Positive Treatment Effects and High Heterogeneity of Hormonal Contraceptive Use on Women’s Sexuality

Data

Analyses were based on data from a German panel study called PAIRFAM (Brüderl et al., 2021; Huinink et al., 2011). It contains information about contraceptive use and women’s sexuality from more than 5,000 women. The longitudinal design consists of annual waves with the first data collection in 2008 and the latest available data from 2022 (wave 14). The ethics committee of the Faculty of Management, Economics, and Social Sciences of the University of Cologne approved PAIRFAM. Huinink et al. (2011) provide a detailed description of the PAIRFAM dataset. In addition, the present manuscript contains supporting information including rmd files and html files for the blind code, the simulation code, planned analyses, data wrangling, and conducted analyses online at https://osf.io/u8ntf/.

The data on which our analyses are based were already available and can be used for scientific purposes; the Leibniz Institute for the Social Sciences (GESIS) grants access to the scientific community. Only one of the authors had previously accessed the PAIRFAM data; JMR was granted access to Release 7.0 (waves 1–7) in 2016 within the context of a different research project but never actually worked with the data beyond an initial screening of the included variables to determine suitability for her research question (birth order effects on personality). Thus, some of the data used to answer this research question had been previously downloaded by one of the authors, but we certify that we have not observed any part of the data relevant to the present research question (Level 3 bias control based on the categorization in Table 1 by Peer Community In Registered Report; https://rr.peercommunityin.org/help/guide_for_authors).

Exclusion Process and Participants

We excluded individuals who did not identify as female. Furthermore, once a woman crossed the age of 50 or reported to be (post-)menopausal, her data (including subsequent waves) were excluded, but previous waves of data collection remained in the analysis. In addition, we excluded all individual waves of data in which participants indicated being in a homosexual relationship or only reported homosexual relationships in the past, were pregnant, trying to become pregnant, gave birth to a child in the last year, were currently breastfeeding, or indicated using the morning-after-pill or an unknown contraceptive method. Besides these exclusion criteria explicitly mentioned in stage 1 of this registered report (see Table S1 in the supplement for details on all deviations from the in-principle accepted registered report stage 1), the implemented models for effects of hormonal contraception on sexuality only used information from one wave if the information from the previous wave was also available (because we used predictors from the previous waves in our models). In addition, we could only include information from participants if (1) hormonal contraception was available in the previous and the current wave (2) all additional predictors were available in the current wave and (3) one of the three outcomes was available in the previous and the current wave. To make these data exclusion steps following from modeling decisions more explicit, we now list them as additional exclusion criteria in stage 2 of this registered report.

In further separate robustness analyses, we additionally excluded waves in which participants indicated that they are sterilized, as well as all subsequent waves of those participants. We also excluded all waves in which participants indicated that their partner is sterilized and all waves in which women indicated using no contraceptive method, ⁴ an intrauterine device as a contraceptive method, or hormonal methods other than the oral contraceptive pill. In addition, we excluded all waves in which women indicated that they had never been sexually active. All exclusion criteria, reasons for exclusion, and excluded observations are summarized in Table S2 in the supplement.

In addition to these robustness analyses, which focused on excluding specific women or waves that might bias the estimates of the originally registered main analysis, we would have liked to conduct exploratory subanalyses based only on women who reported being in a homosexual relationship or who have reported only homosexual relationships in the past (otherwise using the same exclusion criteria as in the originally registered main analysis). While we hoped to gain some initial insight into the potential effects of hormonal contraceptives on sexuality of homosexual women, the sample size after applying our registered exclusion criteria was already too small to perform any meaningful analysis (n = 188, observations = 539). We had the information from the previous and the current wave that was necessary to perform the proposed models for only 62 exclusively homosexual women (213 observations). Of these, only three women reported a switch from non-hormonal to hormonal contraception and no women reported a switch from hormonal to non-hormonal contraception (compared to our registered threshold of 200 homosexual women reporting a switch between hormonal and non-hormonal contraception at least once). As the threshold of a sufficient number of exclusively homosexual women was not met, we did not perform the additional exploratory subanalyses based on this sample.

Variables

All variables, including the predictor variable, potential time-varying confounders, outcome variables, and variables used to investigate treatment heterogeneity are listed in Table S3 in the supplement. The original German item wording can be found here: https://www.pairfam.de/dokumentation/fragebogen/.

The predictor hormonal contraception was based on the items about the contraceptive method; participants were able to report multiple contraceptive methods. Hormonal contraception was coded as 0 if participants indicated that they use no contraceptive method at all. The variable hormonal contraception was also coded as 0 if participants indicated that they use no hormonal contraceptive method and at least one of the following methods: condom; intrauterine device; ⁵ diaphragm, foam, suppository, gel; natural birth control; female sterilization; male sterilization; or withdrawal method, coitus interruptus. In addition, the variable hormonal contraception was coded as 0 if participants were never sexually active in their life, as these participants were not asked about their contraceptive method. The variable hormonal contraception was coded as 1 if participants indicated that they use a birth control pill, mini pill, or other hormonal method (implant, patch, ring), even if they additionally use non-hormonal methods. Exclusion criteria for originally registered main as well as robustness analyses based on the contraceptive method are described above.

