Predictive modeling of religiosity, prosociality, and moralizing in 295,000 individuals from European and non-European populations

The study addresses why moralizing religions exist and, specifically, whether religiosity at the individual level promotes large-scale cooperation and social trust, or whether it reflects an individual strategy under conditions of social mistrust that motivates moralizing others' behavior. Two competing psychological hypotheses are evaluated: (1) the cooperation hypothesis, which posits that belief in rule-enforcing supernatural agents increases compliance with cooperative norms, thereby elevating social trust and large-scale cooperation; and (2) the mistrust–moralization hypothesis, which posits that higher social mistrust leads individuals to endorse religiosity as a means to police others’ behaviors, especially regarding sexual norms. Prior evidence is mixed and often limited by small samples, ecological fallacy from aggregate-level inference, arbitrary inclusion of country-level controls, and lack of out-of-sample validation. This study leverages very large, diverse datasets and stratified k-fold cross-validation to provide robust, generalizable tests of these hypotheses at the individual level.

The paper situates the question within interdisciplinary work on religion's roles in social complexity, moralization, and cooperation. It contrasts group-level accounts arguing moralizing religions sustain large-scale cooperation (e.g., Norenzayan et al., Turchin) with individual-level accounts such as the Reproductive Religiosity Model (Weeden and Kurzban), suggesting religiosity enforces monogamous strategies by moralizing sexual behaviors. The authors highlight methodological issues in prior research: reliance on aggregate comparisons prone to ecological fallacy, mixed individual-aggregate models with arbitrary macro-factor selection, small and non-representative samples, and scarce tests of generalizability. This motivates their individual-level, cross-validated structural modeling approach using large-scale EVS and WVS data.

Data sources: European Values Study (EVS) waves 3–4 (1999–2004) and World Values Survey (WVS) waves 3–6. After exclusions for missingness (respondents with too many missing values or missing age) and removing European countries from WVS for non-European replication, the final samples comprised EVS: 100,599 initially eligible, 100,599 screened, final analytic n=90,539 used in models; WVS: 219,533 after screening, final analytic n=176,323–195,914 depending on model, across 62 countries. Multiple imputation: 20 complete datasets per survey were created using fully conditional specification tailored to data types (predictive mean matching, logistic regression, proportional odds models). SEM framework: For each dataset, three models were specified. Model 1 tests Hypothesis 1 (paths from Religiosity to Social mistrust and to Large-scale cooperation, plus covariance between mistrust and cooperation). Models 2.1 and 2.2 test Hypothesis 2, estimating paths from Social mistrust to Religiosity, Social mistrust to Moralizing (sexual promiscuity in 2.1; free-riding in 2.2), and Moralizing to Religiosity, with indirect effects from mistrust to religiosity via moralizing. Measurement models: Latent constructs and indicators—Religiosity (EVS: six indicators including service attendance, confidence in religious institutions, self-identity as religious, importance of religion and God, religious moral beliefs; WVS: five indicators excluding religious moral beliefs). Social mistrust (EVS: latent with General mistrust composite and Interpersonal mistrust—unwanted neighbor categories; WVS: single composite summing z-scores of binary general mistrust item and interpersonal mistrust across six categories). Large-scale cooperation (EVS: latent from volunteering in collective activities plus five political action behaviors; WVS: latent from three political actions; volunteering not available). Moralizing sexual promiscuity (EVS: latent from six items covering homosexuality, prostitution, abortion, divorce, adultery, casual sex; WVS: four analogous items). Moralizing free-riding (EVS: latent from seven items on benefit fraud, fare evasion, tax cheating, bribery, car stealing, lying, paying cash to avoid taxes; WVS: three analogous items). Covariate: Respondent age included as an adjustment variable in all models. Estimation and fit: Weighted least squares with mean and variance adjustment (WLSMV) to handle non-normality and categorical/ordinal indicators. Reported fit indices: scaled χ², CFI, RMSEA (with 95% CI), SRMR, averaged over iterations and imputations. Software: R with mice for imputation, lavaan for SEM, semTools runMI for pooling estimates. Cross-validation: Stratified 10-fold CV with 100 rounds (1000 total iterations). In each iteration, models are fit on 90% training data; predictive accuracy is assessed by comparing model-implied covariance matrices to observed covariance in the 10% held-out test data via a weighted least squares discrepancy function F, converted to χ² and RMSEA. A permutation test applies the same procedure to test datasets with within-indicator values randomly permuted to compute a predictive accuracy ratio (permuted F divided by real-data F). Lower F and RMSEA and higher predictive accuracy ratios indicate better generalization performance. Models’ performance and parameter estimates are averaged across the 1000 cross-validation iterations.

