Dimensionality and Genetic Correlates of Problem Behavior in Low-income African American Adolescents

Abstract

Objective

Researchers have long observed that problem behaviors tend to cluster together, particularly among adolescents. Epidemiological studies have suggested that this covariation is due, in part, to common genetic influences, and a number of plausible candidates have emerged as targets for investigation. To date, however, genetic association studies of these behaviors have mostly focused on unidimensional models of individual phenotypes within European American samples.

Method

Herein, we compared a series of confirmatory factor models to best characterize the structure of problem behavior (alcohol and marijuana use, sexual behavior, and disruptive behavior) within a representative community-based sample of 592 low-income African American adolescents (50.3% female), aged 13 to 18. We further explored the extent to which three genes previously implicated for their role in similar behavioral dimensions (CHRM2, GABRA2, and OPRM1) independently accounted for variance within factors specified in the best fitting model. Supplementary analyses were conducted to derive comparative estimates for the predictive utility of these genes in more traditional unidimensional models.

Results

Findings provide initial evidence for a bifactor structure of problem behavior among African American adolescents, and highlight novel genetic correlates of specific behavioral dimensions otherwise undetected in an orthogonal syndromal factor.

Conclusions

Implications of this approach include increased precision in the assessment of problem behavior, with corresponding increases in the reliability and validity of identified genetic associations. As a corollary, the comparison of primary and supplementary association analyses illustrates the potential for overlooking and/or overinterpreting meaningful genetic effects when failing to adequately account for phenotypic complexity.

By the end of high school, nearly four out of five adolescents in the U.S. have consumed alcohol, almost half have used marijuana, and nearly one quarter have had four or more sexual partners (Centers for Disease Control, 2012). In addition, up to 25% of adolescents in grades seven through twelve have reported recently engaging in some type of disruptive behavior (e.g., shoplifting, skipping school, and damaging property; Bartlett, Holditch-Davis, & Belyea, 2007). Although these specific risk-taking behaviors are all somewhat developmentally normative, the relative pervasiveness is still quite alarming, particularly when considering that the potential long-term consequences to self and society include substance use disorders, antisocial behaviors leading to incarceration or injury, sexually transmitted diseases, and unwanted pregnancies.

Perhaps even more concerning is the frequency with which these behaviors tend to co-occur in adolescents (McGee & Newcomb, 1992), and the increased likelihood of risk that accompanies each additional occurrence (McGue & Iacono, 2005). These issues have received a great deal of attention in the nearly four decades since Jessor & Jessor (1977) initially posited that frequently observed co-occurrences among a set of problem behaviors (illicit drug use, problem drinking, delinquent behavior, and precocious sexual intercourse; Donovan & Jessor, 1985), reflects an underlying problem behavior syndrome (PBS). Findings from a recent latent class analysis of health risk behaviors within a representative sample of early-adolescents illustrates the continued prevalence of this phenomenon, suggesting that one in four exhibits a behavioral profile characterized by high probabilities of substance use, delinquency, and precocious sexual behavior (Hair, Park, Ling, & Moore, 2009). Moreover, another recent study suggests that once PBS is manifest, one’s ability to desist from specific behaviors becomes increasingly more difficult (Monahan, Rhew, Hawkins, & Brown, 2013).

Notably, Jessor’s operationalization of PBS as problematic, socially undesirable or unconventional behaviors that cause concern and/or elicit social control responses (1977) is generally synonymous with what others have referred to as externalizing behavior or externalizing symptomatology (Seglem, Torgersen, Ask, & Waaktaar, 2015). All of these are broadband constructs reflecting shared variance across a set of more narrowly defined, but threshold free behavioral continua. In contrast, both externalizing psychopathology and externalizing spectra are generally operationalized as overarching constructs consisting either exclusively of disordered behaviors, or some combination of clinically derived disorders and symptoms (Tackett, 2010). Moreover, although models of PBS have included various constellations of behavior over the years, the four original dimensions have remained constant (e.g., see Donovan, Jessor, & Costa, 1991).

The empirical validation of PBS has generally relied on confirmatory factor models in which composite scores for the distinct domains of problem behavior simultaneously load onto a single common factor. As such, the factor reflected shared variability across a broad range of externalizing behaviors, and its structure has been replicated in adolescents and young adults across both U.S. (Donovan & Jessor, 1985; Donovan, Jessor, & Costa, 1988) and cross-national samples (Vazsonyi et al., 2010). Other studies utilizing multiple indicators of domain-specific externalizing problems in place of summary scores have suggested that a higher order structure significantly improves model fit (McGee & Newcomb, 1992; Newcomb et al., 2002). That is, at the item level, behavioral indicators are more highly correlated within versus between domains, so those inter-item relationships are captured via domain-specific first order confirmatory factors, and covariance among the first order factors is accounted for by a second order factor representing the syndromal nature of problem behavior. While much of the hierarchical work on problem behavior has been conducted across diverse samples, this higher order factor structure has also been validated within a sample of African American (AA) adolescents (Resnicow, Ross-Gaddy, & Vaughan, 1995).

Overall, the literature implies that beyond accounting for the shared variance across forms of problem behavior, domain specific factors can also yield important information, suggesting the need to explore both syndromal and unique aspects (Allen, Leadbeater, & Aber, 1994). Toward this end, a more comprehensive and systematic evaluation of the underlying factor structure may help clarify the etiology of problem behavior (Reise, Morizot, & Hays, 2007), as our present understanding is largely contingent upon how these constructs have been operationally organized.

