Machine learning: Men's sexual satisfaction was overall more predictable than women's

Identifying the strongest self-report predictors of sexual satisfaction using machine learning. Laura M. Vowels, Matthew J. Vowels, Kristen P. Mark. Journal of Social and Personal Relationships, January 11, 2022. https://doi.org/10.1177/02654075211047004

Abstract: Sexual satisfaction has been robustly associated with relationship and individual well-being. Previous studies have found several individual (e.g., gender, self-esteem, and attachment) and relational (e.g., relationship satisfaction, relationship length, and sexual desire) factors that predict sexual satisfaction. The aim of the present study was to identify which variables are the strongest, and the least strong, predictors of sexual satisfaction using modern machine learning. Previous research has relied primarily on traditional statistical models which are limited in their ability to estimate a large number of predictors, non-linear associations, and complex interactions. Through a machine learning algorithm, random forest (a potentially more flexible extension of decision trees), we predicted sexual satisfaction across two samples (total N = 1846; includes 754 individuals forming 377 couples). We also used a game theoretic interpretation technique, Shapley values, which allowed us to estimate the size and direction of the effect of each predictor variable on the model outcome. Findings showed that sexual satisfaction is highly predictable (48–62% of variance explained) with relationship variables (relationship satisfaction, importance of sex in relationship, romantic love, and dyadic desire) explaining the most variance in sexual satisfaction. The study highlighted important factors to focus on in future research and interventions.

Keywords: Sexual satisfaction, machine learning, random forests, Shapley values

Discussion

Our results showed that we could predict between 48 and 62% of the variance in sexual satisfaction using a random forest algorithm, up to two to three times more than previous studies even after deleting relationship satisfaction from the model (Byers & Macneil, 2006; Laumann et al., 2006). The algorithm is also explainable because it does not suffer from suppression and cancellation effects or multicollinearity. The results show that using machine learning can help move psychological research into a new era of highly predictive and accurate models that generalize better to the population and have a higher utility in practice (Yarkoni & Westfall, 2017).

The strongest predictors

Because of the importance of sexual satisfaction on relationship quality (Joel et al., 2020; McNulty et al., 2016; L. M. Vowels & K. P. Mark, 2020b) and overall well-being (Davison et al., 2009; Del Mar Sánchez-Fuentes et al., 2014), understanding factors that are the most, and the least, strongly associated with sexual satisfaction is important. This can enable researchers and practitioners to target individuals who may be at a particular risk of poor sexual satisfaction and helps to address factors that are the most likely to induce changes in sexual satisfaction while ignoring those that are the least likely to produce change. Thus, we added to the literature by examining which factors were the most, and least, predictive of sexual satisfaction in two samples.

Several variables that have previously been identified as important predictors of sexual satisfaction were included in the top-10 predictors: relationship satisfaction (Joel et al., 2020; McNulty et al., 2016; L. M. Vowels & K. P. Mark, 2020b), dyadic desire (Kim et al., 2020; Mark, 2012, 2014), romantic love (L. M. Vowels & K. P. Mark, 2020a), sexual communication (Impett et al., 2019), and perception of love and sex (Hendrick & Hendrick, 2002). Importantly, when relationship satisfaction was low, it had up to three times higher impact on the model outcome compared to when relationship satisfaction was high. Furthermore, participants in Sample 1 who viewed sex as an important part of their relationship and those who had sex regularly also had higher sexual satisfaction compared to participants who placed less importance on sex and more on love and had sex less frequently. Similarly, participants who reported a higher frequency of more varied sexual behaviors such as giving and receiving oral sex and mutual masturbation in Sample 2 reported higher levels of sexual satisfaction. These results suggest that frequency and value of sex as well as a more varied sexual repertoire in relationships are important predictors of sexual satisfaction. More varied sexual repertoire is also likely to lead to more satisfying sexual experiences, especially for women given that women have a higher likelihood of orgasm from clitoral stimulation than from intercourse. These results confirm earlier findings using traditional statistical models (Haavio-Mannila & Kontula, 1997; Laumann et al., 2006).

Gender was not an important predictor of sexual satisfaction suggesting that men and women overall had similar levels of sexual satisfaction in both samples which is consistent with some studies (Mark et al., 2018; McClelland, 2011) and inconsistent with others (Laumann et al., 2006). Men’s sexual satisfaction was overall more predictable than women’s. This may be because women’s sexuality is thought to be more complex than men’s (Basson, 2001). There were also some notable differences in the top-10 predictors for men and women. Attachment avoidance was only in the top-10 predictors for women’s sexual satisfaction (18th for men changing the outcome very little). Women who were higher in attachment avoidance reported lower sexual satisfaction compared to women lower in attachment avoidance. Attachment avoidance is associated with fear of closeness and intimacy, which tend to be more strongly tied to sexuality for women than men (Péloquin et al., 2014), which may explain why attachment avoidance was particularly important for women.

