Etiological Overlap Between Sex Under the Influence and Number of Lifetime Sexual Partners

Abstract

Drug and alcohol use is associated with risky sexual behavior (RSB). It is unclear whether this association is due to correlated liabilities (e.g., third variable influencing both traits), or whether substance use during sexual decision-making increases RSB. This study addresses this question by fitting a series of biometrical models using over 800 twin pairs assessed in early adulthood (m= 25.21 years). Measures included an index of sex under the influence (e.g., frequency that drugs or alcohol affect sexual decision making), number of lifetime sexual partners, and a general measure of substance use. Analyses suggest the covariance among these measures is explained by both genetic and environmental correlated liabilities. The overlap was not specific to sex under the influence, but was shared with a measure of general substance use. Models testing necessary but not sufficient parameters for direction of causation suggest that sex under the influence is unlikely to cause an increase in RSB; more evidence for reverse causation was found.

Risky sexual behaviors (RSB) are those that increase one’s risk for contracting human immunodeficiency virus (HIV), other sexually transmitted infections, and unintended pregnancies. Within the United States, there are about 19.7 million new cases of sexually transmitted infections each year (Satterwhite et al., 2013), which translates to an estimated $15.6 billion dollars of direct treatment costs (Owusu-Edusei et al., 2013). Relatedly, approximately 45% of all pregnancies in the United States are unintended (Finer & Zolna, 2016). As a result, there has been widespread public health interest in reducing RSB and related outcomes: from sexual education in public schools, to “safe-sex” campus initiatives, and community based preventative healthcare programs that aim to reduce infections and unintended pregnancies. Drug and alcohol use, specifically sex under the influence of substances, has been identified as a target health risk behavior (Kann et al., 2018) and potential cause of RSB (HHS, NIH, Task Force on the National Advisory Council on Alcohol Abuse and Alcoholism, 2002). Thus, some RSB prevention programs operate on the assumption that reducing drug and alcohol induced impairment will lead to safer sexual practices (Jackson et al., 2012).

Several mechanisms have been put forth to explain why substance use could cause increased risk behavior, including pharmacological disinhibition or expectancy effects. For example, the alcohol myopia hypothesis (Steele and Josephs, 1990) posits that the link between substance use and RSB is due to direct physiological impairment in information processing and decision making, Further, a self-fulfilling prophecy effect (Lang, 1985) can occur when people have strong social expectations about drug and alcohol use. For example, Dermen and Cooper found that among those with strong expectations that drinking alcohol promotes sexual risk taking, alcohol use was negatively correlated with condom use during first intercourse and first intercourse with most recent partner (2000). Associations or expectations between desire to engage in RSB, actual engagement in RSB, and concurrent alcohol use could result from expectations that impairment facilitates sexual encounters (Corbin. Scott, & Treat, 2016). However, it is difficult to disentangle causal explanations from possible alternatives.

While drug and alcohol use is associated with higher rates of RSB on a population level, these behaviors could be correlated for several reasons: 1) drug and alcohol use during sexual encounters may cause people to take more sexual risks, 2) drug and alcohol use during sex is caused by risky sexual behavior, or 3) there are third variables (e.g. environmental or genetic factors) that lead to both substance use and RSB. For example, problem behavior theory (Jessor & Jessor, 1977) and large-scale epidemiological comorbidity studies (Kruger & Markon, 2006) do not require causal explanations to explain this overlap. Rather, correlations between substance use and RSB can be largely explained by a liability towards behavioral disinhibition or “lack of constraint, tendency toward impulsivity, or inability to inhibit socially undesirable or otherwise restricted actions” (Iacono et al., 2008). Alternatively, behavioral disinhibition has been described as predisposition for high novelty seeking, impulsivity, and lack of constraint (Sher & Trull, 1994). In terms of public health, identifying the optimal approach to reduce HIV, sexually transmitted infections, and other RSB-related maladaptive outcomes relies on understanding the causal structure of these related behaviors. For instance, substance use prevention may be useful in its own right, but it may have little effect on overall RSB related outcomes if the association is due to other confounding factors such genes (e.g., a genetic liability towards behavioral disinhibition) or familial environments (e.g., parental monitoring).

While there is a wealth of literature examining the relationship of alcohol (and to a lesser extent other drug use) and RSB, the evidence for causality is mixed (Leigh & Stall, 2008). Several recent meta-analyses of randomized control studies have concluded that exposure to alcohol has a direct effect on RSB intent (e.g., measures of intention to use condoms, likelihood or intentions of having unprotected sex, etc.; Rehm et al., 2012; Scott-Sheldon et al., 2016). While lab-based randomized controlled trials suggest there is a causal effect of alcohol intoxication on theoretical RSB behaviors, event-level based studies exploring whether real world behaviors change due to substance use are less conclusive. Such studies directly assess RSB behavior such as condom non-use during a recent critical incident (e.g., most recent sexual encounter) or track patterns of behavior across many sexual events via participant recorded logs or diaries (Weinhard & Carey, 2000). Meta-analyses at the event level do not find strong associations between substance use and real world behaviors such a non-condom use (Ritchwood et al., 2005). Further, several studies have found that condom use patterns are similar across sober encounters compared to those where cannabis (Walsh et al., 2013) or alcohol (Morrison et al., 2003; Leigh et al., 2008) had been consumed. While some interactions with partner type (e.g., casual vs. regular) have been reported (Leigh, 2002; Brown & Vanable, 2007; Scott-Sheldon, Carey, & Carey, 2010), substance use did not necessarily increase RSB. Rather, strong global level associations which are consistently reported across studies (Ritchwood et al., 2015) may not be specific to sexual encounters, as individuals engaging in RSB may have higher general substance use may use both within and outside of sexual contexts. It is possible that both explanations are at least partially true, in that there is both a direct effect of substance use on RSB and that there are underlying influences that increase both behaviors (i.e., correlated liabilities). Using a quasi-experimental design in which many genetic and environmental confounds are shared between twin pairs (McGue, Osler, & Christensen, 2010), biometrical modeling is a useful tool for exploring the extent to which either of these explanations are supported. That is, with a genetically informed sample, the covariance between these traits can be partitioned into genetic and environmental components (Neale & Cardon, 1992). The resulting patterns of covariance can be used to test assumptions about causality. Specifically, results of multivariate biometric may suggest the presence of genetic or familial-environmental factors that contribute to both variables (Kendler et al, 1992). Using a model comparison approach, twin data can be used to compare the likelihood of alternative models of comorbidity specified by Neale and Kendler (1995). (For patterns of results and simulations of these models, refer to Rhee et al., 2005). Based on previous work, we hypothesize that the relationship between sex under the influence and RSB will show partial confounding with other correlated liabilities. Specifically, we expect that the shared liability will not be specific to sex under the influence but rather to substance use more generally.

