c y (Loeber, 1991; Moffi t t , 1993; Patterson, DeBaryshe, & Ramsey, 1989). This group-based ap- proach is well suited to anal yzi ng questions ab out devel opmental trajectories t hat are i nher ent l y cat- egoricaldo certain types of people tend to have dis- t i nct i ve devel opmental trajectori es? For example, Moffi t t (1993) and Patterson et al. (1989) have argued that early onset of del i nquency is a di st i nct i ve feature of the devel opment trajectories of chr oni cal l y cr i mi nal and antisocial adul ts but not of i n di vi dua l s whose offendi ng is l i mi ted to thei r ado- lescence. The group-based model i ng strategy can test whether the devel opmental trajectories predicted by the theories are actual l y present in the popul ati on. It can also be used to test key predictions such as wheth- er the chronic trajectory is t ypi fi ed by an earlier onset of del i nquent b ehavi or than the adolescent l i mi ted tra- jectory. In contrast, because hierarchical and l at ent curve modeling assume a conti nuous distrib ution of trajectories wi t hi n the popul ati on, these methods are not designed to i dent i fy di s t i nct cl usters of trajecto- ries. Rather they are designed to describe how pat- t er ns of gr owt h vary c o n t i n uo us l y t hr oughout the popul ati on. As a resul t, it is awkwar d to use these methods to address research quest i ons that contrast di s t i nc t categories of developmental trajectories. The group-based modeling approach assumes that the population is composed of a mi xt ur e of di sti nct groups defined by their developmental trajectories. The assumpti on t hat the popul ati on is composed of di sti nct groups is not l i kel y l i t er al l y correct. Unl i ke biological or physi cal phenomena, in whi ch popula- ti ons may be composed of actual l y di s t i nct groups such as di ffer ent types of ani mal or pl ant species, popul ati on differences in developmental trajectories of b ehavior are un l i kel y to reflect such b r i ght-l i ne di ffer ences ( a l t hough biology has a long tr adi ti on of debates concer ni ng cl assi fi cati on; see Appel, 1987). To be sure, taxonomic theories predict di ffer ent tra- jectories of devel opment across sub popul ati ons, but the purpose of such taxonomies is gener al l y to draw at t ent i on to di ffer ences in the causes and conse- quences of different developmental trajectories wi t hi n the popul ati on rather than to suggest that the popula- tion is composed of l i t er al l y di sti nct groups. In thi s respect, devel opment al toxonomies are analogous to prototypal di agnosti c cl assi fi cati ons i n c l i ni c a l psy- chology. U n l i ke classical classifications, prototypal cl assi fi cati ons acknowl edge fuzzi ness i n classifica- t i on, employ polythetic diagnostic criteria, and recog- ni ze wi t hi n-gr oup var i ati on in the prevalence and se- ver i t y of symptoms (Cantor & Genero, 1986; Widiger & Frances, 1985). The group-based modeling strategy is prototypal in design and provides a methodological complement to theories that predict prototypal devel opmental etiolo- gies and trajectories w i t h i n the popul ati on. The mod- el i n g str ategy e xp l i c i t l y recogni zes uncer t ai nt y i n group membership, al l ows an exami nat i on of the im- pact of mul t i pl e factors on prob ab i l i ty of group mem- bership, and anoi nt s no set of factors as necessary and suffi ci ent i n deter mi ni ng group memb er shi p. Wi t h t hi s approach the basic elements of taxonomi c theo- ries can be directly tested: Are the r el at i vel y homo- geneous devel opment al tr ajector i es and eti ol ogi es predicted by the theory act ual l y present in the popu- l at i on? In so doing, the method provides an al ternati ve to usi ng assi gnment r ul es based on sub jecti ve categori- ANALYZI NG DEVELOPMENTAL TRAJECTORIES 141 zation criteria to construct categories of developmen- tal trajectories. Although such assignment rules are gener a l l y reasonable, there are l i mi t at i ons and pi tfal l s attendant to t hei r use. One is that the existence of the various developmental trajectories that under l i e the taxonorni c theory cannot be tested; they must be as- s umed a pr i or i . A r el ated p i t f a l l of c on s t r uc t i n g groups wi t h sub jecti ve classification procedures is overfillingcreation of trajectory groups that reflect only random var i ati on. Second, ex ante specified rules provide no basis for cal i b rati ng the precision of i ndi - vi dual cl assi fi cati ons to the various groups that com- pose th;: taxonomy. The semiparametric, group-based method presented in t hi s article avoids each of these l i mi tati ons. It provides a formal basis for determining the numb er of groups that best fi t s the data and also provides an explicit metric, the posterior prob ab ility of group membership, for eval uat i ng the precision of group assi gnments. Data and Three I l l ustrati ve Examples Thr oughout, key methodological poi nts are i l l us - trated wi t h fi ndi ngs from the anal ysi s of two wel l - known l ongi t udi nal studies. One is the Cambridge Study of Del i nquent Development, whi ch tracked a sample of 41 1 Br i t i sh males from a working-class area of London. Data collection began in 1961-62, when most of the boys were age 8. Cr i mi nal i nvol ve- ment is measured by conviction for cr i mi nal offenses and i s avai l ab l e for al l i n di vi dua l s i n the sample through age 32, wi t h the exception of the 8 i ndi vi du- als who died prior to this age. Between ages 10 and 32 a wea l t h of data was assembled on each i ndi vi dua l ' s psychological makeup, fa mi l y ci rcumstances i ncl ud- ing parental b ehaviors, and performance in school and wor k. For a complete discussion of the data set, see Farrington and West (1990). The other data set is from a prospective l ongi tudi - nal study based in Montreal, Canada. This study has been tracki ng 1,037 White males of French ancestry. Participants were selected in 1984 from kindergarten classes in low-socioeconomic-status Montreal neigh- borhoods. Following the assessment at age 6, the boys and other i nfor mant s were i nt er vi ewed a n n ua l l y from ages 10 to 18. Like the Cambridge study, assessments were made on a wi de range of factors. Among these is physi cal aggression, whi ch was measured on the basis of teacher r ati ngs usi ng the Social Behavior Question- nai re ( Tr emb l a y, Des mar ai s -Ger vai s , Ga gnon, & Charl eb oi s, 1987). The boys were rated on three i tems: ki c ki ng, b i t i ng, and hi t t i ng other chi l dr en; fi ghti ng wi t h other chi l dren; and b ul l yi ng and i nt i mi - dating other children. See Tremblay et al. for fur ther details on t hi s study. Figures 1, 2, and 3 show the results of trajectory analyses based on these two data sets. The software used to esti mate these models is a customized SAS procedure that was developed wi th the SAS product SAS/TOOLKIT. The procedure, whi ch is cur r entl y availab le for the PC pl atform only, is a program writ- ten in the C programming l anguage and is designed to interface wi t h the SAS system to perform model fit- ti ng. It accommodates mi ssi ng data. Thus, i ndi vi dual s wi th incomplete assessment histories do not have to be dropped from the analysis. Also, time between assessment peri ods does not have to b e e qua l l y spaced, and assessment periods do not have to be identical across participants. The software, whi ch is easily used and incorporated i nto the standard SAS package, is avai l ab l e on request and is described in detail in Jones, Nagi n, and Roeder (1998). Figure 1 displays the results of the anal ysi s of the Cambridge data. The response variab le in this anal y- sis is a count, the numb er of convi cti ons for each 2-year period from ages 10 to 31 years. The solid lines represent actual b ehavior, and the dashed l i nes repre- sent predicted b ehavior. 1 Three groups were i denti - fied: A group called "the never convicted" was com- posed of respondents who, but for a few i ndi vi dual s wi t h a si ngl e convi ct i on, were not convicted through- out the observation period. This group accounts for an estimated 71 % of the sampled popul ati on. A second group of i ndi vi dual s who ceased thei r offending (as measured by conviction) by their early 20s was la- beled "adolescent limited" after Moffitt (1993). This group is estimated to constitute 22% of the popula- tion. Fi nal l y, a chronic group was identified and is estimated to make up 7% of the popul ati on. I ndi vi du- 1 Predicted b ehavior is calculated as the expected val ue of the random variab le depi cti ng each group' s b ehavior. Ex- pected values are computed based on model coeffi ci ent es- timates. For the appl i cati on depicted in Figure !, t hi s ex- p ec t a t i on e qua l s t he a n t i l o g of Equa t i o n 1. For t he appl i cati on depicted in Figure 2, it is calculated according to the r el a t i ons hi p provided in footnote 4. For the appl i cat i on depicted in Figure 3, it equals the b i nomi al prob ab i l i ty as computed by Equati on 3. Actual b ehavi or is computed as the mean b ehavior of all persons assigned to the vari ous groups i denti fi ed in esti mati on. As described hereinafter, the assi gnments are based on the posterior prob ab ility of group memb ershi p. 142 NAGI N 2.5 4 - Never-actual -Never-pred. -Adol. Limited-actual -Adol. Limited-pred. -Chronic-actual - Chronic-pred. 10 12 14 16 18 20 22 24 26 28 Age Figure I . Trajectories of numb er of convi ct i ons (Cambridge sample). Adol. = adolescent; pred. = predicted. 