Consistent with previous studies using both traditional analyses (Rubin et al., 2012; L. M. Vowels & K. P. Mark, 2020a) and machine learning (Joel et al., 2020; L. M. Vowels et al., 2021), including partner effects added little additional variance. However, both actor and partner variables were among the top-10 most important predictors. Partner effects alone could also explain around half as much variance as only actor effects. Important partner variables included partner’s sexual satisfaction, romantic love, relationship satisfaction, and dyadic desire. Interestingly, for women, their male partner’s sexual satisfaction was just as important a predictor for their own sexual satisfaction than their relationship satisfaction. This is consistent with several studies finding that women partnered with men tend to answer questions of sexual satisfaction relative to their partner’s satisfaction as much as their own (McClelland, 2011, 2014; Pascoal et al., 2014) and may be due to there being a societal expectation on women to prioritize men’s pleasure. For men, their female partner’s sexual satisfaction only accounted for about third as much change in sexual satisfaction compared to their own relationship satisfaction. These findings suggest that while we may be able to predict actor’s sexual satisfaction relatively well using only their own variables, accounting for both partners’ variables can provide important additional insights.

The present study also provided an important addition to the literature by evaluating which factors were unimportant for sexual satisfaction. Many of the variables that have previously been associated with sexual satisfaction in traditional analyses were less important compared to other predictors. These included variables such as gender, sexual orientation, children, religiosity, attitudes toward sexuality, and mental health (Del Mar Sánchez-Fuentes et al., 2014; Laumann et al., 2006). This suggests that even though differences in demographic variables may be statistically significant in some studies especially when sample sizes are large (e.g., Laumann et al., 2006), this does not mean that the differences are meaningful. In fact, the present study suggests the opposite; couple’s overall relationship and sexual behaviors are more proximal to sexual satisfaction and appear more important than who the person is. Understanding which variables are less related to the outcome is important, so that researchers and practitioners do not waste their time and resources on factors that are less likely to change the outcome.

Implications for research, theory, and practice

The study has several strengths as well as important implications for research, theory, and practice. We used explainable machine learning and cross-validation in which the model performance is tested on unseen data to avoid overfitting and thus improve the generalizability of the results. The code used in the study is readily available and provides a pipeline to relationship researchers to conduct more robust and predictable science. The results showed that dyadic level variables are the most likely to contribute to sexual satisfaction while individual predictors are less important. Furthermore, examining individuals’ perceptions of love and sex (Hendrick & Hendrick, 2002), keeping sex as a central element of relationships, and broadening couple’s sexual repertoire may enhance their sexual satisfaction. Finally, we expect many of these variables to have a bidirectional association with sexual satisfaction meaning that improving one (e.g., introducing more varied sexual behaviors) may produce a positive change in the other (e.g., enhanced sexual satisfaction) which will in turn improve the first variable (e.g., increased desire to try new things).

Limitations and future directions

The study also has several limitations that should be considered when interpreting the results. While the study included many predictors that have been associated with sexual satisfaction in previous research, there are other variables that we did not account for, that predict sexual satisfaction (e.g., responsiveness, self-esteem, personality, sociocultural variables). We also only had access to self- and partner-report measures. Thus, the algorithm could only make the predictions based on the variables that were available in the dataset. Therefore, future research should consider a greater number of individual, relational, and societal factors and include behavioral measures to predict sexual satisfaction. We also used data from two relatively large samples including a large subset of couples, the data were convenience samples and limited in their generalizability; most of the participants were white and well-educated and all participants in Sample 2 were in mixed-sex relationships, albeit nearly half the participants were bisexual. We also did not ask participants about any disabilities which may have contributed to their sexual satisfaction. Therefore, future research is needed to examine predictors of sexual satisfaction in a more representative sample. Random forests are a powerful tool that will take advantage of any correlations and interactions in the data, no matter how non-linear, it cannot be used to estimate causality. However, in the absence of a means to reliably estimate causality when examining factors relating to sexual satisfaction, we believe that using a predictive model is perhaps the best option. There are limitations to the Shapley method which have been discussed elsewhere (Kumar et al., 2020), and the notion that the human-interpretable Shapley model sufficiently explains our model suggests that a simpler model may be adequate to begin with, even if the simpler model is harder to identify (Rudin, 2019).

Furthermore, the data were cross-sectional and therefore we could not examine which predictors may account for the most change in sexual satisfaction over time, or indeed whether sexual satisfaction is predictable over time. Joel et al. (2020) found that they could predict little relationship satisfaction longitudinally. Cross-sectional self-report measures are also prone to shared method variance which results in higher correlation among variables collected at the same point in time. We attempted to overcome some of these issues by testing the models without relationship satisfaction given its high correlation with sexual satisfaction and only using partner effects to predict actor’s sexual satisfaction. The models with relationship satisfaction excluded were still predictive but predicted less variance. The models with partner effects alone could predict nearly 30% of the variance in actor’s sexual satisfaction which is higher than most other previous studies using actor or actor and partner effects. Future longitudinal and behavioral research is needed to understand whether the self-report variables measured in this study are predictive over time or whether behavioral measures could also be predictive. Finally, we examined whether men and women differed in the predictors that were important for their sexual satisfaction and future research could also examine whether the predictors of sexual satisfaction differ by sexual orientation.