Bodily action penetrates affective perception

Are MAMIP effects truly perceptual?

The effectiveness of MAMIP has been tested by Fantoni & Gerbino (2014) under the assumption that the repeated execution of motor actions with a variable perceived degree of comfort generates a transient mood shift consistent with affective adaptation (Wilson & Gilbert, 2008); i.e., with the progressive reduction of action valence paralleling a prolonged induction. Affective adaptation is a convenient umbrella expression, including anecdotal evidence that the valence of objects and events tends to fade away (i.e., tends to become less extreme) during prolonged exposure, as a possible consequence of a shift of the internal reference level involved in the evaluation of perceived objects and events. In the domain of motor actions, the repetition of comfortable reaches should: (i) induce a better mood in the actor; and (ii) make comfortable reaches progressively closer to neutrality (i.e., less discrepant from the shifted internal reference level). Likewise, the repetition of uncomfortable reaches should: (i) induce a worse mood in the actor; and (ii) make uncomfortable reaches progressively closer to neutrality (i.e., less discrepant from the shifted internal reference level).

Transient mood shifts modify the valence of not only bodily actions but also external objects; notably, expressive faces. Mood-congruent perception has been frequently reported in the literature (Bouhuys, Bloem & Groothuis, 1995). Such an effect is consistent with the notion of affective priming (Klauer & Musch, 2003), given that the observer’s state induced by action valence can be treated as an affective prime, capable of pre-activating detectors selectively tuned to face features involved in the expression of congruent emotions. Consider a morphed face that interpolates the facial expressions of happiness and anger. Depending on observer’s mood, the interpolated face will contain features that support (in different proportions, by definition) both mood-congruent and mood-incongruent emotions: for instance, a cheek puffer should be more salient than an upper lip riser for an observer in a positive mood, while the opposite should hold for an observer in a negative mood.

According to a strong version of the facilitation-by-congruency hypothesis, the selective pre-activation of feature detectors by the transiently induced mood might produce two effects: an emotion identification bias (leading to a higher probability of classifying a neutral face as expressing a congruent emotion) as well as an improvement of the detection of a congruent emotion. According to a weak version of the same hypothesis, the transiently induced mood might produce only an emotion identification bias, without a change in sensitivity. This distinction is consistent with the categorization of priming effects proposed by Spruyt et al. (2002). According to the encoding stimulus account of affective priming, action valence acts as an affectively polarized prime that pre-activates the memory representations of affectively related facial features, thus making it easier to encode emotional features belonging to the same valence domain rather than to a different one. Alternatively, according to the response selection account of affective priming action affects only post-perceptual processing by automatically triggering response tendencies that facilitate or interfere with response types. Notice that in both cases one should expect a shift of the point of subjective neutrality due to an unbalance that depends on the valence congruency between the action-induced mood and the facial expression of emotion. Both accounts predict that the likelihood of interpreting a facial expression as angry, for instance, is larger after an uncomfortable rather than comfortable reaching act.

Fantoni & Gerbino (2014) demonstrated the effectiveness of MAMIP using a facial emotion identification (not detection) task. However, according to the above-described distinction and following the criticism of identification data voiced by Firestone & Scholl (2015), results obtained by Fantoni & Gerbino (2014) might simply reflect back-end memory processes rather than the penetrability of perception by internal states.

Furthermore, according to Fantoni & Gerbino (2014) the affective perception of faces is influenced by congruency between valence of performed actions and valence of the target facial expression (mediated by transient mood) as well as arousal. In principle, perceived facial expressions might be influenced by both mood valence (negative after uncomfortable acts vs. positive after comfortable acts) and arousal (higher after uncomfortable than comfortable acts). Independent of congruency between observer’s transient mood and emotion displayed by the target face, higher arousal might have facilitated processing of expressive faces evaluated after uncomfortable actions. The arousal-based hypothesis is consistent with evidence that visual sensitivity can be improved by hyperarousal induced by cold pressor stimulation (Woods et al., 2012; Woods, Philbeck, & Wirtz, 2013). In the experiment by Fantoni & Gerbino (2014), hyperarousal induced by uncomfortable actions might have influenced emotion identification by modulating selective attention (Derryberry & Reed, 1998; Gasper & Clore, 2002; Jeffries et al., 2008). Following Firestone & Scholl (2015), this interpretation would shed doubt on the truly perceptual origin of the effect found by Fantoni & Gerbino (2014), as modulations of selective attention could in principle change the input rather than subsequent processing.

To corroborate Fantoni & Gerbino’s (2014) results, in the present study we asked whether MAMIP effects reported in their paper are truly perceptual and whether evidence can be provided in favor of an encoding rather than response account of affective priming of perceived facial expressions of emotion, as mediated by action-induced mood states.

Rationale and expectations

Experiments were designed to fulfil three goals: (G1) to disentangle the contributions of action valence and arousal to the perception of facial expressions of emotion; (G2) to decide whether action-induced mood congruency affects the valence of perceived expressions through a top-down modulation of perceptual processing, signalled by a modification of emotion sensitivity in the absence of response criterion shifts, or only affects a post-perceptual stage, by modifying the response criterion (i.e., a decision, not perception, threshold); (G3) to evaluate the contribution of action-induced mood congruency relative to the natural mood owned by a participant not performing any movement before the perceptual task.

To accomplish these goals we selected two emotions with opposite polarity in the valence domain (happiness and anger), an objective yes/no task, and a sensitivity experiment (according to the terminology suggested by Witt et al. (2015)) comprising positive (signal + noise, S + N) trials in which the face displayed a target emotion and negative (noise, N) trials in which the face displayed a neutral expression. Observers should be both fast and accurate when making their yes/no judgment on whether the face contained the emotion signal. Such a method minimized the role of cognitive top-down factors, unavoidable in the facial emotion identification procedure used by Fantoni & Gerbino (2014), and allowed us to decide whether MAMIP influenced emotion detection.

To measure sensitivity and bias in a facial emotion detection task we conducted four experiments with stimuli belonging to happiness and anger continua, generated by morphing a neutral face and an emotional face displaying either full happiness or full anger (Fig. 1). Four morphing levels were chosen for each continuum, corresponding to different per cent emotion: 0 (original neutral face), 10, 20, and 30% (Fig. 1B). Participants were tested individually in two successive sessions distinguished by reaches of opposite valence (comfortable/positive vs. uncomfortable/negative), with the ordering of action valences counterbalanced across participants. In doing so, we disentangled the independent contribution of mood congruency, action valence, and arousal, by using as between-subjects factors the temporal ordering of action type (comfortable ⇒ uncomfortable in Experiments 1 and 3 vs. uncomfortable ⇒ comfortable in Experiments 2 and 4) and the valence of the target emotion (positive in Experiments 1 & 2 vs. negative in Experiments 3 & 4). By combining action valence and target emotion valence in a crossover design we planned to test the effect of action/emotion valence congruency on the detection of both positive and negative emotions (Fig. 1A, cells of the two matrices). Specifically, we planned to contrast:

• –

  – in Experiments 1 & 2, happiness detection after congruent comfortable reaches (top-right cell of the left matrix & top-left cell of the right matrix) vs. incongruent uncomfortable reaches (top-left cell of the left matrix & top-right cell of the right matrix);

• –

  – in Experiments 3 & 4, anger detection after congruent uncomfortable reaches (bottom-left cell of the left matrix & bottom-right cell of the right matrix) vs. incongruent comfortable reaches (bottom-right cell of the left matrix & bottom-left cell of the right matrix).