To our knowledge, this is the first twin or family study directly exploring the nature of the relationship between sex under the influence and a measure of RSB in early adulthood. The current study investigates the role of drug and alcohol use on RSB using several self-report measures. We aimed to replicate the phenotypic relationship between substance use during sexual decision making (a composite measure hereinafter referred to as sex under the influence) and RSB—assessed by number of lifetime sexual partners (corrected for age and gender). Lifetime number of sexual partners has been shown to be a robust measure of RSB, and is a strong predictor of sexually transmitted infections in several samples (Karlsson, et al., 1995; Santelli et al., 1998; Sturdevant et al., 2001; Epstein et al., 2013).

We first fit univariate biometric models to test whether it was appropriate to account for mean/threshold gender differences and for the presence of qualitative sex differences (e.g., whether the proportion of variance explained by genetic and environmental factors differed across males and females). Next, we use biometrical modeling to estimate the extent to which the covariance between sex under the influence and number of lifetime partners is genetic or environmental in nature for males and females. Patterns of covariance are then compared to three possible result scenarios, either consistent with causality or an alternative model. Given the global level association between substance use and RSB, we then test whether the overlap is specific to use in sexual contexts (i.e., by fitting a trivariate model that essentially controls for general substance use). Finally, we employ two direction of causation models which provide an additional line of evidence for or against causality. Specifically, we compare the likelihoods of the following two scenarios: 1) where substance use during sex causes higher RSB or 2) increases in RSB cause more substance use during sex (i.e., reverse causation; Neale & Kendler, 1995; Rhee et al., 2005).

Method

Participants

Participants were drawn from the third wave of data collection from the Center on Antisocial Drug Dependence (CADD; PI: J. K. Hewitt), a longitudinal study of adolescent/young adult antisocial behavior and substance use, which includes four genetically-informative samples. The Colorado Longitudinal Twin Study and the Colorado Community Twin Study were included in the primary analysis. Twins in the analysis were in early adulthood (female m=25.24 years, s.d.= 2.50, n=1047; male m=25.18, s.d.= 2.66, n=823) and were representative of Colorado demographics (Rhea, Gross, Haberstick, & Corley, 2006; 2013). While the additional two CADD samples were not included in the primary analysis, all four of the community-based Center samples were used to calculate age- and sex-normed measures of substance use and RSB. The additional samples included the Colorado Adoption Project which follows adoptive children, matched controls, and their families (Rhea, Bricker, Wadsworth, & Corley, 2013; Rhea, Bricker, Corley, DeFries, & Wadsworth, 2013), and the control participants of the Colorado Family Study which is comprised of probands formerly in treatment for adolescent antisocial drug dependence, their siblings, and matched control families (Stallings et al., 2003).

The full sample included 723 same sex twin pairs and 140 opposite sex twin pairs.

Measures

Sexual behavior was assessed using the Modified Risk Behavior Questionnaire (M-RBQ, adapted from Booth, Corsi, & Mikulich-Gilbertson, 2004). Due to the sensitive nature of the items on this questionnaire, the M-RBQ was administered via a computer program in a private room and a “would rather not answer” option was available for all items. While this did result in some missing data, this option was seldom chosen and did not seem to be correlated with other attributes of the participant. Number of lifetime sexual partners was scored from a single item (i.e. “in your lifetime, with how many people (different partners) have you had oral, vaginal or anal sex?”). Responses were scored on a seven-point scale (“none” [0], “one” [1], “two” [2], “three-five” [3], “six-nine” [4], “ten-nineteen” [5], and “twenty or more” [6]). The raw score distribution is reported in Table 1. For the purpose of the models scores are treated as quasi-continuous, after corrected for age using standard regression procedures and then z-scored within sex.

Table 1.

Frequency of raw scores, by gender

	Females	Males
1) General substance use
None	3.8% (n=40)	4.5% (n=37)
1	57.0% (600)	41.2% (339)
2	27.2% (286)	31.1% (256)
3	5.2% (55)	9.1% (75)
4	3.1% (33)	6.2% (51)
5+	3.6% (38)	7.9% (65)
2) Sex under the influence
0	65.6% (n=669)	60.5% (491)
1	7.1% (72)	11.7% (95)
2	10.0% (108)	8.1% (66)
3	21.7% (97)	8.8% (71)
4	7.8% (80)	10.9% (88)
3) Number of lifetime sexual partners
None	8.6% (n=88)	9.0% (74)
1	16.5% (173)	14.5% (118)
2	11.0% (115)	8.4% (69)
3–5	21.7% (227)	22.2% (183)
6–9	19.0% (199)	16.9% (139)
10–19	15.5% (162)	17.7% (146)
20+	5.4% (57)	10.4% (86)

Note. The index of General Substance Use reflects the number of substances for which a participant meets “user” criteria, as defined by the CIDI-SAM. This composite is measured on a six point ordinal scale (“not a user of any substances” [0], “one substance” [1], “two substances” [2], “three substances” [3], “four substances” [4], “five or more substances” [5]). The index of Sex Under the Influence is a composite measure which combined four items of past 12-month co-occurrence of substance use and sexual decision making. Participants with zero or one partners in the past five years were scored as zero [0], as where participants who answered “never” on all items. The composite scores of the remaining participants were split into quartiles [1–4]. This is an ordinal scale where 4 indicates the highest frequency of behavior.

Sex under the influence was a composite of four items (adapted from items used by the Minnesota Center for Twin and Family Research; Iacono, McGue, & Krueger, 2006), which assessed the extent to which substance use co-occurred with sexual decision making. Items assessed the frequency in the past 12 months that participants 1) had alcohol or drug use influence a decision to do something sexual with a partner, 2) did more sexually with a partner than planned due to alcohol or drug use, 3) used drugs or alcohol to feel more comfortable with a sexual partner, or 4) had unprotected sex due to alcohol or drug use. Each item was assessed on a seven-point scale (“never” [0], “one time” [1], “two times” [2], “three-five times” [3], “six-nine times” [4], “ten-nineteen times” [5], or “twenty or more times” [6]). Participants who either had 1) no sexual partners in their lifetime, or 2) only reported one partner within the past 5 years were not assessed on these questions (they skipped out of this assessment), but were coded zero since their drug and alcohol use would have little to no effect on sexual risk taking (Derks, Dolan, & Boomsma, 2004). Additionally, participants who answered “never” on all four individual items were scored zero and were presumed to be low in sex under the influence and general sexual risk taking. Remaining participants were split into quartiles based on their composite scores to form a 0 to 4 ordinal index of sex under the influence, with the 4^th quartile indicating the highest risk (See Table 1).

