Introduction

Life satisfaction, happiness, and health status are fundamental indicators of individuals’ and countries’ well-being. Interestingly, there appear to be systematic gender differences in these outcome measures, with women, on average, reporting higher life satisfaction and happiness1,2,3,4, but lower self-reported health status5,6. Two other key factors in understanding how societies function are trust and political preferences. Trust acts as a catalyst for economic growth and well-being by facilitating informal contracts and cooperation between individuals and groups7,8. Political preferences determine who is elected into power and which policies are implemented in democratic countries. For both trust and political preferences, systematic gender differences have been reported, too, with women typically expressing lower trust levels9,10,11,12 and leaning more toward leftist political views13,14,15,16.

While the gender gap in these five fundamental variables appears well-established, less is known about the variability of these gender differences across various populations and countries. Systematic variation of effect sizes is typically referred to as heterogeneity. Understanding the heterogeneity of effect sizes has become a crucial topic in the social sciences in recent years17,18,19,20,21,22,23,24.

In this study, we analyze survey data from approximately 500,000 participants in the European Social Survey (ESS), which comprises representative samples of the general population from 39 European countries. To confirm earlier findings, we first estimate meta-analytic effect sizes regarding gender differences in generalized trust (henceforth “trust”), political preferences, life satisfaction, happiness, and subjective general health (henceforth “health status”). Since all the data have been gathered through a unified large-scale survey, our meta-analytic estimates are unaffected by publication bias, which can substantially distort effect size estimates in conventional meta-analyses that aggregate results from various published studies25,26,27,28,29.

We then turn to our key research question and estimate population heterogeneity in gender differences across the 39 countries included in our study, which is defined as the variability of true effects across countries while holding the research design and analysis constant. The ESS is an ideal test bed for estimating the extent of heterogeneity in effect sizes due to systematic differences across populations. Systematic evidence regarding population heterogeneity is limited, particularly for general population samples. Due to the variability in research design and analysis, accurately estimating population heterogeneity is challenging. One exception to this is the use of data from preregistered multi-lab replication studies, which have primarily been conducted in the field of psychology. Holzmeister et al.23 reviewed estimates of population heterogeneity for 70 such multi-lab replication studies, reporting small to moderate heterogeneity, consistent with two previous studies that used a subset of the same data19,22.

However, it is ambiguous to what extent these population heterogeneity estimates are generalizable, as they are primarily based on university student samples in Western countries. Another recent estimate of population heterogeneity was provided by Krefeld-Schwalb et al.24, who collected data on four psychological effects across various online populations. They reported substantially larger estimates of population heterogeneity. However, their estimates may be driven more by variability in attention, attrition, and comprehension within and across online populations than by actual variation in population effect sizes30,31. Our study contributes to a better understanding of effects that are fundamentally important for society by estimating population heterogeneity using nationally representative samples.

To reduce researchers’ degrees of freedom and limit bias due to selective reporting32,33,34,35,36,37,38, we strictly follow a pre-analysis plan (PAP) posted on OSF https://osf.io/zaq3n before conducting our analyses. We utilize data from the 11 rounds of data collection available on the ESS webpage (www.europeansocialsurvey.org), covering survey waves from 2002 to 2023. We include all 39 countries surveyed in at least one of the 11 rounds. The ESS is a repeated cross-sectional survey of the countries’ general population aged 15 and older, aiming to collect at least 1,500 observations per country in each round.

Five ESS survey measures served as dependent variables. All outcome measures are scored on a 0–10 scale, except for health status, which is rated on a 1–5 scale. The total sample size across the five outcome measures in our analysis ranges from n = 435,246 to n = 506,871, which implies that our analyses are highly powered to estimate both meta-analytic effect sizes and population heterogeneity in gender differences.

Results

Below, we report the results of our preregistered hypothesis tests. We begin by examining the extent of gender differences for the five outcome variables, which can be thought of as a (conceptual) replication of previously documented gender gaps in the literature. We then focus on the systematic variation in gender differences across the 39 European countries included in our sample. We interpret two-sided p-values smaller than 0.05 as “suggestive evidence” and two-sided p-values smaller than 0.005 as “statistically significant evidence,” in line with the recommendation of Benjamin et al.39. To enhance the interpretability of our results, we present all estimates in Cohen’s d units (see the Methods section for details). In Supplementary Table 1, we report the minimum detectable effect size (MDE) in Cohen’s d units, which we could detect with 90% statistical power at a significance level of 0.5%. The estimated MDEs emphasize that our analyses were highly powered to detect even small effect sizes.

How large are the gender differences?

