ابراز هیجانی به عنوان سیگنال های اجتماعی رد و پذیرش: شواهدی از پارادایم اسناد اشتباه عاطفه

Abstract Inclusion in social groups is vital to human survival and wellbeing. We propose that emotional expressions signal acceptance versus rejection to observers. Based on this idea, we hypothesized that happy facial expressions prime acceptance, whereas angry expressions prime rejection. In six experiments using the Affect Misattribution Paradigm (Payne, Cheng, Govorun, & Stewart, 2005), we tested to what extent observers associate facial expressions (angry, happy, sad, fearful, and neutral) with three different operationalizations of acceptance and rejection (accept/reject, warm/cold, close/distant). A meta-analysis on these experiments revealed that angry expressions were more strongly associated with rejection than other (negative) expressions, and that happy expressions were more strongly associated with acceptance than other facial expressions. Effects were stable and robust at presentation times of 50 ms and higher and were similar across conceptualizations of acceptance/rejection. We discuss implications for theorizing on the social functions of emotions and the processing of emotional expressions.

Results Analytic strategy A two-stage meta-analysis (Simmonds et al., 2005) was conducted to assess the overall effects of the facial expression primes and the impact of the Response Dimension and Prime Duration manipulations on the responses given by participants. In the first stage, effect sizes were estimated for each unique between-subjects condition within each study. This means that, for instance, two sets of effect sizes were obtained from Study 2, which had a 1 (Response Dimension: warm–cold) × 2 (Prime Duration: 17 ms vs 50 ms) between-subjects design: warm–cold × 17 ms; and warm–cold × 50 ms. Slicing up the data in this fashion resulted in a total of 19 datasets, which were treated as separate experiments. This approach is recommended when the effect sizes from one study can be assumed to be independent (Borenstein, Hedges, Higgins, & Rothstein, 2010). This assumption was met because between-subjects manipulation of the moderators of interest ensured that each participant contributed to only one of the effect sizes obtained from a study. Other potential sources of effect size dependency, such as the location and the experimenter, were constant across all experiments, making it unlikely that any two conditions from the same experiment would be more related than two conditions from different experiments. In the second stage of the meta-analysis, the effect sizes obtained from the first stage were aggregated, and the overall impact of the moderators was assessed. To obtain the effect size estimates in the first phase, mixed-effect logistic regression models were fitted to the individual conditions using the lme4 package (version 0.999999.2; Bates, Maechler, & Bolker, 2013) for R (version 3.0.1; R Core Team, 2013). A random intercept for participant was specified. The model estimated the following planned contrasts: 1) between the happiness condition (− 1) and the four other conditions (each coded as 1/4); 2) between the anger (1) and fear/sad conditions (each coded as − 1/2); 3) between the neutral (− 1) and the three negative emotional conditions (each coded as 1/3); and 4) between the fear (− 1) and sad (1) conditions. In those studies that omitted either the fear, or the sad condition (see Table 1), only the first three contrasts were fitted, and the contrast weights were adjusted (e.g., in contrast 1, all conditions except happiness received weights of 1/3 instead of 1/4). Note that contrasts 3 and 4 were specified in this way to obtain an orthogonal set of contrasts. Although no hypotheses were formulated about these contrasts, the fact that they are part of the same model (and therefore adjusted for while estimating contrasts 1 and 2) necessitates us to report them in order to provide a complete picture of the results. In the second stage of the meta-analysis, a random-effects meta-analytic model was fit to the regression coefficients and standard deviations obtained from the first step using the metafor package (version 1.9.1; Viechtbauer, 2010) for R.2 Using the random-effects model instead of a traditional fixed-effects model has the advantage that heterogeneity among effect sizes is allowed, thereby reducing the influence of any within-experiment dependency of effect sizes. The results of the meta-analysis are reported as odds ratios (ORs), followed by the associated 95% confidence intervals. Conventional p-values are also reported where applicable. Note on figures To ensure correct interpretation of the figures (forest plots), some explanation is in order. The figures are divided in two parts. The top part contains the results for each of the individual datasets for which an effect size was estimated. The label (left column, e.g., ‘accept/reject 17 ms (exp 1)’) indicates the Response Dimension (acceptance/rejection) and Prime Duration (17 ms) condition, and the experiment that this dataset originates from (Experiment 1). The right column presents the effect sizes (ORs) and their 95% confidence intervals numerically. In the center, these effect sizes and confidence intervals are graphically presented relative to a reference line set at an OR of 1 (which indicates no difference). Larger blocks denote that the dataset contained more participants, and therefore received more weight in the meta-analysis. Finally, the bottom part presents the meta-analytic summary. If there is no moderation, the overall ‘main effect’ is presented; if there is moderation, the summary is split out for each level of the moderator. Participants and data cleaning Table 2 summarizes the total number of participants in each experiment, and the results of the following checks that ensured data quality. First, participants were excluded from the sample if they could read Chinese, or if they used one of the two response keys in over 90% of the trials. Then, outliers were identified by first determining the distribution of log latencies in each condition within each experiment, and then removing responses with log latencies that deviated more than three standard deviations from the mean log latency in the condition. Finally, participants whose responses were outliers in more than 10% of the cases were dropped from the sample. Without dropping outliers, the analyses produce similar effect