Authors: Frances L. Wang, Lambertus Klei, Bernie Devlin, Brooke S.G. Molina, Laurie Chassin
Categories: Article, Genetic influences, Conduct Disorder, Depression, Temperament, Alcohol Use
Source: Research on child and adolescent psychopathology
Authors: Frances L. Wang, Lambertus Klei, Bernie Devlin, Brooke S.G. Molina, Laurie Chassin
The genetic architectures underlying symptoms of conduct problems and depression have largely been examined separately and without incorporating temperament, despite evidence for their genetic overlap. We examined how symptoms and temperament dimensions were transmitted together in families to identify highly heritable composite phenotypes, and how these composite phenotypes predicted alcohol outcomes in young adulthood. Participants (N=486) were drawn from the third generation of families oversampled for alcohol use disorder in the first generation. Conduct problems, depression, and temperament were reported at 11–19 years old and alcohol outcomes at 18–26 years old. Using principal components of heritability analysis, we found seven highly heritable composite phenotypes, five of which predicted alcohol three characterized by co-occurring conduct problems and depression and two by conduct problems. Novel composite phenotypes that were characterized by both conduct problems and depression showed different types of symptoms, temperament features, and genetic underpinnings. Children manifesting differing composite phenotypes might benefit from distinct treatments based on their unique etiologies.
Alcohol use problems and disorder (AUD) are major public health concerns. Symptoms and features of conduct and depressive disorders, and especially their co-occurrence, have been shown to be robust predictors of alcohol problems and may thus constitute prime targets for preventive intervention (Marmorstein & Iacono, 2001, 2003). Key elements of conduct disorder include aggression and serious violations of certain rules and social norms (DSM-5; American Psychiatric Association, 2013). For major depressive disorder, key elements include feelings of worthlessness and loss of enjoyment, which are often accompanied by changes in sleep, appetite, energy level, and concentration (American Psychiatric Association, 2013). Conduct disorder, major depressive disorder and their co-occurrence are heritable, as are scores summarizing their behaviors and symptoms (Hannigan et al., 2018; O’Connor et al., 1998; Pappa et al., 2015, 2016). Yet, the heritability of the specific symptoms and behaviors corresponding to conduct and major depressive disorders have been understudied. This information could clarify our understanding of the etiologies underlying these problem behaviors in adolescence, which represents a critical step towards enhancing prevention and treatment in the future.
There has been some research on the heritability of symptoms within conduct disorder and within depression. For example, a twin study on the genetic architecture underlying symptoms of conduct problems revealed two latent genetic factors comprised of symptoms of rule-breaking and symptoms of overt aggression (Kendler, Aggen, & Patrick, 2013). For depressive symptoms, three distinct latent genetic factors have been identified, including cognitive/psychomotor, neurovegetative, and mood symptoms (Kendler, Aggen, & Neale, 2013). Another study involving only common genetic variation (i.e., single nucleotide polymorphisms [SNPs]) also found more than one latent genetic factor underlying depressive symptoms, although somewhat different in nature from the factors found by Kendler and colleagues (Thorp et al., 2019). Interestingly, the rule-breaking genetic factor (Kendler, Aggen, & Patrick, 2013) and the cognitive/psychomotor genetic factor (Kendler, Aggen, & Neale, 2013) each showed the strongest associations with AUD relative to the other genetic factors within conduct disorder or depression, respectively. This suggests that understanding the latent genetic factors that underlie symptoms of conduct disorder and depression could inform higher risk pathways to AUD.
Taken together, these results inform which symptoms within conduct disorder or within depression may be influenced by common etiological factors. However, research also suggests that there exist shared genetic influences that explain the co-occurrence between conduct problems and depressive symptoms (Hannigan et al., 2018; O’Connor et al., 1998; Subbarao et al., 2008). Thus, important shared genetic influences could be missed when symptoms of conduct disorder and major depressive disorder are studied separately, as has often been done in the literature.
The Research Domain Criteria (RDoC) initiative and other similar efforts were developed, in part, because studies of individual psychiatric disorders can miss important etiological influences shared by two or more disorders. RDoC encourages etiological studies that do not rely on diagnoses and instead considers psychopathology in the context of major domains of basic human neurobehavioral functioning (Insel, 2014). Two prominent approaches that embody principles of RDoC include the Hierarchical Taxonomy of Psychopathology Consortium (HiTOP; Kotov et al., 2017) and studies that have employed genomic structural equation modeling (SEM; Grotzinger et al., 2019, 2022; Lee et al., 2019; Waldman et al., 2020). These approaches have modeled the hierarchical structure of psychopathology based on either phenotypic or genetic correlations among psychiatric disorders. Both types of studies have found higher-order factors that account for the correlations among psychiatric disorders; e.g., a factor labeled externalizing disorders, which accounts for the tendency for conduct disorder and oppositional defiant disorder to co-occur. These structural models of psychopathology can help elucidate shared and unique genetic influences that operate across and on different types of mental health problems (Waszczuk et al., 2020).