An index of general substance use was created from several measures on the Composite International Diagnostic Interview- Substance Abuse Module (CIDI-SAM; Robins et al., 1988) as a way to capture general participant endorsed substance use. Participants were considered ‘users’ of a drug class (e.g., cannabis, stimulants, sedatives, club drugs, cocaine, heroin /opioids, PCP, hallucinogens, or inhalants) if they reported using a class of drugs more than 5 times in their life. However, participants were considered to be alcohol users if they had ever had more than one drink. These thresholds were determined by the administration algorithms of the CIDI-SAM. Subjects skipped out of drug assessment categories if they did not meet these minimum use thresholds, so sub-threshold use was not recorded by the CIDI-SAM. Since tobacco is not typically thought to play a role in sexual decision making, it was not included in the final computed variable. Given the relative scarcity of scores at the high range of use (Derks, Dolan, & Boomsma, 2005), higher scores were truncated resulting in a six point ordinal scale (“not a user of any substances” [0], “one substance” [1], “two substances” [2], “three substances” [3], “four substances” [4], “five or more substances” [5], see Table 1).

For 92% of the participants’ DNA was extracted from buccal cells through saliva and/or cheek swabs, and zygosity was confirmed by analyzing 11 highly polymorphic short tandem repeat (STR) polymorphisms (Smolen, 2005). For twins yet to provide DNA samples, zygosity was established via repeated tester ratings on a 9-question survey (Nichols & Bilbro, 1966; see Rhea et al., 2013 for details).

Statistical Analyses

All descriptive statistics and composite variables were computed using SPSS version 21 (SPSS, 2012). Phenotypic correlations, twin correlations, and structural equation models were estimated using Mplus version 7.4 (Muthén & Muthén, 1998–2012). Phenotypic correlations were estimated in a multi-group analysis with MPlus using type=complex to account for the non-independence of twin pairs and recover robust standard errors. Twin correlations and structural equation models were estimated using raw data options and employing maximum likelihood estimation (ML) for univariate analysis of the continuous variable and robust weighted least squares (WLSMV) in the case of ordinal variables and multivariate analyses.

Model fit was determined using the following indices: root mean square error of approximation (RMSEA), the comparative fit index (CFI), and the Akaike information criterion (AIC), the latter of which could only be estimated for continuous variables. Nested models were compared using likelihood-ratio χ² difference tests using the difference in degrees of freedom (df) in the full model compared to more constrained models.

Nested models tested 1) significance of sex differences in means/thresholds and path estimates, and 2) significance of each pathway (e.g. comparing full model to a model dropping the path of interest). The DIFFTEST option in Mplus was used in model comparisons using WLSMV estimation to address biases in the χ² test produced by ML estimation for ordinal variables. In rare cases of non-convergence, the significance level of specific path estimates was determined by testing whether the Mplus-computed confidence interval included 0.

Biometrical Modeling

Univariate structural equation models were fit to data for each of the three variables. Using a genetically informative twin sample, these models are able to partition the sources of variance of these variables into additive genetic factors, or sum effect of each individual segregating allele (a²), interactive effects of alleles within loci (referred to as dominance effects; d²), shared or common environmental (c²) factors, and non-shared environmental (e²) factors (See Figure 1). Estimates are derived from the difference between monozygotic (MZ) twin similarity, who share 100% of their segregating alleles (e.g., 100% of both additive and dominance effects) and dizygotic (DZ) twin similarity, who share 50% of their segregating alleles on average (e.g., 50% of additive effects and 25% of dominance effects). The equal environment assumption, which has been empirically supported to be valid for many traits (Conley, Rauscher, Dawes, Magnusson, & Siegal, 2013; LoParo & Waldman, 2014), requires that the magnitude of environmental influences shared between co-twins is the same between MZ and DZ twins. Thus, an estimate of a² and d² can be inferred from the extent to which MZ twin pairs are more similar than DZ twin pairs. Additionally, c² will contribute to similarity in both MZ and DZ pairs, while e² (environmental influences uncorrelated across twins or measurement error) will only contribute to trait variance and not twin resemblance. Finally, the similarity of DZ pairs will approach the similarity of MZ pairs as shared environmental influences increase in magnitude. The similarity of DZ pairs will be much less than the similarity of MZ pairs in the presence of strong dominance effects (e.g. MZ twins shared 100% of their dominance effects, DZ pairs share 25%, on average). With twins reared together there is insufficient information to estimate both d² and c² in the same model. ACE models (including a², c², and e² estimates) are typically selected over ADE models (including a², d², and e² estimates) when DZ twin correlations (rDZ) are more than half of the MZ twin correlation (rMZ), suggesting that presence of shared environment is masking any effects of dominance that may exist. In classical twin modeling (using standardized path models), squared path coefficients (e.g. a², d², c², and e²) provide estimates of genetic and environmental variance components. Path coefficients are regression coefficients and are commonly referred to as influences on the trait.

Figure 1.

Univariate twin model

A bivariate Cholesky decomposition of covariance (e.g., a reparameterization of the correlated liabilities model; Loehlin, 1996; See Figure 2) was used to explore the nature of the relationship between sex under the influence and number of lifetime sexual partners, which follows a similar logic of univariate twin model. Much like variance of a single trait, the covariance between traits can be decomposed into genetic and environmental sources. To the extent that additive genetic effects explain the co-occurrence of these behaviors, we would expect that MZ cross-twin cross-trait correlations to be higher than DZ correlations (e.g., a twin’s sex under the influence score should be more strongly associated with their co-twin’s number of lifetime sexual partners within MZ pairs compared to DZ pairs if there are common additive genetic factors that contribute to both variables). Although bivariate twin models were used to obtain parameter estimates, the magnitude of additive genetic covariance can be estimated from the difference between MZ and DZ pair cross-twin cross-trait correlations. Thus, if cross-twin cross-trait correlations were not significantly different between MZ and DZ pairs we would assume all covariance is due to shared environment. Finally, non-shared environmental factors will capture the covariance between the variables within individuals (e.g., in a case where all covariance is explained by E, there would be no cross-twin cross-trait resemblance for either MZ or DZ twin pairs). As such, the Cholesky model decomposes sources of variance and covariance into multiple A, C (or D), and E factors. In the bivariate Cholesky, the first A, C (or D), and E factors model the covariance between the variables. The secondary A, C (or D), and E factors are orthogonal to the first factors and model the residual variance in the second variable.

