To Blame or Not? Modulating Third-Party Punishment with the Framing Effect

Abstract

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 the “third-party framing effect”. In these experiments, participants observed an anonymous perpetrator deciding whether to keep her/his economic benefit while exposing a victim 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 that perpetrator at their own cost. Our results showed that the participants were more willing to execute third-party punishment under the harm frame compared to the help frame, manifesting a framing effect. Self-reported anger toward perpetrators mediated the relationship between empathy toward victims and the framing effect. Meanwhile, activation of the insula mediated the relationship between mid-cingulate cortex activation and the framing effect; the functional connectivity between these regions significantly predicted the size of the framing effect. These findings shed light on the psychological and neural mechanisms of the third-party framing effect.

Introduction

People punish norm deviation not only when they are the victims (i.e., a second party), but also—and more importantly—when they witness unrelated others being victimized (i.e., a third party) [1, 2]. Across societies, social norms are enforced predominantly through third-party punishment [3, 4]. Nonetheless, a third party’s judgment about whether something is ethical or morally justifiable can be significantly biased by its frame (i.e., verbal description) [5–9]. This phenomenon is called the “third-party framing effect”. For instance, Brockner et al. found that the frame used to describe an organizational layoff (some employees lost their jobs vs some employees kept their jobs) predicts a third party’s trust level of that organization [10]. In the same vein, people believe that a contractor's breach more deserves to be punished when its motivation is framed as “making more money” rather than “avoiding to lose money” [11]. As pointed out by Skarlicki and Kulik, third-party judgments can be manipulated because they rely on information that has been framed and interpreted by its source (e.g., mass media) [12]. Consequently, those favored by the framing context might be freer to perform unethical behavior without the anticipation of being condemned and punished by the public [6]. Regarding the value of third-party punishment in maintaining a society, investigating its susceptibility to framing techniques and the associated neural underpinnings is important.

Generally speaking, “framing effect” means that our judgment of an object or issue is affected by its description [13]. As pointed out by De Martino et al., the framing effect can be interpreted as an effect that is heuristically underwritten by the emotional system [14]. That is to say, positive and negative frames regulate individual preference by evoking different emotional responses [15]. It is therefore not surprising that third-party punishment can be modulated by the framing effect, given that assessing responsibility and determining an appropriate punishment are both largely based on emotional reactions [16]. According to the literature, third-party punishment is mainly driven by (1) empathic feeling for (potential) victims [17, 18], and (2) negative emotions in response to perpetrators’ norm-violating behavior, particularly anger [19–22]. Further, previous studies have demonstrated that both empathy and anger are sensitive to framing techniques [23, 24]. For instance, Tang and Gray recently reported that people’s empathic feeling for an organization increases when this organization is described in an anthropomorphic (human-like) way [25]. Nevertheless, whether empathy and anger play different roles in third-party punishment remains largely unclear.

It is well acknowledged that third-party punishment engages a wide distribution of regions across the whole brain [16, 26–30]; many of these regions are closely associated with emotional processing, supporting the idea that salient emotional responses play a pivotal role in social dilemmas [31–33]. Among these brain areas, we are most interested in those associated with empathy and anger, predicting that they respond to different frames when individuals make third-party punishment.

First, empathy refers to the ability to vicariously share the affective states of others and adopt their point of view [34]. An enhanced empathic response leads to an increased tendency to care about others’ well-being beyond kinship and affinity, and to take altruistic actions including helping and sharing [35, 36]. Different paradigms for empathy research consistently engage the anterior and mid-cingulate cortex (MCC), as well as the medial prefrontal cortex [37–39]. Meanwhile, anger triggered by injustice or other norm violations against others (which should be distinguished from personal anger at being harmed) has a huge impact on moral judgment and decision-making [40–42]. As pointed out by Haidt, moral anger generates a motivation to attack, take revenge on, or humiliate perpetrators and to help victims of unfair treatment [43]. The brain networks of anger mainly include the insula, amygdala, and thalamus, all of which participate in social decision-making when anger is involved [44–47]. For instance, the anterior insula is strongly activated when participants reject unfair offers for themselves or for an anonymous third party [48].

Based on the above knowledge, we explored the potential influence of the framing effect on third-party punishment at both the behavioral and neural levels. Our participants were asked to observe an anonymous perpetrator making a trade-off between income maximization and protecting a stranger from a painful shock. Here, choosing income maximization was described as either “harm to” or “not helping” a stranger in two frame conditions. Then participants could decide whether to punish that perpetrator from a third-party perspective. In one of our recent studies, this social frame manipulation successfully modulated the decision-making pattern from a first-party perspective, such that participants (when playing as perpetrators) were more willing to sacrifice their own benefit under the harm frame than under the help frame [49]. In our opinion, this was because “not harming others” is a stronger moral norm than “helping others” [50, 51]. In this study, we hypothesized that participants (when playing third parties) would make more costly punishments in the harm frame condition than in the help frame condition when they observe other perpetrators’ self-serving choices. This hypothesis was first tested in a behavioral experiment (Experiment 1), in which we asked participants to report their empathic response to victims and angry feelings toward perpetrators in both frame conditions, to determine the roles of empathy and anger. Then in a follow-up experiment (Experiment 2) using the functional magnetic resonance imaging (fMRI) technique, we determined whether the brain regions involved in empathic response and/or anger would be activated more strongly in the harm frame condition. Finally, we investigated the relationship between behavioral and fMRI data across frame conditions.

