Page 46 - Emotions through the eyes of our closest living relatives- Exploring attentional and behavioral mechanisms
P. 46

                                Chapter 2
intercepts per ID and per ID*Session. The dependent variable was reaction time (ms), and we included Congruency (sum coded), Familiarity (familiar versus unfamiliar human model, sum coded) and their interaction terms as fixed factors. We checked which distribution family (gaussian vs. gamma distribution) fit best the data based on the AIC statistic. We checked the assumptions of our models by visually inspecting QQ plots and the residuals plotted against fitted values.
Results
The AIC statistics revealed a better fit for a model with a normal distribution rather than a gamma distribution (AICnormal = 11701, AICgamma = 11756). The model did not show a significant main effect for Congruency (c2(1) = .33, p = .567), nor for Familiarity (c2(1) = .04, p = .840), nor an interaction effect between Congruency and Familiarity (c2(1) = .16, p = .693. See Figure 4, top right. Also see supplements Tables S7.1 and S7.2 for individual averages and further model output).
To substantiate our null-finding, we conducted a Bayesian control analysis. Bayesian analyses have been proposed as a reliable way to establish the strength of evidence against the null-hypothesis when frequentist analyses show non-significant results (Rouder et al., 2009). Specifically, calculating a Bayes Factor (BF) can assist in examining evidence for the null-hypothesis, which is not possible within the frequentist framework (Kass & Raftery, 1995). To do so, we fitted a Bayesian mixed (Gaussian) model using the brms package in R (Bürkner, 2017, 2018). In the model, Congruency, Familiarity and their interaction were defined as fixed factors, with reaction time as dependent variable. Congruency and Familiarity were sum-coded, and we included a nested random intercept (with Session nested within Participant). Priors consisted of a weakly informative Gaussian prior for the intercept (M = 500, SD = 100) and a more conservative Gaussian prior the fixed effect (M = 0, SD = 10). For the random effect and residual standard deviation, we used the default half Student-t priors (with 3 df). We also ran a null model that included the same parameters, excluding the fixed factors and their interactions. For each model, we ran four chains with 4000 iterations (of which 2000 iterations were warmups). Model validity was established by following the WAMBS checklist (Depaoli & van de Schoot, 2017), including trace plots, histograms of the posteriors, Gelman-Rubin diagnostics, and autocorrelation checks. We then calculated an average Bayes Factor01 using 1000 iterations, and found that the mean BF01 = 61.07 (SD = 16.06), indicating very strong
44





























































































   44   45   46   47   48