Figure 2.

Reconfigurations of the bivariate model

Note: a) correlated liabilites model b) Cholesky decomposition model c) direction of causation model

A trivariate Cholesky factorization is a straightforward generalization of the bivariate model: the first factors of a trivariate model load on all variables to model shared covariance. The secondary factors capture only residual variance and covariance between the second and third variables, while the final factor captures the unexplained variance of the final variable.

By modeling a trivariate Cholesky decomposition model with general substance use in the leading position, it is possible to explore the explore the nature of the covariance between sex under the influence and number of lifetime sexual partners which is independent of this potential third variable (i.e., controlling for general substance use). We retain the order of the second and third variables (i.e., sex under the influence, followed by number of lifetime partners) to directly compare any reduction of significance of pathways; however, reordering the second and third variable would result in the same model fit. The resulting model directly tests whether general propensity towards substance use is a direct mediating variable and if so, whether this mediation is of genetic or environmental origin.

As a final test, we also fit two direction of causality (DoC) models. These models can be used to compare the likelihood of a causal model to alternative models, given that certain conditions are met in genetically informed samples. In order to distinguish between two causal models (i.e., drug and alcohol use during sex causes an increase in RSB, compared to the reverse causation where RSB causes substance use during sexual decision making), the two traits must have different modes of inheritance (i.e., the magnitude of genetic vs. environmental influences must vary between traits, which is tested via the univariate biometrical model). While tests of reciprocal causation (both directly cause each other over time) are theoretically possible, large samples and multiple trait indicators are required (Neale, Duffy, & Martin, 1994; Neale, Røysamb, & Jacobson, 2006). Additionally, it is assumed that measurement error for the two traits is uncorrelated between twin pairs (for more information on DoC models, see Heath et al., 1993). DoC models are nested under the bivariate model (See Figure 2); however, rather than allowing A, C (or D), and E cross paths to model the covariance between sex under the influence and number of lifetime sexual partners, all covariance is modeled in a single pathway from trait one to trait two. Model fit is indicative of both the likelihood of causality, as well as used to rule out alterative models (e.g., one where the direction of causation is reversed).

Model Fit and Implications for Causality

Evidence for correlated liabilities is necessary but not sufficient for causal inference. That is, all covariance is explained by either familial (i.e., A or C [or D]) or non-shared environmental (i.e., E) shared factors. Significant factor loadings in the biometrical models could signal the presence of either direct or indirect influences. For example, shared genetic liability can result from 1) biological pleiotropy (e.g., when the same genetic factors have a direct biological influence on both traits independently), 2) mediated pleiotropy (e.g., when one trait is causally related to a second trait so that the genetic factors for the first trait are indirectly associated with the second), and 3) spurious pleiotropy (e.g., the appearance of pleiotropy due to misclassification or ascertainment bias; Solovieff et al., 2013). Similarly, environmental influences will be associated with both traits if: 1) there are direct influences on both traits, or 2) if the first trait causes the second trait, those influences on the first trait in turn will influence the second trait indirectly. However, different assumptions about the likelihood of causality (i.e., indirect effects) can be inferred based on the pattern of results from the biometrical models (See Figure 3).

Figure 3.

Representation of potential scenarios from bivariate Cholesky decomposition models

Note: A) Scenario consistent with causality. B) Scenario not consistent with causality, or causality unlikely C) Scenario not informative about likelihood of causality. Highlight indicates significant path loadings.

The following potential patterns of results support alternative causal models. In the scenario where all covariance between two traits is explained by nonshared environmental influences (i.e., A or C [or D] factors would be non-significant), several inferences can be made (See Figure 3A). Such results could be indicative of causality, in that these are the expected pattern of results for a random exposure model (e.g., when the development of a disease is caused by exposure to a pathogen). Given the likelihood of familial influences on most behavioral traits, this scenario is very rare. In this unlikely case, the possibility of mediated or biological pleiotropy, as well as indirect or direct shared environmental (i.e., shared environmental third variables) can be ruled out. Such results may be suggestive of causality, although nonshared direct influences on both traits and/or correlated error cannot be ruled out, under this scenario. Under a second informative scenario, covariance between two traits would be explained by overlapping genetic and/or shared environmental sources in the absence of significant nonshared environmental overlap (see Figure 3B). In the case of causality, any influence on the first trait should have an indirect effect on the second trait. Given that there are always non-shared environmental influences on behavioral traits, the presence of significant indirect genetic effects (i.e., mediated pleiotropy) or indirect shared environmental influences on both traits would also require the presence significant of indirect nonshared environmental effects. Therefore, this pattern is inconsistent with causality in that the familial influences are likely to be directly influencing both traits. In the final scenario, covariance is explained by both familial and nonshared environmental sources (see Figure 3C). In this case, direction of causation models (DoC) models (Heath et al., 1993) are required to determine both the likelihood of and direction of causation. Given causality, the magnitude of each variance component (i.e., A, C [or D], and E) should be approximately equal in magnitude. Thus, a single parameter represents the magnitude of the A, C (or D), and E factors first trait influencing the second trait via the first. To the extent the covariance explained by these factors differs, constraining these influences to operate via a single indirect path should significantly reduce the model fit as measured by chi-square difference tests and AIC. Poor model fit is inconsistent with causality (i.e., it is unlikely each factor explains the exact amount of covariance in a non causal model).

While the Cholesky decomposition models can be used to infer the likelihood of causality, they are not informative about the direction of the causal effect. While theory should inform the order of the traits in a bivariate model, the reverse model will show a similar pattern of results in terms of correlated liabilities (i.e., the model can be transformed for ease of interpretation; Loehlin, 1996). Therefore, DoC models are needed to establish the likelihood of the casual model compared to that of alternative models (i.e., reverse causation), which is determined by the best model fit. Similar to Cholesky models, goodness of model fit for the DoC models cannot prove causality. Rather, poor fit is evidence against a particular causal model.

Results

Consistent with previous literature (Liu et al., 2015), males reported slightly higher mean scores compared to females on all variables (See Table 1). The modal number of lifetime sex partners was 3 to 5 for both males and females in this age group, but nearly twice as many males reported 20+ partners (10.4%) compared to females (5.4%). Across all twins included in the analysis, sex under the influence was significantly associated with number of lifetime sexual partners (r=.600). Additionally, our measure of general substance use was substantially correlated with both sex under the influence (r=.565) and number of lifetime sexual partners (r=.515). For phenotypic correlations computed separately by gender, see Table 2. Twin correlations for the three variables (see Table 3) were suggestive of genetic influences on all three variables, as MZ correlations were consistently higher than DZ correlations. We determined ACE models were more appropriate than ADE models, which was consistent with patterns identified in reviews of substance use behavior (Sullivan & Kendler, 1999; Stallings, Gizer, & Young-Wolff, 2016) and risky sexual behavior (Harden 2014). Given differences in prevalence of traits and magnitude of twin correlations across male and females, sex limitation models were tested (i.e., formal tests of gender differences were conducted; Neale, Røysam, & Jacobson, 2006).