Materials and Methods

Participants in Experiment 1 (Behavior)

One hundred and one right-handed students were recruited from Shenzhen University to join Experiment 1. Three of them were excluded due to failures in data recording, leaving 98 participants in the final sample [46 males, age: 21.14 ± 0.87 years (mean ± SEM)]. All procedures in this study were in accordance with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. The study was approved by the Medical Ethics Committee of Shenzhen University Medical School. Informed consent was given by all participants before the formal experiment.

Experimental Design and Procedures of Experiment 1

Our task design combined the “social framing” task with third-party punishment. This task was developed and has been successfully applied in our recent studies [49]. Specifically, one participant played the decision-maker (perpetrator) and her/his partner was the pain-taker (victim). In each trial, the perpetrator had to determine the occurrence probability of two possible outcomes: loss of 5 monetary units (MUs) from her/his own payoff while the victim avoided a painful electric shock (i.e., costly helping), or keeping all the payoff while the victim received the shock. The perpetrator determined the occurrence probability of these two outcomes (10%, 30%, 50%, 70%, or 90%) by moving a blue triangle on a 5-point scale [52]. The two outcomes were described in different ways in the two conditions: in the help frame condition, the costly helping outcome was described as “help the other person to avoid a shock”, while the other outcome was described as “do not help the other person”; in the harm frame condition, the costly helping outcome was described as “do not shock the other person”, while the other outcome was described as “shock the other person”.

In the present study, all participants acted as third-party observers and received 5 MUs in each trial of the task. After they observed each of the perpetrator’s choices, participants were given a chance to punish that perpetrator by spending their own 5 MUs, in a way that each MU spent would reduce 2 MUs from the perpetrator’s final payoff (Fig. 1A). Before the formal task, participants were told that the perpetrators and the victims were anonymous volunteers who did not know one another; the interactions between perpetrators and victims were recorded in a previous study, but the perpetrators had not yet received their payoff. Each trial would present the decisions of different perpetrators towards different victims. They were also told that 10% of trials would be randomly selected by the computer program. In the end, the MUs involved in these selected trials would be converted to real monetary bonuses. Accordingly, their follow-up decisions would determine not only their own payoff (fixed baseline payment + bonus) but also that of the perpetrators. In reality, the perpetrators and the victims were all putative players [30]. Each perpetrator’s choices were pseudo-randomly set up, hence the number of trials in each condition was the same.

Fig. 1.

Task design and experimental procedure. A The participant observes a “social framing” task in which a perpetrator decides the probability for a victim to receive a painful electric shock. Then the participant decides whether and how to punish that perpetrator with her/his own income. The blue triangle indicates the perpetrator’s choice. B Schematic of an example trial in Experiment 2. The event for brain-imaging analysis is marked with a red rectangle. ITI, inter-trial interval.

We used a 2 (frame: harm versus help) × 5 (moral level of a perpetrator’s choice, “moral level” for short) within-subject design. Here, the five levels of the moral level factor were: highly pro-helping (90% probability of spending own money to save the victim from a shock), medium pro-helping (70% probability of spending own money), neutral (50%), medium non-helping (70% probability of keeping own money), and highly non-helping (90% probability of keeping own money). Each condition (2 × 5) was repeated twice, resulting in 20 trials through the task.

In each trial, the participant first read the two possible outcomes (which were presented in different ways in the two frame conditions) for 2 s. The position of the texts and images indicating each outcome was counterbalanced across trials (the left or right side of the scale) After that, a blue triangle moved on a 5-point scale to show the perpetrator’s decision for 2 s. Each participant then input numbers with numerical keys on a keyboard to answer three questions: (1) altruistic punishment: “how many MUs would you like to pay for punishing the perpetrator?” (0–5 MUs); (2) empathic feeling for victims: “how unpleasant would the victim feel?” (1: not unpleasant at all to 5: extremely unpleasant); (3) anger toward perpetrators: “how angry do you feel about the perpetrator? (1: not angry at all to 5: extremely angry). The whole task lasted for ~5 min. After the formal task, the participants were debriefed and none of them raised any doubt about whether the experimental setup was real.

Participants in Experiment 2 (fMRI)

To determine our sample size, a priori power analysis was applied using G*Power 3.1 [53]. This analysis revealed that 29 participants were required to reach a good statistical power of 0.85 to detect medium-sized (f = 0.25) effects with an alpha value of 0.05 for a 2 × 5 within-subject analysis of variance (ANOVA). To account for possible dropouts or errors during the experiment, 35 right-handed participants who did not participate in Experiment 1 were recruited from Shenzhen University to join in the fMRI experiment. Four who had excessive head movements >2° in rotation or >2 mm in translation during the scanning were excluded, leaving 31 participants in the final sample (14 females, age: 20.62 ± 1.96 years).

