This markdown document contains code and analyses to accompany the following paper.

Ong, D. C., Zaki, J., & Goodman, N. D. (2015). Affective Cognition: Exploring lay theories of emotion. Cognition, 143, 141-162.

I have appended only the results sections of the paper, side-by-side with the corresponding analyses. The paper itself is much easier to read in the journal-formatted form, here: http://www.sciencedirect.com/science/article/pii/S0010027715300196

Experiment 1 Results

Model selection (in Appendix A) revealed that participants’ emotion ratings were significantly predicted only by three of the seven regressors we initially proposed: amount won, the prediction error (PE), and the absolute value of the prediction error (|PE|) (see also Section 3.3 for a re-analysis with more data). Crucially, PE and |PE| account for significant variance in emotion ratings after accounting for amount won. This suggests that affective cognition is remarkably consistent with economic and psychological models of subjective utility. In particular, emotion inferences exhibited reference-dependence—tracking prediction error in addition to amount won—and loss aversion—in that emotion inferences were more strongly predicted by negative, as opposed to positive prediction error. These features suggest that lay observers spontaneously use key features of prospect theory ( Kahneman and Tversky, 1979 and Kahneman and Tversky, 1984) in reasoning about others’ emotions: a remarkable connection between formal and everyday theorizing. It is worth noting as well that the significant regressors for surprise followed a slightly different pattern from the rest of the other emotions, where the win probability, as well as regret and relief, seem just as important as the amount won, PE, and |PE|.

The aforementioned analysis suggests that amount won, PE, and |PE| are necessary to model emotion inferences in a gambling context and suggest a low dimensional structure for the situation features. Next, we explored the underlying dimensionality of participants’ inferences about agents’ emotions via an a priori planned Principal Component Analysis (PCA). Previous work on judgments of facial and verbal emotions (e.g., Russell, 1980 and Schlosberg, 1954) and self-reported emotions (e.g., Kuppens et al., 2012 and Watson and Tellegen, 1985) have suggested a low-dimensional structure, and we planned this analysis to see if a similar low-dimensional structure might emerge in attributed emotions in our paradigm.

The first principal component (PC) accounted for 59% of the variance in participants’ ratings along all 8 emotions, while the second PC accounted for 16%; subsequent PCs individually explained less than 10% of the variance. The first PC accounted for most of the variance in the emotion ratings, although the second PC accounted for a far lower, but still noteworthy, amount of variance. Full details of the PCA procedure and loading results are given in Appendix A.

Post-hoc exploratory analysis of the first two PCs revealed that the first PC positively tracked happiness and contentment, while negatively tracking all negative emotions; by contrast, the second PC positively tracked the intensity of both positive and negative emotions (Fig. 4A). Interestingly, this connects with classic concepts of valence and arousal, respectively, which feature centrally in emotion science6 (e.g., Kuppens et al., 2012, Russell, 1980 and Schlosberg, 1954). In particular, some theorists view emotional valence as a crucial form of feedback to the agent: positively valenced emotions like happiness signal a positive prediction error—that the agent is doing better than expected—hence, positively reinforcing successful behavior. Conversely, negatively valenced emotions could signal to the agent to change some behavior to achieve a better outcome (e.g., Carver and Scheier, 2004 and Ortony et al., 1988). In line with this, we find that the first PC (“valence”) of emotions attributed by the observer correlated strongly with the PE of the situation (r = 0.737, 95% C.I. = [0.707; 0.764]). Additionally, we find that the second PC (“arousal”) correlated with |PE| (r = 0.280 [0.222, 0.364]; Fig. 4B).

plot of chunk expt1-PCA-plots

## Warning: Removed 2 rows containing missing values (geom_point).