Table 2.

Phenotypic correlations, by gender

	General substance use	Impaired sex	Number of lifetime sexual partners
1) General substance use	1	.427(.038)*	.484(.030)*
2) Sex under the influence	.409(.038)*	1	.655(.029)*
3) Number of lifetime sexual partners	.557(.027)*	.642(.027)*	1

Note. Female estimates shown on the lower diagonal, male estimates on the upper diagonal in bold.

Table 3.

Twin correlations for the three variables by type of twin

	MZ-F	MZ-M	DZ-F	DZ-M	DZ-OS
1) General substance use	.644	.760	.372	.573	.478
2) Sex under the influence	.347	.345	.334	.212	.193
3) Number of lifetime sexual partners	.661	.655	.335	.365	.312
(number of pairs)	254–263	171–176	147–155	129	135–140

Note. F= female, M=male, OS=opposite sex. Polychoric correlations were estimated in MPlus.

Univariate Model: Test of sex limitation

We first tested whether mean/threshold gender differences should be accounted for in our models. For each variable, we tested whether thresholds for ordinal variables (i.e., sex under the influence and number of lifetime sexual partners) and means for the quasi-continuous variable (i.e., lifetime number of partners) could be constrained to be equal across gender in models that also estimated separate path coefficients for males and females. Male and female distributions for general substance use were significantly different, so that thresholds could not be constrained across gender (χ²_diff(5)= 62.29, p< .001). In contrast, thresholds were not significantly different across gender for sex under the influence (χ²_diff(4)= 8.292, p=0.081). We expected no mean differences across gender for number of lifetime partners as it was z-scored within gender, and this constraint did not result in a decrement of fit (χ²_diff(1)= 1.37, p=0.242). Thus, separate thresholds for males and females were used in subsequent models for general substance use, while means and thresholds were constrained across gender for sex under the influence and number of lifetime sexual partners.

Second, we conducted a quantitative sex limitation model that tested whether the A, C, and E factor loadings could be constrained across gender. Constraining these estimates, when there are significant differences in the magnitude of these influences between males and females, will result in significant χ² difference tests in nested model comparisons. While there were no quantitative gender differences in the biometrical factor loadings for sex under the influence (χ²_diff(3)= 5.22, p=0.157) and number of lifetime sexual partners (χ²_diff(3)= 2.24, p=0.523), there was a significant decrement in model fit in the constrained model for general substance use (χ²_diff(3)= 8.05, p=.0449). For general substance use genetic influences were significantly greater for females than for males while shared environment made a greater contribution to variation in general substance use for males.

Influences of the A, C, and E factors are reported as standardized path estimates (i.e. a₁₁, c₁₁, and e₁₁; reported in Table 4). When squared, path estimates represent the proportion of variance explained by a specific influence (e.g. a₁₁² is equivalent to the proportion of variance explained by additive genetic variation, also termed heritability). Heritability was greatest for number of lifetime sexual partners (a₁₁²= .710² = 50%) and general substance use (a₁₁²_male= .612² =37%, a₁₁²_female= .700² = 49%), while sex under the influence was moderately heritable (a₁₁²= .434² = 19%). The greatest evidence of shared environmental influences on a single trait was for general substance use, specifically for males (c₁₁²_male= .622²=37% of variance explained by shared environment, c_11female²= .389²= 15% of variance explained by shared environment).

Table 4.

Standardized path loadings for univariate models

Model		A112	c112	e112
1) General substance use	Female	0.49*	0.15*	0.36*
	Male	0.37*	0.39*	0.24*
	Constrained	0.46*	0.24*	0.30*
2) Sex under the influence	Female	0.02	0.32	0.65*
	Male	0.27	0.07	0.65*
	Constrained	0.19	0.16	0.65*
3) Number of lifetime sexual partners	Female	0.42*	0.15	0.44*
	Male	0.61*	0.04	0.35*
	Constrained	0.50*	0.09*	0.40*

Note. Female and male paths estimated separately are compared to a constrained model with paths estimated jointly.

^*indicates p<0.05.

Bold indicates the best fitting model; fit statistics: 1) χ² = 42.335, df = 49, p=0.7384, RMSEA < 0.001, CFI = 1.000; 2) χ² = 47.446, df = 42, p=0.2603, RMSEA = 0.025, CFI = 0.874; and 3) χ² = 31.198, df = 21, p=0.0704, RMSEA = 0.049, CFI = 0.959, AIC= 4975.576. Estimates do not add to 1 due to rounding.

Bivariate Model: Covariance between sex under the influence and number of lifetime sexual partners

As expected from the lack of gender differences in the univariate models for sex under the influence and number of lifetime sexual partners, constraining path estimates across gender in the bivariate model did not significantly reduce the fit of the model (χ²_diff(9)= 10.425 p=0.317). Though the constrained bivariate model had the best fit (See Table 5), we report paths separately for the sake of completeness as gender differences emerged in the trivariate model.

Table 5.

Bivariate model standardized path loadings by gender

	a11	a21	a22	c11	c21	c22	e11	e21	e22
Female	0.249	0.719	0.000	0.542	0.335	0.007	0.803*	0.339*	0.505*
Male	0.482*	0.695*	0.001	0.287	0.067	0.019	0.882*	0.385*	0.603*

Note. Model depicted for one twin only.

^*indicates p<0.05.

Significance of the cross paths (a₂₁, c₂₁, and e₂₁) are dependent on the size of the loading on the first factor (a₁₁, c₁₁, and e₁₁, respectively). For example, only a small proportion of the covariance would be explained by additive genetics if the first variable had a small heritability- regardless of the magnitude of the cross path. Thus, significant cross paths are of interest because they indicate a source of third- variable confounding between sex under the influence and number of lifetime sexual partners (i.e. confounding by shared genes or environments). Finally, loadings on the final factor represent unique influences not shared with the first variable (e.g. genetic or environmental factors that influence number of lifetime sexual partners but not sex under the influence).