Experimental Design and Procedures of Experiment 2

The task design was the same as in Experiment 1, but the experimental setting was adjusted for fMRI scanning. Before the scan, participants were familiarized with the task by finishing a practice block consisting of 8 trials. In each trial, the participant first read two possible outcomes in different frames for 2 s. After that, a blue triangle moved on a 5-point scale to show a perpetrator’s decision process for 2 s. Each participant then waited for another 4 s before they could press one of two pre-assigned buttons on an MRI-compatible button-box to indicate how many MUs (0–5) he/she would like to pay for punishing that perpetrator. The participant had 2 s to choose the preferred number (which would turn from black to red), the starting point of which was randomized across trials. Finally, the inter-trial interval (ITI) was set at 1–4 s (Fig. 1B). Each condition (2 × 5) contained 24 repetitions and there were 240 trials in total. The 240 trials were equally divided into four runs. The whole experiment lasted for ~1 h.

Neuroimaging Data Acquisition and Preprocessing

We used a Siemens TrioTim 3.0T MRI machine (Siemens Medical Solutions, Erlangen, Germany) for data acquisition. Functional volumes were acquired using multiple-slice T2-weighted echo-planar imaging (EPI) sequences with the following parameters: repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, field of view = 224 × 224 mm², 33 slices covering the entire brain, slice thickness = 3.5 mm, voxel size = 3.5 × 3.5 × 3.5 mm³.

fMRI data were preprocessed in SPM12 (Wellcome Department of Imaging Neurosciences, University College London, UK, http://www.fil.ion.ucl.ac.uk/spm). Images were slice-time corrected, motion-corrected, and normalized to Montreal Neurological Institute (MNI) space for each individual with a spatial resolution of 3 × 3 × 3 mm³. Images were then smoothed using an isotropic 6-mm Gaussian kernel and high-pass filtered at a cutoff of 128 s.

Neuroimaging Data Analysis

General Linear Model (GLM)

Statistical parametric maps were generated on a voxel-by-voxel basis with a hemodynamic model to estimate brain responses. We mainly focused on analyzing participants’ brain signals when they watched the presentation of a perpetrator’s choice (red rectangle in Fig. 1B). Frame presentation, perpetrator’s choice, and participant’s response were included in the GLM at the single-participant level. The six rigid-body parameters were also included in the GLM to exclude head-motion interference. The independent regressors analyzed in the GLM were different frames (harm versus help) under different moral levels of a perpetrator’s choice (highly pro-helping, medium pro-helping, neutral, medium non-helping, and highly non-helping), namely 10 independent regressors.

Regression Analysis

Since we were most interested in the third-party framing effect, we combined four moral levels (medium pro-helping, neutral, medium non-helping, and highly non-helping) to focus on the main effect of frames; as noted above, the highly pro-helping condition was excluded because no framing effect was detected in this condition. For the group-level analysis, we conducted a one-sample t-test using the whole brain as the volume of interest to localize the differences in brain activity between the harm frame and the help frame. Regression analysis was also conducted to explore which brain areas were activated more strongly in the contrast of harm minus help frame, as a function of the third-party framing effect.

Region of Interest (ROI) Analysis

We selected ROIs from the contrast (harm − help) that covaried with the third-party framing effect, MNI coordinates [− 33 6 0] for the left insula, [33 6 9] for the right insula, and [− 6 − 21 45] for the MCC. Besides, to test our hypothesis that the significant interaction effect in the MCC reflected different levels of empathy toward the victim, while insular activation reflected different levels of anger towards the perpetrator under different frames, we selected the MCC ROI ([− 2 23 40]) from a previous meta-analysis concerning empathy [54] and the left insula ROI ([− 37 31 6]) from a previous study about moral anger triggered in social decision-making [55]. The mean parameter estimates from these ROIs were calculated within a sphere (radius = 6 mm) centered in the coordinates for each participant. We then applied two-way ANOVA to the extracted values to determine whether their activation was sensitive to the frame and/or moral level.

Mediation Analysis

To test the mediation model revealed in Experiment 1, we ran a mediation analysis with the difference (harm frame minus help frame) in the activation level of the MCC [− 6 − 21 45] as X, the same difference in the activation of the left insula [− 33 6 0] as M, and the behavioral third-party framing effect as Y.

Effective Connectivity Analysis: Dynamic Causal Modeling (DCM)

We hypothesized that the modulation effect on effective connectivity between brain regions that were sensitive to frame manipulation (i.e., the left insula and MCC; see below) were positively correlated with the behavioral index, that is, the third-party framing effect size.

ROI analysis revealed that only the left insula [− 33 6 0] and MCC [− 6 − 21 45] showed significant differences between the harm and help frames while the right insula [33 6 9] did not. We, therefore, used DCM12 [56] to assess the effective connectivity between the left insula and MCC during the task. The first eigenvariate for the single-participant time courses was extracted from volumes located in the left insula and MCC, using a sphere with a 6-mm radius that was centered at individual maxima and adjusted for the effects of interest. The ROI time series were extracted from the whole-brain activation in the harm and help frame conditions.

