Weight Loss and Sexual Activity in Adult Obese Individuals: Establishing a Causal Link

1 Introduction

A close link between a fulfilling sex life and general life satisfaction is well documented in the literature (e.g. Woloski-Wruble et al. 2010; Schafer et al. 2013; Soysal and Smith 2022). A growing body of literature also demonstrates that a higher frequency of sexual activity is associated with better health. For example, Kloner et al. (2016) and Soysal and Smith (2022) document reductions in cardiovascular events later in life, lower risk of fatal coronary events, prostate and breast cancer, and better quality of life. Impairments in sexual life thus are likely to have a strong negative impact on individual happiness and life satisfaction and likewise deteriorate the general health condition.

One possible reason for limitations in sexual life is obesity; see Larsen et al. (2007) and Sarwer et al. (2018) for reviews of the medical literature focussing on the linkage of obesity and sexual dysfunction such as erectile dysfunction. Though the majority of analyses reviewed by Larsen et al. (2007) and Sarwer et al. (2018) find a positive connection, the evidence is still mixed.^[1] Taking a less physiologically oriented perspective and considering deficits in sexual life beyond sexual dysfunctions, Kolotkin et al. (2006) list ‘lack of enjoyment’, ‘lack of desire’, ‘difficulties with performance’, and ‘avoidance of sexual encounters’ as examples of obesity related impairments of sexual life. Rowland et al. (2017) state that the relationship between obesity and sexual function is complex and multivariate, with at least three different likely pathways: direct effects of adipose tissue, effects of comorbidities, and psychologically driven effects.^[2] In general, besides undesirable general health outcomes, obesity may have a more direct detrimental effect on happiness and life satisfaction that operates through an unsatisfactory sex life.

The non-medical, health economics oriented literature has taken a very different view on the link of body weight to sexual life. This relatively small literature (Ali et al. 2014; Cawley 2001; Cawley et al. 2006; Halpern et al. 2005; Neymotin and Downing-Matibag 2014; Sabia and Rees 2011) focusses on individuals of very young age, adolescents in particular. Due to the focus on this special age group, sexual activity is associated with premature initiation to sex, potentially exerting detrimental effects on adolescents’ later lives. For female adolescents, Sabia and Rees (2008) find a causal and detrimental effect of early sexual intercourse on psychological well-being. Ali et al. (2014) allude to disease and pregnancy risks teenagers take when having early sex, suggesting that a postponed initiation to sexual activities might be a beneficial sideeffect of teenage overweight.

The contribution of the present paper is to (i) conduct an analysis focusing on adult individuals, for whom the above line of argument does not apply and whose sexual behavior, most likely, differs from the behavior of the age group studied by the literature cited above. Unlike the majority of the medical literature, our focus is (ii) on (self-reported) sexual activity rather than specific sexual dysfunctions such as erectile dysfunction. This is important as obesity related limitations in sex life may often be unrelated to any physiological deficit but can possibly be attributed to social and psychological body-weight related factors, such as being less attractive to a potential partner, a lack of self-confidence when dating, or less enjoyment in and desire for sex. Finally, the present analysis contributes to the existing evidence by (iii) establishing a causal effect of weight loss on sexual activity as opposed to a mere correlation in the data.

Causality in the link between body weight and sex life has not always attracted much attention. Halpern et al. (2005), for instance, estimate that the probability for adolescent girls of being involved in a romantic relationship (with or without sexual intercourse) increases by 6–7 percent with one BMI unit less. Using Add Health (National Longitudinal Survey of Adolescent Health; Harris 2013) data,^[3] a US longitudinal school-based survey, Cawley et al. (2006) find that the odds of initiation to sexual intercourse for obese girls are just 32 percent of those of normal-weight girls. However, both analyses do not account for unobserved heterogeneity. Hence, the striking results may well be attributed to unobserved confounding factors, such as personal character traits being relevant for both obesity related behavior (eating habits, physical activity) and mating behavior. Numerous medical studies (see Larsen et al. 2007; Sarwer et al. 2018) likewise focus on the association of overweight and impairments in sex life rather than on a causal relationship.