Simulation

In order to contrast our different analytical approaches, we compared the performance of our models (conceptually summarized in Figure 1) on data simulated under different data generating mechanisms. These simulations are discussed in more detail in the supplement and the implications are mentioned in the section Analysis. Based on these simulation results, to estimate the average treatment effects and treatment effect heterogeneity, we performed adjusted regression analysis without accounting for systematic missingness. In addition, we estimated the average treatment effects based on the IPTW approach accounting for systematic missingness.

Analysis

To answer the question whether hormonal contraceptive use influences women’s sexuality, and to separate these potential causal effects from confounders and attrition effects, we used two different analytical approaches, as outlined in Figure 1. This decision was based on simulations contrasting our different analytical approaches and comparing the performance of our models on data simulated under different data generating mechanisms (for more details on the simulation see the supplement). Based on these simulation results, to estimate the average treatment effects and treatment effect heterogeneity, we performed adjusted regression analysis without accounting for systematic missingness. In addition, we estimated the average treatment effects based on the IPTW approach with stabilized, truncated weights at 1% (Thoemmes & Ong, 2016) accounting for systematic missingness. All planned analyses can be found in form of an rmd file and an html file: https://osf.io/u8ntf/.

All Bayesian models included a random intercept and a random slope for hormonal contraceptive use nested within participants. In addition, each model included information from the previous wave about the outcome, hormonal contraception, and their interaction as predictors. ⁶ In order to be able to estimate the causal relationship between the hormonal contraception and the outcome, we controlled for individual mean levels of hormonal contraceptive use across observations (see Bafumi & Gelman, 2007; Hamaker & Muthén, 2020, for further information); this approach effectively controls for stable confounding influences that work between women (time-invariant confounders). For both models, potential time-varying confounders included linear effects for log transformed net income, educational attainment, and fertility plans; a thin-plate spline effect (Wood, 2003) for age; and a categorical effect for number of children (no children, one child, two children, three or more children).

Furthermore, relationship duration was included as a nested variable. This allowed us to model a linear association with relationship duration which is only informed by women who are in a relationship, while simultaneously including those who are not in the analysis. Technically, we achieved this by including a dummy coded variable for current relationship status (single vs. non-single) and its interaction with relationship duration as a predictor. No main effect of relationship duration was included in the model. Relationship duration for singles was set to −1; this value is arbitrary and does not affect the resulting estimates because when multiplied with the relationship status dummy, relationship duration for singles is dropped from the analysis. In addition, we included two dummy coded variables: one indicating whether a woman started a relationship between the previous wave and the current wave and one indicating whether a woman became single between the previous wave and the current wave.

In the IPTW approach the outcome in the first model was the contraceptive method. The first model results in an estimated weight which was then included in the second model. In the IPTW approach, the effects were additionally weighted for systematic missingness based on weights provided by PAIRFAM. ⁷ Separate analysis for desired sexual frequency, reported sexual frequency, and sexual satisfaction as outcomes were performed. All included variables are listed in Table S3 in the supplement.

To answer the question whether interindividual differences predicted individual treatment effects, we extracted individual treatment effect estimates from the adjusted regression analysis and subsequently correlated them with age and the Big Five personality traits at draw level. Finally, these correlations were averaged across draws (we call this approach “correlate, then average across draws”; Ly et al. (2017) for example use the term “plausible values”). ⁸

To answer the question whether women guide their contraceptive method choices by deciding against hormonal contraceptive methods after experiencing adverse effects, we again used individual treatment effect estimates from the adjusted regression analysis, this time correlating them with the proportion of years using hormonal contraceptives (observations in which hormonal contraceptives were used divided by total number of observations) at draw level. Finally, these correlations were averaged across draws (we call this approach “correlate, then average across draws”). ⁹ This analysis can potentially provide tentative evidence for assortment based on experiences with contraceptive methods.

Additionally, given the possibility of unobserved confounding, we ran sensitivity analysis to estimate how sensitive our results are to hidden bias. We calculated E-values for the effect of hormonal contraception on all outcomes (VanderWeele & Ding, 2017). As VanderWeele and Ding (2017) write “The E-value is defined as the minimum strength of association, on the risk ratio scale, that an unobserved confounder would need to have with both the treatment and the outcome to fully explain away a specific treatment-outcome association, conditional on the measured covariates.“ A large E-value implies that unobserved confounding would need to be relatively substantial to explain away an effect. Conversely, a small E-value implies that even just a little unobserved confounding would be able to explain away the estimated effect. E-values are one of the few approaches to unobserved confounding that can be applied to longitudinal designs (VanderWeele et al., 2020).

All planned analyses can be found in the form of an rmd file and an html file: https://osf.io/u8ntf/. In addition, all models are outlined in the supplement using a simplified readable notation.