30 als in this group offended at a hi gh level thr oughout i he observation period. Figure 2 displays r esul ts based on the Montreal data. In this anal ysi s the response vari ab l e was a psy- chometric scale of physi cal aggression. A four-group model was found to best fit the data. A group called "nevers" is composed of i ndi vi dua l s who never dis- play physi cal l y aggressive behavior to any sub stanti al degree. This group is estimated to make up about 15% of the sample popul ati on. A second group that con- sti tutes about 50% of the popul ati on is best labeled "low-level desisters." At age 6 boys in this group di spl ayed modest levels of physical aggression, but by age 10 they had largely desisted. A thi rd group, con- s t i t ut i ng about 30% of the popul ati on, was labeled ' ' hi gh-l evel near desisters." This group started off scoring hi gh on physi cal aggression at age 6 but by age 15 scored far lower. Not wi t hst andi ng thi s marked decl i ne, at age 15 they continued to display a modest level of physical aggression. Fi nal l y, there is a small chronic group, consti tuti ng ab out 5% of the popula- t i on, who displayed high levels of physical aggression thr oughout the observation period. Figure 3 displays another variant of the analysis of the physi cal aggression data. For thi s analysis the ag- gression data for each assessment period was trans- formed from an intensity scale over the interval 0 to 6 to a binary indicator equal to 1 if the i ndi vi dual dis- played any evidence of physi cal aggression in that period (i.e., if his score was greater than 0). Such a symptom indicator variab le is an example of b i nary dataa type of data that is widely analyzed in the social sciences. It is also the type of data Rindskopf s group-based method was designed to analyze. Three symptom trajectories are depicted in Figure 3. One group, const i t ut i ng about 20% of the popul at i on, never displayed any symptoms. This group is the counterpart of the nevers in Figure 2. A second group i ncl udes about 50% of the popul ati on. It started off wi t h a fairly high probability of di spl ayi ng symptoms of physical aggression, ab out .6, but thi s probability declined rapidly to near 0 by age 15. The final group, whi ch is estimated to account for about 30% of the population, began wi t h a prob ab ility of physi cal ag- gression that is onl y modestly hi gher than that of the second group, but the sub sequent rate of decline was ANALYZI NG DEVELOPMENTAL TRAJECTORIES 143 - Never-actual Never-pred. - Low desister-actual Low desister-pred. High desister-actual High desister-pred. - Chronic-actual -Chronic-pred. Figure 2. Trajectories of physical aggression (Montreal sample), pred. = predicted. far slower. Even at age 15, the probability of their dis- playing symptoms of physical aggression is about 4. Model Figure 4 provides an overview of the model and its key outputs. Model estimation does not require the ex ante sorting of the data among the various groups depicted in Figures 1 , 2, and 3. To the contrary, the data are used to i denti fy the numb er of groups that best fit-, the data and the shape of the trajectory for each group. The data also provide an estimate of the proportion of the population whose measured behav- ior conforms most closely to each trajectory group. In the discussion that follows, I first lay out the general form of the model for any given numb er of groups. Next, I discuss the determination of the opti- mal numb er of groups and the shape of trajectory for each gr oup. I then move to a discussion of how the model' s parameter estimates can be used to compute the fi nal element depicted in Figure 4the probabil- ity that each i ndi vi dual / in the estimation sample belongs to each of the trajectory groups identified in estimation. These capab ilities are illustrated with re- sults from the Cambridge and Montreal data sets. For the Montreal data the censored normal anal ys i s is used for thi s purpose. However, all of these capab i l i - ties are equal l y well employed in the anal ysi s of bi- nary data. Differences in the form of the response variab le in the analyses depicted in Figures 1, 2, and 3a count var i ab l e, a psychometric scale, and a b i nar y vari- ablenecessitate technical differences in the statisti- cal model used in each analysis. For the count data the under l yi ng mi xtur e model b ui l ds from the Poisson distrib ution and its even more general rel ati ve, the zero-inflated Poisson (Lambert, 1993). The Poisson fami l y of distrib utions is wi del y used in the anal ysi s of count data (Cameron & Trivedi, 1986; Parzen, 1960). For the psychometric scale data the under l yi ng model is based on the censored normal di str i b uti on. The censored normal distrib ution is well suited to accommodate a common feature of psychometr i c scale data. Typi cal l y, a sizab le cont i ngent of the sample exhi b i ts none of the behaviors measured by the scale. The result is a clustering of data at the scale mi ni mum. Also, there is us ual l y a smal l er contingent that exhi b i ts all of the behaviors measured by the scale. The result is another cluster of data at the scale 144 NAGI N - Desister-actual - Desister-pred. Non-desister-actual Non-desister-pred. Figure 3. Trajectories of symptoms of physi cal aggression (Montreal sampl e), pred. = predicted. ma xi mum. Fi nal l y, the b i nary logit di stri b uti on is used to model binary data. See Maddala (1983) or Greene (1990) for a thorough di scussi on of the cen- sored normal and binary logit di stri b uti ons. Li ke hi erarchi cal and l atent cur ve modeling, a poly- nomial rel ati onshi p is used to model the l i nk between age and b ehavior. Speci fi cal l y, a quadr ati c r el ati on- ship is assumed. 2 For the Poisson-based model it is assumed that log(\/,) = ft, + ft Age, + ft Age?,, ( 1) where X -J, is the expected numb er of occurrences of the event of interest (e.g., convictions) of s ub j ec t i at ti me / given memb ership in group/ Age,, is subject i ' s age at ti me t, and Age 2 , is the square of subject f s age at t i me r. 3 The model ' s coefficients(3{,, fj',, and ft deter mi ne the shape of the trajectory and are super- scripted by j to denote that the coefficients are not constrained to be the same across the j groups. The condi ti onal prob ab i l i ty of the actual numb er of events, P(^ it \j), gi ven j is assumed to fol l ow the wel l -known Poisson distrib ution. 4 For the censored normal model, the l i nkage be- tween age and behavior is established by means of a latent variab le, >',*,', that can be thought of as measur- Optimal Number of Groups & Trajectory Shapes I Proportion of Population in Each Group Probability that Individual i Belongs to Trajectory Group j Figure 4. Overview of the model. 2 The software package used in est i mat i ng the models reported here allows for esti mati on of up to a cubic pol y- nomi al in age. For ease of exposition I l i mi t the discussion to examples where the ma xi mum order considered is qua- dratic. * A log-linear r el at i onshi p between X ', and age is assumed to ensure that the requirement that \j, > 0 is f ul f i l l ed in model esti mati on. 4 In an even more general version of this model, P(v/,[/) is assumed to fol l ow the zer o-i nfl ated Poisson di st r i b ut i on. See Land, McCal l , & Nagi n (1996), Nagi n and Land (1993), or Roeder, Lynch, and Nagi n (in press) for the devel opment of thi s more general case. ANALYZI NG DEVELOPMENTAL TRAJECTORIES 145 ing potential for engaging in the behavior of inter- estsay, physical aggressionfor i ndi vi dual ;'s age at ti me t gi ven memb ershi p in gr oup/ Agai n, a qua- dratic r el ati onshi p is assumed between > '/ and age: B v*/=(y o +p' I (2) where Age,-, and Age 2 ,, are as pr evi ous l y defi ned and e is a di s t ur b ance assumed to be nor mal l y di stri b uted wi t h zer o mean and constant variance a 2 . The l atent variab le, y* /, is l i nked to its observed but censored counterpart, y,-,, as fol l ows. Let 5 mi n and S max , respecti vel y, denote the mi n i mum and maxi - mum possible score on the measur ement scale. The model assumes y'ii ~ -^mi n 1' y / r ^ mi n ' y /; = y*,' if 5 mm <y* ' <S max , and . V i> = -X mix " }'it -"" ^ ma x- In words, if the l atent vari ab l e, y* ', is less than 5 mi n , it is assumed that observed behavior equal s this mi ni - mum. Likewise, if the l atent var i ab l e, y* /, is greater than S max , it is assumed that observed behavior equals t hi s ma xi mum. Onl y if y* / is wi t hi n the scale mi ni - mum and ma xi mum does y,-, = y* '. 