Page 72 - Demo

Chapter 370et al., 2018) and employed higher-order polynomials in Linear Mixed Models (LMMs). Given the fast changes in EMG activity related to affective states (Van Boxtel, 2010) as well as variations in response shapes (Cacioppo et al., 1988), we did not expect higher-order polynomials to reliable capture signal changes in the two EMG channels within the 4 seconds response window. In previous research on perception of static emotional expressions, EMG data was mostly analyzed over time periods of 1.5 %u2013 2.5 seconds (Bornemann et al., 2012; Hermans et al., 2009; Rymarczyk et al., 2011, 2016; Sato et al., 2008) and, even if longer time windows were looked at, the EMG signal was averaged over time (Kret, Stekelenburg, et al., 2013; Vrana & Gross, 2004). To keep the temporal resolution similar across measures and still allow for a fine-grained description of the EMG time courses, we therefore chose to identify time bins in which the stimulus content affected the signal rather than describing the signal as a whole, similar to the approach of Achaibou and colleagues (Achaibou et al., 2008). The two analysis approaches will be outlined in more detail below.Pupillometry, skin conductance & skin temperature. The time courses of the pupil size data, SCL data and the SKT data were modeled using growth curve analysis (Mirman, 2014) with the nlme package (Pinheiro et al., 2022) in R statistic (R Core Team, 2020). Three separate analyses were done for the three emotional expression modalities (prototypical facial expressions, bodily expressions and subtle facial cues). LMMs were fitted as follows: In order to capture the shape of the signal, firstand second-order orthogonal polynomials were used to model changes in pupil size, and first-, second- and third-order polynomials were chosen for the SCL and SKT models based on visual inspection of the overall shape of the time courses per subject. Within each expression%u2019s modality, emotion category (subtle: cue type) of the stimulus was included as categorical predictor (prototypical facial/bodily: angry, happy, sad, fearful and neutral; subtle: blush, dilated pupils, tears and neutral). Since these predictors of interest were assumed to influence the shape of the signal, interactions with the polynomials were added as fixed-effects to the models. Given the observed individual differences in the overall shape of the time series, a random intercept and random slopes of the polynomials were defined on a subject level. In order to account for autocorrelation between subsequent data points, an autoregressive structure with trials nested in subject as grouping factor was included. The Nelder-Mead technique was chosen as optimization method. Given the complex model structure, we increased the maximum number of