Model Convergence

Using the default settings for brms, a lot of the models had divergent transitions (between 6 and 240 diverging transitions). Therefore, we increased the sampler’s target acceptance rate during Stan’s adaptation period in these models from brms’s default 0.80 to 0.99 (Stan Development Team, 2024). With this increased average proposal acceptance probability, none of the originally registered main models had divergent transitions, except for the adjusted regression analysis with sexual frequency as an outcome (4 divergent transitions), which we considered harmless. In all originally registered main models, the Rhats were at most 1.01, indicating that the 4 chains of the models converged. ¹⁰

Deviations From Stage 1 Concerning the Analysis Sample

In contrast to our registered analysis plan, we decided to report the results from robustness analysis 6 (additional exclusion of women who had never been sexually active) as our main analysis. All results based on the original sample for the originally registered main analysis are additionally presented in the supplement. We did so because we realized that including women who had never been sexually active likely induces bias. This particularly applies to the IPTW approach and less so to the adjusted regression analysis, with the overall consequence that including these women makes results more dependent on modeling choices (i.e., means of covariate adjustment) that should not make a substantive difference. Table S1 in the supplement provides all deviations from the in-principle accepted registered report stage 1 and the reasons for the deviations.

In the originally registered main analysis (and all other robustness analyses), women who had never been sexually active were classified as non-hormonal contraceptive users (even though they were not asked about their contraceptive method) and their sexual frequency was 0 (never). This coding decision appears to systematically bias results, because women who are not sexually active over a longer period of time are incorporated in the analyses as non-hormonal contraceptive users (because by study design, they are not asked questions about contraception usage) with the lowest score for sexual frequency (nobody else was assigned this score). The adjusted regression analysis seems to partly account for this bias by controlling for hormonal contraception in the previous wave, the linear effect of sexual frequency from the previous wave (which was 0 = never as well), and their interaction on the sexual frequency in the current wave. In contrast, in the IPTW approach the control for hormonal contraception in the previous wave, the linear effect of sexual frequency from the previous wave, and their interaction on hormonal contraceptive use in the current wave seems to be unable to account for this bias, presumably because the effect of sexual frequency from the previous wave was modeled as a linear effect. Thus, when including these women, estimates from the IPTW approach were about twice as large as estimated based on the adjusted regression analysis (see the section Robustness of Results for more information). When excluding these women, results are much less dependent on the choice of the data-analytic approach and thus more robust. The new exclusion implies that we should not generalize our conclusions to women who start being sexually active and/or start using hormonal contraception.

For women who are being sexually active for the first time, several factors may make their experience different from that of women who have been sexually active for a longer period of time. Initially, these women may be in the process of exploring and understanding their sexual preferences, desires, and boundaries, which may influence their sexual frequency and satisfaction. These factors may result in a unique pattern of sexual behavior that differs from those with more established sexual routines and experiences. Women who begin using hormonal contraceptives without being sexually active may experience different dynamics than those who are already sexually active. For these women, the decision to use hormonal contraception may be driven by reasons other than birth control, such as managing menstrual cycles, reducing menstrual pain, or addressing hormonal imbalances. Without the context of sexual activity, their experience with contraception may be primarily focused on managing these health concerns.

Description of Samples

In PAIRFAM, 18,912 individuals provided a total of 104,661 observations (i.e., annual assessments of individuals). After applying our predefined exclusion criteria and excluding observations that could not be incorporated into the analyses because of modeling decisions, up to 23,130 observations from n = 5,041 women were eligible for our main analysis (registered as robustness analysis 6) estimating effects of hormonal contraception on sexuality. ¹¹ Due to further missingness in single outcomes, the sample sizes for our main analysis vary between outcomes. In the main analysis estimating the effect of hormonal contraception on sexual frequency 4,403 women with 19,801 observations were included (sexual frequency was only measured in waves 2 to 14); in the main analysis estimating the effect of hormonal contraception on desired sexual frequency 3,057 women with 9,939 observations were included (desired sexual frequency was only measured in waves 7 to 14); in the main analysis estimating the effect of hormonal contraception on sexual satisfaction 4,983 women with 22,622 observations were included. The full exclusion process for the main analysis is outlined in Figure S3 in the supplement. The total eligible sample sizes for estimating effects of hormonal contraception on sexuality for all samples are displayed in Table S4 in the supplement.

Descriptives

Overall, after applying all exclusion criteria n = 5,041 women with 23,130 observations were eligible for our main analysis estimating effects of hormonal contraception on sexuality. ¹² Each woman was observed between 1 and 13 times (the first wave was always excluded because no information from the previous wave was available); with an average of 4.59 observations (standard deviation of 3.46). ¹³ Table S5 in the supplement displays the number of waves (for all categorical variables) or means, standard deviations, and ranges (for all continuous variables) for all included women and all observations.

Table S6 in the supplement shows zero-order correlations of all variables included in the same analysis averaged within women across all available observations. Hormonal contraception measured in the current wave correlated positively with hormonal contraception measured in the previous wave (r = .86 [95% CI: .86, .87]). As is to be expected, both measures (hormonal contraception in the current wave, hormonal contraception in the previous wave) correlated positively with the average frequency of using hormonal contraception (r = .95 [.94, .95] and r = .95 [.94, .95], respectively).

Hormonal contraception in the current wave correlated negatively with income, years of education, relationship duration, completed fertility plans, and age (ordered from weakest to strongest negative correlation, rs ranging from −.11 [−.14, −.08] to −.50 [−.52, −.48]), as well as positively with the end of a relationship in the last year (r = .11 [.09, .14]) and the start of a relationship in the last year (r = .15 [.12, .17]).

