Introduction
Externalizing (EXT) problems, such as aggression, rule-breaking behaviors, and harmful substance use, are characterized by impulsivity and behavioral undercontrol (Krueger et al., Reference Krueger, Hicks, Patrick, Carlson, Iacono and McGue2002). EXT problems in adolescence are prospectively associated with negative life outcomes, such as financial and legal difficulties (Gornik, Clark, Durbin, & Zucker, Reference Gornik, Clark, Durbin and Zucker2022) and emotional problems (Hamdi & Iacono, Reference Hamdi and Iacono2014). While genes are known to play a prominent role (Barr & Dick, Reference Barr, Dick, de Wit and Jentsch2019), EXT problems are also influenced by environmental factors (Fortuin, van Geel, & Vedder, Reference Fortuin, van Geel and Vedder2015; Pinquart, Reference Pinquart2017; Samek & Hicks, Reference Samek and Hicks2014). Furthermore, genetic and environmental factors for EXT problems are closely intertwined, as genetic risks for EXT are often correlated with psychosocial environments (i.e. gene-environment correlation, rGE) (Elam, Lemery-Chalfant, & Chassin, Reference Elam, Lemery-Chalfant and Chassin2023; Kuo et al., Reference Kuo, Salvatore, Barr, Aliev, Anokhin, Bucholz, Chan, Edenberg, Hesselbrock, Kamarajan, Kramer, Lai, Mallard, Nurnberger, Pandey, Plawecki, Sanchez-Roige, Waldman, Palmer and Dick2021). Emerging studies have begun to show that genetic effects on EXT problems may be mediated by environments (Dong, Sasia, & Li, Reference Dong, Sasia and Li2026; Li, Reference Li2019), highlighting the role of psychosocial influences in what are putatively believed to be genetic effects.
That said, the very notion of environmentally mediated genetic effects is complicated, in part because it may also reflect rGEs driven by confounders, such as population stratification, assortative mating, or parental genotypes. While these particular confounds are notoriously challenging to control for in large population-based designs (Pingault et al., Reference Pingault, Allegrini, Odigie, Frach, Baldwin, Rijsdijk and Dudbridge2022), a family-based design provides one robust way to account for these confounds. For example, researchers can estimate rGE effects within families by comparing twins and full-siblings, who are born to the same household, have the same ancestral backgrounds, and are influenced by similar family environments. They also randomly inherit genotypes from the same parents during meiosis, such that PGS variances within a family are independent of environmental confounders. Therefore, family-level confounding effects are mostly accounted for when estimating rGE within families (see Figure 1) (McAdams, Rijsdijk, Zavos, & Pingault, Reference McAdams, Rijsdijk, Zavos and Pingault2021; Selzam et al., Reference Selzam, Ritchie, Pingault, Reynolds, O’Reilly and Plomin2019; Tanksley et al., Reference Tanksley, Brislin, Wertz, de Vlaming, Courchesne-Krak, Mallard, Raffington, Karlsson Linnér, Koellinger, Palmer, Sanchez-Roige, Waldman, Dick, Moffitt, Caspi and Harden2025; Zhou et al., Reference Zhou, Gidziela, Allegrini, Cheesman, Wertz, Maxwell, Plomin, Rimfeld and Malanchini2025). rGEs that persist within families may offer stronger support for a conceptualization where social environments mediate genetic effects on EXT development, whereas rGEs that substantially attenuate when estimated within families may indicate the presence of confounding. As such, our study aims to test whether rGEs for EXT problems are potentially robust to family-level confounds by leveraging a family-based design.
Explanations of rGE between PGS and social environments.
Figure 1. Long description
The flowchart consists of a central column of factors with arrows pointing outward to P G S on the left and Social Environment on the right.
At the top center is a large box labeled Confounding effects that vary between twin and full sibling pairs. Inside this box are three sub-boxes stacked vertically:
1. Parental genotypes and phenotypes e.g. parent personality. Arrows point from this box to both P G S and Social Environment.
2. Population stratification and related confounders e.g. socioeconomic status. Arrows point from this box to both P G S and Social Environment.
3. Assortative mating induces correlations between maternally and paternally inherited genomes. Arrows point from this box to both P G S and Social Environment.
Below this is a second large box labeled Causal effects that vary within and between twin and full sibling pairs. Inside this box is one sub-box:
* Mediating phenotypes e.g. externalizing behaviors, temperament, etc. An arrow points from P G S to this box, and another arrow points from this box to Social Environment.
The P G S box on the left and the Social Environment box on the right serve as the terminal points for all directional arrows.