More recent papers address possible endogeneity by using instrumental variables (IV) estimation. Analyzing data from Add Health and using siblings’ BMI and mother’s obesity status as instruments for the respondents’ body weight, Sabia and Rees (2011) find that for girls aged 14 to 17 one BMI unit more reduces the probability of sexual activity by 3.5 percentage points. In contrast, no significant effect is found for boys. Rather, the point estimate for boys bares the opposite sign in several specifications. Ali et al. (2014) take a similar approach by instrumenting body weight with maternal obesity status. For white girls – but not for blacks – they find that higher body weight or obesity significantly lowers the probability of having had sex or having been involved in a romantic relationship.

Though both papers carefully discuss the validity of the instruments used, one may still doubt whether mothers’ – and other relatives’ – body weight is a valid instrument: body weight as well as sexual behavior – or at least attitudes towards it – may be intergenerationally transmitted from mother to child (e.g. Taris 2000). If so, maternal body weight could influence children’s sex life through mothers’ own sexual behavior, given that body weight matters for sexual behavior. Another concern with respect to instrument validity is the social environment children share with their mothers and siblings. It may strongly influence both, body weight and attitudes towards sex for both mothers and children and, most likely, cannot entirely be controlled for by including covariates in the analysis.

While following Ali et al. (2014) and Sabia and Rees (2011) in relying on instrumental variables estimation, we do not use maternal body weight as instrument. Rather, in our analysis identification is based on exogenous variation in BMI having been artificially induced in a randomized field experiment. More precisely, we use data (Augurzky et al. 2017) originating from a randomized trial that was conducted to evaluate the effectiveness of financial incentives for weight loss (Augurzky et al. 2012, 2018; Eilers and Pilny 2018). In this experiment, individuals in the intervention groups were exposed to financial incentives for losing body weight, while control group members were not. Following Hafner et al. (2021) and Reichert (2015), treatment status can serve as an instrument for the change in body weight. Besides using different instruments and focussing on adult individuals, our analysis materially deviates from Ali et al. (2014) and Sabia and Rees (2011) by only considering obese^[4] individuals. That is, we are interested in whether – even moderate – weight loss in obese individuals makes a difference to their sexual lives.

The identification strategy of the present analysis is related to the approach of Esposito et al. (2004). This medical study also relies on data from a randomized field experiment with obese participants. In this study, the intervention was not exposition to financial incentives but guidance on how to lose weight combined with weight monitoring through monthly group sessions. Esposito et al. (2004) estimate a reduced form effect by comparing post-intervention IIEF (International Index of Erectile Function) scores between the intervention and the control group. The statistically significant advantage in IIEF found for the intervention group nevertheless provides striking evidence for body weight mattering for sexual dysfunctions in obese persons. However, by focusing on erectile dysfunction alone, Esposito et al. (2004) may miss out on other obesity related impairments of sexual life. By considering two different measures of sexual activity, our analysis takes a broader perspective on the link between obesity and sex life.

The remainder of this paper is organized as follows. Section 2 introduces the data and describes the underlying experiment. Section 3 discusses the econometric model and Section 4 presents the estimation results for the basic model. In Section 5, alternative model specifications and estimation results are discussed. Section 6 summarizes and discusses the main findings.

2 Data

2.1 The Experiment

The data (Augurzky et al. 2017) used in this analysis originate from a large scale, four-phase, randomized field experiment that is described in more detail elsewhere (Augurzky et al. 2012, 2018; Eilers and Pilny 2018; Hafner et al. 2021). The experiment was conducted between March 2010 and July 2013. Roughly 700 obese^[5] individuals were recruited in four rehabilitation hospitals in Germany (For details about rehab in Germany, see e.g. Ziebarth 2014; Krauth and Bartling 2017). After a typically three weeks stay, but shortly before discharge, the physician in charge set all participants an individual weight-loss target that they were prompted to realize within four months. Subsequently, the participants were randomly assigned to two incentive groups (incentive 150_1–4 and incentive 300_1–4)^[6] and one control group (incentive 0_1–4). Covariate balancedness across the groups is documented in Augurzky et al. (2018, Table 2). Members of the former two groups were promised € 150 and € 300, respectively, for realizing (or exceeding) their individual weight-loss target within four months, while members of the control group were not exposed to any financial incentives for reducing body weight.