To determine the driving input (matrix C) and modulation effect (matrix B), we fixed the effective connectivity between every two regions as bilateral connections. Our model space consisted of three families that were differentiated by where the modulation effect (matrix B) took place. We then permutated the driving input into each node respectively, or into both of the two nodes, which led to three models in each family (Fig. 5A, B). We finally specified all the nine (3 × 3) models separately for each run and each participant. We then estimated all the models and subjected them to random-effect Bayesian model selection to select the best-fitted family from our model space based on the model evidence [57]. Bayesian model averaging was then used to calculate weighted-model parameters for the winning family.

Fig. 5.

DCM results. A Model structure of all the three families and Bayesian model selection results of the task-fMRI dynamic causal modeling. The families differed in where the modulation effects were. The expected and exceedance probabilities of the three families are shown in the right panels. The red box indicates the winning family (F3). The light grey lines indicate the intrinsic connectivity between nodes and the black lines indicate where the modulation effect was. The arrows of the lines indicate the directions of the connectivity. B The model structure of all the three models (M1–3) in the winning family (F3) and the results of Bayesian model selection. Within each family, the models differed based on the location of the inputs. The expected and exceedance probabilities of the three models are shown in the right panels. The red box indicates the winning model (M2). The dark grey arrows indicate inputs in the DCM model. MCC, mid-cingulate cortex. XP, exceedance probability. RFX, random-effects.

Statistical Analysis

All the data were shown as the mean ± SEM. Since we used a 2 × 5 within-subject design, repeated measures analysis of variance (ANOVA) was applied. If the interaction effect was significant, pairwise comparison was applied to compare every two conditions. Statistically significant difference was indicated as follows: ***P <0.001, **P <0.01, and *P <0.05. The statistical analysis was performed with the software of SPSS version 22 (IBM Corp., Armonk, USA).

Results

Experiment 1

As noted in the Introduction, we focused on whether different frames would bias each participant’s altruistic punishment on the perpetrators’ choice (i.e., third-party framing effect).

First, to determine whether the strength of altruistic punishment was sensitive to the two within-subject factors (i.e., frame and moral level), we applied a 2 × 5 repeated-measures ANOVA to the punishment data (in MUs). The results showed significant main effects of both factors (frame: F_{(1, 97)} = 35.64, P < 0.001, η_p² = 0.27; moral level: F_{(4, 388)} = 164.38, P < 0.001, η_p² = 0.63) as well as an interaction (F_{(4, 388)} = 10.15, P < 0.001, η_p² = 0.10). Post hoc tests showed that when a perpetrator was highly pro-helping, participants’ punishment did not significantly differ between the two frames (harm frame: 0.09 ± 0.03; help frame: 0.08 ± 0.03, P = 0.53); in the other four conditions (medium pro-helping, neutral, medium non-helping, and highly non-helping), participants punished more under the harm frame than the help frame (P < 0.02, Bonferroni corrected; Fig. 2A).

Fig. 2.

Behavioral results of Experiment 1. A The strength of third-party punishment for each moral level in the harm frame and help frame conditions. *P < 0.05, two-sided t-test, n = 98, data are presented as the mean ± SEM. B Results of the mediation analyses.

We then calculated the difference of participants’ punishment between the harm frame and the help frame as an index of third-party framing effect size. Pearson’s correlation analysis revealed that the framing effect sizes in different conditions were significantly correlated with one another (P < 0.002) except when a perpetrator was highly pro-helping, reflecting a ceiling effect (P > 0.13).

Overall, these results revealed a significant third-party framing effect (i.e., altruistic punishment increased more under the harm frame than the help frame) except when a perpetrator was highly pro-helping. In the other four conditions (medium pro-helping, neutral, medium non-helping, and highly non-helping), the third-party framing effect was homogenous. We, therefore, averaged each participant’s framing effect size in these four conditions as a general behavioral index. Also, we calculated the difference in “empathic feeling for victims” as well as “anger toward perpetrators” between the two frames, in the same way as we calculated the third-party framing effect (i.e., harm frame minus help frame). One-sample t-tests revealed that these three indexes were all significantly larger than zero (framing effect size: 0.23 ± 0.04, t₍₉₇₎ = 6.13, P < 0.001; empathic feeling: 0.28 ± 0.04, t₍₉₇₎ = 7.07, P < 0.001; anger: 0.32 ± 0.04, t₍₉₇₎ = 8.43, P < 0.001). Pearson’s correlation analysis revealed that the three indexes were all significantly correlated with one another (r > 0.63, P < 0.001).

We then ran a mediation analysis with the difference in empathic feeling as X, the difference in anger as M, and the third-party framing effect as Y. This analysis revealed a significant full mediation effect of M between X and Y. Normal theory tests showed a significant a path (t₍₉₇₎ = 11.30, P < 0.001) and a significant b path (t₍₉₇₎ = 5.06, P < 0.001). A significant c path (t₍₉₇₎ = 7.86, P < 0.001) became not significant (t₍₉₇₎ = 1.94, P = 0.054) when the mediator was adjusted. The bias-corrected confidence interval was between 0.21 and 0.61 (Fig. 2B).