We used an equal-variance signal detection model (Macmillan & Creelman, 2004) to obtain, for every participant, the individual values of three indices of performance extracted from the proportions of yes responses in emotion present trials (hits) and emotion absent trials (false alarms). Two indices refer to a perceptual stage: sensitivity d′ (one value for each of the three levels of emotion) and absolute threshold AT (defined as the per cent emotion value above which d′ gets positive). The third index, referring to the response stage, is criterion c.

After adaptation to comfortable/uncomfortable reaches under unrestrained body conditions, we measured detection performance at three increasing levels of emotion (10, 20, 30%) for two morph continua (positive vs. negative target emotion), to contrast two hypotheses about possible effects of motor action valence on sensitivity to facial expressions.

H1) According to the facilitation-by-congruency hypothesis (see Fig. 2), higher d′ and lower AT values were expected in congruent than incongruent conditions, independent of action valence per se, as follows: (1) performing comfortable rather than uncomfortable actions should improve happiness detection (Experiments 1 & 2); (2) vice versa, performing uncomfortable rather than comfortable actions should improve anger detection (Experiments 3 & 4). This hypothesis attributes the role of mediator to a transient modification of mood in a direction congruent with the performed action.

H2) According to the facilitation-by-arousal hypothesis, if uncomfortable reaches induce higher arousal in the observer and emotion detection depends on arousal (at least more than on transient mood), higher d′ and lower AT values were expected after uncomfortable actions, regardless of target emotion valence, as follows: performing uncomfortable rather than comfortable actions should improve detection of both happiness (Experiments 1 & 2) and anger (Experiments 3 & 4).

Note that the two hypotheses lead to the same expectations for anger detection (Experiments 3 & 4), but opposite expectations for happiness detection (Experiments 1 & 2). In the present study Action Ordering has been treated as a balancing between-subjects variable, while Target Emotion has been manipulated as a between-subjects factor to avoid a possible carryover effect intrinsic to an experimental design in which the same actors/observers detect emotions of opposite valence in successive sessions.

Furthermore, as depicted in Fig. 2 (panels A, D for happiness detection; panels B, E for anger detection), we considered two alternative possibilities about the MAMIP effect on AT, which was based on three d′ values:

1. additive effect, producing different AT values in congruent (red solid line in Figs. 2A and 2B) and incongruent (cyan solid line in Figs. 2A and 2B) conditions, as a consequence of constant d′ increments/decrements at increasing per cent emotion in the morphed target; an additive effect on d′ would be signalled by significant main effects of morph intensity and of action/emotion valence congruency, in the absence of their interaction, in all four experiments;

2. multiplicative effect, producing the same AT value in congruent (red solid line in Figs. 2D and 2E) and incongruent (cyan solid line in Figs. 2D and 2E) conditions, as a consequence of proportionally increasing d′ increments/decrements as a function of per cent emotion in the morphed target; a multiplicative effect on d′ would be signalled by a significant morph intensity × action/emotion valence congruency interaction in all four experiments.

Note that both hypotheses hold if the assumption of d′ additivity is satisfied beyond MAMIP: The expected increase of sensitivity as a function of per cent emotion in the morph (as calculated by taking the proportion of false alarms in N trials as a common comparison value for the proportions of hits associated to each tested level of per cent emotion in S + N trials) should be approximately linear, with slope β₂ and with negative intercept β₁, thus allowing for the identification of an absolute threshold value above which d′ gets positive even in a baseline inaction condition (grey dotted line in Fig. 2).

Finally, we considered two alternative possibilities about the value of response criterion c, calculated by pooling together hits and false alarms for all three morph intensities, within each block of emotion detection trials following a MAMIP session. We expected that: (1) if MAMIP effects are perceptual (i.e., if bodily actions penetrate affective perception) then the c value should be constant across action/emotion valence congruency conditions associated with mood-congruent sensitivity effects; (2) if MAMIP effects are post-perceptual the c value should vary consistently with observer’s transient mood, even in the absence of an effect on sensitivity, due to a response bias in the direction of the action/mood congruent emotion (e.g., when required to detect anger observers might increase their yes rate after uncomfortable, relative to comfortable, actions).

Participants

Forty undergraduates of the University of Trieste, all right handed, participated in the experiment lasting about one hour and half. All had normal or corrected-to-normal vision, were naïve to the purpose of the experiment, had no history of mood disorder and were not using antidepressant medications at the moment of the experiment. They were randomly assigned to one of the four experiments resulting from the combination of Action Ordering (comfortable first, uncomfortable first) and Target Emotion (happiness, anger). Each experiment included 10 participants distributed as follows across females and males: Experiment 1 (uncomfortable ⇒ comfortable, happiness; mean age = 23.7, SD = 2.5), females = 7; Experiment 2 (comfortable ⇒ uncomfortable, happiness; mean age = 24.3, SD = 2.1), females = 6, Experiment 3 (uncomfortable ⇒ comfortable, anger; mean age = 23.2, SD = 2.1), females = 6; Experiment 4 (comfortable ⇒ uncomfortable, anger; mean age = 23.5, SD = 1.5), females = 7.

The study was approved by the Research Ethics Committee of the University of Trieste (approval number 52) in compliance with national legislation, the Ethical Code of the Italian Association of Psychology, and the Code of Ethical Principles for Medical Research Involving Human Subjects of the World Medical Association (Declaration of Helsinki). Participants provided their written informed consent prior to inclusion in the study. The Ethics Committee of the University of Trieste approved the participation of regularly enrolled students to data collection sessions connected to this specific study, as well as the informed consent form that participants were required to sign. Dataset is available as a Supplemental data file (S1).

Apparatus, stimuli & design

The experimental setting utilized exactly the same Augmented Reality apparatus described in Fantoni & Gerbino (2014). Participants were seated in a dark laboratory in front of a high-quality, front-silvered 40 × 30 cm mirror, slanted at 45° relative to the participant’s sagittal body midline and reflecting images displayed on a Sony Trinitron Color Graphic Display GDM-F520 CRT monitor (19″; 1024 × 768 pixels; 85 Hz refresh rate), placed at the left of the mirror (Fantoni & Gerbino, 2014, Figs. 1B and 1C).

