Gender as a Moderator between Normative Identity Style and Conservative Political Orientation in Chinese Population: A Bayesian Analysis

Overview This dataset supports a study investigating the relationships among normative identity style, empathy, egalitarianism, authoritarianism, and demographic variables (age group and gender) among Chinese adults. The study examines how normative identity style predict authoritarianism? And does this link have gender differences, using Bayesian linear regression implemented in R (rstanarm). Participants A total of 733 participants were recruited from Guizhou Province, China, comprising two distinct age cohorts:Younger group (n = 400),university students aged 18–25 years (M = 21.56, SD = 1.23) from Guizhou University of Finance and Economics. Older group (n = 333):adults aged 45–60 years (M = 51.04, SD = 5.12) recruited from hospitals, banks, and middle schools. The total sample was 37.65% male (n = 276) and 62.35% female (n = 457). Measures Age Group: Categorical variable, 0=Younger group (18-25 years), 1=Older group (45-60 years) Gender: Categorical variable,0=Male, 1=Female Normative Identity Style Inventory-4 (ISI-4; Smits et al., 2008) : 7 items and 5-point (0-4) Egalitarian Sex Role Attitudes Scale (ESRAS; Suzuki, 1991): 12 items and 4-point (1-4) Toronto Empathy Questionnaire (TEQ; Spreng et al., 2009): 10 items and 5-point (1-5) Right-Wing Authoritarianism Scale (RWAS; Zakrisson, 2005): 9 items and 4-point (1-4) Analytic Plan Summary Bayesian linear regression was conducted using the rstanarm package in R (version 4.2.2). The main model examined the effects of normative identity style (NIS), empathy, age group, gender, and the NIS × gender interaction on authoritarianism. Noninformative priors were specified, with mean values of 0.5 or −0.5 assigned to reflect expected positive or negative effects (standard deviation = 40), and the intercept prior was set to N(15, 20²). The auxiliary parameter followed an exponential distribution (rate = 1). The Markov Chain Monte Carlo (MCMC) algorithm ran 4 chains with 10,000 iterations each (first 5,000 discarded as warm-up). Posterior predictive performance was assessed using median absolute error (MAE) and the proportion of observed values falling within their 95% posterior prediction interval, with 10‑fold cross‑validation to avoid overfitting. A model without the interaction term was compared using the expected log‑predictive density (ELPD). Convergence was verified via trace plots, density plots, and autocorrelation plots, with all R‑hat values < 1.01.