RDoC, HiTOP, and genomic SEM represent useful frameworks for understanding psychopathology and provide guidance into fruitful future directions for etiological studies. Although these research frameworks highlight the importance of shared genetic variation underlying multiple psychiatric disorders, it is important to consider how this information could be translated for more practical purposes. For example, it could be important to understand whether there is shared genetic variation underlying specific symptoms of conduct disorder and depression, as individuals exhibiting these symptom clusters will be encountered within patient populations. Moreover, individuals whose symptoms arise from a similar etiological basis may have unique future outcomes or could benefit from unique treatment strategies, which could be important topics for further investigation. Here, we examined how adolescents’ symptoms of conduct problems and depression were transmitted together in three-generational families and how these transmission patterns can be used to form genetically distinct composite phenotypes.
Research shows that distinct constellations of conduct problems and depressive symptoms are uniquely associated with certain temperamental or personality traits, suggesting the importance of examining how temperament dimensions cluster genetically with conduct disorder and depressive symptoms. Temperament dimensions are constitutionally based individual differences in reactivity and regulation (Rothbart & Derryberry, 1981). According to Rothbart’s model, three superordinate dimensions of temperament include effortful control (ability to inhibit a dominant response to perform a subdominant response), negative emotionality, and surgency (approach orientation). Several lower-order factors also underlie these superordinate dimensions in adolescence including, but not limited to, attentional, activational, and inhibitory control (within effortful control), fearfulness and high-intensity pleasure (within surgency), and frustration (within negative emotionality). Temperament dimensions also share many similarities with personality traits. Indeed, researchers have suggested that temperament and personality represent different ways of describing the same basic traits either earlier or later in life, respectively (Shiner & DeYoung, 2013). Specifically, the personality trait of conscientiousness may be analogous to effortful control, neuroticism to negative affectivity, and extraversion to surgency (Shiner & DeYoung, 2013).
With regards to the unique associations of temperament and personality dimensions with symptoms of conduct disorder and depression, Kendler, Aggen, and Patrick (2013) found that only the rule-breaking genetic factor underlying conduct problems showed an association with extraversion. Moreover, relative to the other depression factors discovered, the cognitive/psychomotor factor showed the strongest association with neuroticism (Kendler, Aggen, & Neale, 2013). Additional research showed that low fear and high surgency could each underlie the manifestation of different types of conduct problems (i.e., callous-unemotional vs. reactive aggression; Nigg, 2006). High negative affectivity and low surgency have also been shown to be separate pathways to depressive symptoms (Lonigan et al., 2003; Olino et al., 2011) and high negative affectivity and low effortful control may be separate pathways to co-occurring conduct problems and depression (Lemery-Chalfant et al., 2008; Tackett et al., 2011).
Because temperament dimensions could inform the affective and self-regulatory traits that underlie heritable constellations of conduct problems and depressive symptoms, their inclusion could be invaluable in understanding how symptoms of conduct problems and depression come together genetically. Few studies, to our knowledge, have tested how temperament dimensions cluster genetically with symptoms of conduct problems and depression in adolescence. By conducting these analyses, the present study could uncover unique clinical presentations of a severely impairing form of psychopathology, co-occurring conduct problems and depression, while also understanding the potential affective and regulatory traits that underlie or motivate them (i.e., temperament dimensions). This, in turn, could reveal unique pathways to AUD. Determining how these features cluster together genetically also allows us to understand conduct problems and depression in the context of fundamental neurobehavioral systems, such as negative valence and regulatory systems. Doing so is central to the mission of RDoC and may shed light on the etiology of these mental health problems in adolescence.
We use a method novel to the child psychopathology literature called Principal Components of Heritability (PCH) analysis (Ott & Rabinowitz, 1999) to examine how symptoms of conduct problems and depression and dimensions of temperament are transmitted together in families to identify novel, highly heritable composite phenotypes defined by combinations of temperament features and conduct problem and depressive symptoms. This method has proved useful in the field of psychiatric genetics (e.g., Oualkacha et al., 2012; Wiener et al., 2013) and is novel to the child psychopathology literature. PCH is similar to classical principal components analysis in that it identifies one or more approximately orthogonal, linear combinations of variables (i.e., principal components). However, PCH analysis maximizes the heritability of linear combinations of variables by capitalizing on the genetic correlations among them, which we estimate using pedigree information (consisting of siblings, half-siblings, and cousins) derived from a study oversampled for familial alcohol use disorder (AUD).
We hypothesized that we would find some heritable composite phenotypes that blended across symptoms of conduct disorder and depression and dimensions of temperament. We also expected to find some composite phenotypes that reflected previously discovered latent genetic factors within conduct problems, such as rule-breaking and aggressive symptoms, and within depression, such as cognitive/psychomotor symptoms, neurovegetative symptoms, and mood symptoms. Despite preliminary hypotheses, we acknowledge that our analyses are largely exploratory. We then tested the associations among composite phenotypes and alcohol use outcomes approximately 3–4 years later in late-adolescence/early adulthood to understand whether certain composite phenotypes represented greater risk for alcohol problems over others.