Experiment 2

We performed a 2 × 5 repeated measures ANOVA on altruistic punishment and found that the main effects of both within-subject factors (frame: F_{(4, 388)} = 28.43, P < 0.001, η_p² = 0.49; moral level: F_{(4, 388)} = 187.70, P < 0.001, η_p² = 0.86) as well as their interaction (F_{(4, 136)} = 17.36, P < 0.001, η_p² = 0.34) were significant. Pairwise comparison showed that when a perpetrator was highly pro-helping, participants’ altruistic punishment was not significantly different between the two frames (harm frame: 0.18 ± 0.04; help frame: 0.19 ± 0.05, P = 0.88), showing that no framing effect was detected in this condition; regarding the other four conditions (medium pro-helping, neutral, medium non-helping, and highly non-helping), participants punished more under the harm frame than the help frame (P < 0.01; Fig. 3A).

Fig. 3.

Behavioral results of Experiment 2. A The strength of third-party punishment for each moral level in the harm frame and help frame conditions. **P < 0.01; n.s., not significant; two-sided t-test, n = 31, data are presented as the mean ± SEM. B The third-party framing effect size for each participant, calculated as the average across four moral levels (except “highly pro-helping”) is represented as the solid black line.

As in Experiment 1, these results confirmed the third-party framing effect in four conditions. Accordingly, we then used the average of these conditions to index each participant’s third-party framing effect size, as in Experiment 1. And found that the third-party framing effect size was significantly larger than zero on the group level (0.42 ± 0.42, t₍₃₀₎ = 5.62, P < 0.001; Fig. 3B).

fMRI Results

GLM

Contrasting the brain activation level under the harm frame and the help frame, we found significantly stronger activation in the left precentral gyrus (peak MNI [− 27 − 15 63]), left supplementary motor area (peak MNI [− 9 − 3 69]), and right insula (peak MNI [42 3 0]) during the presentation of a perpetrator’s choice (Fig. 4A, Table 1). The reversed contrast (help frame > harm frame) revealed no significant cluster.

Fig. 4.

Results of brain activation in Experiment 2. A Brain activation of the contrast under the harm frame > help frame. B The contrast of the harm frame > help frame covaried with the individual third-party framing effect at the group level. C ROI analysis for the MCC ROI from Lamm et al. (2011) [54] and left insula ROI from Gilam et al. (2018) [55]. *P < 0.05, two-sided t-test, n = 31, data are presented as the mean ± SEM. D Results of mediation analysis. MCC, mid-cingulate cortex.

Table 1.

Brain activations of the contrast under the harm frame > help frame

Hem	Brain region	BA	Coordinates			Vol	T-value
			(X, Y, Z)
L	Precentral gyrus**	6	− 27	− 15	63	228	5.57
	Postcentral gyrus	4	− 42	− 21	45		4.88
	Postcentral gyrus	3	− 33	− 30	48		4.45
L	Supplementary motor area*	6	− 9	− 3	69	72	5.49
	Supplementary motor area	6	− 6	− 12	54		3.87
R	Insula*	48	42	3	0	72	5.33
	Insula	48	45	9	− 6		5.27
L	Angular gyrus	39	− 39	− 63	48	24	4.91
L	Anterior cingulate and paracingulate gyri	32	− 12	45	12	40	4.72
R	Superior frontal gyrus, medial	10	15	54	9	15	4.56
L	Insula	48	− 36	6	0	35	4.41
	Superior temporal gyrus	21	− 51	0	− 9		4.32
R	Cerebellum		12	− 60	− 24	46	4.32
	Cerebellum	19	21	− 63	− 27		3.90
L			− 21	33	30	15	4.24
L	Middle temporal gyrus	21	− 54	− 18	− 12	45	4.15
	Middle temporal gyrus	21	− 60	− 21	− 3		3.94
R	Middle temporal gyrus	21	63	− 30	− 6	12	4.04
L	MCC and paracingulate gyri	23	− 3	− 12	42	16	4.00

All the above results were significant at P < 0.001, Cluster size ≥ 10 voxels, uncorrected at the voxel level; *, cluster-level family-wise error rate (FWE) correction at P < 0.05; **, cluster-level FWE correction at P < 0.005; Hem, hemisphere; BA, Brodmann area; Vol, volume.

Regression Analysis

The regression analysis showed that the activations in the MNI coordinates [− 33 6 0] for the left insula, [33 6 9] for the right insula, [− 39 − 18 45] for the left postcentral gyrus, [− 6 − 21 45] for the MCC, and [9 − 63 − 18] for the cerebellum positively covaried with the third-party framing effect size (Fig. 4B, Table 2).

Table 2.