Though principally rooted in a multifaceted system of psychosocial factors (Jessor, Van Den Bos, Vanderryn, Costa, & Turbin, 1995), Jessor and colleagues explicitly denote genetics as an important explanatory domain with respect to problem behavior (Jessor, 1987; Jessor, Donovan, & Costa, 1991). Consistent with this perspective, twin and family studies indicate shared sources of genetic liability across a wide range of externalizing problems/disorders. At the latent level, variance decomposition analyses in twin samples provide evidence for common genetic etiologies with respect to alcohol dependence (AD) and various disruptive behavior disorders, including oppositional defiant disorder (ODD), conduct disorder (CD), and antisocial personality disorder (ASPD) (Kendler, Prescott, Myers, & Neale, 2003; Markon & Krueger, 2005), for alcohol and marijuana use (Sartor et al., 2010), and across adolescent conduct problems and risky sexual behavior (Verweij, Zietsch, Bailey, & Martin, 2009). Moreover, genome-wide linkage analyses intimate regions of shared liability for substance dependence (SD) and ASPD (Stallings et al., 2005), and for AD and CD (Dick et al., 2004).

In the present era of large-scale gene finding efforts, genome-wide association studies (GWAS) have provided increasingly detailed information, allowing us to more precisely identify specific candidates for association, from whole genes down to the level of single nucleotide polymorphisms (SNPs). As a result, genes initially identified as candidates for specific disordered behaviors, such as AD (CHRM2, GABRA2) and opioid dependence (OPRM1), have subsequently been implicated as playing a broader role in relation to discrete non-disordered forms of substance use (Corley et al., 2008; Dick et al., 2008; Edenberg et al., 2004; Ehlers, Lind, & Wilhelmsen, 2008; Ittiwut et al., 2012; Kendler et al., 2011; Lind et al., 2008; Luo, Kranzler, Zhao, & Gelernter, 2003; Smelsen et al., 2012; Zhang et al., 2006), disruptive behavior (Dick et al., 2006; Dick et al., 2008), and inhibitory control (Dick et al., 2008; Hendershot, Bryan, Feldstein Ewing, Claus, & Hutchison, 2011), as well as generalized externalizing behavior (Dick, 2007; Dick et al., 2008; Dick et al., 2009; Latendresse et al., 2011).

Unfortunately, despite mounting evidence suggesting (1) shared genetic variance across many common externalizing constructs, and (2) improved measurement and increased validity of externalizing phenotypes and resulting associations when explicitly modeling known multidimensionality, all attempts to identify specific genetic variants associated with externalizing problems/disorders to date appear to rely on independent analyses of unitary phenotypes derived to reflect unidimensional disorders, behaviors, or broad theoretical constructs. Therefore, any meaningful covariation amongst related externalizing dimensions might have been misallocated in extant studies of association between genotypes and unconditional externalizing phenotypes.

In addition, genetic association studies are often conducted within racially homogeneous subgroups. Thus, many studies test solely for associations within the largest group in their sample, typically European Americans (EA) within U.S. samples, or simply include race (a frequently used proxy for differentiating discrete ancestral populations; Risch, Burchard, Ziv, & Tang, 2002) as a covariate in pooled samples analyses (cf. Ittiwut et al., 2012; Kendler et al., 2011; Smelsen et al., 2012). As such, there has been little research specifically exploring genetic associations among AA youths, despite the fact that they experience a number of disparities related to the consequences of problem behaviors.

Herein, we have compared a series of confirmatory factor models characterizing the structure of problem behavior within a representative community-based sample of low-income AA adolescents. The relative value of a superior phenotypic measurement model was assessed by separately regressing factor scores derived from the best fitting model and traditional unidimensional confirmatory factor models commonly used in genetic association studies onto a common set of SNPs. The present study focused on variants spanning three genes previously implicated in association with general and/or specific dimensions of externalizing behavior/disorder: CHRM2, GABRA2, and OPRM1. Lastly, converging/diverging results across distinct models, and broader implications for future genetic association studies with complex behaviors were discussed.

Method

Participants and Procedures

Sample selection and recruitment

The Gene, Environment, Neighborhood Initiative (GENI) constitutes a representative community sample of 592 adolescents aged 13 through 18 and their primary caregiver (see Mustanski et al., 2013). GENI participants were recruited from the Mobile Youth Study (MYS), a community-based, multiple cohort longitudinal study with annual data collection. MYS has been described in detail elsewhere (Bolland, 2007; Park, Lee, Bolland, Vazsonyi, & Sun, 2008) and consists of adolescents and caregivers from predominantly AA, very low-income inner-city neighborhoods in the Mobile, Alabama metropolitan statistical area (United States Census Bureau, 2002: e.g., > 92% AA, > 27% poverty rate among youth). Previous comparisons between MYS participants who did and did not participate in the GENI study suggest no significant differences with respect to risk behaviors, and only trivial demographic differences (Byck, Bolland, Dick, Ashbeck, & Mustanski, 2013).

Participant interviews

The Institutional Review Boards at the University of Alabama, Northwestern University, and Virginia Commonwealth University approved procedures for this study. Participation in GENI involved an approximately two and a half hour interview for both the adolescent and his/her caregiver. Interviews were conducted between March 2009 and October 2011. Written parental consent and youth assent were obtained. Caregivers and adolescents were compensated for participation.

Biological data collection

DNA was obtained via saliva sample using Oragene collection kits, under the supervision of a specially trained interviewer. Samples were coded anonymously and mailed to the laboratory of Dr. Brien Riley at the Virginia Institute for Psychiatric and Behavioral Genetics (VIPBG) at Virginia Commonwealth University, where DNA extraction was performed. In total, DNA samples were obtained from 573 adolescents (50.3% female), representing 96.8% of the total GENI sample. The racial composition within this group was derived via interviewer administered Family Demographic Caregiver Reports, wherein each adolescent was characterized as a member of one of six discrete “Race/Ethnicity” groups: Black (N=566), White (N=2), Mixed Race (N=5), Asian (N=0), Hispanic/Latino (N=0), and Other (N=0).

