Utilizing Trajectory Analyses to Refine Phenotype for Genetic Association: Conduct Problems and the Serotonin Transporter (5HTTLPR)

Abstract

Conduct disorder is a serious, relatively common disorder of childhood and adolescence. Findings from genetic association studies searching for genetic determinants of the liability toward such behaviors have been inconsistent. One possible explanation for differential results is that most studies define phenotype from a single assessment; for many adolescents conduct problems decrease in severity over time, while for others such behaviors persist. Therefore, longitudinal datasets offer the opportunity to refine phenotype.

Methods

We utilized Caucasians first assessed during adolescence from the National Youth Survey Family Study. Nine waves of data were utilized to create latent growth trajectories and test for associations between trajectory class and 5HTTLPR genotype.

Results

For the full sample, 5HTTLPR was not associated with conduct problem phenotypes. However, the short (s) allele was associated with chronic conduct problems in females; a nominally significant gender by 5HTTLPR genotype interaction was noted.

Conclusions

Longitudinal studies provide unique opportunities for phenotypic refinement and such techniques, with large samples, may be useful for phenotypic definition with other study designs, such as whole genome association studies.

INTRODUCTION

Childhood and adolescent conduct problems are common and are predictive of poor outcomes in early adulthood. In one study, researchers demonstrated that children with conduct problems early in life (ages 7–9) were significantly more likely than their peers to have a host of problematic outcomes as young adults including criminal behaviors, antisocial personality disorder, suicide attempts, risky sexual behaviors, teen pregnancy, and inter-partner violence (Fergusson et al., 2005). These associations persisted even after controlling for family socio-economic disadvantage, family instability, and child abuse. In adults, antisocial personality disorder (ASPD) (APA, 1994) is strongly associated with poor physical health/medical illness/accidents/highmortality/unnatural death (Goldstein et al., 2008; Dumais et al., 2005; Erbelding et al., 2004; Black et al., 1996; Martin et al., 1985), unstable interpersonal relationships (Afifi et al., 2006), difficulty maintaining employment (Grella et al., 2003), and imprisonment (Jordan et al., 1996; Fazel & Danesh, 2002).

The high cost to individuals and society from such antisocial behaviors/disorders underscores the importance of studies that attempt to understand etiology. Because antisocial behaviors are heritable (Rhee & Waldman, 2002), genetic association and linkage studies (e.g. Stallings et al., 2005; Dick et al., 2004) represent promising approaches to delineating the biological underpinnings of these disorders.

The serotonin system represents a logical target for genetic candidates of antisocial behavior phenotypes (conduct problems). This system has been linked with aggression through measurement of serotonin metabolites in the cerebrospinal fluid (CSF) (Brown 1979), measurement of central responsivity to administration of serotonin agonists (O’Keane 1992), and measurement of laboratory aggression after manipulation of serotonin levels (i.e. agonist administration or tryptophan depletion) (Cherek & Lane 1999; Dougherty 1999). The serotonin transporter gene (SLC6A4) contains a 43 base pair insertion/deletion polymorphism (5HTTLPR) in the 5′ regulatory region of the gene (Heils et al., 1996) that appears to be associated with variations in transcriptional activity; the long (l-allele) variant has approximately three times the basal activity of the short promoter (s-allele) (Lesch et al., 1996), although this is not a universal finding (Willeit et al., 2001). This functional polymorphism has been linked with markers of serotonin dysregulation such as CSF levels of a serotonin metabolite (Williams 2003), central responsivity to serotonin agonist administration (Reist 2001) and with functional brain differences (Pezawas 2005).

The results of association studies between 5HTTLPR and conduct problem phenotypes have been mixed. Several studies have found associations between the s-allele and conduct problem phenotypes (Hallikainen et al., 1999; Liao et al., 2004; Gerra et al., 2004; Gerra et al., 2005; Retz et al., 2004; Beitchman et al., 2006; Sakai et al., 2006; Haberstick et al., 2006). In other studies the l-allele has been associated with similar phenotypes (Twitchell et al., 2001; Zalsman et al., 2001), the association has varied by sex (Cadoret et al., 2003) or no association has been seen (Reist et al., 2001; Beitchman et al., 2003; Davidge et al., 2004; Sakai et al., 2007).