Whole-brain activations based on group-level regression analysis covaried with the behavioral third-party framing effect

Hem	Brain region	BA	Coordinates			Vol	T-value
			(X, Y, Z)
L	Insula**	48	− 33	6	0	482	6.82
	Insula	48	− 36	− 3	9		6.08
	Insula	34	− 30	0	− 9		5.81
R	Insula**	48	33	6	9	270	6.78
	Temporal pole: superior temporal gyrus	38	54	9	− 9		4.97
	Temporal pole: superior temporal gyrus	48	36	− 3	− 6		4.84
L	Postcentral gyrus**	4	− 39	− 18	45	175	5.92
	Postcentral gyrus	3	− 30	− 30	48		4.66
	Paracentral lobule	4	− 36	− 21	57		4.41
L		21	− 42	− 45	9	41	5.26
		41	− 36	− 42	15		4.40
L	Cerebellum_superior	30	− 21	− 30	− 24	18	5.08
			− 9	− 27	− 27		3.57
L	Middle temporal gyrus	21	− 51	− 6	− 21	31	4.97
L	Mid-cingulate and paracingulate gyri**		− 6	− 21	45	168	4.94
	Mid-cingulate and paracingulate gyri**	23	3	− 9	45		4.66
			− 15	− 33	39		4.25
R	Cerebellum_superior*	18	9	− 63	− 18	92	4.89
	Cerebellum_superior	37	24	− 45	− 24		3.92
L			− 9	− 27	0	43	4.66
R	Paracentral lobule		0	− 24	72	16	4.02
R	Supramarginal	48	54	− 30	27	16	3.93
L	Precentral gyrus	6	− 24	− 12	69	10	3.89
	Precentral gyrus		− 21	− 21	75		3.76
R		24	9	12	33	11	3.80

All the above results were significant at P < 0.001, Cluster size ≥ 10 voxels, k > 10, uncorrected at the voxel level; *, cluster-level FWE correction at P < 0.05; **, cluster-level FWE correction at P < 0.005; Hem, hemisphere; BA, Brodmann area; Vol, volume.

ROI Analysis

Regarding the MCC [− 6 − 21 45], the main effect of frame was significant (F₍₁, ₃₀₎ = 7.86, P = 0.01, η_p² = 0.21) such that the harm frame elicited significantly stronger activation than the help frame (− 5.01 ± 0.65 vs −5.38 ± 0.65). The main effect of moral level approached significance (F_{(3, 90)} = 2.84, P = 0.05). The interaction was significant (F_{(3, 90)} = 4.54, P = 0.01, η_p² = 0.13): pairwise comparison revealed that in the medium pro-helping and neutral conditions, the difference between the two frame conditions was not significant (P > 0.14); in the medium non-helping and highly non-helping conditions, this difference was significant (P <0.01).

Regarding the left insula [− 33 6 0], the main effect of frame was significant (F_{(1, 30)} = 6.88, P = 0.001, η_p² = 0.19), which means that the harm frame elicited stronger activation than the help frame (− 1.04 ± 0.45 vs − 1.28 ± 0.43). The main effect of moral level was significant (F_{(3, 90)} = 7.34, P < 0.001, η_p² = 0.20), suggesting that the activation increased as the perpetrators’ decision became more self-serving. The interaction was not significant (F_{(3, 90)} = 1.03, P = 0.38, η_p² = 0.03). Regarding the right insula [33 6 9], no significant effect was detected (P > 0.24).

Further, we found similar patterns for the MCC [− 2 23 40] extracted from the meta-analysis for empathy [54] and left insula [− 37 31 6] extracted from the literature for anger [55]. Regarding the MCC [− 2 23 40], the main effect of frame was significant (F₍₁, ₃₀₎ = 4.55, P = 0.04, η_p² = 0.13), meaning that the harm frame elicited stronger activation than the help frame (− 0.78 ± 0.65 vs − 1.09 ± 0.63). The main effect of moral level was not significant (F_{(3, 90)} = 2.16, P = 0.09, η_p² = 0.07). Regarding the left insula [− 37 31 6], the main effect of frame was significant (F_{(1, 30)} = 8.67, P = 0.01, η_p² = 0.22), indicating that the harm frame elicited stronger activation than the help frame (− 0.02 ± 0.38 vs − 0.15 ± 0.39). The main effect of moral level was significant (F_{(3, 90)} = 3.30, P = 0.02, η_p² = 0.10), such that the activation increased as the perpetrators’ decision became more immoral. The interaction was not significant (F_{(3, 90)} = 1.07, P = 0.37, η_p² = 0.03; Fig. 4C).

Mediation Analysis

The mediation analysis revealed a significant full mediation effect of M (activation of left insula _{harm frame – help frame}) between X (activation of MCC _{harm frame – help frame}) and Y (third party framing effect). Normal theory tests showed a significant a path (t₍₃₀₎ = 6.71, P < 0.001) and a significant b path (t₍₃₀₎ = 3.04, P = 0.005). The c path was also significant (t₍₃₀₎ = 4.62, P = 0.001), but no longer so when the mediator was adjusted (t₍₃₀₎ = 0.90, P = 0.38; Fig. 4E).

DCM

We further elaborated on the direction of the functional connectivity (FC) between the left insula and MCC, to unravel how different frames modulate these brain regions. The results showed that the family defined by the bidirectional modulation effect that worked on the bidirectional connection between the left insula and the MCC best accounted for the data (Fig. 5A, exceedance probability = 0.99, mean-variance explained 73.74%). Other families exhibited significantly lower probabilities (P < 0.001). In the winning family, the results showed that the model in which the driving input was the MCC was the best-fitted model (Fig. 5B, exceedance probability = 0.56, mean-variance explained 39.20%).