Genotyping

From these samples, a total of 69 single nucleotide polymorphisms (SNPs) were genotyped across CHRM2 (25), GABRA2 (20), and OPRM1 (24) (see Table 1). Genotyping was conducted at the VIPBG using fluorescence polarization detection of template-directed dye-terminator incorporation with appropriate AcycloPrime SNP detection kit for specific polymorphisms (PerkinElmer, Boston) and an automated allele-scoring platform (Van den Oord, Jiang, Riley, Kendler, & Chen, 2003). First, sixteen SNPs previously typed in EA samples by our group were force-included in the tagging set for the purpose of future comparison. Next, TaqMan assays were not available for three previously typed SNPs, so SNPs in complete linkage disequilibrium (LD) with those EA typed SNPs (i.e., R² = 1.0) were selected as proxies. Finally, to complete LD tagging across each of these previously identified regions of genetic risk for externalizing phenotypes, a supplementary set of 50 SNPs was selected from among those identified in the Nigerian Yoruba population of the HapMap project, thus capturing any additional genetic variability existing within individuals of more recent African descent.

Table 1.

Targeted Single Nucleotide Polymorphisms (SNPs)

CHRM2			GABRA2			OPRM1
SNP a	Location b	MAF c	SNP a	Location b	MAF c	SNP a	Location b	MAF c
rs10249472	136049718	.26	rs497068	45945434	.35	rs12205732	154400626	.10
rs6977141	136091394	.35	rs548583	45958101	.32	rs1799971	154402490	.02
rs10954546	136102622	.42	rs526805	45958148	.36	rs553202	154406510	.44
rs11984255	136141629	.42	rs537134	45966025	.25	rs524731	154416785	.12
rs4407820	136161629	.42	rs279871	46000490	.23	rs3778150	154425351	.19
rs6970010	136166257	.37	rs1808851	46006204	.28	rs10457090	154432766	.07
rs12333547	136182323	.25	rs279858	46009350	.22	rs9478503	154434368	.21
rs6967120	136186671	.30	rs279851	46014748	.44	rs589046	154434831	.44
rs10246433	136191614	.38	rs1561774	46015347	.35	rs3778157	154447394	.19
rs978437	136264718	.23	rs11941602	46018451	.21	rs10485057	154454948	.18
rs7782965	136274673	.23	rs279843	46019961	.43	rs562859	154456266	.40
rs7800170	136274860	.34	rs279845	46024480	.26	rs511420	154465725	.21
rs1455858	136282243	.19	rs183960	46025645	.33	rs9322447	154466013	.35
rs1378646	136285541	.23	rs1440130	46028010	.29	rs681243	154469433	.48
rs1824024	136294234	.32	rs279826	46028966	.48	rs504932	154472161	.18
rs324582	136301147	.26	rs279828	46029567	.27	rs512053	154481209	.01
rs2061174	136311940	.48	rs279836	46033827	.23	rs658156	154483218	.42
rs324650	136344201	.32	rs1442060	46060824	.41	rs645027	154483822	.09
rs7779921	136365936	.21	rs1442062	46071833	.30	rs644261	154483943	.14
rs7796601	136444656	.22	rs11503014	46085622	.25	rs613341	154484971	.09
rs832990	136491912	.32				rs616585	154485574	.27
rs11971038	136492737	.31				rs10485058	154486907	.06
rs832988	136492957	.38				rs497315	154489237	.24
rs833018	136522917	.26				rs678122	154491795	.42
rs833016	136525108	.36

Note

^amarkers shown as rs numbers from the dbSNP database;

^blocations are derived from estimates of chromosomal position in the dbSNP database (build 36.3) and expressed in base pairs;

^csample-specific minor allele frequencies.

Genotyping success rates for SNPs in these three genes ranged from 98.5–98.9%, and duplicate genotyping produced concordance rates of 100%. By way of quality control, all genetic data were subjected to a set of inclusionary thresholds. First, individual DNA samples yielding a gene-wise genotyping success rate of less than 80% were deemed unreliable and removed from consideration for inclusion in the analytic dataset. For all remaining data, any individual SNP with a sample-wise genotyping success rate of less than 80% was also excluded from the analytic dataset. Of the 573 participants from whom DNA data were collected, less than 1% of individuals failed to meet the first threshold for any given gene. Moreover, none of the markers genotyped were excluded on the basis of the second criteria. Finally, no SNPs were eliminated from analyses for significantly deviating from within-race calculations for Hardy-Weinberg equilibrium (p ≤ .001). All SNP chromosomal positions and allele identities are depicted with respect to the genomic (+) strand.

Both allelic frequencies and LD structure have been shown to differ substantially between populations of discrete ancestral descent (Lonjou et al., 2003). To guard against bias resulting from population stratification we limited all genetic analyses to the subsets of caregiver-reported “black” adolescents surviving quality control procedures for each gene: N’s = 561 (CHRM2) and 562 (GABRA2, OPRM1). In addition, reported race was validated via principal components analysis using the Admixture software package (Alexander, Novembre, & Lange, 2009), where the average estimated European ancestry among “black” adolescents was less than 20%, and none reached 50%. Haploview (Barrett, Fry, Maller, & Daly, 2005) was used to estimate LD across the full set of genotyped SNPs within each gene (Figure 1). This inter-SNP correlation reflects the degree to which analyses with individual markers represent unique versus non-independent tests of association. SNPSpD (Nyholt, 2004), a web-based test, was used to derive Bonferroni-like correction factors for multiple testing that account for the number of SNPs genotyped and the LD structure between them. In turn, the effective numbers of independent marker loci for our analyses were estimated to be 15 in CHRM2 and OPRM1, and 10 in GABRA2, yielding adjusted significance thresholds (α/n) of .003 and .005, respectively. Based on these critical values, we had adequate statistical power (i.e., .80) to detect genetic main effects in traditional unidimensional factor models of size R² = .02−.025. Moreover, the multidimensional measurement model employed in the present study had the potential to increase this power to detect genetic associations by 20–99% (van der Sluis, Verhage, Posthuma, & Dolan, 2010).