Gene-environment correlations in the EXT literature
Twin studies have shown that 50%–80% of the variation in EXT problems can be attributable to genetic factors within a given population (Hicks et al., Reference Hicks, Krueger, Iacono, McGue and Patrick2004; Krueger et al., Reference Krueger, Hicks, Patrick, Carlson, Iacono and McGue2002). More recently, the field has moved toward identifying loci associated with EXT problems at the individual level through genome-wide association studies (GWASs). A GWAS performed by Karlsson Linnér et al. (Reference Karlsson Linnér, Mallard, Barr, Sanchez-Roige, Madole, Driver, Poore, de Vlaming, Grotzinger, Tielbeek, Johnson, Liu, Rosenthal, Ideker, Zhou, Kember, Pasman, Verweij, Liu and Dick2021) examined the association of single nucleotide polymorphisms (SNPs) with EXT in a pooled sample of 1.5 million individuals, in which a latent genetic factor for EXT was estimated using GWAS summary statistics for problematic alcohol use, lifetime smoking initiation, lifetime cannabis use, general risk tolerance, attention-deficit/hyperactivity disorder, age at first sex, and number of sexual partners. The effects of all SNPs were then aggregated into a polygenic score (PGS), which is calculated as a weighted sum of the dosages of effect alleles an individual carries across their genomes (Purcell et al., Reference Purcell, Neale, Todd-Brown, Thomas, Ferreira, Bender, Maller, Sklar, de Bakker, Daly and Sham2007). A PGS for EXT computed from the Karlsson Linnér study was found to outperform PGSs based on other studies in capturing genetic risks for a spectrum of EXT problems (Deng et al., Reference Deng, Belisario, Doggett, Pigeyre, Pare, Munafò and MacKillop2025), explaining up to 10% of the variance in an EXT factor and 1%–7% of the variance in specific EXT traits across two held-out samples.
While a GWAS is meant to identify primarily genetic variances that are associated with the trait in question (e.g. EXT problems), EXT PGS computed from GWAS signals consistently show a high degree of correlation with environments (Dong et al., Reference Dong, Sasia and Li2026; Kretschmer et al., Reference Kretschmer, Vrijen, Nolte, Wertz and Hartman2022; Kuo et al., Reference Kuo, Salvatore, Barr, Aliev, Anokhin, Bucholz, Chan, Edenberg, Hesselbrock, Kamarajan, Kramer, Lai, Mallard, Nurnberger, Pandey, Plawecki, Sanchez-Roige, Waldman, Palmer and Dick2021), suggesting that GWAS signals may reflect both genetic and environmental variances. rGE may reflect active and evocative effects, where a higher PGS may lead individuals to select into or evoke certain environmental responses. For example, adolescents with higher genetic risks for EXT may affiliate with more substance-using peers (Elam et al., Reference Elam, Chassin, Lemery-Chalfant, Pandika, Wang, Bountress, Dick and Agrawal2017) or elicit harsher parenting responses (Yan, Ansari, & Peng, Reference Yan, Ansari and Peng2021). Meta-analyses of twin studies have indeed shown that many psychosocial environments share a substantial proportion of heritable variance with EXT problems (Hicks et al., Reference Hicks, South, DiRago, Iacono and McGue2009; Kendler & Baker, Reference Kendler and Baker2007; Klahr et al., Reference Klahr, Thomas, Hopwood, Klump and Burt2013; Marceau et al., Reference Marceau, Horwitz, Narusyte, Ganiban, Spotts, Reiss and Neiderhiser2013). Further, these ‘heritable’ environments may in turn impact development down-the-stream, thereby mediating genetic effects on future EXT problems (Elam et al., Reference Elam, Lemery-Chalfant and Chassin2023; Kretschmer et al., Reference Kretschmer, Vrijen, Nolte, Wertz and Hartman2022). In summary, the presence of rGE may suggest that PGS for EXT not only reflects direct genetic effects, but also mediated effects via heritable environmental exposures (Kendler & Baker, Reference Kendler and Baker2007; Li, Reference Li2019; Pingault et al., Reference Pingault, Allegrini, Odigie, Frach, Baldwin, Rijsdijk and Dudbridge2022).
On the other hand, rGE may reflect passive, or confounding, effects. For example, population stratification arises when genetic associations are confounded by ancestry-related differences in the genotype (i.e. differences in allele frequencies). This could inflate correlations among otherwise independent alleles in the genome, even after controlling for genetic principal components (Elhaik, Reference Elhaik2022; Grinde et al., Reference Grinde, Browning, Reiner, Thornton and Browning2024). Assortative mating occurs when individuals select their partners based on similarity or dissimilarity regarding a phenotype. Parental genotypes may also represent a family-level confound in rGE because individuals inherit half of their genotype from each parent, such that apparent genetic associations may inadvertently capture the effect of the parents’ genotype as well. Collectively, these passive/confounding effects are often present in population-based designs and may inflate associations between genes and environmental exposures (Kong et al., Reference Kong, Thorleifsson, Frigge, Vilhjalmsson, Young, Thorgeirsson, Benonisdottir, Oddsson, Halldorsson, Masson, Gudbjartsson, Helgason, Bjornsdottir, Thorsteinsdottir and Stefansson2018; Price, Zaitlen, Reich, & Patterson, Reference Price, Zaitlen, Reich and Patterson2010).