Bayesian model averaging of the winning model was then used to calculate weighted-model parameters, which were then used for statistical analysis. In the help frame condition, the driving input of the MCC was significantly different from zero (t₍₃₀₎ = −2.99, P = 0.01), In the harm frame condition, the driving input of the left insula was significantly different from zero (t₍₃₀₎ = 2.65, P = 0.01). The connectivity from the left insula to the MCC was significantly different from zero (t₍₃₀₎ = 2.15, P = 0.04) and the connectivity from the MCC to the left insula was significantly different from zero (t₍₃₀₎ = 2.43, P = 0.02).

Finally, we calculated the correlations between the modulation effects and the behavior index (i.e., the third-party framing effect) and found that the different modulation effect between the harm frame and help frame conditions from the MCC to the left insula was significantly correlated with participants’ third-party framing effect size (r = 0.409, P = 0.03; Fig. 6).

Fig. 6.

Correlational results. The correlation between the behavioral third-party framing effect and the degree to which the modulatory effect on the connection from the left mid-cingulate cortex (MCC, red circle) to the left insula (green circle) differ between the two frames (harm frame > help frame). r, correlation coefficient, *P < 0.05.

Discussion

Social norms and social order in modern societies are maintained and reinforced through third-party punishment [16, 27]. However, the likelihood of the same behavior being punished by a neutral party may vary according to how that behavior is described, that is, a third-party framing effect. In this series of experiments, we asked participants to observe two putative players interacting in each trial; specifically, a perpetrator chose between “harming” and “not harming” a victim in the harm frame condition, and between “helping” and “not helping” that victim in the help frame condition. These two conditions were objectively but not verbally equivalent, conforming to the definition of frame manipulation [58]. The same framing technique has been successfully applied to evoke a first-party social framing effect in one of our recent studies [49]. Here, in both experiments, we found that the participants’ willingness to punish a perpetrator at their own cost (i.e., third-party punishment) was modulated by frame manipulation, manifesting a third-party framing effect. This effect was robust except in the “highly pro-helping” condition, possibly because it is unreasonable to punish a highly pro-helping perpetrator. The psychological mechanisms of this effect involve empathy and anger (indicated by self-reports from Experiment 1), while its neural mechanisms mainly occur in the MCC and insula (indicated by fMRI data from Experiment 2).

Aside from the tendency to make costly punishment, the self-reported “empathic feeling for victims” and “anger toward perpetrators” were also stronger in the harm frame condition than in the help frame condition in Experiment 1. Moreover, these three behavioral indexes were significantly correlated with one another. To determine the relationship between them, we conducted a mediation analysis and found that anger acted as a full mediator between empathy and the third-party framing effect. This mediation effect indicates that compared to the help frame condition, enhanced empathic feeling for victims in the harm frame condition turned into stronger angry feelings toward perpetrators, and finally manifested as harsher punishment at the behavioral level [59]. These results are consistent with previous reports that angry feelings associated with unfair treatment of another person are evoked by empathic concern about that person but not unfairness per se [41, 60].

Third-party punishment involves the evaluation of: (1) whether a perpetrator has committed a wrongful act (i.e., causality) and (2) whether that perpetrator did it in “a culpable mental state” (i.e., intentionality) [16, 61]. In consideration of these two aspects, punishment magnitude should be proportional to the combined harmfulness of the wrongful act and blameworthiness of the perpetrator [16, 27]. In our task, whether a victim was indeed harmed by a perpetrator’s decision was unknown to participants. Therefore, our participants (as third parties) were punishing an intention rather than its consequence. We suggest that our task generated a third-party framing effect by modulating the participants’ evaluation of intentionality [27]. Specifically, the self-serving option was described as “harming others” in the harm frame condition, thus the participants may assume that the perpetrators were fully aware of the harmful consequence of choosing that option. As pointed out by Haidt and Graham, evolutionary history has shaped maternal brains to be highly sensitive to signals of cruelty and harm [62]; people generally try to prevent or relieve harm to others, which makes the harm/care norm a strong moral restriction [50]. In the harm frame condition, the participants observed that the perpetrators intentionally violated the harm/care norm, therefore they experienced stronger moral outrage that fuels costly punishment [63, 64]. In contrast, choosing “not helping others” in the help frame condition may be perceived as an unintentional violation of the harm/care norm. In short, we suggest that the differences in emotional engagement between frame conditions, and their influence on third-party punishment, should be understood according to the importance of intentionality in moral judgment [65].

