The sensitivity of third party punishment to the framing effect and its brain mechanism

People as third-party observers, without direct self-interest, may punish norm violators to maintain social norms. However, third-party judgment and the follow-up punishment might be susceptible to the way we frame (i.e., verbally describe) a norm violation. We conducted a behavioral and a neuroimaging experiment to investigate the above phenomenon, which we call “third-party framing effect.” In these experiments, participants observed an anonymous player A decided whether to retain her/his economic benefit while exposing player B to a risk of physical pain (described as “harming others” in one condition and “not helping others” in the other condition), then they had a chance to punish player A at their own cost. Participants were more willing to execute third-party punishment under the harm frame compared to the help frame, manifesting as a framing effect. Self-reported moral outrage toward player A mediated the relationship between empathy toward player B and the framing effect size. Correspondingly, the insula (possibly related to empathy) and cerebellum (possibly related to anger) were activated more strongly under the harm frame than the help frame. Functional connectivity between these regions showed strongest weight when predicting the framing effect size. These findings shed light on the psychological and neural mechanisms of the third-party framing effect. Graphic abstract


25
Protests organized by feminists were once often described as "catfights" by American 26 media in the 1960s; as the gender equality movement develops, similar events have 27 been reporting in more favorable words over time (Ashley & Olson, 1998). The above 28 example showcases an attempt to manipulate third-party judgment with framing 29 techniques, which is prevalent in our daily life (e.g., in lawsuits, journalism, and Buckholtz & Marois, 2012), investigating its susceptibility to different frames (i.e., 38 verbal statements) and the associated neural underpinnings would be meaningful. 39 Generally speaking, "framing effect" refers to the phenomenon that our judgment 40 of an object or issue is affected by its description (Tversky & Kahneman, 1981). As 41 pointed out by De Martino et al. (2006), the framing effect could be interpreted as an hippocampus, cingulate cortex, and basal ganglia (Fumagalli & Priori, 2012). 55 Recently, the cerebellum is also suggested to be involved in the processing of  Accordingly, we investigate the changes of these brain networks in response to 61 the framing effect in two (behavioral and brain-imaging) experiments. Specifically, 62 participants observed an anonymous player made a trade-off between income 63 maximization and helping other persons to avoid a painful shock. Choosing economic 64 benefit was described as either a "harm" to, or "not helping," other persons in two 65 frame conditions. Then participants could decide whether to punish that player from a 66 third-party perspective. We hypothesized that participants would make more costly 67 punishments in the harm frame condition than in the help frame condition (see also 68 69 "helping others" (Crockett et al., 2014). Correspondingly, the brain regions involved 70 in empathic response and/or moral outrage should activate more strongly in the harm 71 frame condition. We also examined the relationship between behavioral data and 72 functional connectivity of these regions across frame conditions, so as to improve the 73 knowledge about how third-party punishment is modulated by the framing effect.  Our task design combined the "social framing" task with third-party punishment. The  or keeps all the payoff while player B receives that shock. Player A determines the 94 occurrence probability (10%, 30%, 50%, 70%, or 90%) of these two outcomes by 95 moving an avatar on a 5-point scale. The two outcomes are described in different 96 ways in two conditions: in the help frame condition, the costly helping outcome was 97 described as "help the other person to avoid a shock," while the other outcome was 98 described as "do not help the other person;" in the harm frame condition, the costly 99 helping outcome was described as "give the other person a shock," while the other 100 outcome was described as "do not give any shock."

101
In this study, all the participants acted as a third-party observer (player C) and 102 received 5 MUs in each trial of the task. After they observed each of player A's choice, 103 participants were given a chance to punish player A by spending their own 5 MUs, 104 such that each MU spent would reduce 2 MUs from player A's final payoff (Fig. 1A).
Insert Figure 1 approximately here 115 116 The present experiment employed a 2 (frame: harm versus help) × 5 (moral level 117 of player A's choice ["moral level" for short]) within-subject design. Here, the five 118 levels of the moral level factor were: highly pro-helping (90% probability of costing 119 own money to save player B from a shock), medium pro-helping (70% probability of 120 costing own money), neutral (50%), medium non-helping (70% probability of keeping 121 own money), and highly non-helping (90% probability of keeping own money), 122 respectively. Each condition (2 × 5) repeated twice, resulting in 20 trials through the 123 task.

124
In each trial, the real participant (player C) first observed two possible outcomes  The order of these three questions was counterbalanced across participants. The whole 134 task lasted for approximately 5 mins.