The weigh-in at the end of the weight-reduction phase – and all further weigh-ins – were carried out at assigned pharmacies. As the pharmacy was selected by the experimenter and not by the test person, the location of weigh-in was exogenous to the participants. Shortly after the weigh-in, all participants – irrespective of group membership in the weight-reduction phase and irrespective of success – were prompted to comply with their target weight by the end of a subsequent six months weight-maintenance phase.^[7] Yet, conditional on success in the weight reduction phase, a second randomization took place at the same time. Participants who had lost at least 50 percent of their contractual weight-loss target were randomly assigned to two intervention groups (incentive 250_5–10, incentive 500_5–10) and one control group (incentive 0_5–10). The former two could gain up to € 250 and € 500, respectively, if they achieved their target weight or less ten months after rehab discharge. Members of the control group were not promised any rewards. Individuals who were not successful in the weight-reduction phase were hence effectively assigned to the control group. Six months later, i.e. ten months after rehab discharge, another weigh-in took place under identical conditions. Yet, at this point in time, the incentive scheme was terminated and no further randomization took place. However, all participants were still requested to comply with their target weights and another weigh-in was announced to take place one year later. This final year serves as the follow-up phase. See Figure 1 for the time line of the entire experiment.

Figure 1:

Time line of the four-phase experiment. N indicates sample sizes at the four points in time at which data was collected. The differences in N (177, 106, and 96) are the number of drop-outs in the respective phase of the experiment. Source: Augurzky et al. (2018), subject to minor modifications.

By the end of each of the four phases, the participants had to answer a detailed questionnaire. The final one, being sent 22 months after rehab discharge, is the only one to include questions about the participants’ sex life. Yet, only for participants with even identification numbers – i.e. for a fifty percent random sample^[8] – the questionnaire included these touchy questions. The experimenters wanted to reduce the risk of too many individuals dropping out by asking them – possibly displeasing – questions about their sex lives. However, this concern proved immaterial as the drop-out rate virtually did not differ between individuals with even and with odd identification numbers.

The experiment population was subject to significant sample attrition. 695 individuals started the weight-reduction phase after rehab discharge. 177 participants dropped out during the weight-reduction phase, i.e. they did not show up at the weigh-in after four months, leading to only 518 individuals entering the weight-maintenance phase. In this phase, another 106 individuals dropped out. The follow-up phase was also subject to some sample attrition reducing the number to 316 individuals for whom weight information is available by the end of the experiment. Among these, 174 were asked questions regarding sex life. Only 17 denied such information. This corresponds to a rate of item-non-response of less than 10 percent, which is remarkably low compared to many surveys addressing sexual behavior (cf. Fenton et al. 2001). Attrition usually occurred in terms of just not showing up at the next scheduled weigh-in.

3 Estimation Procedure

The econometric model rests on formulating a likelihood function that jointly considers the three model components (outcome equation, ‘first stage equation’ for weight loss, selection equation) and hence allows for simultaneous estimation by maximum likelihood (ML). The estimation approach incorporates the assumption of joint normality, which makes us use a probit model with an endogenous regressor (Wooldridge 2002, p. 476) augmented by a Heckman-type selection equation. Though the model components are estimated simultaneously, the number of observations effectively contributing to the identification may still be heterogeneous across the components. For more details on simultaneous estimation of multi-equation models with jointly normally distributed errors, see Roodman (2011). This paper also introduces an accompanying software component.^[13] Simple two-stage least squares combined with the selection correction proposed by Olsen (1980) is a fully linear, multi-step – arguably simpler, yet rather inefficient – alternative to joint ML estimation. However, the ordered categorial nature of the outcome sexfrequency_11–22 renders a fully linear model questionable for the present application, making us stick to ML estimation.^[14]

Weak identification has become a major concern in instrumental variables estimation. The usually reported F-statistic of the test on the instruments – and the corresponding rule-of-thump threshold value of 10 – do not, however, have a statistical basis in the context of non-linear IV-models (Denzer and Weiser 2021). Therefore, we follow the procedure as suggested in Denzer and Weiser (2021) and report the Shapiro–Wilk W-statistic as a measure of instrument strength not limited to linear models. The value of this statistic is compared to the critical values (0.994 for 5 percent relative bias, 0.952 for 10 percent relative bias) provided by Denzer and Weiser (2021, see p. 28) in the context of a probit model with endogenous regressors.^[15]