In Experiment 2, we further combined our task with the fMRI technique. While the whole-brain contrast identified several brain regions, the follow-up ROI analysis revealed that the MCC and insula were most important, given that they were both sensitive to the main effect of the frame. In light of the neuroimaging literature, we suggest that MCC and insular activity in the current study reflected individual empathic and angry responses, respectively. As revealed by a meta-analysis conducted by Lamm et al. [54], empathy for others’ pain has consistently activated the MCC in previous studies [66, 67]. Even in patients with congenital insensitivity to pain (i.e., unable to experience self-pain), a normal MCC activation pattern was observed during empathic pain [68]. Accordingly, de Waal and Preston conclude that the MCC represents the affective rather than the sensory component of empathic pain [37]. Walter also suggests that the MCC is linked to the affective-motivational aspect of nociception [39]. Meanwhile, the insula contributes significantly to the experience of anger [44, 69], possibly due to the primary role of this region in brain-heart interactions [70, 71]. Anger-inducing experimental stimuli have been shown to reliably activate the bilateral insula [72–74]. In many neuroimaging studies on prosocial behavior (e.g., third-party punishment, allocating resources equally, and rejecting unfair distributions), insular involvement has been explained as an index of an angry reaction [75–77]. Recently, Sellitto et al. found that higher insular activity indicates stronger sensitivity to others’ losses and more generous choices in a framing task, which is in line with our findings [78]. Our analysis with ROIs defined in previous studies on empathy [54] and anger [55] supports the above idea of linking the MCC with empathy, as well as linking the insula with anger.

To explore whether (and how) the above areas constitute a network underlying the third-party framing effect, we then ran a DCM analysis and found that a connection between the MCC and left insula predicted the behavioral framing effect size. Different roles of the two areas were further clarified by mediation analysis: left insular activation acted as a full mediator between MCC activation and the framing effect, which shared the same structure with that in Experiment 1 (i.e., anger mediated the relationship between empathy and the framing effect). Here, the left lateralization of insular activity might result from the strong approach tendency in angry feelings (for the relationship between brain activation asymmetry and approach/withdrawal motivation, see [79]). In our opinion, these results are also consistent with our interpretation of MCC and insular activations. Consequently, a neuropsychological mechanism of the third-party framing effect was identified: our frame manipulation triggers different degrees of empathic response (associated with the MCC) to victims by framing the intention of norm violation (intentionally harming or avoiding helping other people); this response indirectly drives altruistic punishment by modulating moral outrage (associated with the insula) toward norm violators.

Here, a significant limitation was that we did not ask participants to directly report their empathic and angry responses in Experiment 2, in order to save MRI scanning time. As a result, the association between empathy and MCC activation, as well as that between anger and insular activation, has yet to be directly examined. However, we would like to point out that in Experiment 2, our analysis of the ROIs selected from previous studies about empathy and moral anger showed highly similar effects with the self-reported empathy and anger in Experiment 1, indicating the relationship between the two experiments. Still, our neuroimaging findings could be interpreted in terms of alternative theories. For instance, given that the insula also belongs to the core network of empathy [38, 66], the connection from the MCC to the left insula might indicate that an empathic response influences third-party punishment directly, rather than relying on angry feelings as a mediator [80]. Follow-up research is needed to investigate this possibility.

Overall, this study reveals that third-party punishment is susceptible to decision frames, which may help understand why moral standards and moral actions can be flexible and relativistic [81]. More broadly speaking, our findings enrich the knowledge about the psychological processes of moral judgment and their corresponding neural underpinnings [82]. On the other hand, our findings may broaden the understanding of the framing effect in general. According to previous studies, a stronger framing effect in non-social contexts (e.g., lottery) is associated with higher activation in the amygdala, possibly reflecting aversive emotional responses (e.g., anticipatory anxiety) triggered by heightened risk perception under frame manipulation [14, 83]. Meanwhile, the framing effect under social contexts mainly employs the temporoparietal junction (TPJ), which is a key node of the mentalizing network; TPJ activity may reflect the process of perspective-taking that grounds empathic emotions [49]. To our knowledge, all of these studies have investigated the framing effect from a first-party perspective, that is to say, frame manipulation affects participants’ feelings about the decision outcomes that they would receive. In contrast, frame manipulation in this study modulates vicariously activated moral emotions based on third-party observation [63] and therefore recruits neural circuits distinct from those underlying the classical framing effect. Still, our findings support the idea that the framing effect is essentially an emotional phenomenon regardless of whether it is based on first-party or third-party experience.

Below we propose some additional limitations and future directions for follow-up studies to consider. First, it would be interesting to examine the robustness of our findings when the outcome of perpetrators’ choices is deterministic rather than probabilistic. As we pointed out earlier, participants’ punishment was based on intentional violation but not the resulting outcome in the latter context [65]. Second, unlike our previous research [49], this study did not show a victim’s photograph to the participants before a perpetrator’s choice was revealed. Future studies could investigate the relationship of this factor (which may affect social experience) to third-party punishment. Third, aside from punishing norm violators, another important kind of third-party intervention is to compensate the victims [84]. Investigating the framing effect on third-party compensation would help promote prosocial behavior. Fourth, although we believe that combining the current findings and those of our previous research [49] could help understand the differences between the first-party and the third-party framing effect, this study has not yet directly compared these two kinds of effects. Finally, one of the major threats to altruistic third-party punishment is “diffusion of responsibility”, that is, the presence of bystanders may reduce an unaffected observer’s likelihood to take action [85]. One of our studies showed that diffusion of responsibility inhibits neural activation in the insula [86]. Researchers may explore whether using frame manipulation would counteract the negative influence of diffusion of responsibility.

Conflict of interest

The authors declare no competing interests concerning the subject of this study.