While one explanation for these varied results is that no association exists, another possibility is that utilizing a single time point for phenotypic assessment contributed to inconsistent results. Some individuals, who exhibit antisocial behavior in childhood and adolescence, remit or fall to sub-clinical thresholds, while others suffer lifelong severe antisocial tendencies (Moffitt, 1993). Genetic factors explain much more of population variance for antisocial traits in adulthood than in adolescence (Lyons et al., 1995); this raises the possibility that antisocial behavior persisting into adulthood is more heritable than antisocial behavior limited to adolescence. Utilizing trajectory analyses to refine measurement of conduct problems may help to delineate a more heritable phenotype and improve signal in genetic association analyses. Therefore, we utilized nine waves of data (age span 11–31) from the National Youth Survey Family Study (NYSFS), a longitudinal nationally representative sample, to demonstrate the use of trajectory analyses for conduct problems and to test whether trajectory class could be predicted on the basis of 5HTTLPR genotype. We hypothesized that more persistent antisocial behavior would be associated with genotype. Because of studies suggesting sex differences in conduct problems phenotypes (Zoccolillo, 1993), sex differences in the serotonin system (Jovanovic et al., 2008; Nishizawa et al., 1997; Jonsson et al., 2000) and a sex by genotype interaction in a previous association study of conduct problems (Cadoret et al., 2003), we also sought to explore whether findings varied by sex.

METHODS

Sample

We utilized the National Youth Survey Family Study (NYSFS), a multi-stage probability sample of households in the continental United States (Elliott et al., 1989). Original respondents were ages 11–17 years in 1976 at wave I and were from 1,044 households (n=1,725). All subjects provided written informed consent and the Human Subjects Committee of the Behavioral Research Institute, Boulder, Colorado (1976–1986) and the Human Research Committee of the University of Colorado (other years) approved all study protocols.

Interviews were conducted annually 1977–1981 (waves 1–5), and in three year intervals thereafter 1984, 1987, 1990 and 1993 (waves 6–9). Because of concern regarding cross-race variability in 5HTTLPR allele frequencies, the sample for the analyses presented in this report was restricted to Caucasian respondents. The sample was further restricted to one respondent per household to reduce confounds from non-independence between siblings (siblings share both a genetic background and rearing environment). When making these selections we preferentially chose respondents who were age 11–14 at wave 1 (n=467).

Conduct Problems

At each wave subjects completed a Self-Reported Delinquency (SRD) questionnaire, which inquires about past-year engagement in a range of delinquent behaviors (prevalence) and the number of times those behaviors were exhibited (frequency). It has been found that individuals generally provide reliable and valid responses to delinquency items in self-report surveys (Thornberry & Krohn, 2000). The reliability and validity of the NYS data, including interviews regarding self-reported offending, have been well documented. Test-retest reliabilities generally range from 70–90% and are high for scales involving serious crime such as felony assault 0.81, robbery 0.97 and felony theft 0.99 (Huizinga & Elliott, 1986) and validity of reports of serious offending have also been demonstrated (Elliott & Huizinga, 1989).

We utilized prevalence (whether a behavior was endorsed or not) and binned those questions into seven categories reflecting conduct disorder (American Psychiatric Association, 1994) criteria: (1) Initiating fights, (2) using a weapon, (3) stealing with confrontation, (4) forced sex, (5) vandalism, (6) breaking and entering, and (7) stealing without confrontation. This choice was made to help categorize the broad number of questions in a standardized way and to allow better comparison with other studies utilizing diagnostic criteria. As appropriate questions were not available to adequately represent all DSM-IV conduct disorder criteria, and in the interest of clarity and accuracy, the term “conduct problems” is utilized for phenotypic definitions (see below) rather than “conduct disorder”. For each criterion, if at least one within-category SRD question was endorsed (prevalence), that criterion was counted, such that our conduct problems phenotype had a range of possible scores of zero to seven. For some analyses, the conduct problems phenotype was dichotomized (see Data Analyses). Details on the specific questions included from each wave are provided in the Appendix.