We estimate country-level gender effects in the five outcome measures and determine meta-analytic gender differences using random-effects meta-analyses, aggregating the 39 country-level estimates (primary hypotheses 1a–1e). We illustrate the country-level estimates of gender differences in the five outcome variables in Fig. 1. Detailed forest plots, reporting the 95% confidence intervals of the country-specific estimates, are provided in Supplementary Figs. 1–5. Figure 2a summarizes the meta-analytic effect sizes and their 95% and 99.5% confidence intervals in Cohen’s d units.

We find evidence of a gender difference in the hypothesized direction for all five outcome measures, with statistically significant results for four and suggestive evidence for life satisfaction. However, the meta-analytic effect sizes are small, with d = − 0.061 (se = 0.014; z = − 4.486, p < 0.001) for political preferences (indicating more left-leaning preferences for women), d = − 0.032 (se = 0.010; z = − 3.270, p = 0.001) for trust, d = 0.023 (se = 0.009; z = 2.679, p = 0.007) for life satisfaction, d = 0.039 (se = 0.010; z = 3.875, p < 0.001) for happiness, and d = − 0.110 (se = 0.012; z = − 9.267, p < 0.001) for health status. The mean absolute effect size across the five outcome measures is d = 0.053. On the 0 − 10 scale used for political preferences, trust, life satisfaction, and happiness, the unstandardized effect sizes are − 0.136, − 0.074, 0.049, and 0.075, respectively. For health status, measured on a 1 − 5 scale, the unstandardized effect size is − 0.099.

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Estimated gender differences in the five outcome variables for each country. Plotted are the point estimates and the 95% CIs of the gender difference (0 “male”, 1 “female”) in (a) left-right political preferences, (b) trusting behavior, (c) life satisfaction, (d) happiness, and (e) health status for each of the 39 countries surveyed in the ESS. All point estimates and confidence intervals are standardized based on the pooled standard deviations in the outcome variables and can be interpreted in terms of Cohen’s d units (see Methods for details). Comprehensive meta-analytic forest plots for each of the five outcome variables are presented in Supplementary Figs. 1–5.

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Meta-analytic estimates and heterogeneity estimates of gender differences in the five outcome variables. (a) Plotted are the meta-analytic effect sizes and their 95% and 99.5% CIs, based on random-effects meta-analyses (using the restricted maximum likelihood estimator for τ) of k = 39 countries surveyed in the ESS, for gender differences (0 “male”, 1 “female”) in the five outcome variables. The point estimates and confidence intervals are standardized using the pooled standard deviations of the outcome variables and can be interpreted in terms of Cohen’s d units (see Methods for details). (b) Plotted are the point estimates of τ, i.e., the standard deviation of true effect sizes across the included countries, and their 95% and 99.5% CIs, for gender differences in the five outcome variables. All estimates are reported in Cohen’s d units.

Do gender differences vary across European countries?

Our study’s primary focus was to test for heterogeneity in effect sizes across European countries (primary hypotheses 2a–2e). We evaluate heterogeneity using Cochran’s Q-test and report estimates of τ, I², and H, along with their 95% confidence intervals, to quantify heterogeneity. τ is the standard deviation of the true effect size across the included countries, I² is the fraction of the overall variation in estimates across the countries attributable to heterogeneity, and H is the factor by which the sampling standard error of an individual country needs to be multiplied by to incorporate heterogeneity. We interpret the magnitude of heterogeneity using benchmark levels for small, medium, and large heterogeneity as indicated by I² values of 25% (H = 1.15), 50% (H = 1.41), and 75% (H = 2.00)40,41. We summarize the heterogeneity estimates by plotting the τ estimates, along with their 95% and 99.5% confidence intervals, in Fig. 2b; detailed heterogeneity results are reported in Supplementary Table 2.

We find statistically significant evidence (p < 0.001) of population heterogeneity in gender differences for all five outcome measures. The heterogeneity is large in relative terms, with I² estimates ranging from 89.6% to 95.5% and H estimates ranging from 3.11 to 4.73. This implies that systematic variability across countries explains a large portion of the variation in country-level estimates, and that the standard error of a country-specific gender gap estimate would need to be multiplied by about 4 to account for uncertainty due to population heterogeneity. Notwithstanding, both I² and H are relative measures of heterogeneity that depend on the sample sizes and the precision of the sample standard error of the country-specific estimates. The sample sizes across countries are comparatively large, averaging 12,600 observations per country, which inflates estimates of relative heterogeneity.