Adolescent psychosocial environments
We focus on measured environments during adolescence because it is a developmental stage characterized by heightened neural plasticity and increasing personal autonomy, making it especially salient for studying rGE in the context of EXT problems (Baker, Galván, & Fuligni, Reference Baker, Galván and Fuligni2025; Sebastian, Viding, Williams, & Blakemore, Reference Sebastian, Viding, Williams and Blakemore2010). Genetic factors have been previously shown to influence how adolescents select, evoke, and respond to their environments in the context of EXT problems (Andersen, Steinberg, & Belsky, Reference Andersen, Steinberg and Belsky2021; Dick, Adkins, & Kuo, Reference Dick, Adkins and Kuo2016). The role of parenting has been extensively studied with respect to the development of EXT, with studies showing that higher genetic risk for EXT is associated with less supportive parenting (Pinquart, Reference Pinquart2017; Samek et al., Reference Samek, Hicks, Keyes, Bailey, McGue and Iacono2015). Peer effects are another dimension theorized and studied in relation to EXT. For example, delinquent peer affiliation has been found to be positively associated with genetic risks for EXT (Samek et al., Reference Samek, Goodman, Erath, McGue and Iacono2016). Researchers have also linked broader social contexts, such as school environments, to EXT genetic risks. For instance, Li (Reference Li2019) found that PGS for EXT problems predicted trajectories of substance use and antisocial behaviors both directly and indirectly through perceived school safety and belonging. Crucially, these environmental factors are themselves risk factors for maladaptive outcomes, including the development of EXT problems (Carroll, Klump, & Burt, Reference Carroll, Klump and Burt2023; Monahan, Rhew, Hawkins, & Brown, Reference Monahan, Rhew, Hawkins and Brown2014; Pinquart, Reference Pinquart2017).
Current study
The current study investigated rGEs between PGS for EXT problems and previously well-studied adolescent psychosocial environments in relation to EXT problems (i.e. parental relationships, school connectedness, and delinquent peer affiliation) using data from the National Longitudinal Study of Adolescent to Adult Health (Add Health). We first examined rGE effects in a population sample comprised of genetically unrelated individuals. Next, we estimated the same rGE effects at the within- and between-family levels using a family-based sample of twins and full-siblings. The within-family effects were estimated by comparing twins/siblings within the same families and therefore reflect rGEs when controlling for family-level confounds. The between-family effects were derived from comparisons of family means and therefore resemble estimates of rGE in the general population.
We hypothesized that: (1) in the population sample (i.e. unrelated individuals), higher PGS will be correlated with poorer parental relationships and school connectedness and greater delinquent peer affiliation, (2) at the between-family level (i.e. comparisons between sets of twins/siblings), twins/siblings from a family with higher average PGS will on average experience poorer parental relationships and school connectedness and greater delinquent peer affiliation, (3) at the within-family level (i.e. comparisons of twins/siblings within the same family), twins/siblings with higher PGS will experience poorer parental relationships and school connectedness and greater delinquent peer affiliation than their co-twin/sibling, and (4) within-family rGE effects, where confounding effects are controlled for, will be smaller in magnitude than between-family estimates.
Methods
Participants
Add Health is a longitudinal study of a large sample of youths from the United States who were in grades 7–12 during the 1994–1995 school year. Over 90,000 adolescents were initially sampled from the population through a stratified set of qualifying high schools. 20,745 adolescents were then randomly selected from the previous school sample and administered the in-home interview at Wave I, including oversamples from underrepresented populations. Participants were followed from adolescence to adulthood for five waves of data collection to date: Wave I (1994–1995, grades 7–12, N = 20,745), Wave II (1995–1996, grades 8–12, N = 14,738), Wave III (2001–2002, ages 18–26, N = 15,197), Wave IV (2007–2008, ages 24–32, N = 15,701), and Wave V (2016-2018, aged 33–43, N = 12,300). Information on attrition across waves was documented by Harris et al. (Reference Harris, Halpern, Whitsel, Hussey, Killeya-Jones, Tabor and Dean2019). The full Add Health sample at Wave I was comprised of 50.5% females and 61.45% who self-identified racial-ethnically as ‘White’, 23.17% ‘Black or African American’, 7.64% ‘Asian or Pacific Islander’, 3.57% ‘Native American’, 16.99% ‘Hispanic’, and 9.64% as ‘Other’.Footnote 1 Add Health also included more than 3,000 pairs of individuals of varying genetic relatedness living in the same household, including 289 monozygotic (MZ) twin pairs, 452 dizygotic (DZ) twin pairs, 1,251 full-sibling pairs, 442 half-sibling pairs, 201 cousin pairs, 408 unrelated pairs, and 96 other related pairs (i.e. aunt/uncle, niece/nephew, group homes, spouses, and in-laws).
Our study leverages two subsamples of Add Health participants. First, we formed our admixed population subsample from 7,635 genetically unrelated individuals. Then we formed a family-based subsample of 1,589 individuals from 777 families. These individuals are MZ/DZ twins and full-siblings who were born to and grew up in the same household (Harris, Halpern, Haberstick, & Smolen, Reference Harris, Halpern, Haberstick and Smolen2013).