Model fit: All EVS and WVS models achieved excellent fit (e.g., EVS Model 1 CFI=0.974, RMSEA=0.036; EVS Model 2.1 CFI=0.956, RMSEA=0.052; EVS Model 2.2 CFI=0.970, RMSEA=0.034; WVS Model 1 CFI=0.963, RMSEA=0.045; WVS Model 2.1 CFI=0.956, RMSEA=0.046; WVS Model 2.2 CFI=0.981, RMSEA=0.026). Hypothesis 1 (religiosity increases trust and large-scale cooperation) was disconfirmed. In EVS Model 1, Religiosity predicted higher Social mistrust (standardized path ≈0.29) and lower Large-scale cooperation (≈−0.21). Religiosity explained about 8% of variance in Social mistrust and 4% in Large-scale cooperation (EVS R²). WVS Model 1 showed the same pattern (weaker magnitude; Social mistrust R²≈0.03). Social mistrust and Large-scale cooperation negatively covaried (EVS standardized covariance ≈−0.39). Hypothesis 2 (mistrust increases religiosity via moralization) was supported for sexual moralization (Model 2.1) but not for free-riding moralization (Model 2.2). EVS Model 2.1: Social mistrust strongly predicted Moralizing sexual promiscuity (std ≈0.53), which strongly predicted Religiosity (≈0.53). The direct path from Social mistrust to Religiosity was null (std ≈0.006; ns), indicating full mediation. Indirect effect std ≈0.28–0.33. This model explained 28% of the variance in Religiosity (EVS). WVS Model 2.1 replicated the mediated pattern with weaker effects, explaining 17% of Religiosity variance. EVS Model 2.2: Social mistrust weakly predicted Moralizing free-riding (std ≈0.03), and Moralizing free-riding predicted Religiosity (≈0.20), but Social mistrust also had a positive direct effect on Religiosity (≈0.27); the indirect effect via free-riding was minimal (std ≈0.006). Religiosity R² dropped to 11% (EVS) and to 3% in WVS Model 2.2, and in WVS the mistrust→free-riding path was non-significant. Cross-validation: All models generalized well to out-of-sample folds, with low F and RMSEA on real test data and much poorer fit to permuted data. Predictive accuracy ratios indicated substantial degradation when the test data structure was randomized. Variance-comparison tests showed Models 2.1 were more robust than 2.2 (e.g., EVS Levene’s test statistic 276.22, p<2.2e−16; WVS 185.62, p<2.2e−16). Overall pattern: Religiosity is weakly but robustly associated with higher social mistrust and lower large-scale cooperation, and individual differences in religiosity are well explained by higher mistrust coupled with greater moralization of others’ sexual behavior, but not by moralization of free-riding.

The findings contest the view that, at the individual level, religiosity promotes large-scale cooperation and social trust. Instead, more religious individuals report slightly greater social mistrust and slightly lower engagement in large-scale cooperative actions. The data support an alternative pathway whereby individuals who perceive others as untrustworthy are more inclined to moralize sexual behaviors, and this moralization is strongly associated with higher religiosity. This suggests religiosity may serve, in part, as a strategy to police others’ behaviors in domains relevant to reproductive strategies rather than as a mechanism to enhance generalized cooperation. The results align with broader evidence that religiosity is more prevalent where social trust is lower and institutions are weaker. Cross-validated SEM indicates these associations are robust and generalize across European and non-European samples, reinforcing the reliability of the observed structures and mediated effects. The absence of mediation via moralizing free-riding implies that the moral domain matters: sexual morality appears central in linking mistrust to religiosity, whereas policing free-riding behaviors does not account for religiosity to the same extent.

This study combines very large multinational survey data with stratified k-fold cross-validated SEM to test two competing psychological accounts of moralizing religiosity. Results robustly disconfirm that religiosity increases generalized trust and large-scale cooperation and instead support a pathway in which higher social mistrust fosters stronger moralization of sexual behaviors, which in turn is associated with higher religiosity. Moralization of free-riding does not similarly mediate the mistrust–religiosity link. Contributions include an individual-level, generalizable modeling strategy that avoids ecological fallacy and tests predictive validity. Future research should: (1) refine behavioral measures of large-scale cooperation beyond volunteering and political actions; (2) integrate exogenous determinants (historical, ecological, demographic) to explain more variance in religiosity; (3) examine different social levels of cooperation (e.g., kin, small communities) and the distinction between small-scale and large-scale religious systems; and (4) test links among resources, environmental harshness, reproductive strategies, social trust, moralization, and religiosity at the individual level.

- Measurement validity: Large-scale cooperation indicators (volunteering, political actions) may partly reflect other traits or motivations; volunteering might index broader prosocial dispositions, and political actions can sometimes pursue non-prosocial aims. - Explained variance: Models account for a limited portion of variance in religiosity, indicating important omitted factors (e.g., historical, ecological, institutional variables). - Scope: Analyses focus on modern world religions and large-scale cooperation; results do not address small-scale societies or different cooperation contexts (e.g., kin-based or tightly knit groups). - Level-specific effects: Lack of positive association between religiosity and large-scale cooperation at the individual level does not preclude cooperation at other social levels not captured here. - Data constraints: Some constructs had fewer indicators in WVS than EVS (e.g., mistrust measured as a composite in WVS; lack of volunteering items), potentially reducing measurement precision.