In the best fitting model (constrained across gender), both genetic and environmental factors contribute to the correlations between sex under the influence and lifetime number of sexual partners. Figure 4 shows that there was significant additive genetic (a₂₁=0.670, p= .0164) and nonshared environmental (e₂₁=0.349, p <.001) covariance or overlap (See Figure 4). Though the shared environmental influences on number of lifetime sexual partners were shared entirely with those on sex under the influence (rC=1.00), the parameter of interest was not significant (c₂₁, p=.0744). The proportion of covariance explained by these factors was as follows: A= 57.7%, C= 26.7%, and E= 15.6%, suggesting that much of the covariance between sex under the influence and number of lifetime partners was due to common genetic factors.

Figure 4.

Bivariate model with standardized path loadings constrained across gender

Note: Model fit statistics: χ² = 91.590, df = 80, p = 0.1768, RMSEA = 0.027, CFI = 0.979. Pathway significance determined through model comparison, significant pathway denoted with * (p<0.05).

However, examination of the gender-specific estimates suggests that genetic factors contribute more to this correlation for males, while environmental factors contribute more in females. In females, the parameter signifying genetic overlap in the unconstrained model is not significant (a₂₁= 0.719, p=0.1808). Since sex under the influence was less heritable in females (a² = 0.02), compared to males (a² = 0.27), genetic overlap explains a smaller proportion of the total covariance between sex under the influence and number of lifetime sexual partners in females—despite the larger point estimate for parameter a₂₁. In contrast, shared environmental influences appear to explain a higher proportion of variance for sex under the influence in females; however the specific parameter signifying shared environmental overlap did not reach significance (c₂₁=0.335, p=0.1572). Though path estimates could be constrained, separate path estimates are reported in Table 5 for the sake of completeness, as gender effects emerge in the trivariate model.

It is important to point out that the genetic covariance (as well as the environmental covariance) is the product of path coefficients, in this case a₁₁ × a₂₁. Although the estimate of a₂₁ in females is greater than that for males, the product of the two path coefficients is greater for males than for females. As power to detect significance was reduced by using ordinal variables, it may be useful to examine the estimated proportion of covariance explained by each factor (e.g., for females and males, respectively, A= 55.8% and 67.7%, C= 31.7% and 11.5%, and E= 12.4% & 20.8%).

Trivariate Model: General substance use as a mediator

The trivariate model allows for the decomposition of the residual covariance between sex under the influence and number of lifetime sexual partners after accounting for the common covariance shared between all three variables (See Figure 5). That is, this model allowed us to test whether sex under the influence or impaired sexual decision making due to substance use remained an important predictor or correlate of risky sexual behavior, after controlling for general substance use. Again, we tested the significance of the parameters representing covariance between sex under the influence and number of lifetime sexual partners (i.e., a₂₃, c₂₃, & e₂₃, or the cross paths between the second and third latent factors), as well as the significance of covariance pathways (a₁₂, c₁₂, and e₁₂) between general substance use and sex under the influence.

Figure 5.

Trivariate model with standardized path loadings constrained across gender

Note. Female path estimates shown above, male path estimates are shown in bold. Significant pathway denoted with *(p<0.05). Model fit statistics: χ² = 157.169, df = 148, p =0.2874, RMSEA = 0.018, CFI = 0.995.

We report the constrained model for consistency with the reports of the univariate and bivariate model (See Table 6); however, evidence of sex differences emerged when trying to constrain male and female paths to be the same in the trivariate model (χ²_diff(18)= 44.349, p<.001). To appropriately interpret the model of best fit, males and female patterns should be considered separately.

Table 6.

Trivariate model with standardized path loadings constrained across gender

	11	21	31	22	23	33
a	0.678*	0.307*	0.488*	0.267	0.482	0.282
c	0.491*	0.247	0.186	0.341	0.200	0.001
e	0.547*	0.162*	0.179*	0.794*	0.289*	0.510*

Note. Model depicted for one twin only.

^*indicates p<0.05.

The bivariate results suggested that common genetic factors make an important contribution to the covariance between sex under the influence and number of lifetime partners—particularly for males. After controlling for substance use in general, the genetic covariance between sex under the influence and number of lifetime partners is reduced in males but remains significant. That is, although part of the genetic covariance is mediated through genetic factors in common with substance use in general (a₃₁), additional unique genetic influences contribute to the covariance between sex under the influence and number of lifetime partners (a₂₃). However, nonshared environmental influences specific to sex under the influence and number of lifetime partners (e₂₃) still remain for males when substance use in general is included; thus, causality cannot be ruled out from these results alone.

For females, after controlling for substance use in general, there was no significant covariance specific to sex under the influence and number of lifetime partners. Though the first C and E factors describe covariance shared among all three variables, the path loadings onto sex under the influence were not significant (c₂₁, p= .6930, & e_21, p=.1184). This pattern suggests that rather than a causal relationship, much of the relationship between sex under the influence and number of lifetime partners was mediated through influences in common with substance use in general (a₃₁, c₃₁, e₃₁).

Supplementary analysis: removing opposite sex pairs

It was possible that males and females differ not only in the extent to which genetic or environmental factors influence a trait (i.e., quantitative test discussed above), but that there are different genetic or environmental influences on a trait for males and females (i.e., qualitative sex differences). Qualitative sex difference could occur even if the genetic influences were the same magnitude across genders (e.g. equal hertiabilites across males and females, albeit from different genes). Inclusion of DZ-OS (e.g. opposite sex dizygotic) pairs allows for such tests. If different genes influence female versus male behavior, we would expect that the rDZ-OS (opposite sex twin correlation) would be lower than rDZ-SS (same sex twin correlation). For our three variables, rDZ-OS were not significantly less than rDZ-SS (see Table 3) suggesting that the genes that influences male traits are likely the same genes that influence female traits. Inclusion of DZ-OS twins changes the interpretation of the significance of given path estimates. Notably, when testing the significance of female or male specific paths via model comparisons (e.g. comparing a full model to a model dropping the path of interest), we are essentially fixing both the path to be zero and a portion of the covariance between males and female pairs to be zero. That is, dropping a female (or male) parameter would not only test the significance of a female (or male) loading but also the common pathways across males and females). We re-ran all of the models excluding DZ-OS pairs (i.e., n=135–140 pairs, therefore retaining approximately 80% of the full sample) and estimated paths were of similar magnitude to those reported here, although some significant paths dropped to non-significance due to a reduced sample size (see Supplemental Tables 1–3).

