Below we will first report the effect of stimulus ambiguity and mask on choice and response time data. Next, we will disentangle perceptual and judgmental biases using DDM analyses that explain the descriptive results in terms of parameter changes.
To quantify the effect of facial mask on the interpretation of ambiguous facial expressions, a logistic function was fit on the choice data (Eq. 1; Fig. 2A). For both masked and unmasked ambiguous facial expressions, the proportion unfriendly choices increased as a function of stimulus ambiguity (from happy to angry: see Fig. 2A). For masked faces, there was a small, but significant negative choice bias (β0) reflecting a tendency to choose for friendly more often (b0mdn= -0.22, one-sample Wilcoxon signed-rank test for b0mdn = 0, V = 2793, p < 0.01). No significant bias was found for unmasked faces. Sensitivity (b1) to the stimulus was significantly lower for masked (b1mdn = 6.47) vs unmasked (b1mdn = 10.91) facial expressions (Wilcoxon signed-rank test, W = 9151, p < 0.01).
We tested for significant effects in response times using a 2 (masked vs. unmasked) × 6 (stimulus ambiguity) repeated measures analysis of variance (ANOVA). For response times, there was a significant main effect of emotional ambiguity of the facial expression, with increasing response times for lower ambiguity levels, symmetrical around zero intensity, F(1,135) = 366.4, p < 0.01. The main effect of mask was significant as well, with slower response times for masked stimuli, F(5,675) = 290.5, p < 0.01. In addition, there was a significant interaction effect between the stimulus ambiguity of the facial expression and mask, indicating that the effect of mask was not equally distributed across ambiguity levels, F(5,675) = 44.1, p < 0.01. More specifically, the difference between response times for masked and no-masked facial expression became smaller for the low (-10, 10) ambiguity levels (see Fig. 2B). Furthermore, post-hoc t-test show significant slower response times for masked happy than for masked angry facial expressions with a high (-60 vs 60) or moderate (-40 vs 40) stimulus ambiguity (both ts(135) > 5.4, p < 0.01). No such difference was found for the facial stimuli without a mask. Instead, participants were slower for angry vs happy facial expressions without a mask, with low emotional ambiguity (10, vs -10), t(135) = 3.82, p < 0.01.
Overall, these results of the analyses of choice and response times show that there are small, asymmetrical effects of a facial mask on interpretation of ambiguous emotional expressions. To further quantify these effects, we fitted the DDM to the data allowing us to decompose the effects in the underlying choice parameters.
The RT results in Fig. 2 partly suggest a bias towards unfriendly choices for masked stimuli, showing faster choices for easy (60) and moderate (40) angry masked faces. In contrast, the psychometric data reflect a general loss of sensitivity combined with an unexpected bias to friendly choices in the mask condition. These contradictory findings suggest that facial masks might affect the interpretation of ambiguous emotional faces via different underlying mechanisms. To identify whether bias effects are driven by a judgmental (top-down) or perceptual (stimulus-driven) process, the diffusion model was fitted to both the RT and choice data simultaneously, allowing to disentangle these different types of bias.
For the diffusion-model fits (see Fig. 4 in the methods section for a graphical representation of the goodness of fit), we found that both starting points for masked and unmasked stimuli were different from 0.5, with a median starting point of zMdn(SD)_masked = 0.53(0.07) and zMdn(SD)_unmasked = 0.47(0.05) respectively (Wilcoxon signed-rank test, V > 1490, p < 0.01), indicating a significant judgmental bias towards the unfriendly alternative for masked and towards the friendly alternative for unmasked facial expressions (see Fig. 3A).
Drift-rate varied with stimulus ambiguity, with higher (more negative for happy and more positive for angry) drift-rates for both masked and unmasked emotional expressions, F(1, 135) = 43.4, p < 0.01 (see Fig. 3B). In general, drift-rates were lower (less positive for angry and less negative for happy) for masked facial expressions, indicated by the significant interaction between stimulus ambiguity and mask, F(1,135) = 92.9, p < 0.01. To test for a perceptual bias, we first tested for a linear relationship between (signed) stimulus ambiguity and drift-rate for masked and unmasked facial expressions separately. Regression analyses showed that stimulus ambiguity predicted drift-rate for masked (R2 = .86, F(1, 815) = 4891, p < .001) and unmasked (R2 = 0.90, F(1, 815) = 7193, p < 0.01) facial stimuli. Next, perceptual bias was defined as the vertical shift (e.g., intercept) of the regression lines, representing a possible systematic bias towards either the friendly or unfriendly alternative (see Fig. 3B). The intercept was − 0.068 (t = -2.49, p < 0.01) for masked and 0.25 (t = 8.48, p < 0.01) for unmasked facial expressions, suggesting a small perceptual bias towards friendly choices for masked stimuli and a small perceptual bias towards unfriendly for unmasked facial expressions.
In addition to starting point (z) and drift-rate (v), non-decision time (t0) was free to vary across masked and unmasked conditions as well, to account for possible differences in sensory encoding of the stimulus, prior to the decision process. Non-decision time (t0) was higher for masked (t0Mdn(SD) = 0.51(0.06)) compared to unmasked facial expressions (t0Mdn(SD) = 0.47(0.05)), presumably reflecting a longer time for perceptual encoding, prior to the accumulation process (Wilcoxon signed-rank test, W = 228, p < 0.01).
In sum, our descriptive analyses suggest that masked compared to unmasked faces are judged as more friendly, but that judging masked friendly faces takes more time. Our DDM analyses show that masking a face results in a loss of sensory information and an unfriendly judgmental bias. In contrast, we found a slight friendly perceptual bias as well. This suggests that, although diagnostic cues in masked faces bias our participants towards friendliness via a stimulus-driven process, our participants also have the preconception that masked faces are unfriendly.