5 ' 6 Fi n a l l y , consider the case wher e the measured re- sponse y,,, is b i nar y. In thi s case it is assumed that condi ti onal on memb er shi p in group; the prob ab i l i ty that y,, = 1, ot'ip follows the b i nary logit di st r i b ut i on: (3) 1 +<?" In the Poisson, censored normal , and b inary logit formulations, the parameters defi ni ng the shape of the trajectory$ J 0 . (3',, and (3 ; 2 are left free to di ffer across groups. This fl exi b i l i t y is a key feature of the model because it al l ows for easy i denti fi cati on of popul at i on heterogeneity not onl y in the level of be- havior at a gi ven age but also in its development over time. Fi gur e 5 i l l ust r at es two hypotheti cal possi b i l i - ties. A single peaked trajectoryTrajectory Ais i mpl i ed if p, > 0 and p? <0. Thus, if data col l ecti on began at age 1, the trajectory would i mpl y that for t hi s group the occurrence of the behavior rose steadi l y un t i l age 6 and then began a steady decline. Al t er na- t i vel y, if data collection began at age 6, as was the case in the Montreal s t udy, generally it would be in- appropriate to extrapolate backward to a younger age outside the period of measurement. Thus, for a model 10 11 12 Age Figure 5. Two h y p o t h et i c a l t r aject or i es. A = s i n g l e peaked; B = chronic trajectory. based on data from age 6 onward, t hi s trajectory woul d i mp l y a steady decline in the b ehavior fol l ow- ing the i ni t i a l assessment. Such a trajectory woul d t ypi fy desistance from the b ehavior. The second tra- jectory (B) depicted in Figure 5 has no cur vatur e. Rather it remains constant over age. This trajectory is i mpl i ed if (^ = 0 and |3 2 = 0. If that stable level of the b ehavi or is hi gh, thi s trajectory woul d t ypi fy a group that chronically engages in the behavior. Other interesting possibilities i ncl ude trajectories in whi ch growth is ei ther steadily accelerating or decelerating. The former woul d be characterized by a trajectory in whi ch both (3j and (3 2 are positive and the latter by both b ei ng negative. The trajectories depicted in Figures 1, 2, and 3 are the product of maxi mum-l i kel i hood esti mati on. The l i kel i hood funct i on was constructed as fol l ows. Let the vector y, = {y,,, y/i, , y, T } denote the longi- t udi na l sequence of i ndi vi dua l /"s behavioral measure- 5 The expected val ue of the l at ent variab le y* ' equal s $ n j + (3{ Age,, + (3^ Agef,, whi ch is denoted below by p.v ; . The expected val ue of the measured qua n t i t y, (v',). assumi ng j was observed, is E(Y>,) = V mta S min + P^* !- *!,) + (r (ct > i l i n - 4> ,', u ) + (1 - 4> L* )S max , where * (* ) denotes the c umul a t i ve normal di st r i b ut i on funct i on and <t > m i n and 4> max , respecti vel y, denote < t> '[(S m , n - (iO/cr)] and <l> '[(S max - PX ',)/O-)], wi th corresponding defi ni ti ons for 4> mi n and <j> ma x , where <j> denotes the normal density func t i on. This r el ati on- shi p was used to compute the predicted components of Fig- ur e 1. 6 I use the term latent variable to describe y* ' because it is not f ul l y observed. Thus, my use of the term latent is di ffer ent from that in the psychometric literature, where the term latent factor refers to an unob ser vab l e construct t hat is assumed to gi ve rise to mul t i pl e mani fes t vari ab l es. 146 NAGI N merits over the T periods of measurement, P(Y t ) de- note the pr ob ab i l i t y of Y/ given memb er shi p in group j, and TTj denote the pr ob ab i l i ty of memb er shi p in gr oup / For count data P ! (Y/) is constructed from the Poisson di s t r i b ut i on, for psychometri c scale data from the censored normal di s t r i b ut i on, and for the b i nar y data from the b i nar y logit di st r i b ut i on. If group mem- bership was observed, the sampled i n di vi dua l s coul d be sorted by group memb ershi p and t hei r trajectory parameters estimated wi t h r eadi l y avai l ab l e Poisson, censored normal (t ob i t ), and logit regression software r out i nes (e.g., STATA, 1995 or LIMDEP, Greene, 1991). 7 However, group memb er shi p is not observed. In- deed the proporti on of the popul at i on composi ng gr oup/ iTy, is an important parameter of interest in its own ri ght. Thus, constructi on of the l i kel i hood re- qui r es the aggregation of the J condi t i onal l i kel i - hoods, f' (K/), to form the uncondi t i onal prob ab ility of the data, K,: (4) where P(Y t ) is the uncondi ti onal prob ab ility of ob- ser vi ng i ndi vi dual f s l ongi t udi nal sequence of b ehav- ioral measurements. It equal s the sum across the J groups of the pr ob ab i l i t y of K, gi ven memb er shi p in i i r oupy weighted by the proportion of the popul ati on i n group / The log of the l i kel i hood for the ent i r e sample is t hus the sum across all i n di vi dua l s that com- pose the sample of the log of Equati on 4 eval uated for each i n di vi dua l /. The parameters of interest(3(,, (B 7 ,, |:1 2 , TTJ, and, in addi ti on, cr for the censored nor mal can be estimated by maxi mi zat i on of thi s log l i kel i - hood. As pr evi ousl y mentioned, SAS-based software for accomplishing this task is avai l ab l e on request. For a deri vati on of the l i kel i hood see the appendi x, but i nt ui t i vel y, the estimation procedure works as fol- lows. Suppose unb eknownst to us there were two dis- t i nc t groups in the popul ati on: youth offenders, con- st i t ut i ng 50% of the popul ati on, who up to age 18 have an expected offendi ng rate, X , of 5 and who after age 18 have a X of 1, and adul t offenders, consti tuti ng the other 50% of the popul ati on, whose offendi ng tr ajector y is the reverse of that of the youth offend- er sthr ough age 18 t hei r X = I, and after age 18 thei r X increases to 5. If we had l ongi tudi nal data on the recorded offenses of a sample of i ndi vi dual s from t hi s p o p ul a t i o n , we w o ul d ob ser ve t wo di s t i n c t groups: a cl uster i ng of about 50% of the sample who generally have many offenses prior to age 18 and r el at i vel y few offenses after age 18 and another 50% cl uster i ng wi t h jus t the reverse pattern. Suppose these data were anal yzed under the as- sumption that the r el ati onshi p between age and X was i denti cal across all i n di vi dua l s . The estimated val ue of X woul d be a "compromise" estimate of about 3 for all ages from whi c h we woul d mi s t akenl y concl ude that in t hi s popul ati on the rate of offendi ng is i nvar i ant wi t h age. If the data were analyzed usi ng the approach described here, whi ch specifies the l i kel i hood func- tion as a mi xi n g di st r i b ut i on, no such mathemati cal compromise woul d be necessary. The parameters of one component of the mi xt ur e woul d effecti vel y be used to accommodate (i.e., mat ch) the yout h offend- ing portion of the data whose offendi ng declines wi t h age, and another component of the mi xi n g di st r i b ut i on woul d be avai l ab l e to accommodate the a dul t offender data whose offendi ng increases wi t h age. Model Selection This section addresses two i mpor t ant i ssues i n model sel ect i on: (a) det er mi nat i on of the optimal numb er of groups to compose the mi xt ur e and (b) determi nati on of the appropriate order of the polyno- mi al used to model each group' s trajectory. Here or- der refers to the degree of the pol ynomi al used to model the group' s trajectory, where a second-order trajectory is defined by a quadr at i c equati on, a fi r st - order trajectory is defined by a l i near equation in whi ch (3 2 is set equal 0, and a zero-order trajectory is defi ned by a fl at l i ne in whi ch (}, and (3 2 are set equal to zero. I di scuss these two i ssues in t ur n. One possib le choice for testi ng the opti mal i ty of a specified numb er of groups is the likelihood ratio test. However, the n ul l hypothesi s (i.e., three components vs. more than three components) is on the b oundary of the parameter space, and hence the classical asymp- totic resul ts that under l i e the likelihood ratio test do not hold (Erdfel der, 1990; Ghosh & Sen, 1985; Tit- ter i ngton, Smi t h, & Makov, 1985). The l i kel i hood ra- tio test is suitab le onl y for model selection problems in whi ch the al ter nati ve models are nested. In mi xt ur e models, a k group model is not nested w i t h i n a k + 1 7 In the econometric l i terature censored nor mal regres- sion is gener al l y cal l ed tobit regression after its ori gi nator, James Tob in (Tob i n, 1958). ANALYZING DEVELOPMENTAL TRAJECTORIES 147 group model, and, therefore, it is not appropriate to use the likelihood ratio test for model selection. 