136
As mentioned in the Introduction, we focused on whether different frames would bias  comparison showed that when player A was highly pro-helping, participants' 146 punishment was not significantly different between two frames (harm frame: 0.09 ± 147 0.03; help frame: 0.08 ± 0.03, p = 0.53); in other four conditions (medium pro-helping, 148 neutral, medium non-helping, and highly non-helping), participants punished more 149 under the harm frame than the help frame (ps < 0.015) ( Fig. 2A). 150 We then calculated the difference of participants' punishment between the harm 151 frame and the help frame as an index of third-party framing effect size. Pearson's 152 correlation analysis revealed that the framing effect sizes in different conditions were 153 significantly correlated with each other (ps < 0.002) except when player A was highly 154 pro-helping (ps > 0.126).

155
Overall, these results revealed a significant third-party framing effect (i.e., 156 altruistic punishment was increased under the harm frame than the help frame) except the difference in "empathic feeling for player B" as well as "moral outrage against 162 player A" between the two frames, in the same way as we calculated the third-party 163 framing effect (i.e., harm frame minus help frame). One sample t-tests revealed that 164 these three indexes were all significantly larger than zero (framing effect size: 0.23 ± 165 0.04, t(97) = 6.13, p < 0.001; empathic feeling: 0.28 ± 0.04, t(97) = 7.07, p < 0.001; 166 moral outrage: 0.32 ± 0.04, t(97) = 8.43, p < 0.001). Pearson's correlation analysis 167 revealed that these three indexes were all significantly correlated with one another (rs > 168 0.63, ps < 0.001). 169 We then run a mediation analysis with the difference in empathic feeling as X,   To determine an appropriate sample size, a priori power analysis was conducted using  Experimental design and procedures 198 The task design was the same as in Experiment 1, but the experimental setting was 199 adjusted for fMRI scanning. Before the scanning, participants were familiarized with 200 the task with a practice block consisting of 8 trials. In each trial, the real participant 201 (player C) first observed two possible outcomes in different frames for 2 s. After that, 202 a blue avatar moved on a 5-point scale to show player A's decision process for 2 s.

203
Each participant then waited for another 4 s before they could press one of two 204 pre-assigned buttons on an MRI-compatible button-box to indicate how many MUs 205 (0~5) s/he would like to pay for punishing player A. The participant had 2 s to choose 206 the preferred number (turned from black to red), the starting point of which was 207 randomized across trials. Finally, the inter-stimulus interval was set as 1~4 s ( Fig 1B).

208
Each condition (2 × 5) contained 24 trials and there were 240 trials in total. The 209 experiment, which consisted of four runs of 60 trials, lasted for approximately 1 h.

210
Neuroimaging data acquisition and preprocessing 211 We used a Siemens TrioTim 3.0T MRI machine for data acquisition. Functional   Since we are most interested in the third-party framing effect, we combined four 236 moral levels (medium pro-helping, neutral, medium non-helping, and highly 237 non-helping) to focus on the main effect of frames; as mentioned above, the highly 238 pro-helping condition was excluded because no framing effect was found in this 239 condition. For the group-level analysis, we conducted a one-sample t-test using the 240 whole brain as the volume of interest to localize the differences in brain activity 241 between the harm frame and the help frame. Besides, a regression analysis was also 242 conducted to explore which brain areas were activated stronger in the contrast of 243 harm minus help frame, as a function of the third-party framing effect.

244
Task-dependent FC for predicting the third-party framing effect 245 Regarding the brain network that responded to the third-party framing effect, we 246 tested whether there was sufficient information in the connectivity patterns within this 247 network to predict individual difference in behavioral decisions. To do this, we first 248 defined the regions of interest (ROIs) as brain areas of which the activation level: (1) 249 was significantly higher in the contrast of harm minus help frame, and (2) covaried 250 with the behavioral framing effect size. Five brain regions were found to meet the 251 above two criteria (see the Results section). We then estimated the FC between each 252 pair of the ROIs using psychophysiological interactions (PPI). We then drew spheres 253 (radius = 6 mm) at the coordinates of all the five significant peaks of activity (see 254   Table 3) localized in the overlap between the T-contrast map of harm > help and the 255 regression map that covaried with the framing effect size.  We calculated whole-brain task-dependent FC maps with all the five ROIs as 266 seed regions and then extracted parameter estimates of each FC map within each ROI, 267 so as to obtain a matrix that represented the FC strength between each pair of ROIs 268 for each participant.