Appendix.

Conduct Problems Criteria and Questions:

Criteria	Questions from SRD	Wave
		1	2	3	4	5	6	7	8	9
1. Initiating fights	Been in gang fights	y1v243	y2v280	y3v386	y4v415	y5v468	y6v405	y7v474	y8v515
	Hit or beat someone so badly probably needed a doctor									b34
2. Using a weapon	Aggravated assault - attacked someone with idea of seriously hurting/killing	y1v237	y2v274	y3v380	y4v398	y5v462	y6v399	y7v468	y8v509	b22
3. Stealing w/confrontation	Used force to get money/things from:
	a) Students	y1v273	y2v310	y3v418	y4v501	y5v504	y6v439
	b) Teacher or other adult	y1v275	y2v312	y3v420	y4v510	y5v506
	c) Other people	y1v279	y2v316	y3v424	y4v521	y5v510	y6v441
	d) General question							y7v504	y8v545	b58
	Used force to rob person/business						y6v465
4. Forced sex	Had/tried to have sexual relations with someone against their will	y1v271	y2v308	y3v398	y4v448	y5v482	y6v417	y7v486	y8v527	b38
5. Vandalism	Purposely damaged/destroyed property Belonging to:
	a) School	y1v215	y2v252	y3v358	y4v337	y5v436	y6v381
	b) Other	y1v217	y2v254	y3v360	y4v343	y5v438	y6v385	y7v454	y8v495
	c) Work						y6v383	y7v452	y8v493
	d) General question		y2v248							b8
6. Breaking and Entering	Broke into a building or vehicle (or tried to break in)	y1v289	y2v326	y3v434	y4v547	y5v520	y6v451	y7v512	y8v551	b64
7. Stealing w/o confrontation	Stolen/tried to steal at school	y1v287	y2v324	y3v432	y4v540	y5v518	y6v449
	Stolen/tried to steal (value <$5)	y1v235	y2v272	y3v378	y4v391	y5v460	y6v397	y7v466	y8v507	b20
	Stolen/tried to steal (value $5–50)	y1v285	y2v322	y3v430	y4v534	y5v516	y6v447	y7v510	y8v549	b62
	Stolen/tried to steal (value >$50)	y1v221	y2v258	y3v364	y4v361	y5v444	y6v389	y7v458	y8v499	b12
	Stolen/tried to steal (value >$100)									b78
	Knowingly bought/sold/held stolen goods	y1v223	y2v260	y3v366	y4v368	y5v446	y6v391	y7v460	y8v501	b14
	Stolen motor vehicle	y1v219	y2v256	y3v362	y4v349	y5v440	y6v387	y7v456	y8v497	b10
	Theft from family			y3v396	y4v440	y5v480	y6v413	y7v482	y8v523
	Theft from work						y6v415	y7v484	y8v525	b36
	Embezzled money or funds						y6v463	y7v524	y8v561	b70
	Picked someone’s pocket						y6v461	y7v522	y8v559
	Burglary of a building						y6v467

First table column denotes the 7 conduct problems criteria; Second table column denotes Self-Reported Delinquency (SRD) questions included within each conduct problems criteria; last 9 table columns indicate which SRD questions were included within a criterion by wave. Example, “y1v243” indicates that in wave 1 the “been in gang fights” question is included in the “initiating fights” criteria and “y1v243” identifies the wave 1 codebook variable; w/=with; w/o=without.