We note that, because of our study design, it is possible that genetic effects could be partially confounded with shared environmental effects. We conducted analyses to understand the extent of this potential confounding by comparing solutions between composite phenotypes derived from PCH to those derived in a parallel fashion from phenotypic principal components analysis (PCA). In theory, PCH should be less influenced by shared environment compared to PCA. PCH requires that the key axes of phenotypic variation align with the key axes of variation in pedigree relationships (full sibs, half sibs, and cousins), whereas PCA simply seeks to identify the largest components of variation from the correlation structure of features (items). Thus, differences between PCH and PCA solutions would likely reflect effects of shared environment detected by PCA in sibships that are inconsistent with the patterns of sharing seen in other more distant relative pairs (i.e., half siblings, cousins). We also tested the effect of PCA on alcohol use outcomes to understand the unique contribution of PCH in understanding alcohol risk relative to composite phenotypes purely derived from phenotypic information, as is standard in the field.
Participants were drawn from a three-generational longitudinal study of families oversampled for alcohol use disorder (AUD; Chassin et al., 1992). The original sample consisted of parents recruited to participate using court records, HMO wellness questionnaires, and community telephone surveys. To be included in the AUD sample, at least one biological, custodial parent had to meet DSM-III criteria for lifetime diagnoses of alcohol abuse or dependence or Family History-Research Diagnostic Criteria as reported by the other parent. Demographically matched control parents were also recruited using reverse directories and were included in the study if neither met the study’s criteria for AUD. The participating children of these parents were, on average, 12 years old at the first assessment. These parents and their children (including siblings) were interviewed over 3–6 waves of data collection. These waves spanned early adolescence to adulthood for participating original children (to a mean age of 33.6 years). A detailed description of the original study’s design can be found elsewhere (Chassin et al., 1992). In the fifth wave of data collection, the grandchildren of the original study parents were recruited and they participated in a total of five assessments spanning childhood (Mage=7.5 years) to adulthood (Mage=20.8 years). This study design resulted in a three-generational cohort. The third generation of this cohort (i.e., grandchildren) includes full siblings, half siblings, and cousins.
The grandchildren and their phenotypic data are the focus of this study. They are hereafter referred to as the participants. Other family members were used for estimating the necessary genetic and environmental parameters of the models specified herein. Kinship was identified by an informant to account for familial relationships. Informed consent or assent was obtained from participants.
Participants’ data were drawn from four waves (referred to in this study as T1, T2, T3, and T4) spanning adolescence to young adulthood and were collected through interviews (T1) and telephone or online surveys (T2-T4). Participants were included in analyses involving the computation of principal components of heritability if they were adolescents (11–19 years) at T1 and T2 and had data on conduct problems and depression at T1 and T2. Participants who also had alcohol use data at T3 and T4 were included in analyses involving the prediction of alcohol outcomes.
Participants reported on past six month T1 and T2 conduct problems and depression symptoms using the Achenbach Youth Self-Report Form (Achenbach & Rescorla, 2001) and the Revised Child Anxiety and Depression Scale (RCAD; Chorpita et al., 2000). Items were drawn from the Achenbach DSM-Oriented Conduct Problems, Rule-breaking, Aggression, Anxious/Depressed and Withdrawn/Depressed subscales and the RCAD Major Depression subscale. Thirteen items were excluded because they had <10% endorsement and thus would not provide adequate information for estimating heritability. Eight Achenbach items were excluded because they captured anxiety more than depression and overlapped with information in temperament subscales (i.e., afraid of animals, situations or places; afraid of going to school; afraid to do something bad; nervous/tense; too fearful anxious; self-conscious; worry a lot; too shy/timid). This resulted in 17 conduct problem and 17 depression items (Table 2). Items were dichotomized to reflect having experienced each symptom in the past six months.
At T1, participants reported on eight temperament subscales (activation control, attentional control, inhibitory control, fear, frustration, high intensity pleasure, shyness, and affiliation) of the Early Adolescent Temperament Questionnaire (Capaldi & Rothbart, 1992). Because the high intensity pleasure scale was based on the Zuckerman (1979) construct of sensation seeking (Capaldi & Rothbart, 1992), we use the latter term henceforth. Sensation seeking is a more well-represented construct in alcohol and other research.
Participants reported on alcohol use measures at T3, T4, or both. The age span of assessment during T3 and T4 was 13–26 years. To obtain measures that reflected early adulthood alcohol use, we used the oldest data point from the T3 or T4 assessment that was also collected at age 18 or older. This resulted in alcohol use assessments that ranged from the ages of 18–26.
For the assessment of alcohol problems, 318 participants reported on 14 binary lifetime alcohol consequences and dependence symptoms that were summed to form a measure of alcohol problems. We chose a lifetime measure because alcohol problems tend to occur less frequently than do drinking episodes and their cumulative frequency over time has clinical significance.
We also examined frequency of heavy drinking as an outcome to understand the effects of PCH on alcohol consumption (N=327). At T3, this was defined as the frequency of having 5 or more drinks at one time. At T4, it was defined as 5 or more drinks for men and as 4 or more drinks for women (at one time). Response options 0=never; 1=1–2 times in my life; 2=3–5 times in my life; 3=more than 5 but less than once a month this past year; 4=1–3 times/month this past year; 5=1–2 times/week this past year; 6=3–5 times/week this past year; 7=every day this past year.