269
Linear relevance vector regression (RVR) 270 We selected the RVR due to its high prediction performance in brain-behavior/  Leave-one-out-cross-validation (LOOCV) was used to calculate the prediction 277 accuracy (i.e., Pearson correlation coefficient between the predicted and actual labels).

278
For each round of LOOCV, one participant was designated as the test sample and the 279 remaining participants were used to train the model. The predicted score was then 280 obtained from the feature matrix of the tested sample.

281
The significance level was computed based on 1000 permutation tests. For each 282 permutation test, the prediction labels (i.e., participant's third-party framing effect) 283 were randomized, and the same RVR prediction process used for the actual data was To determine whether our major findings were sensitive to our choices of correlation 296 thresholds for connectivity, we recomputed the FC with different thresholds (0.1) and 297 then re-performed the machine learning analysis.

ROI analysis 299
As we found in the machine learning-based FC analysis, the FC between the left 300 insula and cerebellum was the strongest predictor of the third-party framing effect. We  The ROI time series were extracted from within the whole-brain activation for the 324 harm frame and help frame conditions.

325
To determine the driving input (matrix C) and modulation effect (matrix B), we 326 fixed the intrinsic connection between the two regions as bilateral connections. Our

338
Behavioral results 339 We performed a 2 × 5 repeated measures ANOVA on altruistic punishment and found medium non-helping, and highly non-helping), participants punished more under the 347 harm frame than the help frame (ps < 0.008) (Fig. 3A). As in Experiment 1, these 348 results confirmed the third-party framing effect in four conditions. Accordingly, we 349 then used the averaged of these conditions to index each participant's third-party 350 framing effect size, as we did in Experiment 1 (Fig. 3B).  In this series of experiments, we asked participants to observe two putative 405 players interacted. Specifically, player A chose between "harming" and "not harming" 406 player B in the harm frame condition, and between "helping" and "not helping" 407 player B in the help frame condition. These two conditions were objectively but not 408 verbally equivalent, conforming to the definition of frame manipulation (Rabin, 1998).

409
In Experiment 1, we first examined whether third-party punishment could be 410 modulated by the framing effect at the behavioral level. The results confirm our 411 hypothesis, such that the participants were more willing to punish player A (at the cost 412 of their own benefits) in the harm frame condition than in the help frame condition, 413 resulting in a third-party framing effect. In our opinion, this effect reflected that 414 "harming others" was judged by the participants as a more serious norm violation 415 than "not helping others" (Crockett et al., 2014). The framing effect was robust unless 416 player A was highly pro-helping, which is reasonable because it would be 417 inappropriate to punish player A in this situation. We then calculated the averaged 418 framing effect across different moral levels except "highly pro-helping" and found 419 that it was significantly larger than zero. Likewise, self-reported "empathic feeling to 420 player B" and "moral outrage to player A" were both larger in the harm frame 421 condition than in the help frame condition. Moreover, these three behavioral indexes 422 were significantly correlated between each other. To determine the relationship 423 between these variables, we then conducted a mediation analysis and found that moral 424 outrage acted as a full mediator between empathy and the third-party framing effect.

425
The mediation effect indicates that compared to the help frame condition, enhanced  Our findings about the cerebellum might be surprising, seeing that this brain area 451 has been most often associated with movement-related functions (e.g., motor control)  To explore whether (and how) the above brain areas constitute a network 477 underlying the third-party framing effect, we run an RVR analysis and found that the this study, we showed that third-party punishment could be strengthened when a 506 choice is framed as intentionally harming others. As pointed out by Haidt and Graham 507 (2007), evolutionary history has shaped maternal brains to be highly sensitive to 508 signals of cruelty and harm; therefore, people generally try to prevent or relieve 509 other's harm, thus making the harm/care norm a strong moral restriction. 510 Correspondingly, observing others violating this norm could provoke strong moral 511 outrage that fuels costly punishment (Hartsough et al., 2020;Rothschild et al., 2013).

512
In contrast, when a choice is framed as merely "not helping others," that may be less 513 likely to trigger the harm/care norm in people's minds. Overall, this study reveal that       Note: all the results reported above were significant at p < 0.001, k > 10, uncorrected at the voxel 786 level; * indicates the cluster-level FWE correction at p < 0.05. 787 788 Note: all the results reported above were significant at p < 0.001, k > 10, uncorrected at the voxel 791 level; * indicates the cluster-level FWE correction at p < 0.05. 792 793 Note: all the results reported above were significant at p < 0.001, k > 10, uncorrected at the voxel 795 level. 796