τ measures heterogeneity in absolute terms and is estimated to 0.079 (95% CI [0.063, 0.108]) for political preferences, 0.056 (95% CI [0.046, 0.083]) for trust, 0.049 (95% CI [0.041, 0.076]) for life satisfaction, 0.059 (95% CI [0.046, 0.081]) for happiness, and 0.071 (95% CI [0.057, 0.097]) for health status. The mean τ is 0.063, and all five heterogeneity estimates are remarkably similar. Thus, in absolute terms, the estimates of population heterogeneity appear to be moderate and comparable in size to the median estimate of 0.06 reported for a sample of 46 multi-lab replication studies (measured in Cohen’s d units) in Holzmeister et al.23

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

Estimated distributions of true gender differences across countries in the five outcome variables. Plotted are the estimated distributions of the true gender differences (0 “male”, 1 “female”) in (a) left-right political preferences, (b) trusting behavior, (c) life satisfaction, (d) happiness, and (e) health status across countries based on the estimated meta-analytic effect sizes,, and the estimated standard deviation of true effects, τ. Estimates for θ and τ are based on random-effects meta-analyses using the restricted maximum likelihood estimator for τ (see Fig. 2 for details). The dashed lines indicate the estimated distributions of true effect sizes based on the lower and upper bounds of the 95% CIs of τ, respectively. The percentages of effects falling below and exceeding zero based on the point estimates and 95% CIs of τ are reported in the top-left corners.

In Fig. 3, we plot the distribution of true gender differences across countries based on the meta-analytic means and the estimated τ. The small meta-analytic effect sizes, combined with sizable heterogeneity, imply that a substantial share of the distribution of estimated true gender differences across countries is in the opposite direction of the hypothesized effects. This share is 22.1% for political preferences, 28.4% for trust, 31.6% for life satisfaction, 25.0% for happiness, and 6.0% for health status. This indicates that the estimated gender difference in one country may not accurately reflect the direction of the effect in another country or the meta-analytic mean. The only exception is health status, where only 6% of the estimated distribution of true gender differences across countries contradicts the hypothesis’s directionality.

Do gender differences vary with gender equality and GDP?

In preregistered secondary hypothesis tests, we investigate whether gender differences in political preferences, trust, life satisfaction, happiness, and health status vary across countries’ levels of gender equality and economic performance. We hypothesized that gender differences would diminish as levels of gender equality and economic performance increased. This hypothesis aligns with the convergence theory, which argues that as gender equality increases, social roles, opportunities, and life conditions for men and women become increasingly similar, leading to attitudinal convergence42. The competing compensation/activation hypothesis proposes that as gender equality increases, gender identity becomes more salient politically, leading to greater divergence between men and women43.

To evaluate these conjectures, we estimate meta-regressions based on the random-effects meta-analyses conducted for primary hypotheses 1a–1e, with the Global Gender Gap Index (GGGI)44 and the gross domestic product (GDP) per capita as independent variables (see Methods for details). Importantly, our moderation analyses are purely correlational and should not be interpreted causally.

We report the coefficient estimates for GGGI and GDP per capita in Supplementary Tables 3 and 4, respectively. The estimates of residual heterogeneity after accounting for the variability in either of the two explanatory variables are presented in Supplementary Tables 5 and 6. We graphically illustrate the results of the moderation analyses in Figs. 4 and 5.

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

Moderating effects of countries’ degree of gender equality on gender differences in the five outcome measures. Plotted are the meta-analytic point estimates of the gender difference (0 “male”, 1 “female”) in (a) left-right political preferences, (b) trusting behavior, (c) life satisfaction, (d) happiness, and (e) health status over the countries’ Global Gender Gap Index (GGGI), scaled from 0 (gender inequality) to 1 (gender equality). The marker size is based on the inverse of the estimate’s variance. The solid line and shaded area depict the linear fit and its 95% CI, respectively, based on random-effects meta-regressions of the k = 38 countries included in the analysis sample (using the restricted maximum likelihood estimator for τ). Estimates of the linear fits’ slopes (b) and associated p-values are reported in the bottom-right corners (see Supplementary Table 3 for details). Note that Kosovo (XK) is excluded from the analysis due to a lack of GGGI data.

GGGI is statistically significantly related to gender differences in political preferences and health status. A one-standard-deviation increase in GGGI (indicating greater gender equality) is associated with a 0.047 increase in the gender gap in political preferences (se = 0.011, 95% CI [− 0.069, − 0.025]; z = − 4.150, p < 0.001) and a 0.037 decrease in the gender gap in health status (se = 0.011, 95% CI [0.016, 0.058]; z = 3.438, p = 0.001). Notably, the estimated effect of gender equality on political preferences points in the opposite direction of our hypothesis; further evidence is needed to establish whether and how gender disparities impact differences in political attitudes between genders. As noted above, the direction of the effect stands in contrast to the convergence hypothesis but aligns with the compensation/activation hypothesis.

