Page 62 - Demo

Chapter 360Data analysisTo analyse the data, we used Bayesian ordinal regression to test how the different modalities were correlated with each other, and Bayesian mixed models to explore whether attractiveness ratings were associated with speed-date outcome. All Bayesian models were created in the Stan computational framework and accessed using the brms package (Bürkner, 2017, 2018), version 2.13.5. In all analyses we centered ratings at 4, because this was the middle option. This was done to ease setting priors on the intercept. All models were run with 4 chains and 5000 iterations, of which 1000 were warmup iterations. We checked model convergence by inspecting the trace plots, histograms of the posteriors, GelmanRubin diagnostics, and autocorrelation between iterations (Depaoli & van de Schoot, 2017). We found no divergences or excessive autocorrelation.For the ordinal regressions, which allow the dependent variable to be of the ordinal type (Bürkner & Vuorre, 2019), we specified six models with a cumulative distribution, consisting of the attractiveness ratings for one modality as dependent variables, and attractiveness ratings of another modality as predictor. We added random intercepts for rater and rated individual, and allowed the slope of the predictor to vary by rater. Furthermore, we retained the default priors for the error terms and thresholds, and set conservative Gaussian priors with a mean of 0 and SD of 0.5 for the predictor.To test the relationship between multimodal attractiveness and speed-date outcome, we used Bayesian mixed models with a Bernoulli distribution, with willingness to meet again (yes/no) as response variable. First, we conducted a partial correlation analysis, which contained visual, auditory and olfactory attractiveness each interacting with sex as predictors. Second, we used three independent models with either visual, auditory or olfactory attractiveness as predictor, interacting with sex. This allowed us to see how strong the correlations were per modality when not controlling for the other two modalities. Also, it allowed for a slightly larger sample size per modality, because there were more complete cases. We added random intercepts for participant and dating partner, and allowed slopes for the attractiveness ratings to vary by participant. With regard to priors, we set a conservative Gaussian prior with a mean of 0 and SD of 1 for the intercept. For the predictors, we used conservative Gaussian priors with a mean of 0 and SD of 0.5. For the error terms, we set half-Cauchy priors with a scale of 1.Tom Roth.indd 60 08-01-2024 10:41