Each sexuality variable measured in the current wave correlated positively with its value in the previous wave (rs ranging from .71 [.70, .72] to .82 [.81, .83]). When measured in the current wave, sexual frequency correlated positively with sexual satisfaction (r = .57 [.55, .59]), while desired sexual frequency correlated negatively with sexual frequency (r = −.35 [−.38, −.32]) and sexual satisfaction (r = −.34 [−.37, −.31]). Hormonal contraception in the current wave correlated positively with sexual frequency in the current wave (r = .22 [.19, .24]) and sexual satisfaction in the current wave (r = .16 [.13, .19]), but did not correlate with desired sexual frequency in the current wave (r = −.02 [−.05, .01]). For the remaining covariates measured in the current wave, being in a relationship showed the strongest and most consistent correlations with sexual frequency (r = .44 [.42, .47]), sexual satisfaction (r = .26 [.24, .29]), and desired sexual frequency (r = −.27 [−.30, −.24]).

Descriptive Patterns in Hormonal Contraceptive Use

Of the 23,130 observations of 5,041 women included in our main analysis, 12,321 (53%) were observations in which women reported using non-hormonal contraceptive methods and 10,809 (47%) were observations in which women reported using hormonal contraceptive methods. Over the whole available information (including previous waves), 1,537 women (30%) reported using only non-hormonal contraceptives, 1,335 women (26%) reported using only hormonal contraceptives, and 2,169 (43%) women reported using non-hormonal and hormonal contraceptives at least once. The 2,169 women who used both non-hormonal and hormonal contraceptives reported 2,716 switches (total number of switches across all 5,041 women: mean = 0.54, sd = 0.93, min = 0, max = 7) with 1,584 switches from hormonal to non-hormonal and 1,132 switches from non-hormonal to hormonal contraceptives.

Predictors of Hormonal Contraceptive Use

Figure S4 in the supplement displays the three separate models predicting hormonal contraceptive use in the current wave from, among other predictors, (a) sexual frequency in the previous wave and its interaction with hormonal contraceptive use in the previous wave; (b) desired sexual frequency in the previous wave and its interaction with hormonal contraceptive use in the previous wave; and (c) sexual satisfaction in the previous wave and its interaction with hormonal contraceptive use in the previous wave.

Average Treatment Effects of Hormonal Contraceptive Use on Sexuality

Heterogeneity in Treatment Effects of Hormonal Contraceptive Use on Sexuality

We were also interested in treatment heterogeneity: do the effects of hormonal contraceptive use on sexuality vary between women? Based on the adjusted regression analysis, the standard deviation of the treatment effect (i.e., the standard deviation of the random slope of hormonal contraceptive use on the sexuality outcome) was relatively high compared to the overall standard deviations of the outcomes (see Table S5 in the supplement) for sexual frequency (34% of the overall standard deviation) and sexual satisfaction (29% of the overall standard deviation), but not for desired sexual frequency (20% of the overall standard deviation; see the second column of Table 2).

OPEN IN VIEWER

Table 2.

Heterogeneity in and Reliability of Treatment Effects of Hormonal Contraceptive Use on Sexuality

Outcome	ATE [95% CI]	Standard deviation of the random slope	Number of women with negative point estimates of ITEs	Number of women with positive point estimates of ITEs	Number of women with significant negative ITEs	Number of women with significant positive ITEs	Reliability of the ITE estimates
Sexual frequency	0.30 [0.24, 0.35]	0.51 [0.44, 0.58]	302 (7%)	4,101 (93%)	0 (0%)	31 (1%)	0.15
Desired sexual frequency	−0.06 [−0.12, −0.02]	0.16 [0.04, 0.28]	2,784 (91%)	273 (9%)	0 (0%)	0 (0%)	0.09
Sexual satisfaction	0.29 [0.18, 0.39]	0.77 [0.60, 0.93]	786 (16%)	4,197 (84%)	0 (0%)	45 (1%)	0.17

Taking a closer look at the model-implied point estimates of the individual treatment effects, for all three outcomes, some women experienced negative effects and some women experienced positive effects of hormonal contraceptive use. However, the point estimates of individual treatment effects were estimated very imprecisely, so that only few of them were significantly different from zero when considered in isolation. Putting these pieces together, while we can say that heterogeneity was high, the number of observations per woman was not sufficient to estimate anyone’s treatment effect with certainty (see Table 2). Figure 5 displays selected individual treatment effects (as displaying all of them would not be legible). The figure includes all women who contributed at least 11 observations to the model (7 observations for desired sexual frequency) and reported using non-hormonal and hormonal contraception at least once. The large confidence intervals illustrate the uncertainty in the estimates of the individual treatment effects. To provide an additional quantification of this uncertainty, we computed the reliability of the estimated random slopes. To get an estimate of the unweighted reliability, we subtracted the fraction of the uncertainty of the random slope (reflected by the variance across MCMC draws) by the average between-person variance (reflected by the variance of the random slope) from 1 (see Arslan, 2024, for a detailed description of this procedure and the code that we adjusted for our analysis). The reliability of the random slopes for all outcomes was very low (≤.17, see last column in Table 2), indicating a high uncertainty in the estimates of the individual treatment effects.OPEN IN VIEWER

Predictors of Individual Treatment Effects

Individual treatment effect estimates for sexual frequency and sexual satisfaction correlated ¹⁴ positively with each other (r = .36 [.33, .39], n = 4,346). Individual treatment effect estimates for desired sexual frequency correlated negatively with individual treatment effect estimates for sexual frequency (r = −.24 [−.27, −.20], n = 3,048) and sexual satisfaction (r = −.30 [−.35, −.25], n = 3,034).