GGGI’s moderating effect on political preferences remains robust when controlling for GDP per capita. In contrast, the moderation effect on health status changes from statistically significant evidence to suggestive evidence when controlling for economic performance (see Supplementary Tables 3 and 7). We do not find evidence that gender differences in trust, life satisfaction, or happiness are influenced by country-level variation in gender equality.

GDP per capita is statistically significantly associated with the gender difference in health status. Specifically, a one-standard-deviation increase in GDP per capita reduces the gender gap in health status by 0.036 (se = 0.012, 95% CI [0.013, 0.58]; z = 3.103, p = 0.002). However, the moderating effect of GDP on health status is not robust when accounting for differences in gender equality across countries (see Supplementary Tables 4 and 7). The lack of robustness suggests that GGGI and GDP are likely capturing some of the same variations, which is supported by a statistically significant correlation between the two variables (= 0.490, 95% CI [0.202, 0.700]; t(36) = 3.376, p = 0.002; n = 38).

We do not find evidence of moderating effects of economic performance on any of the other four outcome variables. When accounting for the variability in both variables simultaneously, there is suggestive evidence of a negative association between GDP and the gender difference in happiness. However, since this correlation only reaches the 5% -threshold in the robustness analysis, the result should be taken with a grain of salt.

The residual heterogeneity after accounting for the variance explained by the moderating variables remains statistically significant in all meta-regressions. However, in the regressions that identify a statistically significant moderation effect, the magnitude of the heterogeneity estimates decreases somewhat. Specifically, when accounting for gender equality, the estimate of τ for the gender difference in political preferences decreases by approximately 20%, while the estimate of τ for the gender gap in health status decreases by about 13%. Similarly, accounting for the country-level variability in economic performance reduces the estimate of τ for the gender gap in health status by approximately 11%.

Fig. 5
Fig. 5The alternative text for this image may have been generated using AI.
Full size image

Moderating effects of the countries’ gross domestic product per capita on gender differences in the five outcome measures. Plotted are the meta-analytic point estimates of the gender difference (0 “male”, 1 “female”) in (a) political preferences, (b) trusting behavior, (c) life satisfaction, (d) happiness, and (e) health status over the countries’ gross domestic product (GDP) per capita (based on purchasing power parity), expressed in terms of 100,000$. The marker size is based on the inverse of the estimate’s variance. The solid line and shaded area depict the linear fit and its 95% CI, respectively, based on random-effects meta-regressions of the k = 38 countries included in the analysis sample (using the restricted maximum likelihood estimator for τ). Estimates of the linear fits’ slopes (b) and associated p-values are reported in the bottom-right corners (see Supplementary Table 3 for details). Note that, following our pre-registration, Kosovo (XK) is not included in the analysis.

Non-preregistered robustness tests

As suggested by reviewers, we included some additional robustness tests that were not preregistered. Since Luxembourg is an outlier in terms of GDP per capita, we excluded Luxembourg from our analyses of GDP’s moderating effect. When Luxembourg is excluded, the moderating effects of GDP on both political preferences and health turn statistically significant. For the remaining three outcome variables, the conclusions remain unchanged. When we also control for the GGGI, the moderating effect of GDP remains statistically significant for political preferences, but becomes suggestive for health status. Additionally, the impact of GDP on happiness becomes suggestive when both moderators are considered simultaneously. These results are summarized in Supplementary Table 8. Regarding residual heterogeneity, our conclusions remain unchanged; see Supplementary Tables 9 and 10.

To address the concern that the estimation of gender differences at the country level may confound period shocks or policy changes affecting various countries differently over time, we also conducted meta-analyses aggregating country-year gender gaps, with standard errors clustered by country, instead of country-level observations. The results are reported in Supplementary Table 11. This approach produces similar meta-analytic estimates of the gender gaps; however, evidence of a gender gap in life satisfaction is no longer suggestive. The absolute heterogeneity (τ) remains similar, but the relative heterogeneity, as expected, decreases due to the higher within-study variance with country-year observations.

We also extended the approach of aggregating country-year gender gaps to our moderation analysis to counter the concern that averaging the GGGI and GDP over time risks blurring the potentially varying dynamics across countries. The results of meta-regressions of gender differences on the GGGI and GDP per capita using country-year observations are reported in Supplementary Tables 12 and 13. These findings largely mirror our results based on the primary analyses, although the evidence for the moderating effect of GGGI on health status turns from statistically significant to suggestive (p = 0.006). Additionally, we now find significant evidence of a moderating effect of GDP on political preferences, indicating that the gender gap increases with GDP.