Genotyping
Saliva samples were collected at Wave IV and genotyped using the Illumina Human Omni-1 Quad or the Omni-2.5 Quad BeadChip. Quality control procedures were conducted at the participant level (i.e. removing participants with call rates <90%, poor duplicate concordance, unclear sex chromosomes, identifier abnormalities) and the marker level (i.e. removing variants with call rate <90%, alignment uncertainty, tri-allelic status, duplicate discordance, ancestry-specific Hardy-Weinberg Equilibrium p < 5 × 10−5, and minor allele frequency <0.5%). Genotypes were then merged between arrays and imputed using the 1000 Genomes Phase 3 panel from the GLGC/GIANT protocol. Imputed data containing 9,664,514 markers were available for 9,974 participants, including 2,014 genetically related individuals from 975 families. Participants were partitioned into four genetic similarity groups based on genetic similarity to European, African, East Asian, and Hispanic ancestries, which is defined by distance from the mean of the first two genetic principal components (PCs), as described in Braudt and Harris (Reference Braudt and Harris2020).
Measures
Polygenic scores for EXT (PGSEXT)
PGSEXT was calculated based on a GWAS of EXT behaviors (Karlsson Linnér et al., Reference Karlsson Linnér, Mallard, Barr, Sanchez-Roige, Madole, Driver, Poore, de Vlaming, Grotzinger, Tielbeek, Johnson, Liu, Rosenthal, Ideker, Zhou, Kember, Pasman, Verweij, Liu and Dick2021) that has shown superior predictive performance relative to other GWASs (Deng et al., Reference Deng, Belisario, Doggett, Pigeyre, Pare, Munafò and MacKillop2025). PGSEXT was calculated using PRS-CS (Ge et al., Reference Ge, Chen, Ni, Feng and Smoller2019), which accounts for linkage disequilibrium and optimizes prediction by weighting individual SNP effects in GWAS summary statistics based on external reference panels using Bayesian continuous shrinkage priors. PRS-CS has been shown to have better predictive performance and generalizability across genetic ancestries (Ahern, Thompson, Fan, & Loughnan, Reference Ahern, Thompson, Fan and Loughnan2023). To further control for population stratification, PGSEXT was residualized on the first 20 genetic PCs that were estimated using GCTA (Yang, Lee, Goddard, & Visscher, Reference Yang, Lee, Goddard and Visscher2011) and standardized within each genetic similarity group (Price et al., Reference Price, Patterson, Plenge, Weinblatt, Shadick and Reich2006). To maximize sample sizes, we ran analyses with individuals from all genetic similarity groups while including genetic similarity groups as a dummy-coded covariate.
Parental relationships
Parent report. At Wave I, the same parent reported on their relationship with each adolescent in the household. Six items assessed the closeness, collaboration, trust, antagonism, and overall satisfaction of the relationship (Supplementary Table S1). Parent report of the parental relationship has satisfactory reliability with Cronbach’s α = 0.71.
Child report. At Wave I, participants self-reported on how they were parented as adolescents: five items measured aspects of supportive parenting by each resident parent, including perceived care, closeness, warmth, communication, and overall satisfaction of the relationship (Supplementary Table S2). The maternal and paternal scales both showed high reliability with Cronbach’s αs = 0.85 and 0.88, respectively.
Delinquent peer affiliation
At Wave I, three items assessed how many of the participant’s three closest friends used cigarettes, marijuana, and alcohol. Another item assessed how often the participant took part in a fight where their group of friends fought against another group of people in the past year. These four items (Supplementary Table S3) showed satisfactory reliability (α = 0.71).
School connectedness
At Wave I, six items assessed the participants’ feelings of belongingness, interpersonal connectedness, teacher support, and safety at school (Supplementary Table S4). These items were rated on the same five-point scale and showed high reliability (α = 0.77).
Covariates
Adolescent age and biological sex were cluster-mean centered and covaried out within-family. Family means of age and sex, parental education, family income, and dummy-coded genetic similarity groups were covaried out at the between-family level.
Analysis
Confirmatory factor analysis (CFA)
Psychosocial environments were measured as factor scores because they generally have better psychometric properties than composite scores (McNeish, Reference McNeish2023), are normally distributed, and better suit our analyses. A CFA model was estimated for each environmental variable in the entire genotyped sample using the WLSMV estimator. Item-level missing data were handled by pairwise deletion. Standardized factor scores were computed using the maximum a posteriori method.
Primary analysis
We used Bayesian Structural Equation Modeling (SEM) in Mplus (Muthén & Muthén, Reference Muthén and Muthén1998) with the Gibbs sampler and 30,000 iterations of Markov Chain Monte Carlo draws, which allows for probabilistic interpretations of the posterior distributions. Missing data were handled with Bayesian data augmentation in Mplus (Muthén & Asparouhov, Reference Muthén and Asparouhov2012). We elicited non-informative priors for all parameters because of a lack of empirical precedent. Model convergence was evaluated by visual examination of trace and autocorrelation plots for fixed effect parameters and posterior predictive checking of the chi-square statistic. Because Bayesian models yield posterior probability distributions, we report our results in terms of point estimates (i.e. median of the posterior) and their 95% credible intervals.
