Read on Twitter
Here is a thread summarizing our new #psychology paper "Reconsidering the Duchenne Smile," now in press in the Affective Science journal @affectScience. I worked really hard on this and can't wait to discuss it with you all. [Co-authors: Jeff Cohn, Lijun Yin, LP Morency]1/25
![Here is a thread summarizing our new #psychology paper "Reconsidering the Duchenne Smile," now in press in the Affective Science journal @affectScience. I worked really hard on this and can't wait to discuss it with you all. [Co-authors: Jeff Cohn, Lijun Yin, LP Morency]1/25](https://pbs.twimg.com/media/EsCO2I7XEBYt6Pp.jpg "Here is a thread summarizing our new #psychology paper \"Reconsidering the Duchenne Smile,\" now in press in the Affective Science journal @affectScience. I worked really hard on this and can't wait to discuss it with you all. [Co-authors: Jeff Cohn, Lijun Yin, LP Morency]1/25")
To view a free version of the paper where I did the typesetting to control the layout (i.e., "the director's cut"), see http://psyarxiv.com/397af . To view the official, early access publisher's version of the paper with their typesetting, see http://doi.org/10/ghtfjm 2/25
Intro: There is a longstanding debate about whether a person's emotions can be inferred from their facial behavior. One view argues that certain emotions are reliably revealed by certain facial movements. The hallmark example touted by this view is the Duchenne smile. 3/25
A Duchenne smile (right image below) involves two muscles: the "lip corner puller" (pulls lip corners toward ears) and the "cheek raiser" (raises cheeks and constricts eyes). A non-Duchenne smile (left image below) involves the lip corner puller but not the cheek raiser. 4/25
Ekman, Davidson, & Friedman (1990) argued that Duchenne smiles reveal genuine positive emotion, whereas non-Duchenne smiles are voluntary, false, or miserable smiles. This paper has been very influential in science and public interest. http://doi.org/10/csk95r 5/25
Many claims have been made about the Duchenne smile. Our first goal in the paper was to formalize a subset of these claims in a list of testable hypotheses and then to briefly (given rather tight word limits) review the evidence weighing on these hypotheses. 6/25
The literature seems to support H1 (positive emotion leads to Duchenne smiles) and H7 (Duchenne smiles are viewed as more positive) but contradict H2 (Duchenne smiles can't be counterfeited) and cast doubt on H3 (Duchenne smiles rarely occur without positive emotion). 7/25
The other hypotheses either had mixed support (H4 that eye constriction reliably reveals positive emotion and H5 that eye constriction is the best marker of positive emotion) or had not yet been studied enough (H6 and H8 that these links are context-independent). 8/25
Methods: We thus designed two related studies to further test these hypotheses and advance the literature on this topic. We began with the BP4D+ dataset, which includes video of 140 undergrad participants reacting to emotion-elicitation lab tasks. http://doi.org/10/gfrkr8 9/25
We used expert facial behavior coding to identify and extract all smiles that occurred during the lab tasks. In total, we found 751 smiles from 136 participants. We measured each smile's intensity and duration and the presence and intensity of eye constriction. 10/25
Participants reported their own level of positive emotion in the moment and we also recruited online participants from @Prolific to watch videos of the smiles and rate how much they thought the smiling person was feeling positive emotion. 11/25
Results: We were surprised that *most* smiles were Duchenne smiles, even when no positive emotion was self-reported. As a diagnostic test, eye constriction presence performed rather poorly with a sensitivity of 90% but a specificity of only 20% (PPV=0.50, NPV=0.69). 12/25
We also used more sophisticated statistical methods to explore our hypotheses. Bayesian multilevel regression models were used to predict self-reported positive emotion from each smile characteristic, both as individual (zero-order) and simultaneous (partial) predictors. 13/25
When considered individually, all slopes were significant and positive: higher self-reported positive emotion was associated with more intense smiles, longer smiles, eye constriction presence, and eye constriction intensity. 14/25
When considered simultaneously, however, the slopes for eye constriction were no longer significant. Thus, when smile intensity and duration are already known, learning about eye constriction (i.e., Duchenne smiles) did not help you predict self-reported positive emotion. 15/25
Conditional effects plots do a good job depicting these partial effects (although note that the y-axis is being depicted as a continuous variable here, although it was modeled as an ordinal variable). 16/25
When moderated by emotion-elicitation task, these partial effects also showed a great deal of sensitivity to context. Indeed, Duchenne smiles actually predicted *less* positive emotion in the context of a pain-elicitation task. 17/25
Interestingly, a different pattern emerged when predicting observer-rated positive emotion. All zero-order relationships were again significant and positive, but this time the partial effects of eye constriction were still significant and positive. 18/25
Thus, Duchenne smiles were perceived as more positive than non-Duchenne smiles and more intense eye constriction was perceived as more positive. 19/25
Although again context seemed to matter, with observers perceiving more intense eye constriction as less positive in the pain-elicitation task. 20/25
Discussion: Returning to the list of hypotheses, our work found that H1 (positive emotion often leads to Duchenne smiles) and H7 (Duchenne smiles are perceived as more positive) were supported, but H3, H4, H5, H6, and H8 were all contradicted (H2 was not examined). 21/25
Overall, predicting self-reported positive emotion from smiles was quite difficult, both for an algorithmic approach based on simple characteristics (pseudo-R2 = 0.50) and for human observers watching videos divorced from their broader context (pseudo-R2 = 0.44). 22/25
In conclusion, our data suggests that accurately inferring positive emotion from facial behavior is likely much more difficult than simply looking for the presence or absence (or even intensity) of eye constriction. We need to move beyond this overly simple heuristic. 23/25
It's also important to delineate between work on the *production* of facial behavior and work on the *perception* of facial behavior. Don't conflate their findings! ( http://doi.org/cxrphx ) Finally, Bayesian multilevel models are super flexible and useful in psychology. 24/25
Thanks to the participants, Eva Krumhuber+ @Nate__Haines for providing feedback, @affectScience+Jon Gratch for publishing this, @paulbuerkner, @rlmcelreath, @Dom_Makowski, @strengejacke, @rstudio, @rubenarslan+ @Prolific for developing methods that were integral to this work. /end