Genotyping

Buccal cell DNA samples were utilized for 5HTTLPR genotype determination using a modified protocol (Gelernter et al., 1999; Anchordoquy et al., 2003); isolation, extraction and polymerase chain reaction protocols have been previously described (Anchordquy et al., 2003). Primers for the 43 base pair insertion-deletion polymorphism in the 5′ promoter region were forward, 5′-GGCGTTGCCGCTCTGAATGC-3′ (fluorescently labeled); reverse, 5′-GAGGGACTGAGCTGGACAACCAC-3′ (Lesch et al., 1996). Fragments were amplified using a previously described, modified four-step Touchdown PCR method (Don et al., 1992). Utilizing an ABI Prism® 377 Genetic Analyzer (Foster City, California) and protocols supplied by the company, PCR fragments were run through a 12-cm polyacrylamide gel (4.25%) under denaturing conditions (6M urea). Allele calls were made by two investigators independently.

Data Analyses

Conduct Problems across study waves: We examined within-sex the number of waves that individuals endorsed at least one and separately, at least two, conduct problems criteria (range 0–9; Table 1). Next we examined the mean number of conduct problems at each wave (Table 2) and the proportion of individuals who endorsed at least one conduct problem criterion at each wave (Figure 1).

Table 1.

Number of waves subjects endorsed at least 1 conduct problem criteria and at least 2 conduct problem criteria

	≥One Conduct Problem Criteria			≥Two Conduct Problem Criteria
Waves	Male	Female	Total	Male	Female	Total
0	48	111	159	95	198	293
1	31	51	82	37	30	67
2	30	35	65	33	12	45
3	14	21	35	15	8	23
4	26	10	36	15	3	18
5	17	12	29	6	1	7
6	18	5	23	9	1	10
7	10	4	14	3	0	3
8	15	3	18	0	1	1
9	4	2	6	0	0	0
Total	213	254	467	213	254	467
Mean	3.08	1.51	2.76	1.47	0.43	0.90
SD	2.67	1.97	2.44	1.82	1.04	1.54

All data from the National Youth Survey Family Study; all subjects Caucasian, limiting to one respondent per household (n=467); cell entries represent counts.

Table 2.

Mean number of conduct problems criteria endorsed by wave

Wave	Mean	S.D.	Min	Max
1	0.666	1.086	0	6
2	0.578	0.996	0	6
3	0.576	0.987	0	6
4	0.520	1.043	0	6
5	0.424	0.862	0	6
6	0.319	0.686	0	4
7	0.253	0.611	0	4
8	0.169	0.476	0	4
9	0.133	0.419	0	3

All data from the National Youth Survey Family Study; all subjects Caucasian, limiting to one respondent per household (n=467).

Figure 1. Proportion endorsing at least one conduct problems criteria by wave and sex.

All data from the National Youth Survey Family Study; all subjects Caucasian, limiting to one respondent per household (n=467; 213 male).

Trajectory Analyses

We used latent class trajectory analysis using the procedure PROC TRAJ in SAS 9.0 (Jones et al. 2001). This procedure uses mixture modeling techniques to describe intra-individual change over time as a function of group membership (C_i) where group membership in the K^th class is a post-hoc determination based on similarity in initial levels of conduct problems (intercept) and change in conduct problems over time (slope). Briefly, this model is specified as:

$$f(y)=\sum _{k=1}^{K}Pr(C=k)Pr(Y=y\mid C=k)=\sum _{k=1}^K{p}_{k}f(y,{\lambda }_k)$$ (1)

In this model, the dependent variable is the observed value of conduct problems (0 or 1) at each wave of the study and the dependent variable (conduct problems) is described by y = (y1, y2, y3, ….y9). The primary goal of this model is to estimate p_k which is the probability of belonging to class k as a function of the parameters λ_k which change differentially over time. We use a logit specification of this general trajectory model to predict the likelihood of demonstrating at least one problem behavior at each wave. Because the literature has been inconsistent regarding s- vs. l-allele dominance (Williams et al., 2003; Lesch et al., 1996) and several prior significant association studies with related phenotypes have not made assumptions regarding dominance (Hallikainen et al., 1999; Gerra et al., 2004; Gerra et al., 2005; Sakai et al., 2006) we considered an additive model the most conservative approach. These parameter estimates are presented in Table 3.