Psychosocial environments were modeled as correlated outcomes in the same Bayesian SEM. To test the first hypothesis, psychosocial environments were regressed on PGSEXT and covariates in the population sample (i.e. 7,635 unrelated individuals of Add Health), using multiple regression. The population-level effect therefore reflects the expected difference in psychosocial environments given a one-unit increase in PGSEXT among unrelated individuals in the general population.
To test the second and third hypotheses, a multilevel SEM was estimated with random intercepts for families using the family-based sample (1,589 twins/full-siblings from 777 families), where psychosocial environments were regressed on PGSEXT and covariates at both within- and between-family levels. At the between-family level, psychosocial environments were regressed on family means of PGS. The between-family effect therefore reflects the extent to which twins/siblings from a family with higher average PGSEXT experienced more/less of the environment. At the within-family level, PGS was centered around the family means to reflect a twin/sibling’s deviation from their family mean. The within-family effect therefore reflects the extent to which the twin/sibling with higher PGSEXT in a family experienced more/less of the environment relative to their co-twin/sibling.
Finally, to test the fourth hypothesis, the difference of the between-family effect from the within-family effect for each psychosocial environment was estimated as Δβ = β between-β within. This was evaluated at each Markov Chain Monte Carlo draw and aggregated into a posterior distribution, which allowed us to test whether within- and between-family effects were statistically different in the same multilevel model.
We used a Bayesian analytic framework with non-informative priors for all parameters, thus circumventing the need for multiple-testing correction (Sjölander & Vansteelandt, Reference Sjölander and Vansteelandt2019). We report all tested effects, including those considered null in the frequentist sense.
Results
Descriptive statistics
Table 1 shows descriptive statistics for the population sample and the family-based sample. Supplementary Tables 5 and 6 show descriptive statistics by pair type and genetic similarity grouping. Supplementary Table 7 shows missing data patterns for primary analysis (i.e. SEM).
Sample descriptive statistics
Table 1. Long description
The table compares two groups: Twins and full-siblings (N = 1,589 from 777 families) and Unrelated individuals (N = 7,635).
Key variables and statistics (Twins vs. Unrelated, p-value):
* Age at Wave 1: 16.10 (1.66) vs. 16.08 (1.72), p = 0.815.
* Female sex: 51.3% vs. 52.9%, p = 0.238.
* Genetic similarity group (European): 62.1% for both groups, p = 0.118.
* Self-identified race/ethnicity: Significant differences found in Black/African (18.6% vs. 22.0%, p = 0.012) and Other (8.2% vs. 5.6%, p = 0.001).
* Annual household income: $47,741 vs. $46,179, p = 0.721.
* Parental education: Distribution is similar across both groups, p = 0.988.
* P G S: -0.03 (0.99) vs. 0.00 (0.99), p = 0.210.
* Parent-reported relationship: -0.02 (1.01) vs. -0.01 (0.99), p = 0.979.
* Child-reported relationship: 0.04 (0.97) vs. -0.01 (1.01), p = 0.300.
* Delinquent peer affiliation: -0.04 (0.98) vs. 0.00 (1.00), p = 0.151.
* School connectedness: 0.04 (1.00) vs. 0.00 (1.00), p = 0.043.
I C C (intra-class correlation) for twins/siblings ranges from 26.5% for female sex to 100% for household income and parental education.
Note: Means are estimated for continuous variables with standard deviations in parentheses, while counts and percentages are shown for categorical variables. ICC, intra-class correlation (i.e. proportion of between-family variance over total variance of the variable). ICC is estimated for dichotomous variables by dummy coding. Parental relationships, delinquent peer affiliation, and school connectedness were measured as standardized factor scores estimated in the entire Add Health genotyped sample (i.e. population + related sample combined). Subsample comparison was performed by first randomly selecting one twin/sibling from each family and comparing these adolescents to the unrelated adolescents in the population sample. Continuous outcomes were compared using linear regression, binary outcomes using logistic regression, and categorical outcomes using chi-square test.
a ICCs = 100% because these variables are identical across twins/siblings within the same family.
b Sum of percentages larger than 100% because individuals were allowed to endorse multiple racial-ethnic identities.
c Parental education was measured as the higher of maternal and paternal education levels. In our analysis, ordinal categories of parental education were modeled as a continuous variable.
CFA models
CFA models achieved satisfactory fit for parent-reported (N = 8,736; CFI = 0.991; TLI = 0.986; RMSEA = 0.044) and child-reported parental relationship (N = 9,796; CFI = 0.988; TLI = 0.981; RMSEA = 0.082), delinquent peer affiliation (N = 9,934; CFI = 0.999; TLI = 0.997; RMSEA = 0.029), and school connectedness (N = 9,942; CFI = 0.984; TLI = 0.966; RMSEA = 0.086). The factor structures were consistent with our expectation. For child-reported parental relationships, two highly correlated factors emerged for maternal and paternal relationships, and the prorated mean of maternal and paternal factor scores was therefore used to measure the average quality of parental relationships in the household. Details regarding CFA models can be found in Supplementary Tables S1–S4.