Figure 1.

LD structure within the GENI sample for (a) CHRM2, (b) GABRA2, and (c) OPRM1, with individual coefficients reflecting pairwise R² values.

Measures

Alcohol use

Use of alcohol over the past 12 months was assessed via four self-report items. First, a single dichotomous item inquiring about whether participants had consumed six or more drinks during that period (0 = no, 1 = yes). This item was taken from the Diagnostic Interview Schedule for Children version 4.0 (C-DISC; Shaffer, Fisher, Lucas, Dulcan, & Schwab-Stone, 2008), a widely used instrument for assessing psychiatric diagnoses among adolescents that is administered as a computerized, structured interview by trained lay interviewers (for details of administration, see Mustanski et al., 2013). In addition, three ordinal response items from the AIDS-Risk Behavior Assessment (ARBA; Donenberg, Emerson, Bryant, Wilson, & Weber-Shifrin, 2001) were used to assess the frequencies of use and heavy drinking in the last 12 months (0 = zero days, 1 = one or two days, 2 = monthly or less, 3= two to three days per month, 4 = one or two days a week, 5 = three to five days per week, 6 = every day or almost every day), as well as the average quantity consumed (0 = zero drinks, 1 = one drink, 2 = two drinks, 3 = three drinks, 4 = four drinks, 5 = five drinks, 6 = six or more drinks).

Marijuana use

Use of marijuana during the past 12 months was assessed via two dichotomous items (0 = no, 1 = yes) from the C-DISC: “used”, and “used six or more times”. In addition, overall frequency of marijuana use within the past year was assessed via self-report using a single item from the ARBA. Responses for this item were coded as follows: 0 = zero times, 1 = once or twice, 2 = monthly or less, 3 = two to three times per month, 4 = one or two times a week, 5 = three to five times per week, 6 = every day or almost every day.

Sexual behavior

Lifetime sexual behavior was assessed via five self-reported items from the HIV-Risk Assessment for Sexual Partnerships (H-RASP; Mustanski, Starks, & Newcomb, 2012), a computerized self-administered interview that was adapted from the ARBA to better assess sexual behavior and associated situational and contextual variables at the level of the sexual partnership. First, three items inquired about whether participants had ever engaged in oral, vaginal, and anal sex (0 = no, 1 = yes). Two additional items asked about the number of lifetime vaginal/anal sex partners, and the number of those partners with whom no condom was used during sex. Both of these items used an open response format, and participants who indicated abstinence on initial inquiries regarding vaginal and anal sex were coded zero.

Disruptive behavior

The Youth Self Report (YSR; Achenbach, 1991) was used to assess disruptive behaviors. This battery consists of 112 items, for which participants indicate whether each is “not true”, “somewhat or sometimes true”, or “very or often true” (scored 0, 1, and 2, respectively) with respect to their behavior over the last six months. Of these items, five are considered to reflect a DSM-like oppositional defiant problems subscale (e.g., “I disobey my parents”), and another 15 items are considered to reflect a DSM-like conduct problems subscale (e.g., “I cut classes or skip school”).

Analytic Plan

A series of unconditional confirmatory factor models (see Figure 2) were fit in Mplus (Muthén & Muthén, 1998–2010) using the weighted least squares mean and variance adjusted estimation procedure. One-factor, four-factor, second-order factor, and bifactor models were estimated sequentially, with the four-factor model estimated both with and without covariances between the four factors reflecting problem-specific dimensions. Since the standard chi-square fit statistic (χ²) is sensitive to sample size (Bentler & Bonnet 1980), a relative chi-square ratio (χ²/df; Wheaton, Muthén, Alwin, & Summers, 1977) and several additional omnibus fit indices were considered when assessing overall model fit, including both the Tucker-Lewis index (TLI; Tucker & Lewis, 1973) and root-mean-square error of approximation (RMSEA; Browne & Cudeck, 1993). In general, larger TLI values and smaller relative chi-square ratios and RMSEA values indicate better fitting models. Currently adopted conventions suggest using a relative chi-square ratio smaller than 2.0 (Tabachnick & Fidell, 2007), TLI values ≥ .95, and RMSEA values ≤ .06 denote good fit (Hu & Bentler, 1999).

Figure 2.

Confirmatory factor models used to assess the dimensionality of problem behavior: (Model A) One-factor model, (Model B) Four-factor model, (Model C) Correlated Four-factor model, (Model D) Second-order factor model, (Model E) Bifactor model.

In the set of primary association analyses, partial main effects for the three candidate genes were assessed by simultaneously regressing all latent factors within the best-fitting model onto each measured genotype, while controlling for sex and age. Subsequently, in the supplementary association analyses, unidimensional confirmatory factor structures were separately modeled for all five of the aforementioned general and specific dimensions of problem behavior. As with the latent bifactor structure in the primary analyses, each of these supplementary phenotypes was independently regressed on SNPs from the three candidate genes in order to assess the extent to which disparate models of phenotype robustly or differentially determine the observed genetic associations.