4 Estimation Results for the Preferred Specification

In this section, we discuss estimation results for our preferred model specifications described above. Table 3 displays results for the model variant using sexpartner_11–22 as dependent variable, while Table 4 displays results for the model explaining sexfrequency_11–22. Since the determinants of sex life are likely to differ substantially between genders, we also estimated separate models for males and females. Yet, when using the rather small all-women sample, the ML estimator suffered from severely weak instruments (W-statistics of about 0.1) and convergence issues, which made us abstain from reporting results for that specification. On the contrary, with the outcome sexpartner_11–22 the Shapiro_–Wilk W-statistic exceeds the threshold value of 0.952, both for the pooled and the all-males sample, which argues against weak identification. This does not in the same way apply to the estimations considering the frequency of sexual intercourse as the outcome variable. Yet, the W-statistics still take rather large values of about 0.9; see Table 4.

The estimates for the cross-equation error correlation ρ bmi sex are positive throughout and statistically significant (except for sexpartner_11–22, pooled sample) suggesting that endogeneity of weight loss needs to be taken into account. From a non-technical perspective, this finding conflicts with our earlier reasoning that self-confidence is a major source of unobserved heterogeneity. It rather argues in favor of the desire for physical pleasures driving the correlation. For the estimates of ρ info sex the pattern is less clear. Though also positive and of substantial magnitude, they are only statistically significant (at conventional levels) for the models with binary outcome. But still, the rather strong estimated cross-equation error correlations suggest that selection cannot be ignored. For the outcome sexpartner_11–22 pooling the sample over the sexes is clearly not rejected by a likelihood-ratio test. In line with this, the estimated coefficients exhibit little difference between the all-male and the pooled sample. The corresponding test for the outcome sexfrequency_11–22 is at the margin of rejecting the pooled model. Yet, the estimated coefficients still do not deviate strikingly.

Turning to the question of what makes a sexual relationship more likely, Table 3 indicates that being involved in a sexual relationship is not significantly associated with age. Not surprisingly, those who lived with a partner by the end of the weight-reduction phase are more likely to have a sexual partner 6–18 months later (statistically significant only for the pooled sample). Model estimation also yields a significantly negative association with initial BMI. Yet, this coefficient does not isolate a causal relationship but is likely to capture the influence of unobserved individual heterogeneity. Most importantly, the coefficient of Δ BMI_1–10 is statistically significant (in the pooled sample only at the 10 percent level) and negative for both the pooled and the all-males model. This indicates that – conditional on initial weight and initial relationship status – losing body weight increases the likelihood of being involved in a sexual relationship. The coefficient estimates correspond to marginal effects of substantial size. At its respective maximum value ( β ̂ Δ BMI ⋅ ϕ ( 0 ) ) the corresponding marginal effects are as high as −0.143 (all male) and −0.128 (pooled), meaning one BMI unit reduction in body weight increases the probability of having a sexual partner by more than ten percentage points. The mean marginal effects in the estimation sample are of similar magnitude (−0.112 and −0.099, respectively).

To dig deeper into the interdependence of body weight and sexual activity, we re-estimated the model with an indicator for living together with a partner (together₂₂) as an alternative dependent variable; see Table 6 and Figure 3 in the Appendix. The estimated coefficients were in absolute terms smaller and by all standards statistically insignificant (and of the opposite sign). In other words, it is not the relationship status per se, but its sexual nature that is significantly affected by losing body weight.

In line with this result, we find statistically significant and negative effects of weight gain on the outcome frequency of intercourse, too. Table 4 shows that the point estimates are very similar to their counterparts from using the indicator for a ‘sexual relationship’ as the outcome variable. This implies that – according to these estimates – losing one BMI unit has quite a big effect on the probability of having sex regularly. More specifically, for the most affected individual the effect amounts to 12.6 percentage points (all-males) and 13.2 percentage points (pooled), respectively.

To sum up, our preferred model shows that even moderate weight loss in obese individuals has remarkably strong effects on their sex lives, both at the extensive and the intensive margin. Nevertheless, this result is driven by male individuals. Albeit including females in the estimation sample has little impact on the estimation results, the rather small size of the sub-sample of women does not allow for a reliable analysis of the link between weight change and sex life separately for women.