Discussion

We provide statistically significant meta-analytic evidence supporting four of our five primary hypotheses regarding gender differences in left-right political preferences, life satisfaction, happiness, and health status, as well as suggestive evidence for our fifth hypothesis regarding the gender gap in trust. However, the estimated effect sizes are small, with Cohen’s d estimates ranging from 0.023 for life satisfaction to − 0.110 for health status and a mean absolute Cohen’s d of 0.053. For trust, life satisfaction, and happiness, with absolute effect sizes ranging from about 0.02 to about 0.04, it is open to question whether gender differences of this magnitude are large enough to be considered practically relevant. The estimated gender gap in respondents’ health status is more substantial and appears practically relevant; the gender gap in left-right political preferences may also be large enough to matter in practice.

One strength of our study is that it relies on representative samples of the general population from 39 countries. However, a limitation to consider when interpreting our results and comparing them with estimates reported in the previous literature is that our results may not be generalizable beyond Europe. The gender gaps observed across countries worldwide may differ from our estimates, both in the meta-analytic average and in between-country variability. Although offering a comprehensive overview of the literature is beyond the scope of our research, it appears expedient to compare our findings with studies that estimate gender gaps across a wide range of countries.

Our results on the gender gap in life satisfaction and happiness can be compared to some previous estimates in the literature. Zweig2 examined gender differences in life satisfaction across 73 countries using data from the Gallup World Poll. The average across countries indicates that women reported a life satisfaction rating that is 0.04 points higher than men on a 0–10 scale. Joshanloo and Jovanovic4, utilizing Gallup World Poll data across 166 countries, reported a gender difference in life satisfaction favoring women of 0.05 on a 0–10 scale. Both these estimates are close to our estimate.

Further estimates of gender differences in life satisfaction and happiness are provided by Arrosa and Gandelman3 based on data from the Gallup World Poll, the World Value Survey, and the ESS, covering rounds 1–3. However, only the life satisfaction data from the Gallup World Poll and the happiness data from the ESS utilize a 0–10 scale, as in our study. For life satisfaction, they found an average gender difference of 0.15 favoring women, without controlling for covariates, and 0.30 when accounting for various socioeconomic factors. These estimates are substantially larger than the meta-analytic mean reported in our study. Regarding happiness, they estimated an average gender difference of 0.07 favoring men without controlling for any variables, and 0.14 in favor of women when controlling for socioeconomic factors. Our research indicates a gender difference in happiness of 0.075 favoring women.

Concerning trust, Falk et al.45 compared gender differences across 76 countries as part of the Global Preference Survey (GPS). They introduced a novel survey item to elicit trusting behavior, phrased as “I assume that people have only the best intentions,” rated on a 0–10 scale. Surprisingly, they found that women, on average, reported trust levels about 0.06 Cohen’s d units higher than men. However, recent research by Tannenbaum et al.46 suggests that the trust question used in the GPS may not be a strong indicator. Tannenbaum et al.46 correlated the lost wallet reporting rates from a field experiment in various countries—rates expected to relate to trust—with both a binary general trust item and the GPS trust measure. They found that the correlation for the general trust question was 0.57, while the correlation for the GPS item was only 0.02.

For left-right political preferences, Dassonneville16 examined gender differences over time, measured on a 1–10 scale, across OECD countries, using data from several large-scale survey studies. The study reported that the gender difference remained relatively stable at approximately 0.2 for nine European countries between 2000 and 2020 (as shown in Fig. 1 of the article; the exact estimates are not provided). Our meta-analytic estimate of the gender difference is 0.14 on a 0–10 scale, which aligns with the effect’s direction but is somewhat lower than in Dassonneville16.

While several studies informally compare gender differences across countries, they often do not account for the variability attributable to sampling differences. We provide confirmatory evidence for our five primary hypotheses about population heterogeneity, with statistically significant variability attributable to systematic differences across countries in all five outcome measures. In relative terms, population heterogeneity is high across all five outcome measures, which is at least partly due to the large sample sizes in the included countries, as heterogeneity is measured relative to the average country-level sampling error.

In absolute terms, heterogeneity (τ) appears to be more moderate and comparable in magnitude to previous estimates of systematic variability across populations23. Still, the estimates of τ are large compared to the relatively small meta-analytic effect size estimates. The estimates of the standard deviations of the true gender differences (τ) across the countries exceed the meta-analytic effect sizes of four out of five outcome measures. This implies that for a significant portion of countries, ranging from an estimated 6.0% for health status to an estimated 31.6% for life satisfaction, the true gender difference actually aligns in the opposite direction of the hypothesis. Consequently, the gender difference in one country may not be diagnostic of the direction of gender differences in other countries, let al.one the meta-analytic mean gender gap.