estimates as reported below, but with wider confidence intervals. Significant effects remain significant, with the exception of contrast 2, which becomes marginal. Hypothesis 1. Happy facial expressions and acceptance According to our first hypothesis, happy facial expressions should be more strongly associated with acceptance than other facial expressions. For this hypothesis to be supported, the first contrast should show an OR smaller than 1, which indicates that the likelihood of a rejection-related response is lower (i.e., the likelihood of the opposite acceptance-related response is greater) after a happy facial expression prime than after another facial expression prime. An initial random-effects model without any moderators showed the predicted effect, OR = 0.735 [0.648, 0.835], p < .001. Remaining heterogeneity (Qw[18] = 204.17, p < .001) suggested potential moderators, however, so we examined the influence of both Response Dimension and Prime Duration. The impact of Response Dimension on the contrast between happy and other facial expressions was found to be marginally significant, Qdimension(2) = 5.80, p = .055. Exploring the pattern of moderation indicated that the contrast between happy and other facial expressions was slightly more pronounced for the accept/reject Response Dimension (OR = 0.622 [0.516; 0.750]) than for the warm/cold (OR = 0.769 [0.635; 0.933]) or close/distant (OR = 0.887 [0.704; 1.118]) response dimensions. Because of the weak evidence for moderation, and because the simple effect was in the predicted direction for each Response Dimension, we did not consider the influence of Response Dimension further. For Prime Duration, on the other hand, we found much stronger evidence that it affected the relative likelihood of rejection responses after happy facial expression primes compared to other facial expressions, Qprime(3) = 33.96, p < .001. The results of this analysis (summarized in the bottom part of Fig. 2) indicate that happy facial expressions did not affect the likelihood of a rejection response at Prime Durations of 17 ms (OR = 1.004 [0.847, 1.190]) and 33 ms (OR = 1.131 [0.868, 1.473]), but the hypothesized decrease in likelihood of a rejection-related responses for happy, relative to other facial expressions was found at Prime Durations of 50 ms (OR = 0.620 [0.551, 0.698]) and 67 ms (OR = 0.647 [0.559, 0.748]). Accordingly, inclusion of Prime Duration as a moderator decreased the observed heterogeneity, although not to non-significance: Qw(15) = 56.65, p < .001. In sum, our first hypothesis, that happy facial expressions would be more strongly associated with acceptance than other facial expressions, was supported for Prime Durations of 50 ms and up. Hypothesis 2. Angry facial expressions and rejection. Forest plot of odds ratios for the contrast between happy versus neutral, angry, ... Fig. 2. Forest plot of odds ratios for the contrast between happy versus neutral, angry, fearful, and sad facial expressions and a meta-analytic summary. Figure options Our second hypothesis was that anger would be more strongly associated with rejection than other negative facial expressions. For this hypothesis to be corroborated, the OR of the second contrast should be higher than 1, which indicates that the likelihood of a rejection-related response is greater after angry facial expression primes than after sad and fearful facial expression primes. The initial random-effects without moderators showed the predicted effect (see Fig. 3 for a summary) to be small but reliable, OR = 1.033 [1.001, 1.066], p = .043. There was little evidence of heterogeneity among the effect sizes (Qw[18] = 8.96, p = .961), which suggests that the effects were comparable across conditions. Thus, no further moderation analyses were conducted. In sum, our second hypothesis, that angry facial expressions would be more strongly associated with rejection than other negative facial expressions, was supported. Forest plot of odds ratios for the contrast between angry versus fearful and sad ... Fig. 3. Forest plot of odds ratios for the contrast between angry versus fearful and sad facial expressions and the meta-analytic summary. Figure options Remaining contrasts The final two contrasts that were specified in the models were also subjected to the meta-analytic procedure. Note that no hypotheses were specified for these contrasts. They are reported to provide a complete picture of the results. The relevant forest plots are available as Supplementary materials online. The third contrast tested all negative emotional facial expressions (i.e., angry, fearful, and sad facial expressions) against neutral facial expressions. A model without moderators indicated that negative emotional facial expressions increased the likelihood of a rejection response relative to neutral facial expressions, OR = 1.073 [1.019, 1.129], p = .007. There was substantial heterogeneity among effect sizes, Qw(18) = 33.74, p = .014. Only Prime Duration significantly moderated the contrast, Qprime(3) = 18.46, p < .001. No effects were found at Prime Durations of 17 ms (OR = 0.968 [0.894, 1.048]) and 33 ms (OR = 0.875 [0.740, 1.035]). At Prime Durations of 50 ms (OR = 1.125 [1.067, 1.187]) and 67 ms (OR = 1.140 [1.071, 1.214]), however, negative facial expressions increased the likelihood of rejection-related responses compared to neutral facial expressions. The remaining heterogeneity was non-significant, Qw(15) = 15.28, p = .431. The results for the fourth contrast did not show a difference between the likelihood of a rejection-related response after sad facial expressions, compared to fearful facial expressions (OR = 1.013 [0.953, 1.077], p = .669), with little evidence of heterogeneity among effect sizes (Qw[8] = 9.91, p = .272). Auxiliary analyses In addition to these hypotheses tests, we explored whether the effects described above were moderated by participant sex. We found that it did not moderate any of the contrasts. Then, we explored whether response latencies were affected by the facial expression primes, by running the same analyses as described above, but substituting response for response latency as the dependent variable. Happy facial expressions were responded to slightly faster than all other facial expressions (β = − 0.019 [− 0.037, − 0.002]), but no other differences were found. We are reluctant to interpret this effect, because it is small and does not relate in a meaningful way to the hypotheses.