Direction of Causation Models

The DoC models were nested under the best fitting bivariate twin model (e.g., separate thresholds for sex under the influence, constrained means for number of lifetime partners, and constrained path estimates across males and females; See Figure 6). In the model testing whether sex under the influence has a direct effect on lifetime number of partners, constraining A, C, and E covariance through a single causal pathway resulted in a significant decrement of fit of the model (χ²_diff(2)= 22.629, p<.001). This suggests this direction of causation is unlikely, and the observed patterns of covariance are unlikely to be due to mediated pleiotropy or indirect shared environmental influences. Interestingly, the model testing reverse causation (e.g., whether an increase in lifetime number of partners increased the likelihood of sex under the influence) did not result in a significant decrement of fit (χ²_diff(2)= 2.995, p=.2237). While the general bivariate model could be constrained across genders, we tested whether there were gender differences in the causal pathway. This parameter could be constrained without a significant decrement in fit (χ²_diff(1)= 0.058, p=.2237).

Figure 6.

Direction of causation models with standardized path loadings constrained across gender

Discussion

In this study, we replicated the global level association between substance use and risky sexual behavior (RSB) using a population-based, genetically informative sample. Higher scores on our composite measure of sex under the influence (e.g., how frequently substance use influenced sexual decision making) were associated with higher numbers of lifetime sexual partners. Additionally, we found evidence that people who use more drugs and alcohol generally, also have higher numbers of partners. The positive association between substance use and number of lifetime sexual partners was relatively consistent across males and females, though prevalence of traits and magnitude of twin correlations differed. Thus, before modeling the nature of the overlap using multivariate biometrical models, we first conducted formal tests of gender differences in univariate twin models.

Multivariate twin analyses were used to model the common additive genetic, shared environmental, and nonshared environmental influences shared between sex under the influence and number of lifetime sexual partners (i.e., correlated liabilities). The patterns of results of the bivariate Cholesky decomposition were then compared to three possible scenarios (See Figure 3), in which likelihood that sex under the influence caused an increase in number of lifetime sexual partners could be inferred. In the combined model, where female and male parameters were set to be equal, covariance between sex under the influence and number of lifetime partners was explained by both common additive genetic and nonshared environmental factors (and to a lesser extent, there was marginal evidence for shared environmental overlap). This pattern did not provide strong evidence for or against a causal model. That is, the presence of familial influences did not match the causal model described in Figure 3A. Additionally, the presence of significant nonshared environmental influences makes it impossible to infer causality or distinguish whether the additive genetic overlap is due to biological or mediated pleiotropy (e.g., as described in scenario 2 in Methods). Similarly, we cannot be certain how much of the non-shared environmental covariance is due to direct or indirect influences on both traits, or due to spurious correlated measurement error. Thus, the bivariate model was inconclusive about causality.

We hypothesized that some of the overlap between sex under the influence and RSB may be related to an overall propensity towards substance use generally, so we examined general substance use as a potential mediator. Using a trivariate Cholesky decomposition, we partitioned the covariance further my modeling additive genetic, shared environmental, and nonshared environmental overlap specific to sex under the influence and number of lifetime sexual partners from that shared with general substance use. Reduction in the significance of nonshared environmental overlap between these traits reduces in significance once controlling for general substance use would provide evidence against causality. In this case, it can be assumed that these nonshared environmental influences (i.e., environments that increase the likelihood for substance use both within and outside the context of sexual activity and RSB itself), are reflective of third variable confounding (i.e., direct influences) rather than causation. This pattern was identified in females only, though there are some noteworthy caveats. In the trivariate model the covariance was explained by the first common E factor loading on all three traits, though the path estimate loading on sex under the influence was not significant. This pattern of non-significant overlap may be due to reduced power, as the use of sex limitation models reduces the number of data points contributing to a given path estimate (i.e., the trivariate analyses reduce the effective sample size from [n=1870] to females [n=1047] and males [n=823]). Although all subjects are included in unconstrained and constrained models, the information to estimate sex-specific parameters in a model is limited to data for a given sex. It was also necessary to include two ordinal variables in our study; thus, our models have significantly reduced power compared those using continuous variables (Neale, Eaves, & Kendler, 1994; Neale, Røysamb, & Jacobson, 2006).

A second line of evidence from the trivariate Cholesky decomposition model comes from the first A and C factor loadings (See Figure 5). The presence of familial influences (i.e., additive genetic or shared environmental influences) on all three behaviors reduces the likelihood that these influences are only acting on lifetime number of partners indirectly through sex under the influence. A common additive genetic factor explained covariance between these variables among females, and a similar pattern for shared environmental influences was seen for males. While these biometrical models cannot prove or disprove a causal model, this evidence suggests that correlated liabilities between substance use and number of lifetime sexual partners may not be specific to sexual contexts and that not all of the overlap is due to direct effects through sex under the influence.

In addition to the Cholesky decomposition models, direction of causation (DoC) models were used to test the strength of evidence for causality. These models tested the likelihood that: 1) sex under the influence has a direct effect on lifetime number of partners, and 2) reverse causation, or that RSB actually directly causes more sex under the influence. While there is some work that suggests these behaviors may show a pattern of reciprocal causation (O’Hare & Cooper, 2015), we were unable to test this third theoretical model as our RSB measure did not have multiple indicators. Results suggested that sex under the influence did not have a direct causal effect on lifetime number of sexual partners, rather the reverse pattern better fit the data.

As the aim of the study was to test whether drug and alcohol use during sexual decision-making caused people to take more risks, the reverse causation finding is somewhat unexpected. The underlying mechanisms of this potential causal pathway were not explored in this study and are avenues for future research. For instance, it is unclear whether these effects are driven by sensation seeking (e.g., those with higher numbers of partners may be more motivated to add drugs or alcohol into the situation to enhance the experience) or partner effects (e.g., those with higher RSB are more likely to acquire partners who are more likely to introduce substance use into the sexual situation). There is some evidence that desire to engage in RSB predicts subsequent alcohol use, specifically in contexts that were more likely to lead to the acquisition of a new sexual partner (e.g., at bars and parties; Hone, Testa, & Wang, 2019). Given the dearth of evidence for this causal pathway and difficulty teasing apart global associations from event-level effects, caution is advised.

Importantly, several limitations of the DoC models should be considered when interpreting these results. The model identified a pattern that is necessary but not sufficient for causality, however unmeasured third variables cannot be completely ruled out and should be investigated further. While the reverse causation model fit better than a model where sex under the influenced caused RSB, our results were also indicative of partial confounding. That is, a reconfiguration and equally valid interpretation of the trivariate Cholesky model suggests the effect of RSB on sex under the influence is in part explained by shared risk factors for general substance use.