Table 3.

Latent trajectory analysis: the influence of age, gender, and 5HTTLPR(s) on conduct problem behaviors over time.

	Group 2 vs. Group 1		Group 3 vs. Group 1		Group 4 vs. Group 1
	b	95% CI	b	95% CI	b	95% CI
Intercept	5.76	(2.76, 8.76)*	5.87	(2.95, 8.80)*	−0.31	(−3.79, 3.16)
Age	−0.34	(−0.55, −0.13)*	−0.38	(−0.58, −0.18)*	0.06	(−0.17, 0.28)
5HTTLPR(s)	−0.51	(−1.20, 0.19)	−0.45	(−1.12, 0.22)	−0.77	(−1.51, −0.03)*
Sex=female	−1.41	(−2.47, −0.35)*	−2.81	(−4.04, −1.58)*	−2.86	(−4.38, −1.33)*
5HTTLPR(s)* Sex	1.00	(0.07, 1.92)*	1.12	(0.08, 2.16)*	1.18	(−0.14, 2.51)

All data from the National Youth Survey Family Study; all subjects Caucasian, limiting to one respondent per household (n=467). The model including the 5HTTLPR(s)*Sex interaction (−2ll = 3560.26) significantly improved model fit (Chi-square = 8.38, df= 3, p < .038) compared to the model without it (−2ll = 3568.64). Group 1 = little or no problematic behaviors; Group 2 = low chronic offenders; Group 3 = high chronic offenders; Group 4 = adolescence limited. CI = confidence interval. 5HTTLPR(s) indicates number of s-alleles.

^*indicates p<0.05.

RESULTS

Table 1 shows the number of waves subjects endorsed at least one and separately, at least two conduct problems criteria. A total of 48 men (22.5%) and 111 women (43.7%) reported no problems at any time during the study. Overall, males averaged 3.08 waves where they endorsed at least one conduct problem criterion compared to an average of 1.51 for females (p<0.001); males averaged 1.47 waves with two or more problems compared with 0.43 waves for females (p<0.001).

Table 2 shows the mean number of conduct problem criteria endorsed at each wave of the study. Subjects endorsed on average less than one conduct problem criterion (0.67) at wave 1; by wave 9 that average had declined to 0.13.

Figure 1 shows graphically the proportion of subjects endorsing at least one conduct problems criterion by wave for males and females, separately. For both males and females, on average, there is a steady decline in the problematic behaviors over time. At wave 1, over 45% of male respondents and nearly 30% of female respondents reported at least one problem behavior but by wave 9 of the study, these numbers dropped to 15.5% and 7.1%, respectively.

Trajectory Analyses

The analyses so far do not differentiate between the paths that people may take to attain the same average history. Trajectory analyses look at different patterns of individual growth (change in problem behaviors over time) to model individual risk factors that predict membership in different trajectories. Results from the trajectory analyses identified four latent class trajectories, which parallel trajectories that have been found in previous studies of conduct problems or similar behaviors by Nagin and colleagues (Nagin & Land, 1993; Nagin et al., 1995). These trajectories are presented graphically in Figure 2. As shown earlier, nearly one-third of the sample (31.0 %) reported little to no problematic behaviors at any point in the study (group 1). A second group comprising 41.3% of the sample more closely resembles the average trajectories shown in Figure 1. These respondents (group 2) start out with a probability of endorsing at least one conduct problem criterion of roughly 0.30 but drop to roughly 0.05 by the final wave of the study, corresponding to Nagin and colleagues’ low chronic offenders. The third group comprises 16.9% of the respondents and this group demonstrates a high probability of problem behavior for nearly all nine waves of the study (Nagin and colleagues’ high chronic offenders). Finally, we identify a fourth group, similar to Nagin and colleagues’ adolescence-limited offenders that has a very high probability of problem behavior during the initial 3 waves of the study but their risk of problem behavior drops precipitously thereafter so that by wave 7, they have a risk of problem behavior that resembles groups 1 and 2.