Primary analysis
Figure 2 shows the association between PGSEXT and environmental variables estimated in the population sample and at the within- and between-family levels in the family-based sample.
Estimates of PGS effects on social environments. Note: Point estimates and 95% credible intervals at the within-family level, between-family level, and in the population (i.e. unrelated) sample were shown. P, parent-reported. C, child-reported. *, 95% credible interval does not contain the null, or is statistically significant in the frequentist sense.
Figure 2. Long description
A forest plot with a vertical dashed line at zero on the X-axis, which ranges from negative 0.2 to 0.2. The Y-axis lists four paths from top to bottom. Each path has three data points with error bars: Within Family (red circle), Between Family (blue triangle), and Population (green square). Asterisks indicate statistical significance.
* P G S yields Parental Relationship - P: All three estimates are negative and significant. Within Family is negative 0.091, Between Family is negative 0.096, and Population is negative 0.093.
* P G S yields Parental Relationship - C: Within Family is negative 0.096 (significant). Between Family is negative 0.044. Population is negative 0.049 (significant).
* P G S yields Delinquent Peer Affiliation: Within Family is 0.008 (near zero). Between Family is 0.193 (significant positive). Population is 0.113 (significant positive).
* P G S yields School Connectedness: Within Family is 0.050. Between Family is negative 0.051. Population is negative 0.070 (significant negative).
Population-level effects
Higher PGSEXT was associated with less supportive parental relationship (parent-reported: β = −0.093, 95% Credible Interval [CI] = −0.116 ~ −0.070; child-reported: β = −0.049, 95% CI = −0.071 ~ −0.027), higher levels of delinquent peer affiliation (β = 0.113, 95% CI = 0.092 ~ 0.134), and lower school connectedness (β = −0.070, 95% CI = −0.093 ~ −0.048) in the population sample.
Between-family effects
The between-family effects of PGSEXT in the family-based sample are largely consistent with the estimates in the population sample in terms of effect sizes. Higher PGSEXT is significantly associated with a less supportive parental relationship by parent report (β = −0.096, 95% CI = −0.171 ~ −0.020), and higher delinquent peer affiliation (β = 0.193, 95% CI = 0.130 ~ 0.256). The association of PGSEXT with child-reported parental relationship (β = −0.044, 95% CI = −0.106 ~ 0.018) and school connectedness (β = −0.051, 95% CI = −0.117 ~ 0.014) did not reach statistical significance, although the estimates were in the expected direction. The Bayesian posterior probability was also quite high for a negative association of PGSEXT with child-reported parental relationship and school connectedness (i.e. 91.5% and 93.8%, respectively).
Within-family effects
A robust within-family effect of PGSEXT on parental relationships was observed across informants: parents report less supportive parental relationships with the twin/sibling with higher PGSEXT (β = −0.091, 95% CI = −0.172 ~ −0.010) in the family; and the twin/sibling with higher PGSEXT was more likely to report less supportive parental relationships with their parents (β = −0.096, 95% CI = −0.184 ~ −0.008) relative to their co-twin/sibling. However, PGSEXT was not significantly associated with delinquent peer affiliation (β = 0.008, 95% CI = −0.072 ~ 0.087) or school connectedness (β = 0.050, 95% CI = −0.041 ~ 0.142).
Between-to-within differences
Between- and within-family effects were not significantly different for parent-reported (Δβ = −0.005, 95% CI = −0.115 ~ 0.106) and child-reported parental relationship (Δβ = 0.052, 95% CI = −0.056 ~ 0.160) and for school connectedness (Δβ = −0.101, 95% CI = −0.214 ~ 0.011). However, there was a significant attenuation in rGE from between- to within-family with respect to delinquent peer affiliation (Δβ = 0.185, 95% CI = 0.084 ~ 0.286).
Discussion
By comparing a family-based design alongside a population sample, the current study explored whether rGEs observed between PGSEXT and measured adolescent psychosocial environments are robust to family-level confounds, such as population stratification, assortative mating, and parental genotypes. At the population level, we found robust correlations between PGSEXT and all measured environmental variables (i.e. parent- and child-reported parental relationships, delinquent peer affiliation, and school connectedness). When we estimated rGE at the between-family level, rGEs were generally consistent with those found in the population sample. When examined at the within-family level, however, rGEs were only observed with parental relationships but no other environments. We also found significant attenuation in rGE for delinquent peer affiliation, such that its within-family effect was much smaller than between-family effect. Our results provide new insights into the role confounders play in often-observed but still partially understood rGEs in the psychiatric and behavioral genetics literature.