Additionally, it is possible that the pattern of results observed was related to higher measurement error for the sex under the influence variable compared to number of lifetime partners (Heath et al., 1993). Replication with rigorous measurement is required before further speculation. Though results are suggestive that RSB could contribute to increases in sex under the influence, they provide stronger support for the idea that sex under the influence is unlikely to cause RSB.

Previous studies that used biometrical models to test the role of adolescent alcohol and drug use on other RSB cast similar doubts on a causal model in adolescence. One study found that age of smoking and drinking were non-causal predictors of early age of sexual initiation (e.g., shared familial factors influenced both traits, controlling for conduct disorder symptoms), though tests of early drunkenness could fit a partially causal model (Deutsch et al., 2014). Agrawal et al. used a sample of female twins to test a causal model of adolescent cannabis use (e.g., before age 17) on repeated voluntary unprotected sex in early adulthood (2016). After controlling for related behaviors and covariates (e.g., adolescent drinking and cigarette smoking, early sexual initiation, early puberty, symptoms of conduct disorder, and childhood sexual abuse), some genetic overlap was detected (i.e., genes that influenced early cannabis initiation also influenced adult non-condom use via biological or mediated pleiotropy). As genes did not explain all of the covariance between these behaviors (i.e., there was significant nonshared environmental overlap), the authors did not rule out the possibility of a partial causal role between these behaviors (Agrawal et al., 2016). To our knowledge, there are no genetically informed studies that test a measure of sex under the influence (substance use during sexual decision making) during adolescence. Given that this is a target population for intervention, a logical follow up would be to extend this specific test to younger samples.

There are several ways in which our study differs from past research, which should be considered when interpreting the results. Our sample was community based, therefore the range of behavior was fell largely within a normative range (Liu et al., 2015). Caution should be taken when extending our results to clinical populations or those with higher mean numbers of sexual partners. Further, we assessed our participants in early adulthood when approximately half of the sample reported having a single sexual partner in the past year. In contrast, many of the other investigations have focused on late adolescence. While late adolescence is often a period of radical environmental changes (e.g., leaving parent’s home), early adulthood is particularly interesting in terms of selecting different life trajectories with various degrees of risk (e.g., getting married). Concerns about twin representativeness should be considered, given the importance or social or peer networks (i.e., which twins may share) on the emergence of both substance use and sexual behaviors during development (Steinberg, Fletcher, & Darling, 1994; Friedlander et al., 2007). However, there is no evidence to suggest that this would increase or decrease these behaviors in twins compared to singletons.

The measures and instruments available in our longitudinal twin samples deviated from other more extensive surveys of sexual behavior or event-level studies. When available, variables such as frequency of condom use, risky sexual acts (e.g., those with higher rates of transmission of diseases or infections), or sex with risky partners (e.g., with higher risk of disease or infection, causal partnerships) can be more nuanced predictors of RSB- related negative outcomes.

A trade off with highly contextual variables is error, in that events that are distal (Croyle and Loftus, 1993) or that are frequent in nature (Catania et al., 1990) are more likely misreported than those that are recent and rarer. Many of these contextual variables were only available for a small subset of the sample; thus, we did not believe we would gain any predictive power and potentially increase our error had we included them. Rather, we chose to use number of lifetime partners as a well-validated proxy of RSB. Similarly, our use of a composite measure of sex under the influence is fairly unique. In essence, it is an attempt to measure how much drug and alcohol use influences sexual decision-making, rather than a measure which objectively assesses whether drug and alcohol use led to risker sex than would have occurred if sober. Possible implications of this measure (e.g., subjectivity, contributions to gender differences) are discussed below. Finally, it is possible that a minority of the participants with only one partner in the past 5 years scored as “no risk” would have endorsed some of the sex under the influence items (e.g., had a drink to feel more comfortable with [regular] partner). Consistent with other work (Epstein et al. 2014), we chose to consider these behaviors within the context of a committed relationship as non-risky due to several reasons. While sex under the influence may be risker with a regular partner in some regards (e.g., unplanned pregnancy), on average the risk for contracting or spreading an STI from a regular partner is minimal compared to people with multiple partners. This risk is highest for those with concurrent multiple partners (Morris & Kretzschmar, 1997); however, risk is still elevated for serial monogamists or those with multiple subsequent partners (Corbin & Fromme, 2002). Thus, our sex under the influence measure aimed to measure the frequency of situations where drug and alcohol use potentially led to risker sexual behavior, which is markedly different than when a person in a committed relationship may use drugs or alcohol to feel more relaxed or comfortable with their regular partner. We encourage future investigations using genetically informed designs to use other available measures.

Despite these caveats, the gender differences in all of the models are of particular interest. It is clear that the reasons for using drugs and alcohol before (or during) sexual activity are different across males and females. In males, sex under the influence was more highly heritable and there is substantial genetic overlap specific to sex under the influence and RSB. As such, future research should further explore the biological systems core to these behaviors (e.g. hormonal system, or genetic basis of underlying traits such as sensation seeking). In females, genes may influence both substance use within and outside of sexual activity and increased sexual encounters. While pharmacological disinhibition theories such as the alcohol myopia hypothesis (Steele and Josephs, 1990) could be at play, differential social pressures for men and women to use substances or to limit sexual behavior may also contribute to these gender differences (Beck, Thombs, Mahoney, & Fingar, 1995). It is also possible that males and females could have different substance use expectations or different pressures to report that their behavior was a consequence of alcohol or drug use. For instance, to the extent to which there is social pressure for males to have more partners and females fewer partners and there are strong social expectations about substance use and risky sexual behavior (e.g., in college campuses; there may be some benefit for females, especially, to endorse that their decisions were clouded by drug or alcohol use), we may expect different patterns of sex under the influence and RSB for males and females. As such, our measure may be capturing the frequency in which substance use impaired sexual decision making (presumably, leading to increased risk), capturing a tendency to use substances in order to facilitate sexual encounters, or a capturing a tendency to endorse sex under the influence measures as a form of post-hoc rationalization to remove dissonance or social shame of RSB. Both examples of reverse causation, where intention to have sex precedes substance use (Cooper 2006), may function differently across gender. Further investigation into the specific motivations to endorse individual risky behaviors associated with sex under the influence (e.g., non-condom use while intoxicated) could inform our interpretation of these gender differences.

In sum, findings indicate that claims that drug and alcohol use during sex causes risker sexual behaviors may be overstated. We identified no clear pattern of causality using bivariate twin modeling, and found evidence that a large proportion of the shared liabilities between sex under the influence and number of lifetime sexual partners was attributable to general substance use. DoC models indicated that it is unlikely that sex under the influence caused more RSB, but that the reverse cannot be ruled out. Future research is required to investigate the nature of this potential causal pathway.