Data analysis strategy
For the analyses of the full experimental design, we used the lmerTest (34) extension for lme4 (35). Participant id was used as a cluster variable with random intercept; random slopes were not specified due to non-convergence. For global evaluation of interaction effects in the full design models, we used type III sum of squares F tests. To test specific hypotheses, we conducted contrast analysis: For the rating data, we tested simple effects and for the facial activity data, we tested custom (or focused) contrasts, both via emmeans (36). Bonferroni correction was used to account for multiple testing (results were not substantially different when more sophisticated methods, such as multivariate t, 37, were used). We used two-tailed testing.
As for the moderation effects, due to the explorative nature of the analyses, we wanted to be more conservative and reduced the criterion for statistical significance for moderation analyses with a continuous moderator from .05 to .01.
The core analyses in this study revolve around a mask*targetage interaction, which, overall, remained intact when moderators were included (see supplementary information document uploaded to the submission system or supplementary information folder at https://osf.io/smqpk/?view_only=f2681fb953f54a149238efaf243bb9c0). Thus, we report these analyses as they were pre-registered, but only in the supplementary information. We departed from the preregistration for the mimicry analysis since we could not include type of emotion as a factor as the model did not converge. We therefore conducted separate analyses for each emotion. This decision was supported by the notion that mimicry of the different emotion expressions may follow different mechanisms (38;39). We report the main analyses below, for detailed results and code, see R Markdown (see document uploaded to the submission system or analysis folder, please navigate through floating table of content at top left, at https://osf.io/smqpk/?view_only=f2681fb953f54a149238efaf243bb9c0).
Emotion Perception
The following results for emotion perception are divided into hit rate (emotion recognition performance as the proportion of correctly identified expressions) and the emotion intensity rating for the target scale (e.g., the happiness rating for happy expressions).
Hit Rate
A model with the fixed factors mask (reference: no mask), emotion (reference: happy), and target age (reference: adults) with hits as the dependent variable revealed that all low order effects were qualified by a significant three-way interaction, F(1, 11034) = 55.64, p<.0001. For the estimated marginal means plot, please see Figure 1 (for the observed means plot, please see R Markdown).
Contrast analyses (simple effects) indicated that the recognition rate of sadness was reduced when masks were worn and the difference between masked and unmasked faces was larger for adult targets (85% vs 58%, p<.0001, Cohen’s d=.75) than child targets (67% vs 57%, p<.0001, Cohen’s d=.24). The hit rate for happiness was only slightly reduced for adult expressers (98% vs 91%, p<.0003, Cohen’s d=.19), whereas there was a substantial reduction in hit rate for masked child targets (98% vs 77%, p<.0001, Cohen’s d=.57). For anger, a higher hit rate for masked versus unmasked child targets (82% vs 76%, p<.0007, Cohen’s d=-.18) emerged, with the reverse pattern for adult targets (88% vs 80%, p<.0001, Cohen’s d=.23).
Intensity
For the emotion intensity ratings as well, all lower order effects were qualified by a significant three-way interaction, F(1, 11034)=136.18, p<.0001. For the estimated marginal means plot, please see Figure 2 (for the observed means plot, please see R Markdown).
Contrast analyses indicated that when expressers wore a mask, the target emotion intensity was rated significantly lower for both happy (adult: p<.0001, Cohen’s d=.51, child: p<.0001, Cohen’s d=1.29) and sad faces (adult: p<.0001, Cohen’s d=.09, child: p<.0001, Cohen’s d=.37). For angry faces, the perceived emotion intensity was lower when adults wore a mask (vs no mask; p<.0001, Cohen’s d=.34) but higher when children wore a mask (vs no mask; p<.0001, Cohen’s d=-.20).
Perceived Interpersonal Closeness
For interpersonal closeness, all lower order factors were qualified by a significant three-way interaction, F(1,11034)=45.4, p<.0001. For the estimated marginal means plot, please see Figure 3 (for the observed means plot, please see R Markdown).
Contrast analyses indicated that when child expressers of happiness wore a mask, interpersonal closeness was particularly low (vs no mask; p<.0001, Cohen’s d= 0.81). A similar, smaller, but still significant effect emerged for adult happy (p=.028, Cohen’s d=0.13) and sad (p=.015, Cohen’s d=.14) targets. No significant differences between mask vs no mask emerged for anger expressions.
Facial Mimicry
Facial mimicry was indexed by either (a) a pattern score [the difference between the Mean(AU12 Lip corner puller + AU6 Cheek raiser) and AU04 Brow furrow), for more details, see Methods section] significantly larger than zero or (b) scores for onset and apex, respectively, that are significantly larger than scores for responses to neutral expressions. Happiness mimicry was indexed by a positive pattern score and sadness and anger mimicry by a negative pattern score. Values significantly above zero indicate a matching expression by participants, values around zero indicate no expression, and values below zero indicate a counter-expression.