Within the context of the social sciences, some additional explanation regarding the analytic strategy may be warranted. To start, we elected to test a seemingly large number of SNPs within each gene of interest relative to other candidate gene studies in the social science literature. The choice to do so reflects current practices in genetics (Dick, Latendresse, & Riley, 2011; Tabor, Risch, & Myers, 2002), and is based on issues often overlooked by social scientists. First, robust association with a given phenotype does not imply identification of a functional genetic variant, and even when evidence for the functional relevance of specific genetic loci exists within a given population, it cannot be assumed to generalize across ancestral lines. Moreover, a gene can include multiple causal variants, each contributing differentially to its functioning and the manner in which it influences susceptibility to the phenotype. Therefore, absent knowledge of all causal variants for a specific population within the genetic region(s) of interest, association analyses require enough markers to account for the observed LD structure of ancestral populations most similar to the sample being studied. Although this entails testing, and thus correcting for large numbers of variants, and increasingly so within more genetically heterogeneous samples (e.g., 25 CHRM2 SNPs in the present sample, compared with 9 SNPs genotyped in a commensurate sample of EA adolescents; Latendresse et al., 2011), there is no justifiable shortcut if the goal is to identify potentially meaningful phenotypic influences within sparsely characterized regions of the genome.

Results

Confirmatory Factor Modeling

One-factor model

The first model posited that the covariance among item responses for all 32 indicators of problem behavior (Figure 2, Model A) can be best accounted for by a single underlying common factor. Goodness of fit for this unidimensional model was poor (fit statistics for all confirmatory factor models are summarized in Table 2), as indicated by a large relative chi-square ratio, sub-threshold TLI and RMSEA values (χ²/df = 5.16, TLI = .85, RMSEA = .08).

Table 2.

Fit Statistics for Unconditional Confirmatory Factor Models

Model	χ2	df	χ2/df	TLI	RMSEA (90% CI)	Parameters Estimated
A. One-factor model	2395.51	464	5.16	.85	.081 (.078, .085)	106
B. Four-factor model	2483.11	464	5.35	.84	.083 (.080, .086)	106
C. Correlated four-factor model	793.64	458	1.73	.97	.034 (.030, .038)	112
D. Second-order factor model	778.38	460	1.69	.97	.033 (.029, .037)	110
E. Bifactor model	684.14	432	1.58	.98	.030 (.026, .035)	138

Note: For all chi-squares statistics, N = 590; TLI = Tucker-Lewis index; RMSEA = root-mean-square error of approximation; BIC = Bayesian information criteria.

Four-factor models

The second model posited that the covariance among item responses for the 32 indicators would be best explained by allowing each of the four dimension-specific groups of indicators (i.e., alcohol use items, marijuana use items, sexual behavior items, and disruptive behavior items) to load onto independent (orthogonal) dimension-specific factors (Figure 2, Model B). Goodness of fit for this uncorrelated multidimensional model was also poor, with fit statistics uniformly worse than those of the one-factor model (χ²/df = 5.35, TLI = .84, RMSEA = .08). However, a less restricted version of the four-factor model (Figure 2, Model C) that allowed inter-correlations among the four dimension-specific factors (range = .31 to .67) provided much improved fit, with all indicators surpassing the accepted thresholds (χ²/df = 1.73, TLI = .97, RMSEA = .03). Moreover, since these are nested models, a formal test of the change in chi-square associated with fixing these inter-factor correlations to zero was used to confirm a significant decrement in fit (χ²Δ(6) = 328.47, p < .001).

Second-order factor model

As with the four-factor model, the second-order factor model assumes that covariance among item responses can best be explained by having each indicator load onto one of four dimension-specific factors. However, rather than allowing for correlation between factors, this model posits that the covariance between the four dimension-specific first-order factors can be accounted for by introducing a higher-order general factor (Figure 2, Model D). The second-order factor model provided no discernable improvement in fit over the correlated four-factor model (χ²/df = 1.69, TLI = .97, RMSEA = .03).

Bifactor model

The bifactor model posits that the covariance among item responses can be best explained by a common factor upon which all indicators load, and dimension-specific factors upon which sub groups of indicators load (Figure 2, Model E). In this way, the bifactor model parses inter-item covariation into that which can be explained by a general syndromal factor representing variance that is shared across all problem behavior indicators, and group-specific factors that model all residual covariation (i.e., not accounted for by the general construct) among behavior-specific subsets of items. For a description comparing the bifactor model to various other factor models see Chen, Hayes, Carver, Laurenceau, and Zhang (2012). Though not nested, fit statistics suggest an improvement in fit over the prior model (χ²/df = 1.58, TLI = .98, RMSEA = .03). Within this, the best fitting model, all individual item loadings were significant at p ≤ .001 (Table 3).

Table 3.

Standardized Loadings for Indicators in the Bifactor Model

	Factors
	G	AU	MU	RS	DB
Alcohol Use:In the LAST 12 MONTHS…
have you had six or more drinks?	.78	.51
how many days did you drink alcohol?	.66	.67
how many drinks did you usually have each time?	.64	.71
how many days did you drink 5 or more drinks in a row?	.66	.62
Marijuana Use:In the LAST 12 MONTHS…
have you used marijuana in the last year?	.89		.45
have you used marijuana six or more times?	.86		.47
how many times did you use marijuana?	.86		.41
Sexual Behavior:In your LIFETIME…
have you ever had oral sex	.50			.35
have you ever had vaginal sex	.47			.88
have you ever had anal sex	.25 a			.41
how many people have you had vaginal or anal sex with?	.45			.76
how many partners did you NOT use a condom or rubber with during vaginal or anal sex?	.37			.57
Disruptive Behavior:Now or within the last 6 MONTHS…
i argue a lot	.21				.71
i disobey my parents	.40				.63
i disobey at school	.38				.73
i am stubborn	.21				.40
i have a hot temper	.22				.52
i am mean to others	.20 b				.72
i destroy things belonging to others	.47				.59
i don't feel guilty after doing something i shouldn't	.26				.41
i break rules at home, school, or elsewhere	.39				.72
i get in many fights	.32				.54
i hang around kids who get in trouble	.37				.52
i lie or cheat	.42				.55
i physically attack people	.35				.68
i run away from home	.32				.41
i set fires	.53				.46
i steal at home	.36				.49
i steal from places other than home	.57				.37
i swear or use dirty words	.50				.50
i threaten to hurt people	.54				.61
i cut classes or skip school	.59				.30

Note. G = general problem behavior; AU = alcohol use; MU = marijuana use; SB = sexual behavior; DB = disruptive behavior;

^ap = .007;

^bp = .004; all other loadings were significant at p ≤ .001.