It is intriguing to compare our estimates of population heterogeneity with those found in the previous literature. Holzmeister et al.23 estimated population heterogeneity in effect sizes across 70 multi-lab replication studies, primarily in the field of psychology. They reported a median H of 1.08 and a median I² of 14.5%. In contrast, our study found a mean H of 3.88 and I² of 92.8%. The degree of relative heterogeneity in our study is thus significantly greater. However, the median absolute heterogeneity (τ) of the 46 studies included in Holzmeister et al.23 that reported effect sizes in terms of Cohen’s d was 0.06, which is similar to the mean τ of 0.063 found in our study. Thus, the marked discrepancies in relative heterogeneity estimates across these studies can primarily be attributed to the significantly larger average sample sizes used to estimate country-level effect sizes, which result in substantially lower estimates of the average within-study variance.

Our secondary hypotheses stated that gender differences are moderated by the countries’ degrees of gender equality and economic performance. Contrary to our ex ante conjecture, our test results suggest that increased gender equality widens the gender gap in political preferences. Consistent with this empirical finding, gender equality observed over time has increased in the countries studied, and previous research has shown that the gender gap in political preferences has widened at the same time13,16. Edlund and Pande13 argued that the widening gender gap in political preferences in the US may be linked to the decline in marriage rates. They suggested that marriage tends to redistribute resources from men to women and that lower income levels contribute to more left-leaning political attitudes. However, this explanation based on income differences seems inconsistent with our finding that greater gender equality, which is expected to narrow the gender wage gap, amplifies the gender gap in political views. In contrast, results reported by Hansen and Goenaga48 align with our findings. Using ESS data from 29 countries, they identified a gender gap in conceptions of democracy, which is more pronounced in countries considered more democratic and with higher levels of gender equality.

Our results also correspond with the compensation/activation hypothesis43 regarding divergence but contrast with the convergence hypothesis42. Further investigation of the association between gender equality and the gender gap in political preferences appears to be an intriguing avenue for future research, particularly in identifying potential mechanisms that may drive greater divergence in political preferences as gender equality rises. Welfare state institutions49 may serve as a possible mediator between gender equality and political attitudes. However, these relationships are likely complex, as the literature on cultural modernization suggests that gender equality and value change may interact in intricate ways50.

Furthermore, our moderation analyses suggest that the gender gap in health conditions is inversely related to both gender equality and country-level economic performance. The relationship between the gender gap in health status, gender equality, and economic performance is also an interesting area for further research. If increased gender equality reduces the gender gap in health conditions, it suggests that women’s self-rated health improves, men’s self-rated health worsens, or both.

Concerning the relationship between gender equality and the gender difference in trust, Falk and Hermle51 reported a positive association that our results do not support. They investigated the relationship between the level of gender equality and gender differences in economic preferences across 76 countries. Trust is the only overlapping variable between their and our study, and the survey measures of trust differ. Falk and Hermle51 also found that other gender differences are more pronounced in countries with greater gender equality. This observation aligns with our findings on political preferences but not with our findings on health status; neither of these two variables was considered in their study.

In conclusion, our meta-analytic results tend to align with previously reported findings on gender differences across the five outcome measures in our study. However, the effect sizes are small and may not be considered practically significant, except for health status, where men report health conditions 0.11 higher than women in Cohen’s d units. The observed heterogeneity in effect sizes is large, both relative to the average within-country variability and relative to the meta-analytic effect sizes, indicating that gender differences can be substantive in some countries while potentially reversing in others. This implies that the gender difference observed in one country may not reliably predict the gap in other countries or reflect the overall average across countries. In other words, the generalizability of gender gap estimates based on data from a single country may be low. The moderating effects of the country-level variability in gender equality and economic performance suggest that gender differences are at least partially influenced by cultural and institutional differences among countries. “Hidden moderators” may account for the residual variance between countries24,52,53,54. The association between gender equality and the gender gaps in political preferences and health status opens up intriguing avenues for future research. Additionally, developing stronger theoretical frameworks to explain heterogeneity remains an important area for further investigation.

Methods

We filed a pre-analysis plan (PAP) on OSFhttps://osf.io/zaq3n before conducting any analyses on the ESS outcome variables in this study. The only exception is a previous project that used ESS data to test whether having daughters affects political preferences47, in which left-right self-placement was an outcome and gender was included as a control. However, the project did not estimate meta-analytic gender differences and focused solely on parents with children living in the household, which accounted for approximately one-third of the sample. We acknowledge that there is no way to independently verify whether additional analyses were performed before the PAP was posted, which we recognize may limit its evidentiary value.