5 Robustness Checks

We consider conditioning on success in the weight-reduction phase and simultaneously accounting for sample attrition as the most appropriate approach to dealing with the limitations of the data. This approach, however, results in estimating a rather complex model – which involves a trivariate normal error distribution – using a rather small sample. Therefore, we also estimate several alternative model specifications, which are technically simpler and/or allow for using a larger estimation sample.

More specifically: (i) We estimate the reference model just without a selection equation. (ii) We condition on success in the weight-loss phase not by restricting the sample to successful participants, but by including an indicator for success as control, which is instrumented by the weight-loss phase incentive indicators incentive 150_1–4 and incentive 300_1–4. This specification, which does not involve a selection equation either, allows for using a substantially larger sample than the reference. (iii) We do not condition on success but use the total weight change over both, the weight-loss and the weight-maintenance phase (Δ BMI_1–10) as explanatory variable, which is instrumented by all four incentive indicators incentive 150_1–4, incentive 300_1–4, incentive 250_5–10, and incentive 500_5–10. This specification is rather simple as it does not involve a selection equation and allows using a relatively large sample compared to the reference model. (iv) We estimate a fully linear, one-to-one counterpart to the reference model combining two-stage-least squares with the selection correction suggested by Olsen (1980); cf. Section 3.

Figure 2 graphically displays the results from these alternative models for the coefficient/marginal effect of Δ BMI and contrasts them with the results of the preferred model. Full regression output is provided in the Appendix; see Tables 7–14. Starting with specification (i), Figure 2 (upper left panel) indicates that it makes little difference to the estimated effects of weight loss when attrition is not taken into account. Without a selection equation, we estimate slightly stronger effects on the probability of living in a sexual relationship and slightly smaller effects on the frequency of intercourse. Yet, the confidence intervals overlap heavily for the two specifications. One exemption is the pooled model specification. Here the effect on sexfrequency_11–22 turns statistically insignificant at the 10 percent level for the simpler specification without selection equation. The pattern of weight-change coefficients found for specification (ii) exhibits almost exactly the same pattern found for specification (i); see Figure 2, upper right panel. That is, including a control indicating success in the weight-loss phase makes little difference compared to excluding unsuccessful participants. As before, without modeling attrition from the experiment in the sample that pools men and women, the effect on the frequency of intercourse turns insignificant. Yet, from a statistical perspective the estimated coefficients still hardly deviate from their counterparts in the reference specification. Things are different for specification (iii) (Figure 2, lower left panel) which does not condition on success in the weight-loss phase, considers the weight change over both intervention phases, and uses all incentive-group identifiers as instruments for the latter. Although in qualitative terms the results do not deviate from the reference model, the estimated coefficients become much smaller. For the outcome sexfrequency_11–22, they are clearly statistically insignificant, though the confidence intervals are tighter than for the reference model (presumably because of the substantially larger sample used for estimation). Thus, specification (iii) questions the effect of weight change at least on the outcome frequency of intercourse. However, this specification uses instruments whose validity is questionable by the design of the experiment (cf. Section 2.3). The deviation from the reference may, hence, not indicate a lack of robustness but might just be due to using invalid instruments. In order to make the results of the fully linear multi-step model (iv) comparable to the reference, the lower right panel of Figure 2, depicts (average) marginal effects rather than coefficient estimates. In qualitative terms, the results are similar to their counterpart from the reference model. Yet, the estimated effects are bigger in absolute terms. The most striking difference to reference is, however, the much wider confidence intervals. This lack of precision can presumably be attributed to the rather inefficient multi-step estimation procedure. Based on the linear model, one can hardly make any reliable statement regarding the link between weight loss and sex life. In turn, the results from the linear model neither confirm nor challenge their counterparts from the non-linear specification of reference.

Figure 2:

Grey: reference specification (Tables 3 and 4); black: (i) reference model without selection equation, (ii) model that includes an indicator for success in weight-loss phase as control, (iii) model that does not condition an success in weight-loss phase, (iv) 2SLS combined with linear Olsen (1980) selection correction. Spikes indicate 90 percent confidence intervals; confidence intervals bootstrapped for the linear model. Markers: , outcome sexpartner_11–22; , outcome sexfrequency_11–22. Source: own calculations.