First, and as expected, we found that rGEs between PGSEXT and psychosocial environments are quite ubiquitous when examined at the population level and mostly present when examined at the between-family level. This is also consistent with previous studies that identified rGE using similar PGS and psychosocial environmental variables (de la Paz et al., Reference de la Paz, Mooney, Ryabinin, Neighbor, Antovich, Nigg and Nikolas2023; Kuo et al., Reference Kuo, Salvatore, Barr, Aliev, Anokhin, Bucholz, Chan, Edenberg, Hesselbrock, Kamarajan, Kramer, Lai, Mallard, Nurnberger, Pandey, Plawecki, Sanchez-Roige, Waldman, Palmer and Dick2021; Li, Reference Li2019). We suspect that the slight differences between our findings at the population level and those at the between-family level may be due to demographic differences and smaller sample sizes in the family-based subsample (see Table 1).
When rGEs were examined at the within-family level, we only found rGEs with respect to parental relationships: the twin/sibling with a higher PGSEXT was more likely to have a less supportive parental relationship than their co-twin/sibling. This finding was replicated by both self- and parent-reported data, suggesting that this result is robust to reporter bias. There was also no evidence of confounding, as the between- and within-family effects were not significantly different for parental relationships. One explanation is that PGSEXT may contribute to the development of certain traits and characteristics (e.g. behavior problems, temperament) in adolescents, which in turn evokes less supportive parental relationships (de la Paz et al., Reference de la Paz, Mooney, Ryabinin, Neighbor, Antovich, Nigg and Nikolas2023; Elam et al., Reference Elam, Chassin, Lemery-Chalfant, Pandika, Wang, Bountress, Dick and Agrawal2017; Klahr et al., Reference Klahr, Thomas, Hopwood, Klump and Burt2013; Kretschmer et al., Reference Kretschmer, Vrijen, Nolte, Wertz and Hartman2022; Kuo et al., Reference Kuo, Poore, Barr, Chirico, Aliev, Bucholz, Chan, Kamarajan, Kramer, McCutcheon, Plawecki and Dick2022; Marceau et al., Reference Marceau, Horwitz, Narusyte, Ganiban, Spotts, Reiss and Neiderhiser2013; Wang et al., Reference Wang, Chassin, Lee, Haller and King2017). To illustrate, adolescents with higher PGSEXT may develop behavior problems that are perceived negatively by their parents, or perhaps make parenting more difficult (Akcinar & Baydar, Reference Akcinar and Baydar2016; Patterson, Reference Patterson, Reid, Patterson and Snyder2002). Prior work has also shown that difficult temperaments may contribute to negative parent-offspring dynamic (Ayoub et al., Reference Ayoub, Briley, Grotzinger, Patterson, Engelhardt, Tackett, Harden and Tucker-Drob2019). Genetic susceptibility to EXT may also explain some of the variability in response to parenting-based psychosocial interventions (Reiss & Leve, Reference Reiss and Leve2007).
Unlike what was observed in the population-based sample, the family-based analyses uncovered no rGE effects between PGSEXT and delinquent peer affiliation. Instead, we found significant attenuation of rGE with delinquent peer affiliation from the between- to within-family level, suggesting that family-level confounders may contaminate rGEs observed at the population-level (Bornovalova et al., Reference Bornovalova, Cummings, Hunt, Blazei, Malone and Iacono2014; Klahr & Burt, Reference Klahr and Burt2014; Kuo et al., Reference Kuo, Poore, Barr, Chirico, Aliev, Bucholz, Chan, Kamarajan, Kramer, McCutcheon, Plawecki and Dick2022). One source of confounding could involve broader, system-level influences that are associated with population differences, such as socioeconomic status and regional and neighborhood factors. For example, PGSEXT may be associated with higher delinquent peer affiliation through the effect of poverty, urbanicity, or other social vulnerabilities, which are reflected in PGSEXT due to population stratification and/or assortative mating (Foster et al., Reference Foster, Horwitz, Thomas, Opperman, Gipson, Burnside, Stone and King2017). Additionally, PGSEXT may reflect heritable traits and characteristics of the parents such as parental monitoring, which has been found to negatively associate with delinquent peer affiliation of their children (Elam et al., Reference Elam, Chassin, Lemery-Chalfant, Pandika, Wang, Bountress, Dick and Agrawal2017; Li, Chen, Li, & Deater-Deckard, Reference Li, Chen, Li and Deater-Deckard2015). Finally, family-level psychosocial factors, such as parental psychopathology, are not differentially experienced by twins/siblings in a family. These factors may contribute to the association between genetic risks and delinquent peer affiliation at the population level, but go undetected in our family-based design.
Lastly, with respect to school connectedness, there was no evidence for rGE after confounding effects were controlled for. However, the within-family estimate did not significantly differ from the between-family estimate, such that there was no clear evidence of confounding either. This set of results may reflect imprecise rGE estimates due to the relatively smaller sample size for twins and full-siblings and correspondingly insufficient statistical power to draw robust conclusions.