To explore the correlation between individual treatment effects and age as well as the correlations between individual treatment effects and Big Five personality traits, we used a “correlate, then average across draws” approach. ¹⁵ That is, we correlated individual treatment effect estimates from the adjusted regression analysis with age and the Big Five personality traits at draw level. Finally, these correlations were averaged across draws. They are summarized in the first columns of Table 3.

OPEN IN VIEWER

Table 3.

“Correlate, Then Average Across Draws” Relationships Between Individual Treatment Effects and Age, Big Five Personality Traits, as Well as Years Spent on Hormonal Contraception

	Age	Extraversion	Agreeableness	Conscientiousness	Neuroticism	Openness	Years spent on HC
ITE sexual frequency	r = −.01 [−.04; .01]	r = .0002 [−.03, .03]	r = −.002 [−.03, .03]	r = −.001 [−.03, .03]	r = .01 [−.02, .04]	r = .01 [−.02, .04]	r = .01 [−.02, .03]
	n = 4,403	n = 4,037	n = 4,035	n = 4,037	n = 4,036	n = 4,036	n = 4,403
ITE desired sexual frequency	r = .003 [−.03; .04]	r = −.03 [−.07, .01]	r = .01 [−.02, .04]	r = −.002 [−.04, .03]	r = −.01 [−.04, .03]	r = −.02 [−.06, .02]	r = −.002 [−.04, .03]
	n = 3,057	n = 2,817	n = 2,817	n = 2,817	n = 2,817	n = 2,816	n = 3,057
ITE sexual satisfaction	r = .01 [−.02; .03]	r = −.02 [−.05, .01]	r = −.03 [−.06, .002]	r = −.04 [−.07, −.01]	r = .07 [.04, .11]	r = −.01 [−.03, .02]	r = .003 [−.02, .03]
	n = 4,983	n = 4,571	n = 4,571	n = 4,571	n = 4,570	n = 4,570	n = 4,983

Using this approach, we found two small correlations for which the 95% credible interval excluded zero. These were between the interindividual treatment effects on sexual satisfaction and (a) neuroticism (r = .07 [.04, .11], n = 4,570), meaning that women who scored higher on neuroticism experienced more positive effects of hormonal contraceptive use on sexual satisfaction and (b) conscientiousness (r = −.04 [−.07, −.01], n = 4,571), meaning that women who scored higher on conscientiousness experienced less positive effects of hormonal contraceptive use on sexual satisfaction. While some of them were significant, most of them were rather small in magnitude and no common pattern across sexuality outcomes emerged. Neither neuroticism nor conscientiousness showed significant correlations with the individual treatment effects for the other outcomes.

Contraceptive Decisions and Individual Treatment Effects

We additionally wanted to investigate whether women guide their contraceptive choices based on their experiences. In fact, our analysis of predictors of hormonal contraceptive use already provided evidence pointing into that direction: women with higher sexual frequency and higher sexual satisfaction were more likely to stick with the method they were previously using. However, our registered analysis went further and asked whether the individual treatment effects (i.e., not just whether women are more or less sexually satisfied, but whether hormonal contraceptives have a negative or positive effect on their sexual satisfaction) correlated with the amount of time that women used hormonal contraceptives. Again, we used the “correlate, then average across draws” approach. ¹⁵ That is, we correlated the individual treatment effects with the proportion of years of hormonal contraceptive use at draw level. Finally, these correlations were averaged across draws. None of the “correlate, then average across draws” relationships were significant (see the last column of Table 3).

Descriptive Patterns in Hormonal Contraceptive Use and Predictors of Hormonal Contraceptive Use

Our results suggest that across 5.6 years (mean number of current and previous waves), about one half of the women exclusively reported non-hormonal or hormonal methods, and the other half of the women reported a switch between non-hormonal and hormonal methods. About 60% of these switches were from hormonal to non-hormonal methods and 30% from non-hormonal to hormonal methods. Only a few women reported several switches. Overall, the percentage of women using hormonal methods declined across the course of PAIRFAM, from 2008 to 2021. These results are in line with a representative survey in Germany comparing contraceptive use between the years 2011 and 2018 showing that the use of oral contraceptives (i.e., the most popular form of hormonal contraception) declined by 6 percentage points while the use of condoms (i.e., the most popular form of non-hormonal contraception) increased by 9 percentage points (Bundeszentrale für gesundheitliche Aufklärung, 2018). The most commonly mentioned concerns about hormonal contraception in Western countries are effects on menstruation, physical side effects, mental health effects and effects on sexuality as well as concerns about future fertility or a general wish to use more natural methods (Le Guen et al., 2021). The discontinuation rate for oral contraceptives was around 28% in an analysis of 23 countries from 1990 to 2008 (Ali & Cleland, 2010). The trend away from hormonal contraception is not generalizable to all countries, though. According to the United Nations et al. (2022), oral contraceptives are the most common method in 36 countries (e.g., Algeria, Morocco), while condoms are the most common method in 27 countries (including Germany).