The same 3D visual displays used by Fantoni & Gerbino (2014) were used in the MAMIP reaching phases of the four experiments: (1) a high-contrast vertically oriented random-dot rod (30% density, visible back-surface, 7.5 mm radius, 65 mm height), depicted in Fig. 3A (see for details Fantoni & Gerbino, 2014, Fig. 1A); (2) a virtual red sphere (3 mm diameter) that visually marked the tip of the participant’s index finger in 3D space after the finger departed from its starting position of about 30 mm. Both stimuli were rendered in stereo and were generated using a frame interlacing technique in conjunction with liquid crystal FE-1 goggles (Cambridge Research Systems, Cambridge, UK) synchronized with the monitor’s frame rate, and were updated in real time with head and hand movements (acquired on-line with sub-millimeter resolution by using an Optotrak Certus with one position sensor) so to keep their geometrical projection always consistent with the participant’s viewpoint.

Each reaching session was preceded by the same calibration procedure used in Fantoni & Gerbino (2014), and detailed in Nicolini et al. (2014). In particular, the positions of the index tip and of the eyes (viewpoint) were calculated during were calculated during the system calibration phase using three infrared-emitting diodes firmly attached on the distal phalanx and on the back of the head. This was needed to ensure a correct geometrical projection of the 3D visual displays (virtual rod and index tip) according to the shifts of different body segments. Specifically, head movements updated the participant’s viewpoint to present the correct geometrical projection of the stimulus in real time. A custom made Visual C++ program supported stimulus presentation and the acquisition of kinematic data associated to the reaching phase, as well as the recording of yes/no responses relative to the facial emotion detection task (left/right keys of the computer keyboard) and RTs.

The simulated egocentric depth of the rod along the line of sight was manipulated according to empirical data by Fantoni & Gerbino (2014), but see also Mark et al. (1997). Fantoni & Gerbino (2014) asked observers to rate the discomfort of 50 reaches whose depth was randomly varied across trials in the entire [0.65–1.00] range of arm length using a 0–50 discomfort scale adapted from the pain scale by Ellermeier, Westphal & Heidenfelder (1991). Following their results the simulated egocentric depth of the rod axis along the line of sight was randomly chosen within the [0.65–0.75] range for the comfortable-reaching block and within the [0.90–1.00] range for the uncomfortable-reaching block, relative to the arm length of each participant.

Furthermore, a physical rod (equal in shape to the virtual one) placed behind the mirror that fully occluded it was attached to a linear positioning stage (Velmex Inc., Bloomfield, NY, USA). The position of the physical rod was matched to the egocentric depth of the simulated rod on a trial-by-trial basis, with submillimiter precision (straight-line accuracy = 0.076 mm) with the Velmex motorized Bslides assembly (Bloomfield, NY, USA), so that real and virtual stimuli were perfectly aligned. This provided participants with a fully consistent haptic feedback as the red sphere marking the index tip reached the illusory surface defined by the constellation of random dots shaping the virtual rod exactly when the participant’s finger entered in contact with the real rod. Furthermore to ensure consistent vergence and accommodative information, the position of the monitor was also attached to a Velmex linear positioning stage that was adjusted on a trial-by-trial basis to equal the distance from the participant’s eyes to the virtual/real object that should be reached during the reaching block. Synchronizing stimulus presentation with the motorized linear positioning stage system allowed us to randomly manipulate the distance of the reaches in a rather continuous way (step resolution = 0.076 mm) over the depth ranges used in our study; i.e., in both comfortable ([0.65–0.75] the arm length) and uncomfortable ([0.75–1.00] the arm length) reaching blocks.

The facial stimulus set included the same 8 characters (four Caucasian males and four Caucasian females) that Fantoni & Gerbino (2014) selected from the Radboud University Nijmegen set (Langner et al., 2010): namely, female models 1, 4, 14, 19; and male models 20, 30, 46, 71. For each of the 8 selected characters of the Radboud database (included in Fantoni & Gerbino (2014), Fig. 1) we utilized color photographs displaying faces expressing two basic emotions, happiness and anger, and the corresponding neutral faces (all stimuli produced a high agreement with intended expressions in the validation study). A neutral-to-angry and a neutral-to-happy continua were first generated for each of the 8 characters, morphing the fully angry/happy face and the neutral face in variable proportions, in 5 per cent steps, using MATLAB software adapted from open source programs. For every two pairs of facial images we selected about 75 key points. The software generated a synthetic image containing a specified mixture of the original expressions, using a sophisticated morphing algorithm that implements the principles described by Benson & Perrett (1999). As in Marneweck, Loftus & Hammond (2013) and Fantoni & Gerbino (2014), we identified corresponding points in the two faces, with more points around areas of greater change with increasing emotional intensity (pupils, eyelids, eyebrows, and lips). Then, as depicted in Fig. 1B, for every character we selected three target stimuli for the neutral-to-angry continuum and three target stimuli for the neutral-to-happy continuum, corresponding to the following morph intensities: 10% emotion (90% neutral), 20% emotion (80% neutral), 30% emotion (70% neutral). Neutral stimuli were the original “no emotion” faces of the 8 selected characters. All images were aligned for facial landmarks and masked by an oval vignette hiding hair and ears presented on a black surround, the vignette being centered on the screen and occupying a visual size of 7.5° × 10.7° at the viewing distance of 50 cm.

The target emotion was positive (happiness) in Experiments 1 & 2 vs. negative (anger) in Experiments 3 & 4. Each experiment included two random sequences (one for each MAMIP condition, comfortable vs. uncomfortable) of 32 facial images resulting from the product of 8 characters × 4 morph intensities (including the neutral face corresponding to 0% emotion in the morph).

The complete 2 × 2 × 2 × 4 mixed factorial design shown in Fig. 1A included the two between-subjects factors called Action Ordering (comfortable ⇒ uncomfortable vs. uncomfortable ⇒ comfortable, Fig. 1A, top row) and Target Emotion (positive vs. negative, Fig. 1A, matrix rows) and the two within-subjects factors called action/emotion valence Congruency (congruent vs. incongruent, Fig. 1A, matrix columns, depending on encoding in each cell) and Morph Intensity (0, 10, 20, 30% emotion in the morph, Fig. 1B).

Procedure

As shown in Fig. 3, our procedure included two sessions, each composed by an induction phase (MAMIP reaching phase) followed by a test phase (emotion detection task).

Statistical analyses

Following Knoblauch & Maloney (2012), all indices of signal detection performance (both perceptual and decision based) were calculated by applying a generalized linear model (glm) with a probit link function to the whole set of binary responses. Individual triplets of d′ values associated to the three combinations of hits (yes responses to 10, 20, 30% morph, respectively) and false alarms (yes responses to 0% morph) were extracted using a glm with variable intercept β₁ and slope β₂ for every participant, reaching session, emotion, and action ordering. We encoded signal presence/absence as a discrete variable (1 for trials with morph intensity >0; 0 for trials with the original neutral face) and reparametrized each individual Gaussian function fit in term of its slope (corresponding to the difference between hit and false alarm rates on the probit scale), or d′.