Analyses were conducted using R Studio (RStudio Team, 2020) and blupf90 (Misztal et al., 2014). The data analyzed were familial relationships among participants and the distribution of phenotype values of the participants in these families. Preliminary heritability estimates for each conduct problem and depression item (using repeated measures of each at T1 and T2) and each temperament subscale were computed using function mmer in sommer (Covarrubias-Pazaran, 2016). We imputed missing data for a small number of cases (Supplementary Text). For PCH analyses, we used gibbs3f90, thrgibbs1f90 and thrgibbs1f90b from the blupf90 package (Misztal et al., 2014). These functions use a Gibbs sampler to estimate (co)-variance components and their highest probability density region (the Bayesian equivalent of the frequentist’s confidence interval). The Supplementary Text provides a full description of the mathematics underlying PCH estimation. PCH describe how phenotypes segregate in families. Although the expectation is that the PCH will largely map onto genetic variation affecting variation and covariation of phenotypes, it is possible that environmental factors contribute to or alter genetically induced covariances and thus the PCH.
A standard approach to PCH analysis is to analyze all variables jointly. Our study, however, included 34 binary conduct problem and depression variables with measurements at T1 and T2 (which itself requires special consideration, such as repeated measures components), as well as five semi-continuous temperament variables measured at T1. Joint estimation of over 1,500 parameters, as required of such a model, presents two parameter estimates could be unstable and the computational time would be substantial and possibly infeasible (Table S5 in the Supplementary Text). Thus, we explored the feasibility of several alternative approaches to modeling the data (Tables S1–S5, Figures S1–S2 in the Supplementary Text). Based on results, we implemented a three-phase approach that we found would yield more, or similarly stable estimates and greater computational efficiency than fitting the full joint model (Figure 1). In all phases, variables were standardized prior to analyses.
In consideration of repeated measures data, we combined identical conduct problem and depression items from T1 and T2. We did so by estimating the first PCH for each pair of identical conduct problem or depression items (referred to as item-PCH). To estimate PCH, the Gibbs sampler was used to obtain estimates and their 95% highest probability density (HPD) for the variance components using a bivariate mixed model, in which the observations at T1 and T2 were the two outcomes, sex and age (T1 or T2) were covariates, and the sample was the random effect.
We next reduced the data in a way that respected correlations among related conduct problem items, depression items, and temperament traits while considering their shared genetics. We implemented a pairwise analysis of the 34 item-PCH from phase 1 and the five T1 temperament variables to estimate variance components. After combining all pairwise results (Supplementary Text), we used cluster analysis (hclust in R; Müllner, 2013) of the genetic correlation matrix to obtain clusters based on cutting the cluster tree at 0.70 (cutree in R; see Figure 2 which shows the variables that clustered together). We used the PCH transformation to obtain the lead (first) PCH for the variables represented within each cluster to form “cluster-PCH.”
Next, we wanted to understand how all the features came together based on shared genetics to form highly heritable composite phenotypes. This differs from cluster-PCH, which each only encompass a small number of the most highly genetically correlated conduct problem items, depression items, or temperament traits. We estimated “final” PCH based on cluster-PCH. We estimated the genetic and residual variance components for a multivariate mixed model with all cluster-PCH as outcomes, sex and average age for T1 and T2 as covariates, and the subject as the random effect. After obtaining the (co)-variance components, PCH were determined. Then, using a single trait mixed model with sex and average age at T1 and T2 as covariates and a random effect for samples, the heritability of these PCH were estimated. “Final PCH” were retained for subsequent analyses if heritabilities>0.50.
For alcohol problems and heavy drinking, we fit zero-inflated Poisson models using the glmmTMB package, due to the presence of many zeros (non-endorsements) for alcohol problems and heavy drinking (see Table S6: Hist Alc). Nuclear family was included as a random effect, age at the alcohol use assessment and sex as covariates, and PCH1–7 simultaneously as predictors.
To determine whether PCH convey novel information over composite phenotypes created based on phenotypic correlations, we conducted parallel PCA analyses with conduct problem, depression, and temperament variables and examined associations with alcohol outcomes. All raw observations were standardized, adjusted for sex and age, and re-standardized. Scaled residuals of identical pairs of T1 and T2 conduct problem and depression items were combined using the princomp function in R. The first PCA for each pair of identical items was retained (item-PCA; analogous to phase 1 of PCH analyses). Because phase 2 (clustering of variables) was not essential for PCA estimation, we calculated final-PCA using princomp in R based on item-PCA and temperament items. Heritabilities and prediction of alcohol outcomes for final-PCA were estimated using the same procedures described for final-PCH.
