The effects of Botulinum toxin on the detection of gradual changes in facial emotion

When we feel sad or depressed, our face invariably “drops”. Conversely, when we try to cheer someone up, we might tell them “keep your smile up”, so presupposing that modifying the configuration of their facial muscles will enhance their mood. A crucial assumption that underpins this hypothesis is that mental states are shaped by information originating from the peripheral neuromotor system — a view operationalised as the Facial Feedback Hypothesis. We used botulinum toxin (BoNT-A) injected over the frown area to temporarily paralyse muscles necessary to express anger. Using a pre-post treatment design, we presented participants with gradually changing videos of a face morphing from neutral to full-blown expressions of either anger or happiness and asked them to press a button as soon as they had detected any change in the display. Results indicate that while all participants (control and BoNT-A) improved their reaction times from pre-test to post-test, the BoNT-A group did not when detecting anger in the post-test. We surmise that frown paralysis disadvantaged participants in their ability to improve the detection of anger. Our finding suggests that facial feedback causally affects perceptual awareness of changes in emotion, as well as people’s ability to use perceptual information to learn.

to Table 1. Mean percentage of trials used for analysis per group, type of change and session based on trials with a detection response time not higher than 25 seconds. Standard deviations are shown. When a detection response exceeded 25s of reaction time, the trial was discarded for analyses. Only responses given based on time accuracy were retained.

Frequentist equivalent of a priori hypothesis testing for detection RT
Note that the frequentist equivalent yield similar results as the Bayesian analysis. In order to accurately calculate this we adjusted the error term as estimated from the error term (MS-error) of the factor Session from the overall ANOVA. Contrary to what we expected, this analysis failed to show any difference between sessions for the BoNT-A group, t(11) = 0.29, p = 0.78, d = 0.063. The same analysis, however, revealed a significant difference between sessions in RTs for the detection of anger in the control group, t(11) = 6.136, p < 0.0001, d = 0.71, (Fig.   2) . The Control group showed shorter reaction times in detecting an emotional change of anger at S2 compared to S1.

Identification and confidence RT
A full factorial 2(Group) x 2(Session) x 2(Emotion) repeated measures ANOVA on log RTs for identification was A series of similar analyses were performed to analyse reaction times of the confidence ratings task.
A full factorial 2(Group) x 2(Session) x 2(Emotion) repeated measures ANOVA on log RTs for confidence ratings was conducted revealing a significant main effect of Session F(1,22) = 17.02, p < 0.001, η2 = 0.43 and a . See tables S3-S6. Table S3. Mean reaction times (Log RT) and standard deviations for the identification responses of anger. Table S4. Mean reaction times (Log RT) and standard deviations for the identification of happiness. Table S5. Mean reaction times (Log RT) and standard deviations for the confidence ratings over identification of anger. Table S6. Mean reaction times (Log RT) and standard deviations for the confidence ratings over identification of happiness.
Mean Log RTs for the identification of Anger

Control-S2
3.23 0.24 Mean Log RTs for confidence ratings of Anger

Control-S2
3.05 0.28 Mean Log RTs for confidence ratings of Happiness

Control-S2
3.04 0.30   Table S9. Mean percentage of trials where a correct identification was given, per group, type of change and session. Standard deviations are shown.

Confidence ratings
Gamma correlations were computed to gauge "relative meta-accuracy" that is, the extent to which participants are able to discriminate between their own correct and incorrect decisions. Gamma correlations are non-parametric correlation coefficients that relate individual trial-by-trial confidence ratings and correct versus incorrect responses (1 or 0) in the identification task.
To determine if confidence level differed overall, we performed paired samples t-tests for each group between each session for both types of changes. All confidence ratings were transformed into proportions. This analysis failed to show any significant difference between S1 and S2 for both types of changes (all p-values > 0.3). We then proceeded to the assessment of relative meta-accuracy. One constraint of the analysis is that participants need to show variability in their first-order accuracy and in their second-order confidence judgments.
In this respect, given that participants where highly accurate identifying changes of happiness, and that most participants reported being highly confident in these discriminations, no gamma correlations are reported for happiness.
Thus, we report gamma correlations for the detection of anger only for those participants for whom it was possible to compute the index. Unfortunately, for the BoNT-A group, gamma correlations could only be computed for 8 participants, whereas for the control group, it was possible only for 9. These gamma correlations (G) where first compared to zero, so as to first assess if participants were better than chance at discriminating correct vs. incorrect decisions. Interestingly, while at S1, for the BoNT-A group, G is not significantly greater than zero, t(7) = 1.55, p = 0.165, at S2, results show a significant difference compared to the null, t(7) = 2.93, p = 0.022. By contrast, for the control group, we find a difference between G and zero at S1, t (8)  suggest that not only did both groups of participants remain equally confident of their identification decisions across time, but also that the BoNT-A treatment had no effects on how sensitive participants were to their accuracy on a trial-by trial basis at S2 compared to S1.