Coherence and divergence in autonomic‐subjective affective space

Abstract A central tenet of many theories of emotion is that emotional states are accompanied by distinct patterns of autonomic activity. However, experimental studies of coherence between subjective and autonomic responses during emotional states provide little evidence of coherence. Crucially, previous studies investigating coherence have either adopted univariate approaches or made limited use of multivariate analytic approaches by investigating subjective and autonomic responses separately. The current study addressed this question using a multivariate dimensional approach to build a common autonomic‐subjective affective space incorporating subjective responses and three different autonomic signals (heart rate, skin conductance response, and pupil diameter), measured during an emotion‐inducing task, in 51 participants. Results showed that autonomic and subjective responses could be adequately described in a two‐dimensional affective space. The first dimension included contributions from subjective and autonomic responses, indicating coherence, while contributions to the second dimension were almost exclusively of autonomic covariance. Thus, while there was a degree of coherence between autonomic and subjective emotional responses, there was substantial structure in autonomic responses that did not covary with subjective emotional experience. This study, therefore, contributes new insights into the relationship between subjective and autonomic emotional responses, and provides a framework for future multimodal emotion research, enabling both hypothesis‐ and data‐driven testing.


Supplemental materials SM1: Additional Methods
List of IAPS stimuli used in the current study (P.Lang & Bradley, 2007; P. J. Lang, Bradley, & Cuthbert, 1997).Note that IAPS stimuli cannot be shared publicly, below is the list of the stimuli used in the current study.

SM2. Details of power estimations
Estimates of the degree of coherence between subjective affective experience and autonomic response in the literature have been heterogeneous, ranging from small to moderately large effect sizes (Kreibig, 2010;Norman et al., 2016, Siegel et al., 2018).Based on data from our previous studies (Cuve et al., 2022, Gaigg et al., 2018), we assumed an average effect size of approximately .4 (Cohen's d, upper bound for a small effect).For a target power of 80% and an alpha of <.05, it results in approximately 50 participants (see power contours A & B).This means that for a population where a true effect is zero (C left), we would detect the target effect about 80% of the time (C right density).Note however that the final sample reflects a practical choice, as, in practice, power will vary with each predictor and a range of model optimisation decisions (e.g.estimating slopes or not), the optimality of which cannot be determined apriori and often needs to be driven by the data at hand (see Bates, Kliegl, Vasishth & Baayen, 2015).However, given that most effect sizes from previous studies are derived from aggregated (averaging) approaches, which are less powerful than the multilevel approach used here, we were confident that our sample estimation should provide enough power for the current analysis.

SM2: Temporal analysis of pupil diameter
There was a possibility that the relationship between arousal and pupil diameter may evolve over time, as suggested by recent studies showing distinct temporal components of the pupil response to emotional states (Kinner et al., 2017).To investigate this, linear mixed models were fitted to predict pupil size from subjective ratings of arousal and valence, their interaction, and image statistics (brightness, contrast) at different time points (n = 60, each bin being an average of 10 time points = 100 ms).Given the rapid pupillary light reflex (PLR) expected with visual stimuli, significant relationships between subjective arousal, valence and pupil are only plausible after the PLR (~ 1 second).
Significant time clusters were therefore defined as adjacent time bins where the corrected pvalue remained < .05after the PLR.The results of these analyses were consistent with the previous aggregated analyses (aggregated across time, not across participants or trials -see Figure 2-B).Specifically, there was a large effect of brightness (Mean effect size0.3-6 s = -1.02,p < .001).There was also a significant effect of subjective arousal, with more subjectively arousing stimuli associated with larger pupil dilation (Mean effect size1-6s = 0.45, p < .05,corrected).The interaction between subjective arousal and valence had a small but significant effect (Mean effect size 1-3s = 0.03, p < .05,corrected), suggesting that early changes in pupil diameter following the light reflex are differently modulated by arousal and valence.As in the aggregated analysis, there was no significant main effect of valence (Mean effect size = 0.02, p > .05,corrected) suggesting that pupil diameter is primarily associated with arousal.Optimal solutions had two dimensions that captured covariance patterns in autonomic and subjective response to emotion induction.However, these results suggest that reduced coherence between subjective and autonomic responses is likely related to individual differences between participants, because analyses that aggregated responses by participant explained less variance than those that aggregated responses by stimulus (averaged across participants) -see Figure SM4.1.

SM5. Classification and Clustering analysis details
A supervised Random forest classifier was used to classify stimuli category (labels derived from the IAPS validation, Lang & Bradley, 2007) from the autonomic-subjective affective space dimensions.The classifier training set consisted of 60% of data points which was validated in held-out validation dataset making up 20% of the sample data, and test set of 20% of the data, all randomly partitioned.The number of decision trees was optimised according to out-of-bag error validation, from a maximum of 100 trees.Random forest analysis were conducted using functions in the R package randomForest (Liaw & Wiener, 2002).
Unsupervised data driven clustering was conducted using k-means.The optimal number of clusters (k) was tested by running the k-mean algorithm with k = 2 to k = 10 and computing the within-cluster sum of squares (WCSS) and plotting the results using the elbow method (the selected model being the model with the lowest WCSS).AIC and BIC were also used as additional criteria to validate k, and the final solution was evaluated using the amount of variance explained by the cluster solution as well as a silhouette score (ranging from -1 to 1 and describing the degree of cluster cohesion) with positive scores indicating better cluster cohesion and separation and negative scores or a score of zero suggesting no detectable cluster.Kmean analysis were conducted using base R (Team, 2013) k-mean function.The significance of the relationship between classification and cluster results were analysed using a chi-square test.Additional analysis and visualisation were conducted using JASP and FactoMiner functions (Lê, Josse, & Husson, 2008;Love et al., 2019), ggplot2 (Wickham & Wickham, 2016) and palates from RColorBrewer (Neuwirth & Brewer, 2014).

Figure S1 .
Figure S1.Illustration of effect sizes and power estimation as a function of the sample (effect

SM4:
Additional MFA results MFA sensitivity analysis conducted on aggregated data (by participant * stimulus class: positive, negative, neutral; by stimulus id and by participant id) produced consistent results with those reported in the paper -see Figure SM4.1.

Figure SM4. 1
Figure SM4.1 Additional MFA results from sensitivity analysis.Left column reflects results for data averaged by participant * stimulus category (averaged across stimulus within each category: negative, positive, neutral).Middle column shows results for data aggregated by stimulus (averaged across participants).Right column shows results for data aggregated by participant (average across stimuli id and category).