Participants were, on average, 12.8 and 14.2 years old at T1 and T2 (N=486) and 20.5 years old when reporting on alcohol outcomes (N=318–327). Forty-eight percent identified as female. The majority of participants identified as “White/not Hispanic” (58%) or “Hispanic” (27.8%; Table 1). On average, their parents’ highest level of education achieved was “some college.” Participants were nested within 193 three-generational families (53.8% familial AUD). Each three-generational (extended) family (i.e., including grandparental, parental, and grandchild generations) had on average 9.8 members (range=4–57) and 2.5 third-generation children (range=1–24; i.e., participants who included siblings, half-siblings, and cousins). Those included versus excluded did not significantly differ on study variables after multiple testing corrections.
On average, participants endorsed 5.0 (SD=4.1) of 17 conduct problems at T1 and 4.0 (SD=4.1) at T2, with the most frequently endorsed at both assessments being, “argue a lot” and “stubborn.” On average, participants endorsed 4.2 (SD=3.5) of 17 depression symptoms at T1 and 3.3 (SD=3.6) symptoms at T2, with the most frequently endorsed at both assessments being, “secretive/keep to myself” and “tired.” At the substance use assessment, 44% of participants had experienced at least one lifetime alcohol problem and 48% of participants drank heavily at least 1–2 times in their lives. Estimated heritabilities of three temperament subscales (inhibitory control, shyness, and affiliation) were not statistically significant (p>0.05) and these variables were dropped (Table S6: Preliminary h2).
PCH analyses proceeded in three phases as outlined in the Methods section and visually depicted in Figure 1.
After estimation of item-PCH from T1 and T2 conduct problem and depression items pairs, we found that, for the majority, responses at T2 were weighted more heavily than responses at T1 (Table S7). Expected heritabilities for item-PCH were always higher than for T1 or T2 items alone.
Based on estimated variance components for item-PCH and temperament variables, we found that observed heritabilities of item-PCH (Table S8: h2 (Phase 2)) were similar to expected heritabilities calculated in phase 1. After performing cluster analysis of their genetic correlation matrix, we produced 13 clusters, each consisting of items whose average pairwise correlations ranged from 0.47–0.88. Eight clusters combined across conduct problems and depression, three across depression and temperament, and two across conduct problems and temperament (Figure 2). A leading PCH was estimated from the variance components of within-cluster items to yield 13 cluster-PCH. Observed heritabilities of cluster-PCH ranged from 0.46–0.67 (Table S9: h2 clusters (Phase 2)).
Using variance components across cluster-PCH to estimate the final PCH, we derived thirteen PCH (Table 3). Seven surpassed the heritability threshold of 0.50. These final PCH were used in subsequent analyses (Table S9: Final PCH weights (Phase 3) and h2 final PCH (Phase 3)). Heritabilities of final PCH ranged from 0.55–0.93. Unlike the original 39 variables, distributions of final PCH were continuous and roughly normal (Supplementary Text). Estimated pairwise correlations between PCH were essentially zero.
To guide interpretation of final PCH, we examined how they were associated with conduct problem and depression item-PCH and temperament subscales (Figure 3). Items correlated with final PCH >0.20 were given highest relative importance for interpretation. Higher levels of PCH1 (h^2^=0.93) were correlated with items across all domains, including conduct problems (e.g., “don’t feel guilty,” “swear”), depression (e.g., “tired,” “don’t want to move”) and riskier temperament features (e.g., lower activational and attentional control; hereafter referred to as dysregulated co-occurring problems). Although PCH2 (h^2^=0.89) was correlated with conduct problems (e.g., “rather hang out with older kids,” “try to get a lot of attention”) and riskier temperament features (e.g., sensation seeking), it was also correlated with several adaptive temperament traits (e.g., higher attentional and activational control and lower frustration; hereafter referred to as sensation seeking, proactive conduct problems). Higher levels of PCH3 (h^2^=0.87) were correlated with features across all domains including conduct problems (e.g., “rather hang out with older kids,” “moods change suddenly”), depression (e.g., “rather be alone,” “feel too guilty”), and riskier temperament (greater frustration, high-intensity pleasure, and lower attentional control; hereafter referred to as negative affect related co-occurring problems). Higher levels of PCH4 (h^2^=0.75) were predominantly correlated with conduct problems (e.g., “hang around kids who get in trouble,” “lie/cheat”) and lower activational control (i.e., temperament), as well as one depression item, “trouble sleeping” (hereafter referred to as defiant conduct problems). Higher levels of PCH5 (h^2^=0.74) were predominantly correlated with higher levels of conduct problems (e.g., “swear,” “stubborn”; hereafter referred to as verbal aggression conduct problems). Higher levels of PCH6 (h^2^=0.68) were correlated with conduct problems (e.g., “mean,” “swear”) and depression (e.g., “sleep more than most kids,” “overtired”), but less so with temperament (hereafter referred to as sleep dysregulated co-occurring problems). Higher levels of PCH7 (h^2^=0.55) were correlated with conduct problems (e.g., “hot temper,” “stubborn”), depression (e.g., “secretive,” “sad/empty”) and temperament traits of higher frustration, lower attentional control and lower sensation seeking (hereafter referred to as negative affect, risk-averse co-occurring problems).