Generalizing upon this statistical technique and following Marneweck, Loftus & Hammond (2013) we further extracted two global indices of detection performance by fitting the same glm to the entire set of yes responses as a function of signal presence/absence. The two indices were global d′ for the perceptual component of the model and response criterion c [as given by $-\frac{\left(2{\beta }_1+{\beta }_2\right)}{2}$] for the decision component. The c index provides a measure of response bias independent of sensitivity to facial expressions of emotion, as needed to draw conclusions relevant to our second goal (G2). Individual c values indicate how far the criterion used by the observer to deliver a yes response departs from the optimal decision rule (i.e., equal false alarm and miss rates), with negative c values indicating an unbalance in favor of yes over no responses. On the average, a negative c value was expected as a consequence of the unbalanced [S + N]/[N] ratio (with only 1/4 of trials displaying a neutral face).

To provide an additional measure of a possible mood-congruent effect on sensitivity to facial emotions, revealed by our detection task, we also analyzed AT values as derived from intercept and slope values estimated by a linear mixed-effect (lme) model with participants as a random effect and morph intensity as the continuous covariate of individual d′ triplets applied to each of the four experimental groups of participants separately.

The negative sign of the slope between individual d′ triplets and per cent target emotion was used as exclusion criterion, given that AT values computed from an indirect relationship between d′ and morph intensity were statistically meaningless: this led to the removal of one participant from the analysis of data from Experiment 1 (out of the total of 40).

Distributions of individual values of performance indices d′, global d′, AT, c were analyzed using a step-wise procedure that contrasted linear mixed-effect (lme) models of increasing complexity (Bates et al., 2014), depending on the number of fixed effects, modelled by the factors of our experimental design (action/emotion valence Congruency, Target Emotion, and Action Ordering). Models were fitted using Restricted Maximum Likelihood. Participants were treated as random effects so to control for the individual variability of emotion detection performance. We followed Bates (2010) and used this statistical procedure to obtain two-tailed p-values by means of likelihood ratio test based on χ² statistics when contrasting lme with different complexities (for a discussion of advantages of a lme procedure over the more traditional mixed models analysis of variance see Kliegl et al., 2010). We used type 3-like two tailed p-values for significance estimates of lme’s fixed effects and parameters adjusting for the F-tests the denominator degrees-of-freedom with the Satterthwaite approximation based on SAS proc mixed theory. Among the indices that have been proposed as reliable measures of the predictive power and of the goodness of fit for lme models we selected the concordance correlation coefficient rc, which provides a measure of the degree of agreement between observed and predicted values in the [−1, 1] range (Vonesh, Chinchilli & Pu, 1996; Rigutti, Fantoni & Gerbino, 2015). Post-hoc tests were performed using paired two sample t-tests with equal variance. As measures of significant effect size, depending on the statistical analysis, we provided Cohen’s d, the coefficient of determination r², and/or r_c.

Evidence from the distributions of yes responses

Figure 4 illustrates the average percentages of yes (“Emotion present”) responses (and SEMs) together with the best fitting cumulative Gaussian (averaged across participants as modelled through glm) as a function of per cent emotion for the two levels of action/emotion Congruency: congruent (red) vs. incongruent (cyan). Panels A, C show detection data from Experiments 1 and 3, following uncomfortable-comfortable MAMIP sessions; panels B, D show detection data from Experiments 2 and 4, following comfortable-uncomfortable MAMIP sessions. Emotions to be detected were happiness for Figs. 4A and 4B and anger for Figs. 4C and 4D.

A preliminary statistical analysis revealed that the glm procedure used to extract our indices of detection performance provided a very good fit to our dataset and was robust enough to describe the metric of responses obtained in our four experiments. In all tested conditions (Experiments 1–4) the best linear fit describing the relationship between individual predicted and individual observed yes percentages was a line with unitary slope and null intercept accounting for a large percentage of variance, as shown in Table 1.

The graphs in Fig. 4 clearly show the effect of action/emotion congruency on sensitivity. The pattern of responses is fully consistent with the facilitation-by-congruency hypothesis and at odds with both a facilitation-by-arousal hypothesis and explanations based on action valence per se. In all panels the increase of yes responses as a function of per cent emotion is well described by two positive halves of a sigmoid, with the red curve fitting the data from the block in which the valence of the target emotion was congruent with the valence of the reaches (Figs. 4A and 4B: happiness detection after comfortable reaches; Figs. 4C and 4D: anger detection after uncomfortable reaches) vs. the cyan curve fitting the data from the block in which the valence of the target emotion was incongruent with the valence of the reaches (panels A, B: happiness detection after uncomfortable reaches; panels C, D: anger detection after comfortable reaches).

Consistently with the facilitation-by-congruency hypothesis, yes percentages monotonically increased with morph intensity, with the rate of increase being larger (indicating higher sensitivity) for the red mood-congruent curve than the cyan mood-incongruent one, in all tested conditions: the yes percentage was indeed smaller in the mood-congruent (red points) than mood-incongruent (cyan points) conditions for small (determining an overall lower false alarm rate) but not large values of per cent emotion in the morph (determining an overall higher hit rate), in both anger and happiness detection tasks. This was confirmed by the results of the lme analysis revealing that the pattern of average “happy” percentages (t = 31.7, df = 158; p = 0.00; r² = 0.86, 95% CI [0.82, 0.90], r_c = 0.93, 95% CI [0.90, 0.95]) determined a significant main effect of Morph Intensity (F_3,140 = 278.4, p < 0.001), and a significant Morph Intensity × Congruency interaction (F_3,140 = 4.078, p = 0.008). The pattern of average “angry” percentages was similarly distributed (t = 31.7, df = 158; p = 0.00; r² = 0.80, 95% CI [0.74, 0.85], r_c = 0.88, 95% CI [0.85, 0.91]), though not determining a fully significant Morph Intensity × Congruency interaction (F_3,133 = 2.10, p = 0.10), probably because of the overall noisier pattern of responses: the consistency of “angry” percentages (average individual variability of yes percentages quantified by average standard error of the mean values) being worse in the anger (0.0563) than in the happiness (0.035) detection task (t = 2.867, df = 30, p = 0.007, d = 1.04). The interaction patterns rising from both distributions of yes percentages were due to the negative congruency gain (Per cent yes congruent–Per cent yes incongruent) at 0% emotion (Experiments 1 and 2, Figs. 4A and 4B: −0.125 ± 0.034, t = −3.649, df = 140, p < 0.001; Experiments 3 and 4, Figs. 4C and 4D: −0.088 ± 0.040, t = −2.221, df = 133, p = 0.028), combined with a positive or null gain for positive at 10–30% emotion.

Notably, the higher sensitivity for detecting happiness after comfortable rather than uncomfortable reaches (Experiments 1 and 2, Figs. 4A and 4B) is consistent with facilitation-by-congruency but not facilitation-by-arousal, given that an arousal-based improvement in emotional face processing should occur after the uncomfortable MAMIP session (requiring a higher level of motor activation/arousal than the comfortable). Furthermore, the pattern of yes responses depicted in Fig. 4 rules out any explanation based on action valence per se. Reaches of opposite valence led to similar improvements of emotion detection performance, consistently with action/emotion valence congruency: in Experiments 1 and 2 (Figs. 4A and 4B) happiness detection was improved after a comfortable (not uncomfortable) MAMIP session; while in Experiments 3 and 4 (Figs. 4C and 4D) anger detection was improved after an uncomfortable (not comfortable) MAMIP session. Confirming previous results (Fantoni & Gerbino, 2014), detection of facial emotions was improved by the congruency between the valence of the inducing actions and the valence of the target emotion.

