Partner choice does not predict prosociality across countries

Why does human prosociality vary around the world? Evolutionary models and laboratory experiments suggest that possibilities for partner choice (i.e. the ability to leave unprofitable relationships and strike up new ones) should promote cooperation across human societies. Leveraging the Global Preferences Survey (n = 27,125; 27 countries) and the World Values Survey (n = 54,728; 32 countries), we test this theory by estimating the associations between relational mobility, a socioecological measure of partner choice, and a wide variety of prosocial attitudes and behaviours, including impersonal altruism, reciprocity, trust, collective action and moral judgements of antisocial behaviour. Contrary to our pre-registered predictions, we found little evidence that partner choice is related to prosociality across countries. After controlling for shared causes of relational mobility and prosociality - environmental harshness, subsistence style and geographic and linguistic proximity - we found that only altruism and trust in people from another religion are positively related to relational mobility. We did not find positive relationships between relational mobility and reciprocity, generalised trust, collective action or moral judgements. These findings challenge evolutionary theories of human cooperation which emphasise partner choice as a key explanatory mechanism, and highlight the need to generalise models and experiments to global samples.

We focus on models that include relational mobility as the only predictor, but these can be generalised to include additional predictors as fixed effects.
In Study 1, we model prosociality as the outcome variable (Pro), relational mobility as the country-level predictor variable (Rel), random intercepts and slopes for different prosociality items in the Global Preferences Survey (item; altruism, positive reciprocity, and trust), and random intercepts for participants (part) and countries (country).
To deal with spatial and cultural non-independence between countries, we allow separate random intercepts for countries to covary according to geographic (G) and linguistic (L) proximity matrices. This is similar to the approach employed in phylogenetic general linear mixed models, which deal with the non-independent structure in model 'residuals' by including a pre-computed covariance matrix specifying the relationships between species (Villemereuil & Nakagawa, 2014; see also here). In addition to these random effects, we include a residual random intercept over countries to capture country-specific effects that are independent of geographic and linguistic relationships with other countries.
We also model relational mobility with measurement error by including standard deviations (Rel SD ) from observed latent variable means (Rel OBS ). This ensures that the uncertainty in the measurement of relational mobility from previous research is propagated into this model.
where ζ parameters represent ordinal intercept cutpoints,β represents the slope fixed effect, other α and β parameters represent random intercepts and slopes, τ parameters represent standard deviations for random effects, R represents the correlation matrix for the item random effects, and λ, κ, and z represent the mean, standard deviation, and standardised latent values for the relational mobility measurement error model.
In brms, this model is written as follows: Relational mobility Someone accepting a bribe d Figure S8 . Posterior predictions from a Bayesian multilevel ordinal regression predicting moral justifiability of different scenarios from country-level relational mobility, controlling for environmental harshness and subsistence style and including a quadratic effect for relational mobility. Higher numbers on the y-axis indicate lower justifiability ratings for each scenario.
Lines and shaded areas indicate median posterior regression lines and 95% credible intervals.

Table S3
Results from power analysis simulations. For each analysis, we simulated multiple datasets with various effect sizes (slopes) for relational mobility and, as a measure of power, determined the proportion of models fitted to these datasets that returned significantly positive slopes ( p < 0.05). We manipulated the effect sizes until analyses returned around 80% power. For effect size thresholds in regression, see Funder & Ozer (2019). For effect size thresholds in logistic regression, see Chen, Cohen, and Chen (2010 (Hu & Bentler, 1999;MacCallum et al., 1996).  (Hu & Bentler, 1999;MacCallum et al., 1996).