8 Given these problems with the use of the likelihood ratio test for model selection, we follow the lead of D'Unger, Land, McCall, and Nagin (1998) and use the Bayesian information criterion (BIC) as a basis for selecting the optimal model. For a given model, BIC is calculated as follows: BIC = log(L) - 0.5* logOz)* (/t), (5) where L is the value of the model's maximized like- lihood, n is the sample size, and k is the number of parameters in the model. Kass and Raftery (1995) and Raftery (1995) have argued that BIC can be used for comparison of both nested and unnested models under fairly general circumstances. When prior information on the correct model is limited, they recommended selection of the model with the maximum BIC. Note that BIC is always negative, so the maximum BIC will be the least negative value. In even more recent work, Keribin (1997) demonstrated that BIC identi- fies the optimal number of groups in finite mixture models. Her result is specifically relevant for the mix- ture models demonstrated here. For insight into the usefulness of BIC as a criterion for model selection, consider its calculation. The first term of Equation 5 is always negative. For a model that predicts the data perfectly it equals zero. As the quality of the model's fit with the data declines, this term declines (i.e., becomes more negative). One way to improve fit and thereby reduce the first term is to add more parameters to the model. The second term extracts a penalty proportional to the log of the sample size for the addition of more parameters. Thus, on the basis of the BIC criterion, expansion of the model by addition of a trajectory group is desirable only if the resulting improvement in the log likelihood exceeds the penalty for more parameters. As Kass and Raftery (1995) noted, the BIC rewards parsimony. For this application the BIC criterion will tend to favor models with fewer groups. Table 1 shows BIC scores for models with varying numbers of groups. For the Cambridge data the pat- tern is seemingly very distinctBIC appears to reach a clear maximum at three groups. I say "seemingly," because, without a concrete standard for calibrating the magnitude of the change in BIC, it is difficult to calibrate what constitutes a distinct maximum. Kass and Wasserman (1995) and Schwarz (1978) have provided such a standard. Let B tj denote the Bayes factor comparing models i and j, where for Table 1 BIC-Based Calculations of the Probability That a "j" Group Model Is the Correct Model for Different Numbers of Group s of Quadratic Trajectories Data set Cambridge No. of groups 2 3 4 5 BIC -1583.43 -1552.62 -1569.42 -1586.21 Probability correct model .00 1.00 .00 .00 Montreal BIC -7325.26 -7289.52 -7289.27 -7292.54 Probability correct model .00 .43 .55 .02 Note. BIC = Bayesian information criterion. this application model i might be a two-group model and j a three-group model. The Bayes factor measures the odds of each of the two competing models being the correct model. It is computed as the ratio of the probability of i being the correct model to j being the correct model. Thus, a Bayes factor of 1 implies that the models are equally likely, whereas a Bayes factor of 10 implies that model i is 10 times more likely t han/ Table 2 shows Jeffreys's scale of evi- dence for Bayes factors as reported by Wasserman (1997). 9 Computation of the Bayes factor is in general very difficult and indeed commonly impossible. Schwarz (1978) and Kass and Wasserman (1995), however, have shown that e BIC/ ~ BIC ' is a good approximation of the Bayes factor for problems in which equal weight is placed on the prior probabilities of models i and / On the basis of this approximation, for the Cambridge 8 The problem is most easily illustrated with an example. The likelihood ratio test is computed as minus two times the difference in the likelihood of the two nested models. This statistic is asymptotically distributed as chi-squared with degrees of freedom equal to the difference in number of parameters between the nested models. For our mixture model, the degrees of freedom is indeterminate, because a group can become superfluous in two ways. One is by the proportion of the population in that group, itj, approaching zero. Alternatively, the three parameters defining the trajec- tory of one group can collapse onto those for another group. What then is the appropriate degrees of freedomone or three? 9 Jeffreys was an early and very prominent contributer to Bayesian statistics. 148 NAGI N Table 2 Jeffreys's Scale of Evidence for Bayes Factors Bayes factor I nterpretati on B ;; <1/10 1/10 < B,j < 1/3 <B ij < 1 1 < BJJ < 3 3 <B]j< 10 B,, > 10 1/3 Strong evidence for model j Moderate evi dence for model j Weak evidence for model j Weak evi dence for model /' Moderate evi dence for model /' Strong evi dence for model ; Note. Adapt ed from "Bayesi an Model Sel ecti on and Model Av- er agi ng" by L. Wasserman. 1997, Wor ki ng Paper No. 666, Car ne- gie Mel l on U n i ve r s i t y , Depar tment of Statistics. Copyr i ght 1997 b y L. Wasser man. Adapted w i t h per mi s s i on. data the odds of the three-group model compared wi t h :he two- or four -gr oup model far exceed 1000 to 1. Thus, according to Jeffr eys' s scale t hi s is very strong evi dence in favor of the three-group model . 10 Schwarz (1978) and Kass and Wasserman (1995) have also provided a related metric for comparing more than two models. Let p f denote the posterior pr ob ab i l i t y that model j is the correct model, where in general j is greater than 2. They show t hat p : is rea- -onab l y approximated b y the fol l owi ng: (6) where BI C ma x is the ma xi mum B1C score of the mod- el s under consideration. Also shown in Table 1 are the probabilities, as c omp ut ed us i ng Equat i on 6, that the models w i t h var yi ng numb er s of groups are the true model. Wi th t h i s BIC-based pr ob ab i l i t y appr oxi mat i on, for t he Camb r i dge data the pr ob ab i l i t y of the thr ee-gr oup model is near 1. Consider now the BIC scores for the models esti- mated wi t h the Montreal data. The four -gr oup model has the best BIC score, but the pr ob ab i l i t y of its b eing the correct model, .55, is far less than 1. Although the four -gr oup model is far more l i kel y than the fi ve- gr oup model, the three-group model is a close com- peti tor. Its pr ob ab i l i t y is .43. On the b asis of Jeffr eys ' s scale, the four-group model is strongl y preferred to the fi ve-gr oup model . The odds ratio in favor of the four -gr oup model is 26 to l (i . e. , .55/.02). However , t h: edge compared w i t h the thr ee-gr oup model is s l i g h t wi t h an odds ratio of only 1.28 (i.e., .551.43). Thus, for models in whi ch each trajectory is described by a f ul l quadratic specification, the three- and four- group models fit the data about equal l y w el l . The Montreal anal ysi s i l l us t r at es a si tuati on in whi ch the mechanical application of the BIC model selection criterion does not resul t in an unamb i guous determi nati on of the "best" model. This b ri ngs me to the issue of determination of the appropriate order for model i ng the trajectory of each group that makes up the mi xt ur e. The best model w i l l not necessarily in- vol ve a mi xt ur e in whi ch all trajectories are of the same order (e.g., quadr ati c). In pr i nci pl e one could exha us t i vel y expl ore all possible comb i nati ons of or- ders for a model. In practice thi s approach is general l y not practical. For a four -gr oup model alone, there are 81 (i . e. , 3 4 ) possib le comb i nat i ons of trajectori es models, the numb er of possi b i l i ti es in a four-group model is 256. Further, a f ul l search woul d require est i mat i ng models over var yi ng numb ers of groups. To reduce the numb er of alternatives, an anal yst wi l l gener al l y have to use knowledge of the problem do- mai n to l i mi t the model search process. For instance, in the Montreal example sub stanti ve considerations suggest that an improved, more par si moni ous model shoul d restrict the numb er of parameters used to de- scribe the never and chronic trajectories. Description of a never t r aj ect or y does not r equi r e a t hr ee- parameter model : a zero-order model wi t h a "very negative" intercept should suffice. Indeed the large standard errors (not reported) associated wi t h the pa- rameters of the never trajectory strongly suggest that the trajectory is overparameterized. Table 3 shows the pr ob ab i l i ti es that var i ous forms of three- and four-group models are the correct model. We refer to models in whi ch the trajectories for all groups are quadr ati c as Typ e A models. Three- and four -gr oup models in whi ch the never trajectory is described by a zero-order model are labeled Typ e B models. Also shown are three- and four-group models in whi ch, in addi ti on, the chronic trajectory is de- scribed by a zero-order model (Typ e C). The i nfor - mat i on was pr ompted b y the t heor i es of Moffi t t (1993) and Patterson et al. (1989), which predict the exi stence of a s mal l group of chr oni cal l y anti soci al i ndi vi dual s . The BIC-based prob ab ility cal cul ati ons provi de strong support for the four-group, Type C model in Table 3. Compared wi t h the other opti ons 10 Usi ng the BIC criterion, D' Unger et al. (1998) and Roeder, Lynch, and Nagi n (in press) showed that a four- group model is best. In those analyses the model is fit based on the more general zero-inflated Poisson. ANALYZI NG DEVELOPMENTAL TRAJECTORIES 149 Table 3 BIC-Based Calculation of Probability of Correct Model for Models Combining Quadratic and Nonquadratic Trajectories: Montreal Data /W) (7) No. of groups \ '\ A .00 .00 Model type B .00 .07 c .00 .93 Note. : :1IC = Bayesi an i n f o r ma t i o n c r i t er i o n ; Model Type A = all trajectori es quadr at i c; Model Type B = one s i n gl e parameter t r a j ec t or y, r e ma i n i n g qua dr a t i c ; Model Type C = two s i ngl e pa- r ameter t r aj ect or i es , r e ma i n i n g qua dr a t i c . listed in the table, the probability of this being the correct model is .93. The closest competitor is the four-group. Type B model. However, the posterior pr ob ab i l i t y of it being the correct model is only .07. As the Montr eal anal ysi s i l l ustr ates, the determi na- tion o: the opti mal numb er of groups is not al ways a c l ea r -c ut process. Al t h o ug h Ba yes i a n s t a t i s t i c i a n s have made i mpor tant strides in devel opi ng theoreti- cal l y grounded criteria for model sel ecti on, the r esul t s are recent and not yet wi del y used. Further, as i l l us - trated by the Montreal example, the model search pro- cess may also be guided in part by nonst at i st i cal con- siderations. Thi s i njects a degree of sub ject i vi t y i nto model selection. St i l l , it is i mportant