There was relatively strong stability in the broad choice of contraceptive method, as indicated by the large predictive contributions of average frequency of hormonal contraceptives and use of hormonal contraceptives in the previous wave. In addition, women were more likely to switch their contraceptive method if sexual frequency and sexual satisfaction were low.

Beyond these method- and sexuality-related predictors, the current relationship situation predicted the choice of contraceptive methods. Women in a relationship, especially women who started a new relationship in the last year, were more likely to use hormonal contraceptive methods. Beyond this initial increase in hormonal contraceptive use after the start of a relationship, relationship duration was a negative predictor of hormonal contraceptive use. In addition, ending a relationship negatively predicted the use of hormonal contraception. Besides the current relationship situation, women who had one child (and potentially two children, even though this result was not robust across all analyses) were more likely to use hormonal contraception compared to women who had no children. In line with these results, we found some (albeit non-robust) evidence that women who said that they had completed their fertility plans were more likely to use hormonal contraceptive methods.

We also found some non-robust evidence for negative effects of income and education on hormonal contraceptive use. These results potentially indicate that women with a higher income and more education were less likely to use hormonal contraceptives. As these effects are rather small and did not emerge across models, they need to be interpreted cautiously. It is quite possible that the effects of income and education on hormonal contraceptive use differ depending on the details of the health care system and how accessible it renders such contraceptives (e.g., is a prescription required? Does health insurance cover the cost of the necessary appointments and of the contraceptives?).

Overall, the choice of contraceptive method appears to be relatively stable and changes in contraceptive methods are mainly explained by low sexual frequency and sexual satisfaction as well as the current relationship and family situation.

Average Treatment Effects on Sexuality

We found strong support for positive average treatment effects of hormonal contraception on sexual frequency and sexual satisfaction across all analytical approaches and all robustness analyses. When women use hormonal contraceptives, they have a higher sexual frequency and a higher sexual satisfaction then when they use non-hormonal contraceptives, and these effects are robust to the inclusion of the potential confounding effects discussed in the previous section. These results are in line with results from an earlier project based on cross-sectional data that aimed to disentangle, to some extent, causal effects of hormonal contraceptives from selection effects by controlling for potential confounders and estimating the sensitivity to unobserved confounders (Botzet et al., 2021). Applying these methodologies to longitudinal data in the current project advances our understanding of the causal nature of the effect by addressing reverse causality and attrition besides confounding as potential alternative explanations. Based on the performed sensitivity analysis, unobserved confounding would need to be relatively strong to fully account for the estimated effects of hormonal contraceptive use on sexual frequency, desired sexual frequency, and sexual satisfaction.

Our results are in line with existing correlational evidence finding positive effects of hormonal contraception on sexuality concerning sexual functioning (e.g., Oranratanaphan & Taneepanichskul, 2006), libido (McCoy & Matyas, 1996), sexual satisfaction (e.g., Caruso et al., 2005), and, most strongly, sexual frequency (e.g., Caruso et al., 2005; McCoy & Matyas, 1996). Most importantly, our results are also in line with two other studies finding support for positive effects of hormonal contraceptive use on sexuality in longitudinal datasets (Blumenstock & Barber, 2022; Ott et al., 2008). Taking the available evidence together supports the hypothesis that the use of hormonal contraception positively affects women’s sexuality; potentially through several mechanisms, including, for example, overcoming the fear of unwanted pregnancy during sexual activity (Blumenstock & Barber, 2022) and the resolution of painful or troublesome gynecologic disorders (Both et al., 2019).

In contrast, our results are not consistent with existing experimental evidence that hormonal contraceptive use negatively affects sexual functioning (e.g., Læssøe et al., 2014) and libido (e.g., Lee et al., 2017; Lundin et al., 2018; Zethraeus et al., 2016), as well as sexual activity, arousal, pleasure, orgasm, and lubrication (Smith et al., 2014). Therefore, based on the correlational data, we found no support for the hypothesis that the intervention in the endocrine system through the use of hormonal contraceptives (Stomati et al., 1998) negatively affects women’s sexuality (Both et al., 2019).

The evidence concerning negative average treatment effects on desired sexual frequency (our available measure for libido) is neither constant across analytical approaches nor robustness analyses. It should therefore be interpreted very cautiously, especially because the sample size of available observations was considerably smaller (desired sexual frequency was only measured from wave 7 to wave 14). Even if we take the potential negative effect of hormonal contraceptive use found in the adjusted regression analysis at face value, its interpretation would be ambiguous because of the way desired sexual frequency was measured (“If it was only up to you, would you like to have less or more sexual intercourse compared to the last three months?”). Because the sexual desire item was phrased relative to sexual behavior, we have to interpret this result in light of the positive effect of HC on sexual frequency. So, potentially, desire is constant, but because frequency is lower among non-HC users, our desire outcome showed a negative effect. Overall, the evidence of an average treatment effect of hormonal contraceptive use on desired sexual frequency seems not very robust and very inconsistent across studies (e.g., Botzet et al., 2021; Lee et al., 2017; Lundin et al., 2018; McCoy & Matyas, 1996; Zethraeus et al., 2016). If an average treatment effect of hormonal contraceptive use on libido exists, it is probably relatively small and large sample sizes are needed to detect it.