Evidence from the distributions of sensitivities, thresholds and response bias

Conclusions from the previous analysis on yes percentages were corroborated by lme statistics on indices of detection performance. These quantitative analyses also allowed us answering two major questions about the facilitation-by-congruency effect induced by MAMIP: (1) is it an additive or multiplicative effect?; (2) is it a perceptual or post-perceptual effect?

Action/emotion valence congruency improves happiness detection but not response bias: against arousal

Figure 5 shows the average d′ triplets (Figs. 5A, 5E), AT (Figs. 5B, 5C, 5F, 5G) and c (Figs. 5D, 5H) for the two reaching sessions in Experiments 1 (Figs. 5A, 5B, 5C, 5D, uncomfortable ⇒ comfortable) and 2 (panels E, F, G, H, comfortable ⇒ uncomfortable). Independent of Action Ordering (panels A vs. E) the distributions of average d′ values for happiness detection as a function of per cent target emotion were in strong agreement with an additive (not multiplicative) facilitation-by-congruency effect, as confirmed by the lme analyses of Experiments 1 and 2, with participants as random effects, and emotion Congruency (comfortable action ⇒ happy expression vs. uncomfortable action ⇒ happy expression) and Morph Intensity (10, 20, 30% emotion) as fixed effects. A separate analysis of data from Experiments 1 and 2 follows.

In Experiment 1 (Fig. 5A), happiness detection increased linearly with an lme estimated rate of about 2.6 ± 0.26 d′ units every 10 per cent increment in the morph (F_1,48 = 138.56, p < 0.001), and with a constant d′ increment of about 0.84 ± 0.36 units in congruent (comfortable-happy) over incongruent (uncomfortable-happy) conditions (F_1,48 = 5.49, p = 0.023). Only 50 reaching acts distributed over 10 min, with a slightly different depth extent (average depth difference between comfortable and uncomfortable reaches = 17.56 cm ± 0.18), produced systematic changes in the detection of subtle variations of happiness. This is consistent with the idea that reaches with positive (i.e., comfortable) but not negative (i.e., uncomfortable) valence pre-activate detectors selectively tuned to emotional facial features with the same valence (i.e., happiness), which in turn facilitate the performance in Congruent relative to Incongruent conditions.

Remarkably, no improvement of fit was found (${\chi }_1^2=0.18$, p = 0.67) with a separate slope lme including the Morph Intensity × Congruency interaction (F_3,56 = 43.44, p < 0.001, r² = 0.74, r_c = 0.85, 95% CI [0.76, 0.90]), relative to an equal slope lme not including the interaction (F_2,57 = 66.05, p < 0.001, r² = 0.74, r_c = 0.85, 95% CI [0.76, 0.90]). This is a proof that the facilitation-by-congruency effect produced by MAMIP on happiness detection followed an additive rather than multiplicative trend in the d′ by morph intensity space.

This is corroborated by the post-hoc analyses (one tailed paired t test) showing that any 10% increment in morph intensity (i.e., from 10 to 20 or from 20 to 30) produced a significant d′ gain (see Table 2). These d′ gains were similar in magnitude in incongruent-uncomfortable (t = 0.78, df = 9, p = 0.45, d = 0.52) and congruent-comfortable conditions (t = 0.78, df = 9, p = 0.45, d = 0.52).

Table 2:

Summary of post hoc analyses for Experiments 1 and 2.

		Per cent happiness range	Mean d′ gain and SEM	Paired t test
Experiment 1	Incongruent	[10%–20%]	2.41 ± 0.65	t = 3.80, df = 9, p = 0.004, d = 2.53
		[20%–30%]	2.57 ± 0.73	t = 3.94, df = 9, p = 0.003, d = 2.62
	Congruent	[10%–20%]	2.77 ± 0.75	t = 4.11, df = 9, p = 0.003, d = 2.76
		[20%–30%]	2.58 ± 0.81	t = 3.62, df = 9, p = 0.005, d = 2.23
Experiment 2	Incongruent	[10%–20%]	3.37 ± 0.82	t = 4.51, df = 9, p = 0.001, d = 3.00
		[20%–30%]	2.15 ± 0.81	t = 2.86, df = 9, p = 0.019, d = 1.90
	Congruent	[10%–20%]	3.47 ± 0.72	t = 5.32, df = 9, p < 0.001, d = 3.55
		[20%–30%]	2.06 ± 0.77	t = 3.09, df = 9, p = 0.013, d = 2.06

DOI: 10.7717/peerj.1677/table-2

The additive facilitation produced by MAMIP on happiness detection performance is further corroborated by the significant facilitation-by-congruency effect on: (1) global happiness sensitivity (F_1,9 = 9.46, p = 0.01, d = 1.94), which was larger after the comfortable (global d′ = 1.76) than uncomfortable (global d′ = 1.20) reaching session, 95% CI [0.15, 0.97]; (2) the absolute threshold for happiness (F_1,9 = 5.79, p = 0.03, d = −1.52), given that, as shown in Fig. 5C, the AT value was smaller after participants completed a comfortable (AT = 4.6 ± 1.2%) than uncomfortable (8.1 ± 1.4%) MAMIP session, 95% CI [0.21, 6.77].

As shown in Fig. 5D, the response bias was clearly influenced by the unbalanced [S + N]/[N] trial ratio. As expected, the 3:1 ratio produced an overall negative bias; i.e., a dominance of “Emotion present” over “Emotion absent” responses in both Congruent (−0.31 ± 0.13, t = −2.714, df = 15.04, p = 0.016, d = −1.40) and Incongruent conditions (−0.52 ± 0.10, t = −4.603, df = 15.04, p < 0.001, d = −2.37). However, unlike sensitivity, the bias was independent of action valence (comfortable vs. uncomfortable). To control for the perceptual vs. post-perceptual locus of the MAMIP effect, according to G2, the same lme analysis conducted on perceptual indices of happiness detection (global d′ and AT) was conducted on response criterion c. The MAMIP effect was not replicated. The main effect of Congruency on c was non significant (F_1,9 = 3.0, p = 0.12).