The main model included the fixed factors segment (reference: neutral expression/still face), mask (reference: no mask), and target age (reference: adult) to predict the positive expressivity score for happy expressers and the negative expressivity score for sad expressers and angry expressers, respectively.
Happy Expressions
All lower order effects were qualified by a significant mask*segment*targetage interaction, F(1,9297)=6.82, p=.001. For the estimated marginal means plot, please see Figure 4 (for the observed means plot, please see R Markdown). Comparisons of estimated marginal means to zero indicated that for unmasked expressers, mimicry was present at both the onset (adult: CI95[0.02, 0.27], child: CI95[0.13, 0.38]) and the apex of the expression (adult: CI95[0.3, 0.52], child: CI95[0.40, 0.65]), whereas for masked expressers, mimicry was significantly different from zero only at apex and only for adult expressers, CI95 [0.11, 0.36]. The facial reaction at the neutral segment was not different from zero, indicating that there were no facial reactions to neutral faces.
For adult expressers, custom contrasts revealed no significant differences between unmasked and masked expressers, neither at onset nor at apex (at apex, the difference was just not significant, estimate=0.16, adjusted p = 0.059, unadjusted p = 0.03, Cohen’s d=.14). For child expressers, the response to unmasked expressers was significantly higher than toward masked expressers for onset (child, onset: estimate=0.25, p<0.001, Cohen’s d=.23) and apex (child, apex: estimate=0.50, p<0.001, Cohen’s d=.45).
Regarding the comparison of onset and apex vs the neutral segment, facial positivity was higher at onset and apex than at neutral for unmasked expressions (adult, onset: estimate: 0.18, p=.022, Cohen’s d=.16; child, onset: estimate: 0.30, p<.0001, Cohen’s d=.28; adult, apex: estimate: 0.43, p<.0001, Cohen’s d=.39; child, apex: estimates: 0.58, p<.0001, Cohen’s d=.52) but for masked expressions this was the case only for adult expressers at apex (adult, apex: estimate: 0.30, p<.0001, Cohen’s d=.27).
In sum, mimicry of happy expressions emerged consistently for both adult and child unmasked expressers. Whereas adult masked expressers were still mimicked at the peak of expressivity, mimicry was entirely absent for masked child expressers.
Sad and Anger Expressions
For both sad and angry expressions, no interaction effects emerged, that is, mimicry was not differentially affected by masks for either adult or child expressers. For sadness, a significant effect of targetage emerged, F(1,9340)=11.23, p=.001, such that sadness in child faces was mimicked more strongly than in adult faces.
For both sadness and anger and in both target age groups, a pattern congruent with mimicry emerged (sadness: segment main effect, F(1,9335)=9.55, p<.0001); anger: segment main effect, F(1,9255)=14.80, p<.0001; see R Markdown for statistical analyses and plots).
The Mediating Role of Emotion Perception in the Mask-Mimicry Relationship
Kastendieck et al. (3) reported that differences in the mimicry of masked vs unmasked happy expressions were mediated by differences in the perceived intensity of happiness. These findings could be replicated for both adult and child faces, as shown in Figure 5 (overall: mediation proportion 55.6%). A separate analysis showed that the effect was twice as strong for child (63%) than for adult (28.4%) faces. This result supports the notion that facial mimicry of happiness is dependent on the appraisal of the face as showing happiness as suggested by the Emotional Mimicry in Context Theory (4,5). Moreover, it supports the view that adult perceivers had more trouble recognizing happiness in masked child faces, possibly due to the absence of eye wrinkles. Since masks did not reduce the mimicry of sad and angry faces, no mediation models were assessed.
Exploratory analysis: Emotion Perception
The main analyses of the emotion ratings revealed reductions in perceived intensity for masked faces (except for anger in masked child faces). This raises the question of whether this implies that overall intensity ratings are lower as found by Hareli et al. (40) or whether participants rated other emotions instead. Hess and Kafetsios (41) distinguish in this context between accuracy (the rating of the correct emotion) and bias (the ratings on all other rated emotions). The intensity ratings reported above correspond to accuracy. Ancillary analyses were conducted to assess bias.
Higher overall bias (i.e., average across all non-target scales) for masked (vs unmasked) faces was found for child happiness expressions and for adult sad expressions. Similar but smaller mask-no mask differences were significant also for the remaining factor level combinations except anger in children. These findings point to reduced signal clarity of masked faces, which may lead to errors in ratings.
Specifically, participants rated masked faces more negatively (i.e., mean of negatively valenced emotions sadness, anger, fear, and disgust), again except in the case of anger in children. Notable were, in particular, higher ratings of anger for happy masked child faces.