Figure 2. Latent trajectory classes: probability of endorsing ≥ 1 conduct problem criteria across nine waves of the National Youth Survey.

All data from the National Youth Survey Family Study; all subjects Caucasian, limiting to one respondent per household (n=467). Group 1 = little or no problematic behaviors; Group 2 = low chronic offenders; Group 3 = high chronic offenders; Group 4 = adolescence limited.

Association Tests Across Latent Class Trajectories

Results from the trajectory regressions are presented in Table 3. According to these results, the 5HTTLPR s-allele is positively and significantly (b = 1.00, p < 0.036) associated with an increased risk of belonging to group 2 (average group) compared to group 1 (the non-offender group) for females but not males (note the significant 5HTTLPR*sex interaction in Table 3). This same pattern is evident in females for group 3 (persistently high risk of problem behavior) compared to group 1 and again, there is no effect of 5HTTLPR genotype for males. The genotype*gender interaction is not significant (b = 1.18, p<0.072) for group 4 compared to group 1. By calculating the exponential of the regression coefficients in Table 3 e^{(5HTTLPR(s) + 5HTTLPR(s)*sex)}, we show that each s-allele increases the odds for females group 2 vs. 1, and group 3 vs. 1, membership by 65% and 95%, respectively.

DISCUSSION

The current study provides two important contributions. First, it demonstrates the use of trajectory analyses as a phenotype for genetic association studies. Second, the analyses identified conduct problem trajectories from mid-adolescence through young adulthood similar to crime/delinquency trajectories found in other samples.

Genetic Association Analyses

We utilized longitudinal assessment to help refine our conduct problems phenotype by utilizing trajectory analyses that take into account an individual’s change in problem behavior over time; this approach might reasonably be expected to reduce measurement error seen with a single assessment. The longitudinal approach applied here supports a nominally significant sex by 5HTTLPR interaction. The 5HTTLPR s-allele was associated with an increased risk for conduct problems for females in this sample (group 2 vs. 1 and 3 vs. 1).

Comparison of group 4 vs. 1 (Table 3) does not support a significant sex by genotype interaction (p=0.07), but the parameter estimate is similar to those demonstrated for other comparisons. Given that group 4 is our smallest trajectory class, this non-significant finding may be explained by power limitations.

Disparate findings by sex for serotonergic regulation are not unprecedented. Sex differences in the serotonin system have been demonstrated for: 5HTT binding potential (Jovanovic et al., 2008), serotonin synthesis (Nishizawa et al., 1997) and levels of serotonin metabolite in cerebrospinal fluid (Jonsson et al., 2000). Sex differences have also been demonstrated in serotonergic responsivity (Manuck et al., 1998), and some work has suggested sex differences in the link between serotonergic responsivity and both impulsivity (Soloff et al., 2003) and aggression (Manuck et al., 1999). Regarding 5HTTLPR specifically, one laboratory study supported a sex-specific gene by environment interaction linking 5HTTLPR genotype and aggressive behaviors (Verona et al., 2006), and Cadoret et al., (2003) demonstrated sex differences in the association between 5HTTLPR and conduct problems. Such gender differences for 5HTTLPR are also supported for other phenotypes (Smits et al., 2008; Brummett et al., 2008a; Brummett et al., 2008b; Brummett et al., 2008c). Still the study results must be interpreted with caution; for example, the direction of our genotype by gender interaction is in the opposite direction of that seen by Cadoret et al., (2003).