As depicted in Figs. 5E, 5F, 5G and 5H, the facilitation-by-congruency effect was strikingly similar in Experiment 2 (uncomfortable ⇒ comfortable). Again, the pattern of happiness detection performance shown in Fig. 5E was optimally accounted for by an equal slope lme (F_2,57 = 71.85, p < 0.001, r² = 0.79, r_c = 0.88, 95% CI [0.81, 0.92]), consistent with an additive facilitation-by-congruency effect induced by MAMIP, not by a separate slope lme (F_3,56 = 47.06, p < 0.001, r² = 0.79, r_c = 0.88, 95% CI [0.81, 0.92]), consistent with a multiplicative facilitation by congruency effect induced by MAMIP (${\chi }_1^2=0.0$, p = 0.99). The lme analysis revealed a similar, though not significant (F_1,48 = 1.46, p = 0.23), constant increment of sensitivity due to action/emotion valence congruency (0.23 ± 0.30), and a similar linear modulation of sensitivity by the per cent happiness in the morph (β₂ = 2.76 ± 0.20, F_1,48 = 179.96, p < 0.001). This was confirmed by post-hoc paired t-tests. As in Experiment 1, any 10 per cent increment in morph intensity (i.e., from 10 to 20 or from 20 to 30) produced a significant d′ gain (see Table 2), with the d′ gains being similar in magnitude in incongruent-uncomfortable (t = 0.78, df = 9, p = 0.45, d = 0.52) and congruent-comfortable conditions (t = 0.78, df = 9, p = 0.45, d = 0.52). Finally the average difference between d′ values for congruent and incongruent conditions was almost constant at increasing per cent happiness in the morph, with the performance gain due to congruency for 10, 20, and 30 per cent happiness in the morph being equal to: 0.20 ± 0.14 (t = 1.81, df = 9, p = 0.09, d = 1.20), 0.29 ± 0.087 (t = 3.83, df = 9, p = 0.004, d = 2.55), and 0.20 ± 0.088 (t = 2.68, df = 9, p = 0.025, d = 2.68), respectively.

The different facilitation-by-congruency effect sizes revealed by the lme analyses in Experiments 1 and 2 were likely due to the unbalanced temporal ordering of reaching sessions. Participants in Experiment 2 were indeed less experienced with both the detection task and the face set after the comfortable (congruent) than uncomfortable (incongruent) MAMIP session, given that the comfortable condition occurred first.

Surprisingly, despite the fact that the effects of action/emotion valence congruency and learning were in opposite directions in Experiment 2, thus reducing the performance difference induced by the two reaching sessions, we found the same significant facilitation-by-congruency effect observed in Experiment 1 on both global d′ (F_1,9 = 8.09, p = 0.02, d = 1.79) and AT (F_1,9 = 8.45, p = 0.02, d = 1.84), even in the absence of significant shifts in response bias c (F_1,9 = 0.96, p = 0.37). As shown in Figs. 5F and 5G, action/emotion valence congruency increased global d′ by about 0.24 d′ units, 95% CI [0.05, 0.43], and decreased AT by about 0.83 per cent emotion, 95% CI [0.05, 0.43]. Again, similar, though negative, c values were obtained in congruent (−0.32 ± 0.10, t = −2.85, df = 11.52, p = 0.015, d = −1.68) and incongruent (−0.40 ± 0.12, t = −3.53, df = 15.04, p = 0.004, d = −2.10) conditions (Fig. 5H), with an average response bias change due to congruency of about (−0.077, 95% CI [−0.26, 0.10]).

Is there an effect of Action Ordering on happiness detection?

To test the possible role of Action Ordering we compared the patterns of d′, AT and c in Experiment 2 directly to those of Experiment 1, including Experiment as an additional fixed effect in lme analyses.

In a first lme analysis we thus asked how the relationship between individual d′ values and morph intensities were affected by Congruency and/or Action Ordering. In this model (t = 19.66, df = 118, p = 0.00, r² = 0.77 95% CI [0.68, 0.83], rc = 0.86, 95% CI [0.81, 0.90]) d′ resulted to be positively affected by Morph Intensity (β₂ = 0.27 ± 0.015, F_1,98 = 315.19, p < 0.001) and Congruency (estimated d′ gain = 0.53 ± 0.24, F_{1, 98} = 4.73, p = 0.03): other main effects or interactions were not statistically significant (${\chi }_5^2=3.08$, p = 0.68), with the intercept of the equal slope lme model, for the mood congruent condition, being negative (β₁ = −1.49 ± 0.37, df = 113.56, t = −4.02, p < 0.001, d = 0.754). A second lme model on global d′ (t = 10.60, df = 38, p = 0.00, r² = 0.75, 95% CI [0.57, 0.83], rc = 0.79, 95% CI [0.68, 0.86]) revealed a significant main effect of facilitation by congruency of about 0.4 ± 0.1 d′ units (F_1,18 = 15.94, p < 0.001); while neither the effect of Experiment (F_1,18 = 0.0020, p = 0.96) nor the Congruency × Experiment interaction (F_1,18 = 2.49, p = 0.13) were significant. Similar results were obtained on ATs, given that facilitation-by-congruency, again, resulted to be the only significant factor affecting the performance (estimated AT after comfortable-congruent = 5.24 ± 0.82 per cent emotion; estimated AT after uncomfortable-incongruent = 7.40 ± 0.84 per cent emotion; F_1,18 = 8.51, p = 0.001).

Does MAMIP affect response bias in happiness detection?

In a final lme analysis, we asked whether these effects occur at the level of response or stimulus encoding. Following Signal Detection Theory (Macmillan & Creelman, 2004) a pure perceptual effect is supported by the full independence between results on sensitivity and response bias. Specifically, the effect of congruency measured on the perceptual indices of performance (d′ and AT), should result to be absent when measured on the decision index of performance (c). This is what we found, as an lme model on c revealed no significant main or interaction effects: no significant decrement of fit was indeed found when contrasting a full lme model including Congruency, Experiment, and their interaction as fixed effects with a baseline lme model with no fixed effects (r_c slightly decreases from 0.83, 95% CI [0.74, 0.89] to 0.76, 95% CI [0.65, 0.84]; ${\chi }_3^2=4.8$, p = 0.18).

In the present investigation, therefore, there is no evidence that learning and arousal contribute to the perceptual response beyond what emotion intensity and the congruency between action and emotion valence can explain.

Action/emotion valence congruency improves anger detection but does not modify response bias: against valence per se

The results of Experiments 1 and 2 corroborated the idea that the facilitation-by-congruency effect induced by MAMIP was additive, robust, independent of learning and arousal, and perceptually-based: the processing of positive emotional features gets more salient after comfortable (congruent) then comfortable (incongruent) reaches. However, the result cannot be generalized to the entire affect domain, being specific for positive emotions (i.e., happiness), and cannot be interpreted univocally given that, in principle, it might have been produced by action valence per se: with happiness detection being facilitated by a comfortable action sequence, regardless of the correspondence between the valence of the mood induced by the action sequence and the emotion to be detected. If the MAMIP effect is general and independent from action valence then anger detection performance in Experiments 3 and 4 should be facilitated by uncomfortable actions, thus producing an action/emotion congruency effect similar to the one observed in Experiments 1 and 2.