to recognize that use of the BIC criterion for model selection adds a very s i gni fi c a nt degree of st at i st i cal ob ject i vi t y to the determination of the opti mal numb er of groups. Thi s provi des an i mpor t ant check against the spurious i n- terpretation of random fl uc t ua t i ons in the data as re- fl ecti ng systematic patterns of b ehavi or . Calculation and Use of Posterior Gr oup Memb ership Prob ab ilities It is not possible to determine defi ni t i vel y an i ndi - vi dua l ' s group memb er shi p. However, i t i s possib le to cal cul ate the pr ob ab i l i ty of his or her memb er s hi p in the var i ous groups that make up the model. These pr ob ab i l i t i es , the poster i or p r ob a b i l i t i es of gr oup memb er shi p, are among the most useful products of t he group-b ased model i ng appr oach. Sp ec i f i c a l l y , based on the model coeffi ci ent estimates, for each i ndi vi dua l i the pr ob ab i l i t y of memb er shi p in group j is c a l c ul a t ed on the basis of the i n di vi dua l ' s l o n g i t u- di nal pattern of b ehavi or , K,. We denote t hi s prob ab il- ity by Pij'iy,-). It is computed as fol l ows : where P(Yj\j) is the estimated pr ob ab i l i t y of observing ;"s actual b ehavi or al trajectory, K,, gi ven member- s hi p in j, and fr ; is the esti mated proportion of the popul at i on in group / The qua nt i t y P(K, -l y) can be cal cul ated postestimation based on the ma xi mum- l i kel i hood esti mates of the trajectory parameters. The posterior pr ob ab i l i ty cal cul at i ons provi de the researcher wi t h an ob jecti ve b asis for assi gni ng i ndi - vi dua l s to the development trajectory group that best matches t hei r b ehavi or . I ndi vi dua l s can be assigned to the group to whi ch their posterior memb er shi p prob- a b i l i t y is l argest. On the b asis of t hi s ma xi mum pos- terior pr ob ab i l i ty as s i gnment r ul e, an i n di vi dua l from the Camb ri dge sampl e wi t h a cl uster of offenses in adolescence b ut none thereafter, i n al l l i kel i hood, w i l l be assigned to the adolescent l i mi t ed category. Al t er - n a t i ve l y , an i n d i vi d ua l who offends at a compar- at i vel y hi gh rate t hr oughout the observation period w i l l l i kel y b e assigned to the chroni c group. Table 4 shows the mean assi gnment prob ab i l i ty for the Montreal and Camb ridge data. For exampl e, in the Camb r i dge data the mean chr oni c group posteri or pr ob ab i l i t y for the 21 i n di vi dua l s assigned to t hi s group was very hi gh.95. The counterpart average for the 77 i n di vi dua l s assigned to the adolescent l i m- ited group is similarly large.94. For the Montreal data, cl assi fi cati on cer t ai nl y i s not so hi gh b ut s t i l l seems r easonab l y good. Across the four groups it ranges from .73 to .93. One i nfor mat i ve use of the posterior pr ob ab i l i t y- based cl assi fi cati ons is to create profi l es of the ''av- erage" i n d i vi d ua l fol l owi ng the trajectory character- ized by each group. Table 5 shows summar y stati sti cs on i ndi vi dua l characteristics and b ehavi or s of each group for both data sets. The profiles conform wi t h l ongs t a ndi ng f i n di n g s on r i sk and protective factors for anti soci al and problem b ehavi ors. In the Cam- bridge data those in the chr oni c gr oup, on average, were most l i kel y to have lived in a low-income house- hold, to have had at least one parent wi t h a c r i mi na l record, to have been sub jected to poor par enti ng, and to have di spl ayed a h i g h pr opensi ty to engage in r i s ky acti vi ti es. Conver sel y, the never group was lowest on these ri sk factors. The pattern is s i mi l a r for physi cal aggression. Members of the chr oni c group had the l ea s t w e l l -e duc a t e d pa r ent s a nd most f r e que n t l y scored in the lowest quar t i l e of the measured IQ dis- t r i b ut i on of the sampl e, whereas the nevers were hi gh- 150 NAGI N Table 4 Average Assignment Probability Conditional on Assignment by Maximum Probability Rule Group Assigned group Never Adolescent l i mi t ed Chronic Never Adolescent limited Chronic Never Low desister High desister Chronic Never .73 .00 .00 .00 Cambridge data .94 .03 0 Montreal data Low desister .27 .93 .09 .00 .06 .94 .05 High desister .00 .10 .84 .24 .00 .03 .95 Chronic .00 .00 .06 .76 est on these protective factors. Further, 90% of those in the chronic group failed to reach the ei ghth grade on schedule, and 13% had a j uveni l e record by age 18; onl y 19% of the nevers had fal l en b ehi nd by the ei ghth grade, and none had a juveni l e record. In be- tween are the low-level and hi gh-l evel desisters, who themsel ves order as expected on these characteristics and behaviors. Aside from pr ovi di ng a basis for group assignment, the posterior probabilities can provide an objective criterion for selecting subsamples for fol l ow-up data collection. For example, in an anal ysi s based on the Montreal data, Nagin and Tremblay (in press) found that i ndi vi dual s fol l owi ng the chronic physical ag- gression trajectory depicted in Fi gur e 3 displayed heightened l evel s of violence at age 18, controlling for chronic opposition and hyper acti vi ty. Conversely, it was found that chronic opposition and hyperactivity did not predict violence at age 18, control l i ng for chronic phys i cal aggression. Nagi n and Tremb lay concluded that chronic physi cal aggression was a dis- tinct risk factor for later violence. A fol l ow-up study that is currently under way at the time of this wr i ti ng was devised to examine whether self-regulatory pro- cesses were a possible explanation for the fi ndi ngs. The study involves a series of laboratory assessments of selected i ndi vi dual s in the Montreal study. I ndi - vi dual s were recruited for thi s study based on thei r posterior probabilites of membership in the various trajectory groups for physi cal aggression, opposition, Table 5 Group Profile Group Var i ab l e Low household income ( %) Poor parenti ng (%) High risk t aki ng (%) Parents wi th cr i mi nal record (%) Years of school mother Years of school father Low lQ a (%) Completed eighth grade on ti me (%) Juven i l e record (%) No. of sexual partners at age 17 (past year) Never Camb ri dge data 16.4 17.4 21.4 17.4 Montreal data Never 1 1 . 1 11.5 21.6 80.3 0.0 1.2 Adolescent 35.1 33.8 42.9 42.9 Low desister 10.8 10.7 26.8 64.6 2.0 1.7 limited Hi gh desister 9.8 9.8 44.5 31.8 6.0 2.2 Chronic 55.6 51.9 74.1 63.0 Chronic 8.4 9.1 46.4 6.5 13.3 3.5 1 Measured IQ score is in the lowest quar t i l e of the sample. ANALYZI NG DEVELOPMENTAL TRAJECTORIES 151 and hvper act i vi t y. The purpose was to recruit strate- gically to ensure that the sample included persons wi t h di sti nct comb i nati ons of development in exter- nal i zi ng b ehaviors (e.g., hi gh pr ob ab i l i ty of chronic phys i cal aggression but low pr ob ab i l i ty of chronic opposi ti on). The posterior prob ab i l i ti es provide objec- tive, quant i fi ed criteria for such strategic sampling. Statistically Linking Group Membership to Covariates What i ndi vi dual , f a mi l i a l , and envi r onmental fac- tors di s t i ngui s h the populations of the various trajec- tory groups? Are the factors consistent wi th received theory on developmental trajectories? What statistical procedures are most appropriate to test whether such factor s di s t i ngui s h among t he trajectory gr oups? Questons such as these fol l ow nat ur al l y from the i denti fi cati on of di sti nct groups of developmental tra- jectories. The profiles reported in Table 5 are a fi r st step in addressing these questions but indeed onl y a b egin- ni ng. First, the profiles are simply a collection of uni - variate contrasts. For the purpose of constructing a more parsimonious l i st of predictors or for causal in- ference, a mul t i var i at e procedure is required to sort out redundant predictors and to control for potential con- found-. Second, the profiles are based on group iden- ti fi cati ons that are prob ab ilistic, not certain. Conven- t i onal st at i st i cal methods to test for cross-group differences, such as F- and chi-square-based tests, as- sume no classification error in group identification (Roeder et al., in press). Thus, in general they are techni cal l y inappropriate in this setting." The mi xtur e model of developmental trajectories comprises two basic components: (a) an expected tra- jectory given memb ership in group j and (b) a prob- ab i l i t y of group memb ership denoted by IT ; . Thus far, the dKcussi on has focused on the former component. The di scussi on of the test for factors that di sti ngui sh groups turns our focus to the latter component. The profi l es in Table 5 show that the trajectory groups are composed of i ndi vi dua l s who di ffer in sub stanti al and predictable ways. Provision of the capacity to test for mal l y for whether and by what degree such factors di st i ngui sh the groups requires that the model be gen- eralized to allow itj to depend on characteristics of the i ndi vi dua l . Heretofore, ir ; has been described as the proportion of the population fol l owi ng each trajectory group /. Equi val ent l y, it can be described as the prob- ab i l i ty that a r andoml y chosen i ndi vi dua l from the population under study belongs to trajectory group j. By al l owi ng T^ to vary with i ndi vi dual characteristics, it