In sum, our findings suggest that PGS-related rGE effects, though commonly observed in population-based studies, may offer different interpretations. Some rGEs (e.g. with parental relationship) may reflect the effect of offspring genotype on the environment. This possibility is consistent with the idea that genetic effects may be mediated by psychosocial environments (Elam et al., Reference Elam, Lemery-Chalfant and Chassin2023). Other rGEs (e.g. with delinquent peer affiliation), however, may reflect a passive/confounding pathway, where offspring PGS and their environments are spuriously associated due to family/population-level confounders. Such confounding effects are challenging to control for in large population-based designs. As such, we advocate for the use of rigorous, family-based designs that account for population- and family-level confounders in conjunction with polygenic methods. For instance, our approach, which compares population, between-family, and within-family PGS-environment associations, could be extended to investigating other heritable psychiatric traits and to other family-based, genotyped cohorts. Because parental genotypes were not measured in this sample, our within-family design is limited to the offspring and controls for the combined influence of confounders (i.e. parental genotype, assortative mating, and population stratification). Parsing out specific sources of confounding may be an important future direction. As an example, extended family data with parent and offspring genotypes can be leveraged to isolate the specific contribution of parental genotype.
We acknowledge several other limitations to our study. First, we operationalized parental, peer, and school environments as unidimensional constructs. Not only are these constructs not necessarily unidimensional, but there are also other environments across development that could have been examined. Next, we acknowledge that our within-family design may be subject to some residual confounding effects – such as due to bias in reporting processes. Twins and siblings in the same family may also imitate, contrast, or otherwise influence each other through family dynamics, leading to under-/over-estimation of within-family effects (Carey, Reference Carey1986; Jensen & McHale, Reference Jensen and McHale2017). Additionally, our family-based design limited us to a relatively small within-family subsample, contributing to a greater degree of uncertainty in estimating rGE effects relative to our larger population subsample of unrelated individuals. This is further complicated by our use of PGS, which generally has modest effects. Our design therefore cannot rule out the possibility of true rGE effects that operate through aspects of genetic liability not captured by the current PGS. Finally, our family-based model also controls for factors that are shared by twins/siblings and may fail to capture true rGE effects that act on shared family environments. Future research should aim to replicate these findings using other family-based designs, such as with genotyped parent-offspring trios.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726106242.
Acknowledgments
We thank the staff members and participants of Add Health for their contributions. Add Health is directed by Robert A. Hummer and funded by the National Institute on Aging cooperative agreements U01 AG071448 (Hummer) and U01 AG071450 (Aiello and Hummer) at the University of North Carolina at Chapel Hill. Waves I-V data are from the Add Health Program Project, grant P01 HD31921 (Harris) from Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Add Health GWAS data were funded by Eunice Kennedy Shriver National Institute of Child Health and Human Development Grants R01 HD073342 (Harris) and R01 HD060726 (Harris, Boardman, and McQueen). Add Health was designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill. We acknowledge the Externalizing Consortium and its Principal Investigators (Danielle M. Dick, Philipp Koellinger, K. Paige Harden, Abraham A. Palmer), Lead Analysts (Richard Karlsson Linnér, Travis T. Mallard, Peter B. Barr, Sandra Sanchez-Roige), and Significant Contributors (Irwin Waldman). The Externalizing Consortium has been supported by the National Institute on Alcohol Abuse and Alcoholism (R01AA015416 – administrative supplement to DMD), and the National Institute on Drug Abuse (R01DA050721 to DMD). Additional funding for investigator effort has been provided by K02AA018755, U10AA008401, P50AA022537 to DMD, as well as a European Research Council Consolidator Grant (647648 EdGe to Koellinger). The content is solely the responsibility of the authors and does not necessarily represent the official views of the above funding bodies. The Externalizing Consortium would like to thank the following groups for making the research possible: 23andMe, Add Health, Vanderbilt University Medical Center’s BioVU, Collaborative Study on the Genetics of Alcoholism (COGA), the Psychiatric Genomics Consortium’s Substance Use Disorders working group, UK10K Consortium, UK Biobank, and Philadelphia Neurodevelopmental Cohort.
Author contribution
Conceptualization: L.D., J.J.L.; Data curation: L.D.; Formal analysis: L.D.; Funding acquisition: J.J.L.; Methodology: L.D.; Project administration: L.D.; Resources: J.J.L.; Supervision: J.J.L.; Visualization: L.D.; Writing – original draft: L.D.; Writing – review and editing: L.D., J.J.L.
Funding statement
James J. Li was supported by grants from Wisconsin Alumni Research Foundation, the National Institute of Mental Health (R01 MH134039, R01 MH128371) and in part by a core grant to the Waisman Center from the Eunice Kennedy Shriver National Institute of Child Health and Human Development (P50 HD105353).
Competing interests
The authors had no conflicts of interest with respect to authorship or publication of this article.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. This study was approved by the University of Wisconsin-Madison Institutional Review Board (UW-Madison IRB# 2015-1189). Consent to participate was obtained by the Add Health study.
Transparency and openness
Our analyses were pre-registered at https://osf.io/3s5tx. Restricted-access data from the Add Health study are available at: https://addhealth.cpc.unc.edu/data/#restricted-use. Analysis codes are available at https://github.com/lichendong/2024_within_fam_rge. We report how we determined all data exclusions, all manipulations, and all measures in the current study.
Open access