Genetic Association Analyses

While accounting for sex and age, conditional main effects for three candidate genes were assessed by sequentially including individual SNPs as independent variables within the best fitting phenotypic model: the bifactor model. Because this model constrains the covariances between factors to be orthogonal, and indicators load only onto the general factor and one additional dimension-specific factor, this approach yields SNP specific information about prediction of variation in a general factor, as well as variation in unrelated behavior specific factors. The results of regressing the five latent factors on each of the SNPs are summarized in Table 4 in terms of partial R² coefficients and p-values for all significant associations.

Table 4.

Significant Main Effects of Markers, by Gene and Analysis

	Primary Analyses					Supplementary Analyses
Models:	Bifactor Model					Unidimensional Confirmatory Factor Models
CHRM2	G	AU	MU	SB	DB	G	AU	MU	SB	DB
rs10249472					.014(.017)	.016(.004)				.021(.003)
rs6977141
rs10954546
rs11984255					.014(.036)
rs4407820					.026(.037)
rs6970010					.027(.033)
rs12333547
rs6967120
rs10246433
rs978437							.007(.050)
rs7782965							.014(.029)
rs7800170						.006(.033)	.016(.018)
rs1455858							.015(.038)
rs1378646							.013(.047)
rs1824024
rs324582
rs2061174
rs324650				.021(.027)	.021(.044)
rs7779921
rs7796601								.023(.030)
rs832990		.016(.009)	.002(.016)
rs11971038									.011(.024)
rs832988			.022(.026)	.024(.028)
rs833018		.038(.035)
rs833016
GABRA2	G	AU	MU	SB	DB	G	AU	MU	SB	DB
rs497068
rs548583
rs526805
rs537134									.011(.028)
rs279871									.013(.035)
rs1808851
rs279858				.027(.036)					.017(.025)
rs279851					.023(.009)
rs1561774					.014(.049)
rs11941602
rs279843				.026(.026)
rs279845
rs183960					.010(.024)
rs1440130
rs279826
rs279828
rs279836
rs1442060					.033(.002)
rs1442062				.022(.041)
rs11503014
OPRM1	G	AU	MU	SB	DB	G	AU	MU	SB	DB
rs12205732
rs1799971
rs553202
rs524731
rs3778150
rs10457090
rs9478503
rs589046
rs3778157									.012(.009)
rs10485057
rs562859
rs511420						.009(.047)
rs9322447
rs681243				.033(.010)
rs504932				.028(.026)
rs512053
rs658156				.036(.004)
rs645027
rs644261
rs613341		.051(.005)		.021(.035)
rs616585			.031(.050)	.034(.030)
rs10485058
rs497315
rs678122				.046(.002)					.008(.020)

Note. G = general problem behavior; AU = alcohol use; MU = marijuana use; SB = sexual behavior; DB = disruptive behavior. Coefficients reflect R² change attributed to a genetic marker and associated p-values (in parentheses). Bold face type denotes effects maintaining significance after accounting for the estimated number of independent marker loci. Markers are shown as rs numbers from the dbSNP database and ordered by chromosomal position as denoted in build 36.3.

CHRM2

There were no significant SNP effects of CHRM2 on general problem behavior, but two SNPs each were linearly associated with alcohol use, marijuana use, and sexual behavior factors. Among these associations, only one marker (rs832990) approached the adjusted significance threshold for multiple testing, and that was in relation to alcohol use (R² = .016, p = .009). With respect to disruptive behavior, five markers were significantly predictive before correction, including three correlated SNPs (rs11984225, rs4407820, and rs6970010 with ranges: R² = .014−.027, p = .033−.037).

GABRA2

No SNPs in GABRA2 were predictive of general problem behavior, alcohol use, or marijuana use. However, three generally distinct SNPs were significantly associated with the sexual behavior factor, and four other SNPs with disruptive behavior, including one surviving (rs1442060: R² = .033, p = .002), and another marginally significant (rs279851: R² = .023, p = .009) following the Nyholt-adjusted Bonferroni correction described in the methods section.

OPRM1

There were no genetic associations with general problem behavior or disruptive behavior, one modest association with marijuana use, and one association with alcohol use that remained marginal after correcting for multiple testing (rs613341: R² = .051, p = .005). In addition, sexual behavior was robustly associated with five SNPs representing a single LD block (ranges: R² = .021−.036, p = .004−.035), including one remaining marginal following correction (rs658156: R² = .036, p = .004), and another adjacent SNP that survived a Nyholt correction (rs678122: R² = .046, p = .002).

Supplementary Genetic Association Analyses

In contrast to the models described above, genotype-based analyses of similar behaviors typically explore associations between one or more genetic variants and a single phenotype of interest; either dimension-specific or broad-based (i.e., incorporating indicators of multiple dimensions of behavior into a single composite or factor score) in nature. In order to determine how genetic influences on more commonly constructed phenotypes for the same set of externalizing behaviors might compare with findings based on the bifactor model, we supplemented our primary analyses by separately regressing five independently derived latent factors (i.e., five separate unidimensional confirmatory factor models, each reflecting a set of indicators from a discrete column in table 3) onto each of the SNPs. Aside from the methods by which the factors were derived, these supplementary analyses were conducted in the same manner as the conditional bifactor analyses. Partial R² and p-value coefficients for significant supplementary associations are presented alongside those of the primary model, in Table 4, and are summarized below.