is possible to test whether and by how much a specified factor affects prob ab ility of group member- ship controlling for the level of other factors that po- t ent i al l y affect TT ; . Let X j denote a vector of factors measuring i ndi - vi dual , fami l i al , or envi ronmental factors that poten- t i al l y are associated wi th group membership, and let Tfj(X j) denote the probability of membership in group j gi ven X j. For a two-group model the logit model is a nat ur al candidate for model i ng group membership probability as a functi on of * ,-. For this special case, we need only estimate Tr ; U,) for one group, say Group 1, because irfjc,) = 1 - TT,(J:,), where e *f> For the more general case, in whi ch there are more than two groups, the logit model generalizes to the mul t i nomi al logit model (Maddala, 1983): (8) where the parameters of the multinomial logit model, 0y, capture the impact of the covariates of interest, *,-, on probability of group membership. Wi thout loss of generality, 9 ; for one "contrast" group can be set equal to zero. The coefficient estimates for the remaining groups should be interpreted as measuring the impact of covariates on group memb ership relative to the contrast group. Table 6 illustrates the application of the extended model to the Montreal data. Results are reported for a four-group model in whi ch probability of group mem- bership is related to four variables: one parent havi ng less than a ninth-grade education, both parents havi ng less than a ninth-grade education, havi ng a mother who began childb earing as a teenager, and scoring in the lowest quarti l e of the measured IQ of the sampled boys. The first panel of Table 6 shows coefficient estimates and r statistics. For this example, the never- convicted group serves as the contrast group. Con- 1 ' Roeder et al. (in press) explored the impact of classi- fication error on statistical inference. One fi ndi ng is not surprising. When classification is hi ghl y certain, as is the case in the Cambridge data, errors in i nference usi ng con- ventional methods are smal l . 152 NAGI N Table 6 The Imp act of Parental Education. Low IQ, and Teen Onset of Motherhood on Group Membership Probabilities: Montreal Data Group Vari ab l e-condi ti on Never Mul t i n o mi a l l ogi t coeffi ci ent s Constant Low educati on" parent Low educati on" both parents Low IQ b
Teen mom c
Predicted membership probabilities based No r i sk factors Low educat i on" both par ents o n l y Low lQ h onl y Teen mom 1 ' o n l y Low desi ster Hi gh desister Chr oni c ( w i t h stati sti cs gi ven i n par entheses) 1. 04(6. 36) 0.44 ( 1 .46) 0. 18(0. 48) 0.07 ( 0. 22) 0.89 (2.03) on mul ti nomi al logit .20 . 15 . 15 .08 Low educat i on" b ot h parents and low IQ h and teen mom c .03 -0.02 (-0.10) 0.69(2.28) 0. 67( 1. 83) 0.85 (2.85) 1. 10( 2. 57) model coefficient .57 .51 .46 .58 .29 estimates .20 .29 .35 .25 .44 -2. 14 (-4.49) -0.1 9 (-0.30) 0. 98( 1. 73) 0. 92( 1. 83) 2.27 (3.89) .02 .05 .04 .09 .25 Parent has less than a n i n l h -g r a de educ a t i on. Measur ed I Q score i s i n the l owest qua r t i l e of the sampl e. Pa r t i c i p a n t ' s mot her b egan c h i l db e a r i n g as a t eenager . sider fi r st the parental l ow-educati on variab les. Rela- t i ve to the never group, h a vi n g one or more poorly educated parents does not s i gni fi c a nt l y increase the pr ob ab i l i t y of memb er shi p in the l ow-l evel desister gr oup (a = .05, one-tailed t est ). However, pr ob ab i l i t y of memb er shi p i n the hi gh-l evel near-desister and chroni c groups is s i g n i f i c a n t l y increased if both par- ent s are poorly educated. Low IQ has a si mi l ar impact on group memb er shi p pr ob ab i l i t y. Compared wi t h the never group, pr ob ab i l i ty of memb er shi p in the l ow- l evel gr oup is not s i gni fi c a nt l y affected, whereas prob- a b i l i t y of memb er shi p i n the two hi gher aggression trajectori es i s s i gni fi cant l y increased. Fi n a l l y , the teen n'iom var i ab l e is associated wi t h a s i gni fi c a nt increase i n al l group memb ershi p pr ob ab i l i t i es r el a t i ve to the never group. The second panel of Table 6 shows cal cul ati ons of the predicted prob ab ilities of group memb er shi p based on the coeffi ci ent esti mates. The c a l c ul a t i ons wer e performed b y s ub s t i t ut i ng the coeffi ci ent esti- mates i nto Equat i on 8 and then comput i ng gr oup memb er s hi p pr ob ab i l i ti es for assumed va l ues of x/. The cal cul at i ons show t hat if none of the ri sk factors are present, the pr ob ab i l i ty of the never or l ow-l evel groups is .77 and the pr ob ab i l i t y of the chroni c gr oup is onl y .02. The largest single factor affecti ng the chr oni c group memb er shi p pr ob ab i l i t y i s the teen mom var i ab l e. For the case in whi ch an i n di vi dua l has none of the other risk factors except for a mother who began chi l db ear i ng as a teenager, the comb ined prob- ab i l i t y of the never and the l ow-l evel desister group decl i nes to .66 and the pr ob ab i l i t y of the chr oni c group increases to .09. Al so reported is a hi gh-r i sk scenario in whi ch three risk factors are presentboth par ents poorly educated, low IQ. and teen mother- hood. In t hi s case the predicted pr ob ab i l i t y of the chroni c group is .25, an order of magni t ude larger than the no-ri sk-factor case. Al so, the predicted prob- a b i l i t y of the hi gh-l evel near-desister group of .44 is double the prob ab i l i ty for the no-r i sk-factor case. Si mul t aneous estimation of the group-speci fi c tra- jectory parameters and the group memb er shi p prob- ab i l i t i es condi ti onal on i n di vi dua l factors is easi l y ac- compl i shed wi t h the above-referenced software. The effi ci ency of esti mati on is also r el at i vel y good. Al- t hough esti mati on t i me w i l l depend on data set size, number of groups and covariates, and the speed of the computer , computati on time is generally 5 to 10 mi n. For anal yses in whi ch hypotheses are wel l formed, such comput at i on ti me is i nconsequenti al . However, in analyses in whi c h hypotheses are less w el l formed or analyses that are gener al l y expl oratory, much more effort w i l l b e put into mode! fi t t i ng. I n these ci r cum- stances comput at i on ti mes of 5 to 10 min per run may be very cumb ersome. Two al t er nat i ves for hi ghl y effi ci ent exploratory model fi t t i ng are recommended. Both are two-stage approaches wi t h the same fi r st stagei denti fi cati on ANALYZI NG DEVELOPMENTAL TRAJECTORIES 153 of the best fi t t i ng model in terms of numb er of groups. The fi r st stage i nvol ves a search for the best fi t t i ng model. This search is conducted wi t hout covariates di s t i ngui s hi ng group memb ershi p. One al ternati ve for the second-stage anal ysi s makes use of the group memb er shi p assi gnments based on the ma xi mum pos- terior pr ob ab i l i ty r ul e. Mul t i nomi a l logit models are then esti mated r el at i ng group as s i gnment to i ndi - vi dua l -l evel factors. A second al ter nati ve for the sec- ond-stage anal ysi s is to regress gr oup memb ership pr ob ab i l i ti es on the i ndi vi dua l -l evel factors of i nter - est. Both of these approaches for i dent i fyi ng candi - date factors that di s t i ngui s h gr oup memb er shi p are hi ghl y effi ci ent and easily conducted wi t h conven- ti onal stati sti cal software. Personal experi ence has shown that the resul ts of an exploratory anal ys i s con- ducted by ei ther of these methods provide a good gui de for speci fi cati on of the proper model in whi ch trajectories and the i mpact of covariates on memb er- shi p pr ob ab i l i t i es are j oi nt l y esti mated. 12 Comparative Advantages and Disadvantages of Group-Based Modeling In t h i s ar ti cl e I have described a group-based mod- el i ng strategy for a na l yzi ng developmental trajecto- ries. Al t er na t i ve methods gener al l y treat the popul a- tion di s t r i b ut i on of devel opment as conti nuous. The two l eadi ng exampl es of t hi s c ont i nuous model i ng strategy are hi er ar chi cal model i ng and l atent curve mo del i n g . ' 3 The group-based and conti nous modes of anal ys i s each have t hei r di st i nct i ve strengths. Growth cur ve model i ng, whether i n the hi er ar chi cal or l a t ent var i ab l e t r adi t i on, is designed to i dent i fy average de- vel opmental tendencies, to calib rate var i ab i l i t y ab out the average, and to expl ai n that var i ab i l i t y in terms of covariates of interest. By contrast, the mi xt ur e mod- el i ng strategy is designed to i dent i fy di st i nct i ve, pro- t ot ypal devel opmental trajectories wi t hi n the popul a- t i o n , to c a l i b r a t e t he p r o b a b i l i t y of p o p ul a t i o n memb er s fol l owi ng each such trajectory, and to relate those pr ob ab i l i t i es to covariates of i nterest. Ra udenb us h (in press) offered a val uab l e perspec- t i ve on the types of problems for whi ch these two broad cl asses of model i ng strategies are most sui t ab l e. He observed, "In many studi es it is reasonable to as- sume t hat all par t i ci pant s are gr owi ng according to some common funct i on but that the growth param- eters va r y in magni tude" (p. 30). He offered chi l dr en' s vocab ul ar y growth curves as an exampl e of such a growth process. Two