We found evidence for strong positive effects of hormonal contraceptive use on sexuality when women are not randomly assigned to one contraceptive method but make free (and potentially to some degree informed) decisions about their contraceptive method.

Taking the evidence for positive effects of hormonal contraceptive use and a high heterogeneity together, this makes it very likely that women in our sample, who experience adverse effects of hormonal contraceptive use, are more likely to discontinue using hormonal contraceptives. Therefore, the estimated effect of hormonal contraceptive use on sexuality in the current study, does not recover the average treatment effect across all women, but is closer to the average treatment effect on the treated.

Differences in Results Based on Modelling Approach

To our surprise, for the outcome of sexual frequency, we found significant differences between the estimated average treatment effects depending on whether we used adjusted regression analysis or the IPTW approach. Estimated average treatment effects were slightly larger for the IPTW approach compared to the adjusted regression analysis (see Figure 3). As registered in stage 1 of this registered report, adjusted regression analysis was treated as the main analysis and the IPTW approach was treated as a robustness analysis for identifying the average treatment effect. Both analytic approaches supported positive average treatment effects of hormonal contraceptive use on sexual frequency, but the question that still remains is how strong this effect actually is.

As outlined in the introduction, both approaches try to estimate the causal effect of a treatment on an outcome and they come with different advantages and disadvantages. ¹⁶ Our adjusted regression analysis relies on the key assumption that the relationships between the covariates and the sexuality outcome are modeled appropriately (Gutman & Rubin, 2017), while the IPTW approach relies on the key assumption that the relationships between the covariates and hormonal contraceptive use are modeled appropriately. For both assumptions, there are reasons to doubt that these assumptions hold true. Considering the regression analyses, relationship duration could affect sexual frequency in a non-linear manner in the form of a sudden drop in sexual frequency after a certain relationship duration, biasing the results. Considering the IPTW approach, education could affect the choice of contraceptive methods in a non-linear manner in the form of a sudden increase in hormonal contraceptive use after a certain degree of educational attainment, biasing the results. We cannot discern which modeling approach is more biased here, and thus we cannot clearly favor one over the other. Maybe more importantly, both approaches still rely on the key assumption of no unobserved confounders, even if our sensitivity analyses suggest that such confounders would need to be fairly influential to completely explain away the estimated effects.

As we cannot decide which of the two approaches correctly retrieves the underlying causal effect of hormonal contraceptive use, we would like to use our results as an example to urge researchers to consider different analytical approaches when estimating causal effects and to lay out all of the tested and untested assumptions as well as the specification of the underlying model.

Individual Treatment Effects

Besides the average treatment effects, we were interested in the heterogeneity of effects when estimating individual treatment effects. For both sexual frequency and sexual satisfaction we found support for relatively strong heterogeneity in individual treatment effects, ranging from negative to positive effects in line with Graham (2019) summary of anecdotal evidence. This finding questions the idea that hormonal contraceptive use affects sexuality only through a uniform and universal biological pathway.

Where could such heterogeneity come from? From a biological perspective, our findings are consistent with the idea that women have varying sensitivity to hormones (Kiesner, 2017) resulting in heterogeneous responses to hormonal contraceptive use. The heterogeneous pattern could also be explained by non-universal psychological mechanisms. For example, there might be interindividual differences in the expectations and worries about the effectiveness of contraception and the likelihood of unplanned pregnancy (Both et al., 2019). This might result in differential effects of hormonal contraceptive use on sexuality.

At the same time, we found no common pattern across sexuality outcomes for relationships between individual treatment effects and potential predictors of individual treatment effects (including age and Big Five personality traits). Unfortunately, the high uncertainty around individual treatment effects as indicated by the low reliability makes it impossible to rule out any of the predictors. Future studies interested in the effects of hormonal contraceptive use on sexuality should further investigate the heterogeneity in individual treatment effects and explore their relationship with other markers of interindividual differences.

What could these markers of interindividual differences be? First, self-reported side effects (i.e., side effects mentioned by women) might be positively related to experienced side effects (i.e., individual treatment effects). This could provide a first test of whether interindividual differences in self-report questionnaires reflect individual treatment effects at all. Second, instead of a general measure of neuroticism, fear of unwanted pregnancy might be positively related to individual treatment effects, with women with a higher fear of unwanted pregnancy showing positive effects of hormonal contraceptives on sexual frequency and sexual satisfaction. Third, general attitudes toward the use of hormonal contraceptives might also be related to individual treatment effects, with women who are very skeptical about the use of hormonal contraceptives (but who still use them, e.g., for other health reasons) showing negative effects of hormonal contraceptives on sexual frequency and sexual satisfaction. Finally, we would hypothesize that hormone sensitivity (Kiesner, 2017) would be negatively related to individual treatment effects, with women with higher sensitivity experiencing more negative effects of hormonal contraceptives on sexual frequency and sexual satisfaction.