This expectation closely matches results of Experiments 3 (uncomfortable ⇒ comfortable) and 4 (comfortable ⇒ uncomfortable) shown in Fig. 6. The distributions of average d′ resulting from each unique combination of morph intensity (10, 20, 30%), emotion and reaching sessions shown in Figs. 6A and 6E are in strong agreement with an additive (not multiplicative) facilitation-by-congruency effect: in both action ordering conditions (uncomfortable ⇒ comfortable in Fig. 6A; comfortable ⇒ uncomfortable in Fig. 6E) d′ increases linearly as a function of per cent anger in the morph (coded on the x-axis), and is aided by congruency (coded by colors with congruent-uncomfortable in red, and incongruent-comfortable in cyan) by an almost constant amount at increasing per cent anger in the morph values.

This observation was confirmed by an lme analysis testing the effects of Morph Intensity, Action Ordering, and action/emotion Congruency on d′ triplets. The analysis revealed that Morph Intensity (β₂ = 0.181 ± 0.0175; F_1,93 = 107.25, p = 0.00), and action/emotion Congruency (estimated d′ gain due to congruency = 0.648 ± 0.29; F_1,93 = 5.14, p = 0.024) were the only factors affecting anger sensitivity (t = 14.30, df = 112, p = 0.00, r² = 0.64, 95% CI [0.53, 0.74], rc = 0.77, 95% CI [0.69, 0.83]). No further main effects or interactions were observed (${\chi }_5^2=4.74$, p = 0.45). This equal slope lme model best fitting our data set had a negative intercept in the mood congruent condition (β₁ = −1.35 ± 0.47; df = 88.25, t = −2.90, p = 0.005, d = 0.64). As shown in panels Figs. 6A and 6E, anger detection sensitivities increased linearly with a similar lme estimated rate in Experiments 3 and 4 of about 1.67 ± 0.24 (Fig. 6A) and 1.98 ± 0.35 (Fig. 6E) respectively, every 10 per cent increment in the neutral-to-angry morph continuum (F_2,85 = 0.401, p = 0.67), with an average d′ gain due to congruency that remained constant at increasing morph intensities in a direction consistent with an additive facilitation-by-congruency hypothesis in both experiments (F_1,85 = 0.174, p = 0.677): the estimated d′ gain due to congruency measuring about 0.77 ± 0.31 units in Experiment 3 (Fig. 6A, t vs 0 = 2.48, df = 8, p = 0.037, d = 1.76) and 0.58 ± 0.22 units in Experiment 4 (Fig. 6E, t vs 0 = 2.38, df = 9, p = 0.041, d = 1.59).

We further confirmed the direct increase of d′ as a function of morph intensity through congruent and incongruent conditions by post-hoc paired one tailed t-tests. As summarized in Table 3, in Experiment 3, as anger increased from 10 to 20% d′ significantly increased by about 0.83 ± 0.44 d′ units and 1.47 ± 0.65 d′ units, after comfortable-incongruent and uncomfortable-incongruent reaches respectively; similar d′ increments due to a 20 to 30% anger in the morph growth were observed after comfortable-incongruent (t vs 0.83 ± 0.44 = 1.87, df = 9, p = 0.093, d = 1.24) and uncomfortable-congruent reaches (t vs 1.47 ± 0.65 = 1.10, df = 9, p = 0.30, d = 0.73). Post-hoc tests on d′ values from Experiment 4 were strikingly similar, confirming that, independent of Action Ordering, Congruency did not affect the rate of increase of d′ over morph intensity, which is consistent with a rather general additive (not multiplicative) facilitation-by-congruency effect. As by in Experiment 3, also in Experiment 4 any 10% increment in the morph intensity (i.e., from 10 to 20%, or from 20 to 30%) produced a significant d′ gain (see Table 3). Furthermore, a similar d′ gain was induced for 10 to 20% and a 20 to 30% anger increments in the morph by both the congruent-uncomfortable (t = 1.81, df = 8, p = 0.11, d = 1.28) and the incongruent-comfortable (t = 0.79, df = 8, p = 0.45, d = 0.55) sequence of reaches.

Table 3:

Summary of post hoc analyses for Experiments 3 and 4.

		Per cent anger range	Mean d′ gain and SEM	Paired t test
Experiment 3	Incongruent	[10%–20%]	0.83 ± 0.44	t = 2.29, df = 9, p = 0.048, d = 1.52
		[20%–30%]	2.18 ± 0.78	t = 3.21, df = 9, p = 0.011, d = 2.14
	Congruent	[10%–20%]	1.47 ± 0.65	t = 2.69, df = 9, p = 0.025, d = 1.79
		[20%–30%]	2.17 ± 0.86	t = 2.93, df = 9, p = 0.017, d = 1.95
Experiment 4	Incongruent	[10%–20%]	1.55 ± 0.75	t = 2.53, df = 8, p = 0.035, d = 1.79
		[20%–30%]	1.71 ± 0.80	t = 2.59, df = 8, p = 0.032, d = 1.83
	Congruent	[10%–20%]	1.40 ± 0.68	t = 2.49, df = 8, p = 0.037, d = 1.76
		[20%–30%]	3.25 ± 0.79	t = 4.63, df = 8, p = 0.002, d = 3.27

DOI: 10.7717/peerj.1677/table-3

The additive facilitation produced by MAMIP on anger detection performance in both action ordering conditions produced a pattern of global sensitivity enhancement and absolute threshold decrements due to congruency that closely resembles the one observed in Experiments 1 and 2. Global d′ after the uncomfortable (congruent) reaching session outperformed those after the incongruent-comfortable reaching session by: 0.24 d′ units, 95% CI [−0.02, 0.50] (estimated global d′ = 0.55 and 0.79, after comfortable and uncomfortable reaching sessions, respectively), in Experiment 3 (F_1,9 = 4.18, p = 0.05, d = 1.3), and 0.35 d′ units, 95% CI [−0.01, 0.71] (estimated global d′ = 0.78 and 1.13, after comfortable and uncomfortable reaching sessions, respectively), in Experiment 4 (F_1,9 = 8.10, p = 0.02, d = 1.8). These increments in sensitivity due to congruency were statistically similar across action ordering conditions as confirmed by an lme model including the Experiment as fixed effect (t = 13.82, df = 36, p = 0.00, r² = 0.84, 95% CI [0.71, 0.91], r_c = 0.88, 95% CI [0.81, 0.93]). The model indeed revealed a significant main effect of facilitation by congruency of about 0.29 ± 0.09 d′ units (F_1,17 = 9.3, p = 0.007), but not of the Experiment (F_1,17 = 1.90, p = 0.19), and not of the Congruency × Experiment interaction (F_1,17 = 0.35, p = 0.56).