di s t i nc t i ve features of such de- vel opment al processes are (a) they are gener al l y monotonicthus, the term growthand (b) they vary r egul ar l y wi t hi n the popul ati on. For such processes it is n a t ur a l to ask, "What is the t ypi cal pattern of growth wi t hi n the popul ati on and how does thi s t ypi - cal growth pattern vary across popul at i on members?" Hierarchical and l atent curve model i ng are specifi- cal l y designed to answer such a questi on. Raudenb us h (in press) also offered an exampl e of a devel opment al pr ocessnamel y, depr essi ont hat does not gener al l y change monot oni cal l y over ti me and does not vary r egul ar l y thr ough the popul ati on. He ob served, "It makes no sense to as s ume t h a t ever yone is i ncr eas i ng (or decr easi ng) in depres- sion. . . . many persons wi l l never b e hi gh i n depres- si on, others wi l l al ways b e hi gh, whi l e others w i l l become i ncr eas i ngl y depressed (p. 30). For problems such as t hi s , he recommended the use of a mul t i no- mi al -t ype method such as that demonstrated here. The reason for t hi s recommendation is that the develop- ment does not vary regul arl y across popul ati on mem- bers. Instead trajectories vary gr eat l y across popul a- ti on sub groups both in terms of the level of b ehavi or at the outset of the measurement period and in the rate of growth and decl i ne over time. For such prob lems a model i ng strategy designed to i dent i fy averages and expl ai n va r i a b i l i t y ab out that average i s far less us eful than a group-based strategy designed to i dent i fy dis- ti ncti ve cl uster s of trajectories and to cal i b r ate how characteri sti cs of i n di vi dua l s and thei r ci r cumst ances affect memb er shi p in these clusters. Summary and Discussion This ar ti cl e has demonstrated a semi par ametr i c, group-based approach for a na l yzi ng devel opmental 12 Al t hough ei ther of the two-stage procedures provides an effi ci ent approach for exploring the potenti al i mpact of covari ates on gr oup memb ershi p, they are not a s ub s t i t ut e for a fi nal anal ysi s based on the j o i n t l y esti mated model. Roeder, Lynch, and Nagi n (i n press) have found t hat the fi r s t described two-stage procedure in par t i cul ar tends to overstate the st at i st i cal s i gni fi cance of covar i ates on group memb er s hi p. The prob ab le reason is t ha t the procedure ig- nores the unc er t a i nt y about gr oup as s i gnment s . u Recent advances i n l at ent cur ve model i ng have adapted the convent i onal assumpt i on of a c ont i nuous di s t r i b ut i on of gr owt h curves to accommodate the group-based approach described here. For an excel l ent summar y of t h i s advance in l at ent cur ve model i ng, see Mut hen (i n press). 154 NAGI N trajectories. Techni cal l y, the model is simply a mi x- ture of prob ab i l i ty distrib utions that are s ui t ab l y speci- fied to describe the data to be anal yzed. Three such di st r i b ut i ons were i l l ust r at ed: the Poisson, whi ch is sui tab l e for anal yzi ng count data; the censored nor- mal , whi ch i s sui tab l e for a na l yzi ng psychometric scale data wi t h cl uster s of ob servati ons at the scale ma xi mum or mi n i mum or both; and the b i nary logit, whi c h is sui t ab l e for a na l yzi ng b i nar y data. For maxi - mum fl exi b i l i t y the parameters that characterize the trajectory of each group are specified to vary fr eel y across groups. This al l ows for sub stanti al cross-group differences in the shape of trajectories. The exampl es used to i l l us t r a t e the method were selected to demonstrate four of the method' s most important capab i l i ti es: (a) the capab i l i ty to i denti fy rather than assume di st i nct i ve developmental trajec- tories; (b) the c a pa b i l i t y to estimate the proportion of the popul at i on best appr oxi mated by the vari ous tra- jectori es so i dent i fi ed; (c) the capab i l i t y to relate the probability of memb er shi p in the var i ous trajectory groups to char act er i st i cs of the i ndi vi dua l and his or her circumstances; and (d) the capab i l i t y to use the posterior pr ob ab i l i ti es of group memb ershi p for vari- ous other purposes such as sampl i ng, creati ng profiles of group memb er shi p, or ser vi ng as regressors in mul - l i var i ate stati sti cal anal yses. The model i ng strategy described here has onl y re- cent l y been developed. Oppor t uni t i es for ext ensi on ab ound. Three ext ens i ons seem par t i cul ar l y worth- whi l e. Fur ther devel opment of approaches for decid- ing on the best fi t t i ng numb er of groups woul d be ext r emel y val uab l e. Al t hough BIC scores are ver y useful for this purpose, in addition it would be hel pful t o have measures of the goodness of fit b etween act ual and predicted trajectories. A second us eful extensi on i nvol ves the j oi nt esti mati on of tr ajector i es of differ- ent b ehavi or sfor exampl e, j oi nt est i mat i on of psy- chological wel l -b ei ng in chi l dhood and empl oyment s t at us over adul thood. This woul d provide the capac- i ty to exami ne the r el at i onshi p between the develop- ment patterns of di sti nct b ut l i kel y l i nked b ehavi ors. A thi rd val uab l e extensi on i nvol ves devel opi ng diag- nosti cs for cal i b r at i ng the degree to whi ch develop- ment al trajectories cl us t er i nt o di st i nct groups or var y r egul ar l y according to some speci fi ed parametri c dis- t r i b ut i on (e.g., mul t i var i at e nor mal ). Nagi n and Trem- b lay (in press) described another rationale that is not di scussed here for the group-based model i ng strat- egyto avoid maki ng strong and gener al l y untestab l e assumpti ons ab out the popul ati on di stri b uti on of de- velopmental trajectories. The semiparametric mi xture can be thought of as approxi mati ng an unspecified but l i kel y cont i nuous di stri b uti on of population heteroge- nei t y in developmental trajectories. In so doing, a standard procedure in nonparametric and semipara- metric statistics of approxi mati ng a conti nuous distri- b ut i on by a discrete mi xt ur e is adopted (Follman & Lambert, 1989; Heckman & Singer, 1984; Li ndsay, 1995). A di ffi cul t but val uab l e research effort woul d be to develop metrics for cal i b rati ng the degree to whi ch trajectory groups are actually closely approxi- mat i ng a specified continuous di str i b uti on. References Appel , T. A. (1987). The Cuvier-Geoffroy debate: French biology in the decades before Darwin. New York: Oxford Uni ver s i t y Press. Br yk, A. S., & Raudenb ush, S. W. (1987). Application of hi erarchi cal l i nea r models to assessing change. Psycho- logical Bulletin, 101, 147-158. Br yk, A. S., & Raudenb ush, S. W. (1992). Hierarchical lin- ear models for social and behavioral research: Ap p lica- tion and data analysis methods. Newb ur y Park, CA: Sage. Cameron, A. C., &Tr i vedi , P. K. (1986). Econometric mod- els based on count data: Comparisons and appl i cati ons of some estimators and tests. Journal of Ap p lied Economet- rics, I, 29-53. Cantor, N., & Genero, N. (1986). Psychiatric diagnosis and nat ur al categorization: A close anal ogy. I n T. Mi l l on & G. Kl er man (Eds.), Contemp orary directions in p sycho- p athology (pp. 233-256). New York: Gui l for d Press. Caspi, A. (1998). Personality development across the l i fe course. In W. Daom (Series Ed.) & N. Eisenb erg (Vol . Ed.), Handbook of child p sychology: Vol. 3. Social, emo- tional, and p ersonality develop ment (5th ed., pp. 311- 388). New York: Wi l ey. D' Unger , A., Land. K., McCal l , P., & Nagin, D. (1998). How many latent classes of del i quent/cri mi nal careers? Resul t s from mixed Poisson regression analyses of the London, Phi l adel phi a, and Racine cohorts studies. Ameri- can Journal of Sociology, 103, 1593-1630. Erdfelder, E. (1990). Deterministic devel opmental hypoth- eses, prob ab i l i sti c rules of mani festati on, and the anal ysi s of fi ni t e mi xt ur e di str i b uti ons. In A. von Eye (Ed.), Sta- tistical methods in longitudinal research: Time series and categorical longitudinal data (Vol. 2, pp. 471-509). Bos- ton: Academic Press Farri ngton, D. P., & West, D. J. (1990). The Cambridge study in del i nquent devel opment: A prospective l ongi t u- ANALYZI NG DEVELOPMENTAL TRAJECTORIES 155 di nal study of 411 males. In H. Kerner & G. Kaiser (Eds ), Criminality: Personality, behavior, and life his- tory i p p . 115-138). New York: Springer-Verlag. Fol l mar i , D. A., & Lambert, D. (1989). Generalized logistic r egr essi on by nonpar amet r i c mi xi n g . Journal of the American Statistical Association, 84, 295-300. Ghosh, J. K., & Sen, P. K. (1985). On the asymptotic per- formance of the log-likelihood rati o stati sti c for the mi x- ture model and related results. In L. M. LeCam & R. A. Ol shen (Eds.), Proceedings of the Berkeley Conference in Honor of Jerzy Neyman and Jack Kiefer (Vol . 