We also did not find significant correlations between the estimated individual treatment effects and the number of years of hormonal contraceptive use over the course of PAIRFAM, which we would interpret as preliminary evidence against potential assortment based on experiences with contraceptive methods. Again, the interpretability of this analysis is limited by the low reliability of the individual treatment effects. Furthermore, the analysis would have at best provided indirect evidence for assortment. Another analysis we conducted can be used to answer a slightly different but related question, namely whether generally low sexual frequency or low sexual satisfaction (which may or may not be attributable to hormonal contraception) predict switches in contraceptive methods. This analysis provides another type of indirect evidence for assortment: In the models predicting hormonal contraceptive use in the current wave, women were more likely to switch their contraceptive method if they experienced lower levels of sexual frequency and sexual satisfaction. However, another recent study using a similar analysis found no support for women’s experiences of lower levels of sexual frequency and sexual satisfaction predicting switches in contraceptive methods (Draxler et al., 2024). We believe that a better empirical test of the underlying idea would require more extensive longitudinal data. In such data, one could use all but the last observation per individual to estimate the individual treatment effect, and then test whether this effect predicts the contraceptive method reported in the last observation of the individual.

Contributions to Personality Psychology

Human sexuality has long been treated as peripheral in personality psychology, in part because sexuality-related descriptors were excluded from early lexical studies that laid the foundation for the Big Five personality traits (Saucier & Goldberg, 1996; Schmitt & Buss, 2000). Yet, individual differences in sexuality abound and accumulating evidence demonstrates systematic associations between human sexuality and personality traits (e.g., Bourdage et al., 2007; Schmitt & Shackelford, 2008), underscoring that this boundary is artificial. Against this backdrop, our study contributes not only to knowledge about the psychological effects of hormonal contraceptives, but also to understanding the interplay between human sexuality and personality.

In addition, our analyses illustrate how robust analytical strategies can be applied to large-scale longitudinal panel data to address complex causal questions in personality research—an issue with which the field has grappled with for some time (Atherton et al., 2021). A key element of this approach is the explicit definition of estimands, which ensures that theoretical research questions are directly connected to statistical analyses (Lundberg et al., 2021). By focusing on the average treatment effect, we were able to provide a clear causal contrast and evaluate its robustness across different modeling assumptions. Longitudinal data played a central role in strengthening causal inference, as repeated observations within the same women allowed us to control for time-invariant confounders and to examine changes in contraceptive use over time. This design enabled us to investigate switches between hormonal and non-hormonal contraceptive methods within individuals, thereby reducing biases that are inherent in purely cross-sectional analyses. Importantly, the comparison of analytical approaches also underscored how methodological choices can shape substantive conclusions: while both strategies pointed to positive effects of hormonal contraceptive use on sexual frequency, the estimated magnitude of the effect differed between methods, highlighting the value of triangulating evidence across multiple approaches.

Taken together, our work contributes to the broader debate on how to draw causal inferences from observational data in personality science. Rather than treating causality as a simple binary, it is useful to conceptualize causal inference as a continuum of plausibility (Janson, 2012) and as inherently probabilistic rather than deterministic (Parascandola, 2011). Within this perspective, our study illustrates how carefully chosen estimands, longitudinal designs, and complementary analytical strategies can move personality research beyond mere associations and toward more credible causal claims. In doing so, we provide both a framework for future studies in this area and a guide for applying these methods to similar research questions in personality psychology.

Limitations

Our current study was a registered report applying two different analytical approaches and six robustness analyses based on a large panel dataset combining all the relevant information about contraceptive use, covariates, and sexuality. By examining change over time within individuals, we reduced the number of assumptions necessary for causal identification and held many of the potential confounding factors that vary between individuals constant. In addition, the assumed underlying causal network was transparently outlined before performing analyses and the assumptions were questioned in the discussion. Therefore, our study aimed to provide an estimation of the effect of hormonal contraceptive use on sexuality in a naturalistic setting where women decide which contraceptive method to use.

Nevertheless, our study comes with a few limitations. First, the two modeling approaches we used to estimate the average treatment effect differed substantially in their estimations of the average treatment effect questioning the robustness of our results and limiting their generalizability. In future research focusing on identifying causal effects based on correlational data, combining different analytical approaches could be beneficial to determine a range of possible effects (and might additionally inform us about systematic differences between these analytical approaches). Second, our outcome variables were only measured with one item each and often deviated from items in existing literature (especially in the case of desired sexual frequency as a measure of libido). Future research could try to implement more nuanced measures (e.g., solitary and dyadic libido; masturbation frequency) or measure sexuality in a more extensive way (e.g., sexual functioning, or using a diary method). And third, the time interval between observations (one year) was potentially too long to capture all of the switches in contraceptive methods that happened between waves. As we had no indicator of the time between switch and reporting of outcomes, we are unable to conclude whether the observed effects appear only shortly after changing contraceptive methods or after some time has passed. Diary studies (e.g., Blumenstock & Barber, 2022; Ott et al., 2008) investigating very short time frames found support for positive effects of hormonal contraceptive use on sexuality as well; more research focusing on an intermediate time frame (e.g., one month, one menstrual cycle) would be helpful to understand the effects of hormonal contraceptive use on sexuality further.