In zero-inflated Poisson regression, it is assumed there are two populations of participants, individuals who are “not at risk” and those who are at risk for alcohol problems or heavy drinking. The count of the subjects not at risk is assumed to follow a binomial distribution; this estimate of structural zeroes also establishes the fraction of participants not at risk. For those at risk, the realization of their count of alcohol problems or heavy drinking is assumed to follow a Poisson some subjects have a random realization of zero, others have one or more problems, as reflected by the estimated Poisson parameter. For participants of average age at assessment of their alcohol problems and with average levels of all PCHs, we estimated that about 88.9% of the observed zeroes for alcohol problems were structural, or true, zeroes and the remaining 11.1% of observed zeroes were random zeroes analyzed as part of the “at risk group.” Similarly, under the same parameters, about 93% of the observed zeroes for heavy drinking were structural zeroes and the remaining 7% were random zeroes.
We found that higher levels of dysregulated co-occurring problems (PCH1) and defiant conduct problems (PCH4) predicted a greater probability of being at risk (versus not at risk) for heavy drinking and higher levels of negative affect-related co-occurring problems (PCH3) and sleep dysregulated co-occurring problems (PCH6) predicted a greater probability of being at risk (versus not at risk) for alcohol problems. Higher levels of dysregulated co-occurring problems, defiant conduct problems, and verbal aggression conduct problems (i.e., PCH1, PCH4, and PCH5) predicted greater numbers of alcohol problems among those at risk (Table 3).
The early adulthood alcohol problems (N=318) and heavy drinking (N=327) variables had smaller sample sizes relative to the number of people for whom PCH were estimated (N=486). A larger sample size could be generated by including alcohol variables for individuals younger than 18 years (N=438 for alcohol problems; N=435 for heavy drinking). To determine the effect of excluding individuals who were younger than the age of 18, we conducted parallel analyses including them. Results were very similar, with the exception that PCH6 no longer predicted a greater probability of being at risk for alcohol problems (Supplementary Text).
Given our sampling framework, we sought to understand whether the participants differed substantially from a nationally representative sample in terms of conduct problems and depression. We found that our sample showed similar levels of endorsement for conduct problems and depression symptoms relative to the National Household Survey on Drug Use Abuse, 1994 (United States Department Of Health And Human Services, 1997; Figure S4 in the Supplementary Text). Moreover, we found that grandparent AUD, and therefore ascertainment, did not account for substantial proportions of the variance in key study variables, including the final estimated PCH and alcohol outcomes, when examined using linear models (R^2^=0.002–5.8%; Table S6: Familial AUD). Still, due to its ascertainment two generations ago, as well as its ethnic and socio-economic composition, heritability estimates presented here may be unique to the current study population.
After deriving principal components (PCs) from conduct problem, depression, and temperament variables, we examined the top seven PCs to mirror PCH results. Higher levels of PCA1 and PCA2 predicted greater numbers of alcohol problems among those “at risk” and PCA1 predicted a greater probability of being at risk for alcohol problems. Higher levels of PCA6 predicted a greater probability of being at risk for heavy drinking (Table S6: PCA ALC).
PCs and PCH showed several differences. PCs were moderately heritable (0.40–0.56; Table S6: h2 PCA), whereas PCH were highly heritable (0.68–0.93). Correlations among PCs and PCH ranged from nearly zero (0.01) to moderate, suggesting they captured different phenotypes (0.55; Table S6: r PCH PCA). The patterns of correlations between input variables and PCs were generally not similar to the patterns of correlations between input variables and PCH (Supplementary Text). Thus, combining variables based on phenotypic similarity (PCA) vs. genetic similarity (PCH) yields different information.
Shared family environment is a potential confounder for our estimates of PCH, with a specific concern that estimates could be biased upwards due to shared environment. Under such a scenario, we reasoned that all heritability estimates of related phenotypes that were obtained from our sample would be biased similarly. To evaluate this potential confounder, we compared the heritability of conduct problems and depression as estimated by prior behavioral genetic studies (e.g., twin, adoption designs) to similar constructs in our sample, namely, the DSM-oriented conduct problem and depression subscales within the Youth Self Report (Achenbach & Rescorla, 2001). In our sample, the heritability of conduct problems was 0.48(0.15), which was quite close to the heritability of conduct problems estimated for children (h^2^=0.46) and adolescents (h^2^=0.43) in a meta-analysis of twin and adoption studies (Rhee & Waldman, 2002). Similarly, the heritability estimate for depression symptoms in our sample was 0.13(0.15; Mage=12.8, range=11–17.8 years). This heritability estimate aligns with the heritability of depression symptoms from a prior twin study of young children, which found 0.03 (CI: 0.00–0.10) for 12 year olds, and is smaller than estimates for older ages, 0.37 (CI: 0.18–0.47) for 14 year olds and 0.45(CI: 0.35–0.54) for 16 year olds (Lamb et al., 2010).
