Page 98 - Emotions through the eyes of our closest living relatives- Exploring attentional and behavioral mechanisms

Chapter 4
Within a three-second trial, bonobos on average looked 1.93 (SD =0.78) seconds to the target and distractor combined (raw, unweighted values) when bonobo stimuli were displayed, and 2.04 seconds (SD = 0.72) when human stimuli were displayed. Thus, as some trials were more reliable than other trials, and to account for variation in overall attention to the target and distractor during the trial, we also calculated a weight for each trial. We calculated the weight by dividing the sum of the looking duration to the target and distractor by the average looking duration to the target and distractor per participant. The weight gives more importance to trials in which the participant paid more attention to the stimuli, and less importance to trials where participants were relatively inattentive. Weights were added to our models for all measures of interest (M = 1, SD = 0.38, range [0.01 – 1.69]).
We used zero-one-inflated beta (ZOIB) regression to account for 0’s, 1’s, and the data between the range [0 , 1]. For our measure of interest, proportional looking duration to emotional stimuli (PLDemotion) across trials, we ran three separate models. In the first model, we examined whether the PLDemotion was higher than 0.5, i.e., whether participants look more than 50% of the time to emotional stimuli. In the second model, we assessed whether the PLDemotion differed between Species displayed on the stimuli (i.e., human or bonobo). In the third model, we zoomed in on the specific emotion categories and examined whether there was an interactive effect between Species and Emotion Category on the PLDemotion.
In all of our models, we used weakly informative priors, specifically a student-t (default) prior (df = 3, M = 0, SD = 2.5) for the standard deviation coefficient, and a normal distribution (M = 0, SD = 1) for all other coefficients. Species and Emotion Category were treatment (dummy) coded. For each model, we report the median estimate coefficient, together with the 89% credible interval (either the Highest- Density Credible Interval [HDI; a “summary credible interval” for the posterior distribution] or the 89% Highest-Posterior Density [HPD; the shortest possible credible interval]). For comparisons between conditions, we report the odds ratio (OR). We also report the probability of direction (pd), which indicates the certainty that an effect goes in a specific direction (Makowski et al., 2019c, 2019a).
To establish model convergence, we followed the guidelines set out in the WAMBS checklist by de Paoli & van de Schoot (2017). We assessed trace and autocorrelation plots, the Gelman-Rubin diagnostic values (convergence indicated by a value close to 1), and density histograms for the posterior distributions. We conducted all of our analyses using RStudio (v. 1.4.1106, R Core Team, 2020) and the package brms (Bürkner, 2017, 2018).
96