A similar result was obtained on individual per cent anger value above which a difference between the original neutral faces and the faces morphed with anger gets just noticeable; i.e., AT (shown in Figs. 6B and 6F for Experiments 3 and 4, respectively). Again, the lme analysis (t = 6.63, df = 36, p = 0.00, r² = 0.55, 95% CI [0.31, 0.63], r_c = 0.61, 95% CI [0.45, 0.74]), revealed that the only significant factor affecting ATs across experiments was Congruency (F_1,17 = 12.35, p = 0.003), which produced an amount of AT decrement that was similar, though significant, in both Experiment 3 (7.1, 95% CI [1.15, 13.03] per cent anger in the morph; F_1,9 = 7.27, p = 0.02, d = 1.7, Fig. 6C) and Experiment 4 (7.75, 95% CI [0.00, 15.50] per cent anger in the morph; F_1,8 = 5.6, p = 0.03, d = 1.57, Fig. 6G).

Does MAMIP affect response bias in anger detection?

A final lme analysis on response bias c demonstrated that the above discussed effects of MAMIP on anger detection performance were purely perceptual as those observed on the happiness detection task (in Experiments 1 and 2). Again, no significant decrement of fit was indeed found when contrasting a full lme model including Congruency, Experiment and their interaction as fixed effects with a baseline lme model with no fixed effects at all (r_c slightly decreases from 0.97, 95% CI [0.94, 0.98] to 0.96, 95% CI [0.92, 0.97]; ${\chi }_3^2=7.0$, p = 0.08). We confirmed this by post-hoc analyses showing that the response bias change due to congruency was statistically unreliable in both Experiment 3 (−0.16, 95% CI [−0.32, 0.01], t = −2.01, df = 9, p = 0.075; Fig. 6D) and 4 (−0.038, 95% CI [−0.22, 0.14], t = −0.48, df = 8, p = 0.64; Fig. 6H).

Symmetry of the action/emotion congruency effect

Our experiments univocally demonstrate that responses in our facial emotion detection task are, perceptually (not cognitively) determined by displayed emotion intensity, as well as by the congruency between bodily motor acts and emotion valence. The facilitation produced by emotional congruency, in addition of being independent of the arousing effect of goal-directed reaches (as demonstrated in Experiments 1 and 2), generalizes to the negative domain of affects (as demonstrated in Experiments 3 and 4) being thus also independent of action valence per se. Emotional congruency indeed facilitates the performance when the detection task is preceded by both a positive-comfortable action sequence (coupled with a positive emotion to be detected as in Experiments 1 and 2), and a negative-uncomfortable action sequence (coupled with a negative emotion to be detected as in Experiments 3 and 4).

However, a major difference between our two sets of experiments is revealed by a strong emotion detection anisotropy consistent with an overall lower global sensitivity (average global d′ of 1.49, 95% CI [1.33, 1.64] vs. 0.80, 95% CI [0.63, 0.97] in Experiments 1 & 2 vs. 3 & 4, respectively; Welch two sample t = 6.00, df = 75.11, p = 0.00, d = 1.36). This caused also a higher, though not significantly, threshold (average AT was 6.32, 95% CI [5.08, 7.56] vs. 8.49, 95% CI [5.85, 11.12] in Experiments 1–2 vs. 3–4, respectively; Welch two sample t = −1.51, df = 52.841, p = 0.13, d = 0.34) in the anger rather than in the happiness detection task. Such a bias was in line with results of the preliminary baseline experiment and with growing evidence in the emotion perception literature showing that realistic faces, as those used in the current experiments, often give rise to a happiness (rather than the more often found anger) detection advantage relative to both angry (Becker et al., 2011; Becker et al., 2012; Juth et al., 2005) and sad (Srivastava & Srinivasan, 2010) faces.

In order to reconcile our data across emotions with different valence and to test (consistently with G3) the degree of symmetry of the facilitation-by-congruency effect, we recoded the four lme relationships describing the covariation of d′ as a function of Morph Intensity and action/emotion valence Congruency in our dataset in the generic emotion detection Cartesian space introduced in Fig. 2 (with d′ for happiness detection on the y axis and d′ for anger detection on the x axis). Consistently with an additive facilitation-by-congruency effect, the best fitting lme models resulting from the average d′ by morph intensity relationships observed in Experiments 1 & 2 and Experiment 2 & 3 respectively, were both additively modulated by morph intensity and by congruency independently from action ordering. This gave rise to the following set of four lme regression lines:

1. a couple of regression lines for happiness detection with a common slope β_{2 happy} of about 0.27 and intercepts equal to −1.49, for the action/emotion comfortable congruent condition (β_{1 happy comfortable}), and −2.02, for the action/emotion uncomfortable incongruent condition (β_{1 happy uncomfortable});

2. a couple of regression lines for anger detection with a common slope β_{2 anger} of about 0.181 and intercepts equal to −1.35, for the action/emotion uncomfortable congruent condition (β_{1 anger uncomfortable}), and −2.00, for the action/emotion comfortable incongruent condition (β_{1 anger comfortable});

Entering the parameters of these four sets of lines into Eq. 1 (pairing them appropriately following the procedure discussed in the Rationale and expectation section) allows recoding them into the emotion detection space. As shown in Fig. 7 this procedures defines two parallel lines with slope β_{2 G} = 1.47: the blue solid line, standing for the average performance after comfortable reaches with β_{1 G} = 1.465, and the violet solid line, standing for the average performance after uncomfortable reaches with β_{1 G} = −0.025. Importantly, the position of the grey dot representing the average d′ in the baseline inaction condition relative to line of comfortable and uncomfortable reaches in the emotion detection space revealed an almost symmetric facilitation by congruency effect induced by MAMIP. The average y coordinate of the baseline inaction point (1.26 ± 0.159), indeed, did not deviate significantly, along the happiness dimension, from the corresponding coordinate of the point along the line bisecting the space in between the comfortable and uncomfortable lines (0.935), standing for a perfectly symmetric effect (Welch two sample t = 1.4543, df = 172, p = 0.1477, d = 0.22), while it was: (1) significantly smaller than the corresponding coordinate of the point along the comfortable line (1.68), standing for an asymmetric effect fully in charge of the uncomfortable reaches (Welch two sample t = 1.98, df = 172, p = 0.049, d = 0.32), and significantly larger than the corresponding coordinates of the point along the uncomfortable line (0.19), standing for an asymmetric effect fully in charge of the comfortable reaches (Welch two sample t = 4.75, df = 172, p < 0.001, d = 0.72).

The average coordinate of the baseline inaction point along the anger dimension (0.15 ± 0.049) instead deviates from the corresponding coordinates of all three lines: the line of perfect symmetry (Welch two sample t vs. 0.37 ± 0.049 = 3.20, df = 170, p = 0.002, d = 0.50), the line of comfortable reaches (Welch two sample t vs. − 0.14 ± 0.049 = 4.05, df = 170, p < 0.001, d = 0.62), and the line of uncomfortable reaches (Welch two sample t vs. 0.87 ± 0.049 = 10.45, df = 170, p < 0.001, d = 1.60).

Present results demonstrate that both comfortable and ucomfortable actions impact the perception of facial expression. However, while the facilitation-by-congruency effect on happiness detection is balanced across different types of actions, anger detection is facilitated by action induced mood congruency in a slight unbalanced way, with uncomfortable reaches being slightly more effective than comfortable reaches.