Although it has been established that conduct problems and depression are genetically correlated (Hannigan et al., 2018; O’Connor et al., 1998; Subbarao et al., 2008), it remains unclear how this critical information can be translated for more practical purposes. Specifically, we sought to examine whether there was shared genetic variation underlying specific symptoms of conduct problems and depression and dimensions of temperament, because this could provide more clinically pertinent information about a high-risk group of youth. We found seven highly heritable (h^2^=>0.50) composite phenotypes (PCH) that blended across multiple construct categories. Moreover, we found that these composite phenotypes differed in content and in their prediction of alcohol outcomes relative to parallel composite phenotypes derived using PCA.
It is useful to compare results from PCH analyses to those from PCA, because PCA composites are more reflective of the constructs used in most studies on adolescent psychopathology (i.e., derived from phenotypic correlations). We found that PCH and PCA composites clearly represented different constructs, as correlations among them were moderate at best. Moreover, the seven most heritable PCH blended features across more than one construct category. This shows that features of conduct problems, depression, and temperament come together based on genetics in ways that do not reflect the original constructs’ boundaries. In contrast, the majority of PCA composites (four of seven) were represented by symptoms aligning with conduct problems only, depressive symptoms only, or temperament only, mirroring common conceptualizations of psychopathology. Thus, PCH analyses could yield novel phenotypes that provide new etiologic insights in future RDoC-oriented research because their constituent features are selected based on common genetic influences. This could represent an interesting new direction for RDoC-related research. Indeed, researchers have noted that RDoC research still heavily relies on symptom profiles derived from traditional diagnostic categories, which do not capture the level of complexity that the framework intended (Ostlund et al., 2021). Future researchers could apply the PCH method to other pedigree-based datasets or large datasets with measured genetics.
We also found that four PCH composite phenotypes represented co-occurring conduct problems and depression, whereas only one PCA composite represented this co-occurring phenotype. Thus, unlike results from PCA analyses, phenotypes derived from PCH analyses show how individuals with co-occurring conduct problems and depression can have differing etiologies. Our results extend the literature, which has generally considered co-occurring problems as a unidimensional phenotype when attempting to understand etiologic risk factors. As co-occurring problems have been linked to particularly severe adulthood psychopathology and impaired functioning (Diamantopoulou et al., 2011; Fanti & Henrich, 2010), our results importantly call for a greater understanding of the distinctiveness of these co-occurring problem phenotypes. Along the same vein, several different PCH composites represented conduct problems. Taken together, results show that PCH analyses can parse heterogeneity inherent within traditional diagnostic categories while also accounting for comorbidity among disorders, which are both known obstacles in understanding the etiology of mental health problems (Insel, 2014).
A greater number of PCH composite phenotypes also predicted alcohol outcomes (five) relative to PCA composites (three). This suggests that reorganizing symptoms based on shared genetic influences could increase understanding of unique pathways to alcohol problems. Thus, relative to phenotypically derived phenotypes, PCH composites could more specifically pinpoint characteristics that create the greatest risk for problematic outcomes. This again illustrates the potential value of PCH analyses in RDoC-oriented research, in which one of the central goals is to increase the specificity of prognosis.
Finally, it is worthwhile to discuss the factors that may drive different solutions derived from PCH analyses versus PCA. Although it is not possible to know definitively, some of the difference likely traces to how shared family environment influences covariation. PCA seeks to identify the largest components of variation from the covariance structure of features (items), whatever the source of variation. Nuclear family environment and shared genetics would be natural drivers. By contrast, PCH analyses requires that the key axes of phenotypic variation align with the key axes of variation in pedigree relationships (full sibs, half sibs, and cousins). For these reasons, we suspect that the differences observed between PCA and PCH composite traits most likely reflect the impact of shared family environment. Additional support for this conjecture comes from post-hoc heritability analyses of conduct problems and depressive symptoms in our sample of extended families. Neither heritability estimate was larger than that found for comparable phenotypes in prior twin and adoption studies, suggesting that family environment is not a major confounder for our results.
Further understanding the nature of the PCH composites discovered in this study could provide insights into future directions for etiological research. Regarding heritable composite phenotypes reflecting co-occurring symptomatology, we found that conduct problems and depressive symptoms loaded highly onto four of the seven PCH composites (PCH1, PCH3, PCH6, PCH7; Figure 3). The most highly heritable composite phenotype (PCH1) reflected conduct problem symptoms involving antagonism towards others and depressive symptoms related to lethargy that may be driven by low levels of effortful control (i.e., dysregulated co-occurring problems). This PCH composite predicted a greater number of alcohol problems among those at risk and it also predicted those who were at risk (vs. not at risk) for heavy drinking. Thus, youth who exhibit symptoms encompassed by the dysregulated co-occurring problems composite phenotype may be prime targets for alcohol prevention or intervention.
The two next most heritable “co-occurring problem” phenotypes predicted a relatively less severe alcohol outcome than dysregulated co-occurring problems, that is, being at risk (vs. not at risk) for alcohol problems. Thus, individuals with high levels of these composite phenotypes may also be worthwhile candidates for alcohol prevention. The first of these composite phenotypes reflected depressive symptoms relating to sleep problems and conduct problems and was not strongly related to any temperament traits (sleep dysregulated co-occurring problems). The second of these composite phenotypes reflected many conduct problem, depression and temperament features directly related to negative affectivity (negative-affect related co-occurring problems; e.g., “moods change suddenly,” “feels too guilty,” high levels of frustration). Finally, the least heritable co-occurring problems phenotype (negative affect, risk averse co-occurring problems) also reflected many conduct problem, depression, and temperament features related to negative affectivity but, perhaps because it uniquely reflected low levels of sensation seeking, was not related to any alcohol outcomes.