6 Discussion and Conclusions

This paper analyzed whether weight reduction in adult obese individuals has an impact on their sex lives, namely whether or not they live in a sexual partnership and how frequently they have sex. Based on data generated by a randomized field experiment, where randomization provides an exogenous source of variation in body weight exploited for identification, we find that obese males are substantially more likely to be involved in a sexual relationship when losing weight. This effect is of substantial magnitude, amounting to about 14 percentage points for the most affected individual in our preferred model. In the preferred specification, the results for the frequency of sexual intercourse are very similar. Yet, in some alternative specifications, the effect on the frequency of intercourse is not statistically significant. This applies in particular to estimation samples not restricted to males. Moreover, weak identification seems to be more of an issue if the intensive margin of sex life is analyzed.

In contrast to those findings for men, the analysis does not yield any significant and reliable effect for women. This can be explained by the small number of women in our data that seriously hampers the empirical analysis. Another explanation might be gender differences in sex life. According to Baumeister et al. (2001) there is ample empirical evidence for men – as compared to women – having more frequent and more intense sexual desires. The strength of sexual motivation, also referred to as sex drive, is for instance reflected in the desired number of partners and the desired frequency of intercourse, which are closely related to the outcome variables of the present study. The stronger male sex drive may facilitate detecting a causal effect of weight loss for men. Based on an extensive literature survey, Petersen and Hyde (2010) state only small gender differences in many sexual behaviors. But casual sex and attitudes toward casual sex are among the exceptions to this general finding. Men report a more active sexual behavior or express more permissive attitudes towards them than women do. This diverging predisposition could be another explanation for the significant effects found for men but not for women, concerning the probability of being involved in a sexual relationship and the frequency of sexual intercourse.

For many individuals, a fulfilling sex life is an important facet of general life satisfaction and happiness as well as better general health. Therefore, our key results may provide additional motivation for obese males to reduce overweight, similar to tobacco education where preventing impairments in sexual life is an important argument for smoking cessation (Linnebur 2006). From a health policy perspective, improved sexuality might be seen as one of the benefits of weight loss. We cannot say much about the mechanism behind fulfilling sex and better mental and physical health, but scientists in public health and medicine (see e.g. Soysal and Smith 2022) argue that sexual activity can be considered as a form of physical exercise (Smith et al. 2019), so those who engage in regular sexual activity are likely to enjoy the mental and physical benefits of a physically active lifestyle. Moreover, during sexual activity or at the peak of intercourse, endorphins are released, endogenous opioid peptides that act as neurotransmitters and produce a feeling of pleasure or bliss (see Ponzi and Dandy 2019). As argued by Soysal and Smith (2022), circulating endorphins and their association with higher natural killer cell activity have also proved to prevent lung infections and to play an important role in ameliorating lung cancer and many other diseases (Ponzi and Dandy 2019; Zhang et al. 2021).

Moreover, our key result provides further evidence for substantial social costs of obesity that are not necessarily linked to pathological health problems. This raises the question about the channels through which the established effects operate, which cannot be answered on basis of the data used in this study. One possibility is recovery from sexual dysfunction on which the medical literature has focussed. In their review Sarwer et al. (2018), for instance, stress that weight loss, whether from bariatric surgery or less intensive interventions, is associated with significant and clinically meaningful changes in sexual functioning and relevant reproductive hormones. Yet, given the strong effects of a moderate weight reduction found in the data, other channels likely play a role. Greater self-confidence when initiating sexual contact, improved attractiveness to potential sexual partners, feeling more comfortable with their own body, and more desire for sex may act as such channels. One limitation of the present analysis is that it does not allow for disentangling the effect of a reduced body weight from the effect of the process of weight reduction. The strong effects found may therefore rather be attributed to being successful in meeting a major individual challenge than to weight loss itself. One possible route to address this question is to evaluate the effects of bariatric surgeries on sexual activity.

While the existing non-medical literature has a strong focus on the effect of overweight on young individuals’ sex lives, our result explicitly refers to adult obese individuals. This can be interpreted in terms of another desirable consequence of weight reduction. In fact, the argument of overweight having the desirable side effect of counteracting premature initiation to sex, put forward elsewhere in the literature, does not apply given the age group considered in the present analysis. Our results rather provide further evidence for obesity imposing restrictions on private lives, but also for even moderate weight reduction mitigating these limitations.