Sample and inclusion criteria

We use data from the European Social Survey (ESS), including all 11 rounds (covering the years 2002–2023) posted on the ESS homepage (www.europeansocialsurvey.org) at the time of the posting of this study’s PAP. We include data from all 39 countries that have participated in at least one of these 11 rounds: Albania (AL), Austria (AT), Belgium (BE), Bulgaria (BG), Croatia (HR), Cyprus (CY), Czechia (CZ), Denmark (DK), Estonia (EE), Finland (FI), France (FR), Germany (DE), Greece (GR), Hungary (HU), Iceland (IS), Ireland (IE), Israel, Italy (IT), Kosovo (XK), Latvia (LV), Lithuania (LT), Luxembourg (LU), Montenegro (ME), Netherlands (NL), North Macedonia (MK), Norway (NO), Poland (PL), Portugal (PT), Romania (RO), Russian Federation (RU), Serbia (RS), Slovak Republic (SK), Slovenia (SI), Spain (ES), Sweden (SE), Switzerland (CH), Turkey (TR), Ukraine (UA), and United Kingdom (GB). At the time of posting the PAP, data were available for 25 of the 31 countries that participated in round 11 of the ESS, and these data were included in our study for this round.

The ESS is a repeated cross-sectional survey of representative samples from a country’s general population aged 15 and older, which aims to collect at least 1,500 observations per country in each round (except for smaller countries with a population of fewer than 2 million people, 15 years and older, where the goal is to collect at least 800 observations). According to the ESS website (www.europeansocialsurvey.org), the data have mainly been collected through face-to-face computer-assisted personal interviews.

For each of the analyses detailed below, we restrict the samples to respondents with valid records for (i) gender (“female” or “male”), (ii) age (“year of birth”), and (iii) interview year, including all respondents who answered the survey question corresponding to the particular outcome measure. Since item non-response varies across the five dependent measures, the sample sizes differ somewhat across analyses. In particular, the sample sizes for the analyses of the five outcome variables described below range from 435,246 (for left-right political preferences) to 506,871 (for health status). The number of observations per country and outcome measure varies between 811 and 27,900, with the average across the five dependent variables ranging from 11,160 to 12,997 participants per country. Note that the significant variation in the number of observations across countries does not pose a challenge for estimating the meta-analytic effect sizes. In random-effects meta-analysis, the weights assigned to each estimate depend on the sample size, so countries with fewer observations receive lower weights when aggregating effect size estimates.

Outcome measures

The following survey questions serve as dependent variables (outcome measures A–E). The survey questions have been identical throughout all rounds of the ESS. The phrasing of the questions listed below is based on source questionnaire files available on the ESS data portal (https://ess.sikt.no/).

  1. A.

    Left-right political preferences (variable “lrscale” in the ESS).

In politics, people sometimes talk of ‘left’ and ‘right’. Using this card, where would you place yourself on this scale, where 0 means the left and 10 means the right?

  1. B.

    Trust (variable “ppltrst” in the ESS).

Using this card, generally speaking, would you say that most people can be trusted, or that you can’t be too careful in dealing with people? Please tell me on a score of 0 to 10, where 0 means you can’t be too careful and 10 means that most people can be trusted.

  1. C.

    Life satisfaction (variable “stflife” in the ESS).

All things considered, how satisfied are you with your life as a whole nowadays? Please answer using this card, where 0 means extremely dissatisfied and 10 means extremely satisfied.

  1. D.

    Happiness (variable “happy” in the ESS).

“Taking all things together, how happy would you say you are?” (0 ”extremely unhappy” to 10 “extremely happy”).

  1. E.

    Health status (variable “health” in the ESS).

“How is your health in general? Would you say it is…” (1 “very good,” 2 “good,” 3 “fair,” 4 “bad,” and 5 “very bad”).

Survey items A through D are measured on a scale of 0 to 10. For questions B through D, low values on the scales indicate negative self-reports, while high values indicate positive self-reports. The outcome variable E is measured on a scale of 1 to 5. We reverse-coded outcome measure E so that 1 indicates “very bad” and 5 indicates “very good,” aligning the coding’s directionality with measures B–D.

Independent and control variables

The key independent variable in all country-level analyses is the respondents’ gender, which is included in the household module of the ESS (variable “gndr”). In our analyses, gender is coded as 1 for “female” and 0 for “male.” All analyses control for the respondent’s age and the square of their age. The respondents’ age was determined as the year of the interview minus the respondent’s year of birth (variable “yrbrn”). Additionally, we account for survey round fixed effects in all analyses.