2, pp. 789-806). Monterey, CA: Wadsworth. Goldstein, H. (1995). Multilevel statistical models (2nd ed. ). London: Edward Arnold. Greene. W. H. (1990). Econometric analysis. New York: Macmi l l an. Greene. W. H. (1991). LIMDEP User' s Manual (Version 6.0) ' computer software], Bel l por t, NY: Econometrics Software, Inc. Heckrriin, J., & Singer, B. (1984). A method for mi ni mi zi ng the i npact of di st r i b ut i onal assumpti ons in econometric models for duration data. Econometrica, 52, 271-320. Hol yoak, K., & Spellman, B. (1993). Thi nki ng. Annual Re- view of Psychology, 44, 265-315. Jones, B. L. , Nagi n, D. S., & Roeder, K. (1998). A SAS p rocedure based on mixture models for estimating devel- op mental trajectories (Wor ki ng Paper No. 684). Carnegie Mellon Uni ver s i t y, Department of Stati sti cs. Kandel . D. B. (1975). Stages in adolescent i nvol vement in drug use. Science, 190, 912-914. Kass, R. E., & Raftery, A. E. (1995). Bayes factor. Journal of th,- American Statistical Association, 90, 773-795. Kass, R E., & Wasserman, L. (1995). A reference Bayesian test :or nested hypotheses and its relationship to the Schwarz criterion. Journal of the American Statistical Association, 90, 928-934. Ker i b i n. C. (1997). Consistent estimation of the order of mixture models (Wor ki ng Paper No. 61) . Uni ver s i t e d' Evr y -Va l d' Essonne, Laboratorie Ana l ys e et Proba- b i l i t e. Lambeu, D. (1993). Zero-inflated Poisson regression, with an appl i cati on to defects in manufactur i ng. Technomet- rics, 34, 1-13. Land, K., McCall, P.. & Nagin, D. (1996). A comparison of Poisson, negative b i nomi al , and semi par ametr i c mi xed Poisson regression models wi th empi ri cal applications to cr i mi nal careers data. Sociological Methods & Research, 24, 387^40. Land, K., & Nagi n, D. (1996). Micro-models of criminal careers: A synt hesi s of the cr i mi nal careers and l i fe course approaches via semiparametric mixed Poisson models wi t h empirical applications. Journal of Quantita- tive Criminology, 12, 163-191. Lindsay, B. G. (1995). Mixture models: Theory, geometry, and ap p lications. Haywar d, CA: I ns t i t ut e of Mat hemat i - cal Statistics. Loeber, R. (1991). Questions and advances in the study of devel opment al pat hways . In D. Ci cchetti & S. Toth (Eds.), Models and integrations. Rochester Symp osium on develop mental p sychop athology ( Vol . 3, pp. 97-115). Rochester, NY: Uni ver si t y of Rochester Press. MacCal l um, R. C., Ki m, C., Mal ar key, W. B., & Kiecolt- Glaser, J. K. (1997). St udyi ng mul t i var i at e change usi ng mul t i l evel models and l atent curve models. Multivariate Behavioral Research, 32, 215-253. Maddala, G. S. (1983). Limited dep endent and qualitative variables in econometrics. New York: Cambridge Uni - versi ty Press. Markman, E. M. (1989). Categorization and naming in chil- dren: Problems of induction. Camb r i dge, MA: MIT Press. McArdle, J. J., & Epstein, D. (1987). Latent gr owth cur ves w i t h i n devel opmental str uctur al equati on models. Child Develop ment, 58, 110-133. Mer edi th, W., & Ti sak, J. (1990). Latent curve analysis. Psychometrika, 55(1), 107-122. Moffi t t , T. E. (1993). Adolescence-limited and life-course p er s i s t ent ant i s oci al b eha vi or : A devel op ment a l t ax- onomy. Psychological Review, 100, 674-701. Mut hen, B. O. (1989). Latent variab le model i ng in hetero- geneous popul ati ons. Psychometrika, 54(4), 557-585. Mut h en , B. O. (i n pr ess). Second-gener at i on s t r uc t ur a l equat i on model i ng wi t h a comb i nati on of categorical and cont i nuous latent variables: New opportunities for l atent cl ass/l atent cur ve model i ng. In A. Sayers & L. Col l i ns (Eds.), New methods for the analysis of change. Wash- i ngt on, DC: American Psychol ogi cal Associ ati on. Mut hen, B. O., & Curran, P. (1997). General l ongi t udi nal modeling of i ndi vi dua l differences in exper i mental de- si gn: A l at ent var i ab l e framework for anal ysi s and power es t i mat i on. Psychological Methods, 2, 371-402. Nagi n, D., Farri ngton, D., & Moffi tt, T. (1995). Life-course trajectories of di ffer ent types of offenders. Criminology, 33, 111-139. Nagi n, D., & Land, K. (1993). Age, cri mi nal careers, and popul a t i on heterogeneity: Specification and estimation of a nonparametric, mi xed Poisson model. Criminology, 31, 327-362. 156 NAGI N Nagi n, D., & Tremb lay, R. E. (in press). Trajectories of b oys' physi cal aggression, opposition, and hyper acti vi ty on the path to physi cal l y violent and nonvi ol ent j uveni l e del i nquency. Child Develop ment. Parzen, E. (1960). Modern p robability theory and its ap p li- cations. New York: Wi l ey. Patterson, G. R., DeBaryshe, B. D., & Ramsey, E. (1989). A devel opment al perspective on antisocial b ehavior. Ameri- can Psychologist, 44. 329-335. Rafter y, A. E. (1995). Bayesi an model selection in social research. Sociological Methodology, 25, 111-164. Raudenb ush, S. W. (in press). Toward a coherent fr ame- work for comparing trajectories of i n di vi dua l change. In A. Sayers & L. Col l i ns (Eds.), New methods far the analysis of change. Was hi ngt on DC: Ameri can Psycho- logical Association. Ri ndskopf, D. (1990). Testing devel opment al models us i ng l at ent class anal ysi s. In A. von Eye (Ed.), Statistical methods in longitudinal research: Time series and cat- egorical longitudinal data (Vol . 2, pp. 443^469). Boston: Academic Press. Roeder, K., Lynch, K., & Nagi n, D. (in press). Model i ng uncer tai nty in l atent class memb er shi p: A case study from cr i mi nol ogy. Journal of the American Statistical Associa- Schwarz, G. (1978). Esti mati ng di mensi ons of a model. An- nals of Statistics, 6, 461-464. STATA 4.0 [Computer software]. (1995). College Station, TX : STATA Corporation. Ti t t er i n g t o n , D. M., Smi t h , A. F. M., & Madov, U. E. (1985). Statistical analysis of finite mixture distributions. New York: Wi l ey. Tobin, J. (1958). Estimation of relationships for limited de- pendent variables. Econometrica, 26, 24-36. Tr emb l ay, R. E., Desmar ai s-Ger vai s, L., Gagnon, C., & Charleb ois, P. (1987). The preschool b ehavi or questi on- nai r e: St a b i l i t y of its factor structure b etween cul t ur e, sexes, ages and soci oeconomi c classes. International Journal of Behavioral Develop ment, 10, 467-484. Wasserman, L. (1997). Bavesian model selection and model averaging (Wor ki ng Paper No. 666). Carnegie Mellon Uni ver s i t y, Department of Statistics. Widiger, T. A., & Frances, A. (1985). The DSM-Hl person- al i t y disorders. Archives of General Psvchiatry, 42, 615- 623. Wi l l et t , J. B., & Sayer, A. G. (1994). Us i ng covariance structure anal ys i s to detect correlates and predictors of i n d i vi d ua l change over time. Psychological Bulletin, 116, 363-381. Appendix Derivation of Likelihood As described in the main text, the form of the like- lihood for each i ndi vi dual i is where P(Y-) is the unconditional probability of ob- servi ng i ndi vi dual /' s l ongi t udi nal sequence of behav- i oral measurements Y t , P'(Y,) is the pr ob ab i l i ty of Y/ jnven memb ershi p inj, and TT ; is the pr ob ab i l i t y of j. Thus, the l i kel i hood for the entire sample of N i ndi - vi dua l s i s N L = For gi ven j, conditional independence is assumed for the sequential realizations of the el ement s of Y/, y,, over the T ages of measurement. Thus, T where p i(y it ) is the prob ab i l i ty di stri b uti on funct i on of v,, gi ven membership in group / For the censored normal, /^'(v,-,) equal s ^mi n P X il (T ' \ a for S, nin := y,, < ax' and where cj> and <J> are, respectively, the densi ty functi on and c umul a t i ve di str i b uti on funct i on of a normal ran- ANALYZI NG DEVELOPMENTAL TRAJECTORIES 157 dom variab le wi th mean ^x it = (jj, + $\ Age,, + P^Agef, and standard deviation cr, and S min and S max are, respecti vel y, the scale mi n i mum and maxi mum. For t he Poisson-based model, /%,,)=- more general form of the Poisson-based model that makes use of the zero-inflated Poisson distrib ution. Fi nal l y, for the b i nary logit-based model, Al so, see Land, McCall, and Nagi n (1996), Nagi n and Land (1993), or Roeder, Lynch, and Nagi n (in press) for a deri vati on of the l i kel i hood for the st i l l Received June 15, 1998 Revi si on received October 16, 1998 Accepted Januar y 21, 1999 Members of Underrepresented Groups: Reviewers for Journal Manuscripts Wanted If you are interested in reviewing manuscripts for APA journals, the APA Publications and Communications Board would like to invite your participation. Manuscript reviewers are vital to the publications process. As a reviewer, you would gain valuable experience in pub lishing. The P&C Board is particularly interested in encouraging members of Underrepresented groups to participate more in this process. If you are interested in reviewing manuscripts, please write to Demarie Jackson at the address below. Please note the following important points: To be selected as a reviewer, you must have published articles in peer-reviewed journals. The experience of publishing provides a reviewer with the basis for preparing a thorough, objective review. To select the appropriate reviewers for each manuscript, the editor needs detailed information. Please include with your letter your vita. In your letter, please identify which APA journal you are interested in, and describe your area of expertise. Be as specific as possible. For example, "social psychology" is not sufficientyou would need to specify "social cognition" or "attitude change" as well. Reviewing a manuscript takes time. If you are selected to review a manuscript, be prepared to invest the necessary time to evaluate the manuscript thoroughly. Write to Demarie Jackson, Journals Office, American Psychological Association, 750 First Street, NE, Washington, DC 20002-4242.