Other important considerations when interpreting our findings should include: (1) We did not genotype the L_G allele, which is thought to be functionally similar to the s-allele (Smolka et al, 2007). (2) Our phenotypic definition represented only a sub-set of DSM-defined conduct disorder criteria. However, in some instances, this was advantageous as behaviors usually restricted to childhood/adolescence were excluded (i.e. truancy, run away/curfew violations). (3) Any association study that includes examination of several phenotypes or phenotype definitions, and several potential hypotheses, must include the caveat that nominal significance may underestimate the probability of false positive findings. As has been shown, even the highly plausible gene by environment interactions between this locus and environmental stressors are not well supported when rigorous standards of replication are applied (Munafo et al, 2008).

Trajectory analyses

Multiple researchers have utilized longitudinal datasets to examine conduct problem trajectories and early factors associated with trajectory classes (Shaw et al., 2003). Such studies have found that, for males, childhood conduct problems predict later violent and non-violent offenses in early and late adolescence (Broidy et al., 2003). Many studies have focused on demonstrating antisocial trajectories from childhood to adolescence (Brame et al., 2001; Broidy et al., 2003; Shaw et al., 2003); relatively fewer studies have examined trajectories from adolescence to adulthood, as we do here.

Comparisons with prevalence estimates for antisocial personality disorder in nationally representative studies (i.e. 3.6% - Compton et al., 2005) suggests that our persistent conduct problems trajectory (16.9% in figure 2) likely includes many people who would not meet diagnostic threshold for antisocial personality disorder diagnosis. Based on these trajectory analyses there exist individuals with less severe (i.e. sub-clinical) but persistent antisocial behavior.

Some previous work has also demonstrated that adult antisocial behavior (without childhood conduct disorder) occurs at high frequency in both restricted populations (i.e. female drug users) (Cottler et al., 1995) and in the general population (Compton et al. 2005). Inconsistent with this, we did not demonstrate an adult onset antisocial behavior trajectory (i.e. adult antisocial behavior). One possible explanation is that those with adult antisocial behavior often exhibit antisocial tendencies in childhood and adolescence that is sub-threshold for diagnosis of conduct disorder.

In figure 2 we demonstrate that most subjects (~72%) had a low probability (~30% or less) of endorsing any conduct problems at any wave, suggesting that the delinquent behavior considered here is not the norm. This is consistent with many previous studies (Odgers et al., 2007; Broidy et al., 2003).

Finally, we incorporated tests for genetic association with latent trajectory analyses and found that for females the number of s-alleles was related to increased risk for chronic conduct problems. More specifically for females the number of s alleles was associated with membership in group 2 (low chronic offenders) compared with group 1 (little or no problematic behaviors) and group 3 (high chronic offenders) compared with group 1 (little or no problematic behaviors).

Our trajectory results should be considered within the context of the following limitations. Compared to other established methods, trajectory analysis is a relatively recent arrival where some of the first established models appeared in the early 1990s (Nagin and Land, 1993). As such, there is no established “gold standard” for growth modeling. An alternate method is provided by Muthén & Muthén (2007) using their software Mplus. An important advantage to their method is the ability to model growth factor variances and covariances. That is, heterogeneity in the development of conduct problems may be described as discrete classes (the PROC TRAJ approach) but there may also be important variation within these discrete classes (the Mplus approach) that provides additional leverage to test competing theoretical models. Future research in this area may consider extending this work to include both between and within class variation.

Future Directions

We conducted trajectory analyses to refine phenotype for genetic association analyses. A cross-sectional measurement of antisocial behavior in a community sample of adolescents would not allow easy identification of desisting and persisting antisocial cases. Given that some individuals meeting criteria for conduct disorder will upon testing one-week later not meet criteria (Cottler et al., 1998), and that individuals with small incentives can markedly alter their scores on measures of psychopathy (Rogers et al., 2002), measurement over multiple time points allows one approach to improve phenotypic definition of conduct problem phenotypes. Although examining a single candidate gene as we do here represents a limited, hypothesis-driven approach to genetic association, the same approach to refine phenotypes could be utilized for genome wide association studies.