The relevance of CHRM2 appeared to differ in the unidimensional models. Most notably, (1) a cluster of correlated SNPs emerged in relation to alcohol use, none overlapping with those reflecting a discrete LD block in the primary analyses (ranges: R² = .007−.016, p = .018−.05); (2) two SNPs were associated with general problem behavior, one remaining marginally significant (rs10249472: R² = .016, p = .004) post correction; and (3) only the most significant marker associated with disruptive behavior in the bifactor model remained associated, with a sub-threshold p-value (rs10249472: R² = .021, p = .003). GABRA2 and OPRM1 were, in contrast, somewhat less associated with unidimensionally derived phenotypic factor scores. Again, three GABRA2 SNPs were modestly associated with general problem behavior before correction, but no other associations emerged. Likewise, in OPRM1 only one marker was associated with general problem behavior, and two with sexual behavior, one of which remained marginally significant following the Nyholt correction (i.e., rs279851: R² = .012, p = .009).

Discussion

The rapid technological advancements and continued reduction in costs that typify today’s era of genomic information have resulted in a proliferation of genetic association studies of both binary and complex behavioral phenotypes. As this body of findings grows at an almost exponential rate, it has become increasingly necessary to characterize the extent to which individual effects (or lack thereof) generalize beyond a particular operationalization of a construct and the majority populations typically being studied. The present study attempts to do this with respect to problem behavior, first by establishing the structural nature, and then by examining relationships between the emerging constructs and a set of candidate genes previously shown to influence one or more of the specific problem behavior dimensions, or related externalizing constructs, all within a sample of AA adolescents.

Model fitting comparisons suggest that the bifactor model best accounted for the associations among the item level data in the present sample. Two important implications follow from this finding. First, this indicates the need to recognize and simultaneously assess both general and specific dimensions of problem behavior. Second, unlike how general problem behavior is typically operationalized as the shared variability across a set of correlated items or factors reflecting multiple dimensions of behavior (e.g., Figure 2, models A and D, respectively), the bifactor model implies an orthogonal structure among the factors, with each item serving as an indicator of the general syndrome and one specific behavioral dimension. This is consistent with recent evidence characterizing the structural nature of associations among dimensions of externalizing psychopathology (Martel, von Eye, & Nigg, 2010). So, just as it was suggested decades ago, individual behavior problems are linked, though perhaps not in the hierarchical manner proposed by those initially exploring the structural nature of problem behavior via latent variable models (e.g., McGee & Newcomb, 1992).

Given the extant literature on shared genetic liability across domains of problem behavior, it seems natural to question whether these observed phenotypic correlations are due, in part, to the pleiotropic effects of an associated variant (McCarthy & Hirshhorn, 2008), or conversely, whether common genetic influences are due simply to shared phenotypic variance. Unfortunately, both GWAS and candidate gene studies have generally, if not exclusively, focused on associations between single variants and rather grossly defined stand-alone phenotypes. In the case of GWAS, this was largely because it was thought not to be feasible to increase phenotypic complexity, given logistical computing constraints. However, just as we are now attempting to extend the reach of the candidate gene approach with respect to increasingly complex phenotypes, so too are statistical geneticists actively working toward multivariate phenotypic extensions within the GWAS framework (O’Reilly et al., 2012; van der Sluis, Posthuma, & Dolan, 2013).

It is interesting to note that both CHRM2 and GABRA2, two genes previously implicated via unidimensional models of general externalizing behavior in EA and other diverse samples that statistically control for race (Dick et al., 2008; Dick et al., 2009; Latendresse et al., 2011), showed no evidence of genetic liability with respect to a general factor when fitting a multidimensional bifactor model to data from this sample of AA youths. In fact, the absence of (1) robust associations with the general problem behavior factor across correlated SNPs and (2) effects that survive the Nyholt-adjusted Bonferroni correction suggests that these genes may not contribute, at least directly, to the manifestation of a problem behavior syndrome within this sample.

In contrast, all of the candidate genes were associated with the more narrowly defined specific dimensions of problem behavior within the bifactor model. Of particular interest were two variants representing a seemingly novel LD block in CHRM2. Despite there being no prior evidence of phenotypic associations with either of these SNPs, one (rs832990) remained marginally associated with alcohol use in AA youths, suggesting a potentially important region for follow up in subsequent studies.

Likewise, three variants spanning GABRA2 displayed modest associations with sexual behavior. Although none had previously been examined in relation to this dimension of behavior, all have been associated with alcohol phenotypes. However, no associations between GABRA2 and alcohol use were evidenced in this study, perhaps due to lower rates of drinking in this population relative to EA adolescents. Whatever the case, these and surrounding variants should serve as targets in follow up studies.

In addition, two largely independent SNPs within GABRA2 remained either marginally or significantly associated with disruptive behavior after correction, one of which was previously associated with symptoms of conduct disorder in adolescents from the Collaborative Study on the Genetics of Alcoholism sample (Dick et al., 2006). Thus, this particular finding may generalize across adolescent populations, irrespective of ancestral background.