In secondary and robustness analyses, we additionally account for the variability in two country-level variables: (i) the Global Gender Gap Index (GGGI) and (ii) the countries’ gross domestic product (GDP) per capita. Data on the GGGI has been obtained from the World Economic Forum44. The GGGI is based on 14 indicator variables and benchmarks the current state and evolution of gender parity across four dimensions: economic participation and opportunity, educational attainment, health and survival, and political empowerment44. Higher index values indicate more gender equality. The variable entering our main analysis was constructed as the arithmetic mean of a country’s GGGI from 2006 (the first year for which GGGI data are available) to 2024.

Data on the countries’ GDP per capita has been obtained from the World Bank (https://data.worldbank.org/). Particularly, we use data on the countries’ GDP per capita based on purchasing power parity (PPP) in constant 2021 international $. To facilitate the interpretability of coefficient estimates, the GDP per capita is expressed in terms of $100,000. For our main analysis, we use each country’s average GDP per capita from 2002, the start of round 1 of the ESS, to 2024, the end of round 11.

In the analyses involving GGGI or GDP per capita, one country, Kosovo (XK), is excluded because it is not included in the Global Gender Gap Report44. In the PAP, we noted that data on Kosovo’s GDP per capita were not available either, but this turned out to be incorrect. Still, we exclude Kosovo from the tests of the moderating effect of economic performance, adhering to our preregistered exclusion criteria. Including Kosovo in the meta-regressions of gender differences in the outcome measures on the countries’ GDP per capita yields nearly identical results to those reported in the paper for all five dependent variables, without altering any conclusions regarding suggestive or statistically significant evidence.

To facilitate the interpretability of the associations between gender differences and GGGI and GDP per capita, respectively, we express the coefficient estimates in terms of one standard deviation changes in the moderating variables in the main text. The standard deviations of GGGI and GDP per capita (in terms of $100,000) across the 38 countries included in the analysis sample are 0.050 and 0.228, respectively.

Hypotheses and tests

We categorized our hypotheses into three groups: primary hypotheses, secondary hypotheses, and robustness tests. When evaluating hypothesis tests, we interpret two-sided p-values of less than 0.05 as “suggestive evidence” and those of less than 0.005 as “statistically significant evidence,” following the recommendations of Benjamin et al.39.

The foundation for our primary and secondary hypothesis tests consists of country-level estimates of gender differences across the five outcome variables. To obtain these estimates, we regressed each dependent measure on (i) gender, (ii) age, and (iii) age squared, accounting for (iv) survey round fixed effects, separately for each of the 39 countries included in our sample. Standard errors were estimated based on a heteroskedasticity-consistent covariance matrix using the HC1 estimator.

In estimating these 39 × 5 = 195 regression models, we followed the recommendation by Kaminska55, weighing observations by the variable “anweight” (analysis weight), defined as the product of “pspwght” (post-stratified design weight) and “pweight” (population size weight), to account for differences in population size and varying selection probabilities across countries, as well as nonresponse, noncoverage, and sampling error related to the post-stratification. Hence, the linear regressions were estimated using weighted least squares.

The primary and secondary hypotheses detailed in the Supplementary Information pertain to meta-analytic effect size estimates of country-level gender differences and tests of effect size heterogeneity. For each dependent variable, we aggregated the country-level estimates of the gender difference, along with the corresponding robust standard errors, in a random-effects meta-analytic model using the restricted maximum likelihood estimator for τ56. The meta-analyses and meta-regressions were conducted using the metafor package57 (v-4.6.0) in R58 (v-4.3.3). The confidence intervals for heterogeneity measures (τ, I², and H) were estimated using the Q-profile method59, implemented with the confint() function provided by the metafor package.

To facilitate the interpretability of our results, we report all estimates in Cohen’s d units. Cohen’s d estimates were obtained by dividing the estimated (unstandardized) effect sizes, standard errors, and standard deviation in true effect sizes across countries (τ) by the pooled standard deviation of the respective outcome variable. The pooled standard deviation was calculated as the square root of the weighted average variance across the included countries, using the number of observations entering the hypothesis test per country as weights. The pooled standard deviation is 2.230 for left-right political preferences, 2.311 for trust, 2.122 for life satisfaction, 1.903 for happiness, and 0.897 for health status.

Minimum detectable effect sizes

As part of our results, we report the minimum detectable effect sizes (MDEs) for which we have π = 90% statistical power at the α = 5% significance level and the α = 0.5% level for primary hypotheses 1a–1e and secondary hypotheses 1aGGGI–1eGGGI and 1aGDP–1eGDP (but not for the robustness tests of these secondary hypotheses). The MDEs have been determined as [Φ(1 − α/2) + Φ(π)] · se, where Φ denotes the cumulative standard normal distribution function, and se denotes the standard error of the meta-analytic effect size or the meta-regression coefficient. The MDEs are reported in Cohen’s d units.