The presence of several unique composite phenotypes underlying co-occurring conduct problem and depressive symptoms may have clinical implications and certainly provide directions for future research. There currently exist few standardized protocols for treating co-occurring problems. Thus, providers practicing in the community tend to focus on treating externalizing symptomatology among youth who present with co-occurring problems (Milette-Winfree & Mueller, 2018). Our findings highlight how this common approach to treatment may not be effective for many youths who present with co-occurring conduct problems and depression given the presence of unique types of this condition. Moreover, knowledge of these novel composite phenotypes could potentially clarify tailored treatment targets based on their distinct etiologic or temperamental risk factors. For example, individuals exhibiting high levels of the dysregulated co-occurring problems phenotype may benefit more from interventions targeting self-control, whereas sleep may be a better target for individuals with high levels of the sleep dysregulated co-occurring problems phenotype. Similarly, individuals exhibiting high levels of either of the negative-affect related co-occurring problems phenotypes (i.e., risk-averse vs. not risk-averse) would likely benefit from interventions to control negative emotions. Results also suggest that it may be possible to more precisely identify youths who need alcohol prevention relative to current approaches, which commonly base these decisions solely on the presence of certain symptoms or traits, like internalizing, externalizing, or impulsive behaviors. It will be important to replicate these composite phenotypes in independent datasets and conduct further research on their early correlates before such steps are taken.
In addition to heritable composite phenotypes reflecting co-occurring conduct problems and depression, we found three additional composite phenotypes that reflected conduct problem and temperament features only. One of these composite phenotypes was characterized by conduct problems reflecting defiance against authority (defiant conduct problems) and predicted both a greater number of alcohol problems and being at risk (vs. not at risk) for heavy drinking. This finding is consistent with prior research showing that there was a distinct rule-breaking genetic factor underlying conduct problems and that this factor was most strongly associated with alcohol dependence relative to other factors underlying conduct disorder symptoms (Kendler, Aggen, & Patrick, 2013). The second of these composite phenotypes reflected verbal aggression symptoms of conduct problems and it predicted a greater number of alcohol problems. The third reflected conduct problem symptoms that may be instrumental due to high levels of effortful control and high levels of sensation seeking (sensation seeking, proactive conduct problems). Interestingly, this composite phenotype was not related to any alcohol outcomes. Similar to prior research, our results show that there exist different latent genetic factors that underlie symptoms of conduct disorder and that they may portend unique future outcomes (e.g., Kendler, Aggen, & Patrick, 2013).
This study has limitations to acknowledge. We found no heritable composite phenotypes primarily represented by depressive symptoms. This may owe to our examination of inheritance patterns in families oversampled for AUD, which is most tightly linked to conduct disorder in adolescence (Chassin et al., 2013), or to the young average age of participants. Results may not be generalizable to samples not ascertained on AUD or to more racially or ethnically diverse populations, although our sample showed similar endorsement rates for conduct problems and depression relative to a nationally representative survey. Ideally, we would have used larger and more pedigrees in analyses to show that results are replicable and are not overfitted to the data. The outcome measure, heavy drinking, was measured slightly differently for participants depending on their sex and selected substance use assessment. However, results were similar when using a heavy drinking measure that was the same for all participants (Supplementary Text). Unfortunately, we also did not have prospective prediction of alcohol problems as we used a lifetime measure. Relating to measurement, we only used adolescents’ self-reports in this study. Although this could introduce reporter bias as an explanation for results, we decided to use adolescents’ reports because they may be more accurate for capturing covert conduct behaviors and internal depressive symptoms (Thomas et al., 1990).
Our study also assumes segregation of phenotypes in families is due to Mendelian assortment of genetic variation; because of our study design, however, it is possible that genetic effects are partially confounded with shared environmental effects. Thus, like other genetically informative designs such as SNP-based studies (Grotzinger & Keller, 2022) and twin studies (Felson, 2014), we depend on a set of assumptions that might not be a perfect reflection of the study sample, resulting in somewhat inaccurate genetic estimates. For this reason, it will be important for future research to evaluate whether these composite phenotypes continue to be heritable when tested using other study designs. Although it was beyond the scope of this study to examine whether PCH differed by sex, we also plan to examine this in future work.
In conclusion, results shed light on highly heritable, novel collections of conduct problem, depression, and temperament features in adolescence that showed unique etiological underpinnings and differentially predicted alcohol outcomes. Our novel analyses demonstrate the value of conceptualizing mental health disorders in the context of fundamental neurobehavioral systems, such as temperament, which is key to the mission of RDoC (Insel, 2014). Findings highlight the utility of re-organizing psychiatric symptoms based on their shared genetic influences.