With respect to OPRM1, alcohol use in this sample was associated with a single SNP deriving from an LD block that includes variants previously associated with alcohol dependence in EA adults (Zhang et al., 2006), suggesting the possibility that this effect might generalize to AA adolescents (and possibly even AA adults). As such, OPRM1 should be a targeted for further exploration in relation to alcohol phenotypes within AA populations. Even more surprisingly, there was robust evidence of association with sexual behavior across the gene, with two uncorrelated variants remaining significant after correction. Previous studies have implicated OPRM1 as playing a role in social affiliation (Curley, 2011), as well as dispositional and neural sensitivity to social rejection (Way, Taylor, & Eisenberger, 2009), both suggesting a similar mechanism by which this gene might influence sexual behavior. In addition, among those variants associated with sexual behavior, one SNP derives from the same LD block as a variant previously associated with heroin dependence among EA (Levran, Awolesi, Linzy, Adelson, & Kreek, 2011), and another was previously associated with subjective responses to alcohol in adults of American Indian descent (Ehlers et al., 2008).

What then should be made of genetic associations identified via unidimensional models in the literature as compared to those found here? To start, the fact that both significant and robust associations were generally only observed in the primary, but not supplementary analyses at least suggests the possibility that the extant literature on genetic associations with complex behaviors may be complicated by unobserved correlations between and across behavioral dimensions. In addition, a comparison of findings from the primary and supplementary analyses illustrates how significant associations can go undetected when ignoring the complex etiology of these clinically relevant phenotypes.

While the ultimate test resides in replication, there are not yet any comparable studies that have used an extended multidimensional approach. However, a number of issues have been shown to influence reliability in traditional genetic association studies, including phenotypic complexity. In considering numerous multidimensional approaches, statistical geneticists suggest that relative to conventional analyses, latent variable models have the potential to limit multiple comparisons while yielding more precise estimates (Liu, Jorgenson, & Witte, 2005), thereby improving our ability to detect “true” genetic associations by significantly increasing statistical power (Kim, Sohn, & Xing, 2009; van der Sluis, Verhage, Posthuma, & Dolan, 2010).

Notably, developmental scientists have suggested that correlations between various etiological levels of influences (e.g., biological, contextual, social) may be more pronounced in high-risk samples (Cairns & Cairns, 1994; Farmer, Quinn, Hussey, & Holahan, 2001), and that such correlated constraints can propagate misconceptions or stereotypes regarding the risk for developing problem behaviors in racial minority populations. Thus, in as much as genes function to regulate complex behavior, we have to be particularly aware of the potential differences attributable to ancestry. Whereas many studies of genetic association merely control for racial differences, the present sample is among the rare few that explores associations strictly within AA adolescents, a population that is largely understudied with respect to the genetic etiology of behavioral outcomes. Thus, specific associations observed in this study are relatively novel, a fact that would likely remain so even if we had limited our analyses to a more traditional set of unidimensional models. Yet, it warrants mentioning that some degree of ancestral variation has been shown to exist within this sample, and could serve as a potential source of population stratification. Even so, these multidimensional models should serve as the basis for subsequent replication efforts in AA samples, as well as the impetus for a more thoughtful standard for the modeling of conceptually related phenotypes in genetic association studies irrespective of the demographic makeup.

Among this study’s principal limitations is the fact that it was statistically underpowered to reasonably conduct and compare the proposed analyses (i.e., tests of association between three distinct genes and the emerging phenotypes) by sex and/or age, particularly given the large number of SNPs required to adequately characterize variability in genomic regions void of known functional variants in AA populations. So, although prior studies suggest, for example, that development and rates of substance use behaviors (Brady & Randall, 1999), etiology of sexual behavior (Collins, Sutherland, & Kelly-Weeder, 2012), and levels of disruptive behavior problems (Lahey et al., 2000) may vary by sex and/or age, we were only able to statistically control for these factors in our the present analyses. This further highlights the need to identify sufficiently powered independent samples wherein the effects of these genes on complex behaviors can be both compared across distinct populations and replicated within other AA samples, as we strive to expand our understanding of susceptible, yet understudied populations.

Another somewhat related limitation is that of implied invariance within the measurement model that we selected and retained for the analyses. That is, observed indicators may function differentially with respect to age and/or sex, and failure to account for these differences might limit construct validity among the latent externalizing factors, thereby increasing the likelihood of observing spurious genetic associations and/or suppressing meaningful ones. Subsequent research efforts should, therefore, explore potential sources of measurement noninvariance within the bifactor model via the analysis of differential item functioning, and specifically with respect to age and sex.

Notably, clinicians are able to best treat individuals when the etiology of their behavior/disorder is more fully understood. As such, the implications of the present study are two-fold: First, refinement of the phenotypic measurement model increases the precision with which we assess behaviors of interest, and assessment is paramount within the treatment paradigm, especially when individuals present with symptoms reflecting multiple, correlated dimensions. Second, any increased precision in the measurement of behavior further increases the reliability and validity of subsequent associations between genes and the behavior and identifying genes implicated in a given phenotype can serve as an important catalyst with respect to advancing our understanding of the biological processes that underlie it. In turn, researchers will be better prepared to hypothesize about the potential genetic mechanisms by which known risk environments come to influence the phenotype. These are particularly relevant with respect to the biological substrates of problem behavior(s) in AA adolescents, about which relatively little is known.

In conclusion, we provide evidence to support the structural value of modeling both individual and collective dimensions of problem behavior. We further demonstrate the utility of multidimensional models in the identification of genetic variants involved in the etiology of externalizing problems, and illustrate the ramifications of selecting simpler but poorer fitting models on our capacity to identify and interpret these effects. Phenotypic complexity remains one of the more overlooked obstacles to the replication of genetic associations, and our ability to adequately account for this complexity may serve as a proxy for our success. In addition, we provide initial evidence implicating this set of candidate genes in specific behavioral dimensions that were not detectable at the level of an overarching problem behavior syndrome. While some associations appear to be consistent with findings previously identified in EA or racially diverse samples (e.g., GABRA2 and disruptive behavior; Dick et al., 2006), others are more novel effects that may provide insight into previously unexplored mechanisms: most notably the association between OPRM1 and sexual behavior.