Who initiates punishment, who joins punishment? Disentangling types of third‐party punishers by neural traits

2.1. Participants and procedure

We recruited 104 students from the University of Bern for participation in this study in the role of third‐party punishers. Two of these participants had to be excluded due to artifacts in the anatomical brain data. Thus, we analyzed a final sample of 102 participants (46 female, 56 male, average 23.34 years old with SD 2.89). Students of economics, psychology, and social sciences were excluded from participation to reduce the possibility of prior knowledge of the concept of third‐party punishment. We recruited participants for one academic year. The goal was to reach a sample size of about 80–100 subjects since two recent methodological studies on fMRI research concluded that sample sizes close to N = 80 (Geuter, Qi, Welsh, Wager, & Lindquist, 2018) or N = 100 (Turner, Paul, Miller, & Barbey, 2018) are required to reliably and reproducibly recover brain regions with medium effect sizes. Data were analyzed after the collection was complete. All participants in the role of third‐party punishers were right‐handed, nonsmokers, and reported no history of psychological disorders or neurological or cardiovascular disease. The study was approved by the local ethics committee and participants signed informed written consent prior to the participation in the study. Subjects received a show‐up fee of 40 Swiss francs and the points that they earned during the experiment with a currency conversion rate of 2 points (p) = 1 CHF. The experiment was programmed in z‐tree (Fischbacher, 2007) and subjects were recruited via ORSEE (Greiner, 2015). Neuroimaging (structural brain data, see below) and behavioral data collections were conducted in different sessions. Behavioral data collections were conducted in a behavioral laboratory with 24 interconnected computer terminals. Participants were randomly assigned to cubicles where they could make their decisions in isolation from others. Before starting the experiment, control questions ensured participants' understanding of the game.

2.2. Paradigm

The study used a newly developed version of a third‐party punishment paradigm. The third‐party consisted of four subjects (potential punishers) who observed an interaction between two other subjects (first mover and second mover) who played a sequential prisoner's dilemma (PD) game. Notably, the first and second mover in the PD were present in the laboratory and were paid according to their decisions and the decisions of the third‐parties punishers (see below). The first and second mover in the PD both received an initial endowment of 5 points. Then, the first mover could decide whether to transfer 5 points to the second mover or to keep all the points. The second mover learned about the first mover's decision and then had to decide him/herself whether to transfer back 5 points or to keep all points (see Figure 1a, right side). Notably, all transferred points are multiplied by two, for example, if both players transfer points, both receive 10 points in total. If only one person transfers points, the transferring person would receive 0 points while the person keeping the points would receive 15 points. If neither person transfers points, both keep 5 points.

The four potential punishers learned about the underlying structure of the interaction between the first mover and second mover in the PD. Each potential punisher received an initial endowment of 14 points. Their task was to decide whether they want to deduct 4 points from the second mover, in case the first mover decided to transfer points and the second mover kept the points, that is, the first mover cooperated, while the second mover defected. More specifically, they were asked: “Do you want to deduct 4 points from the second mover?” As a novel feature of the paradigm, each potential punisher could decide between the following three options that allowed separating the three predicted punishment types: I deduct points (Option 1, i.e., to punish independently); I join the deduction of points (Option 2, i.e., to punish conditionally); I do nothing (Option 3, i.e., not to punish, see Figure 1a, left side).

Critically, the four potential punishers were aware of the following consequences of the different options. They knew that 4 points would be deducted from the defecting second mover in the PD if at least one person chose Option 1 (i.e., to punish independently). This punishment was not additive and one subject choosing Option 1 was sufficient for punishment to be implemented, regardless of the number of punishers. Implementing punishment yielded a total cost of 3 points, but this total cost would be shared equally among all players who participate in the punishment act, that is, subjects who chose either Option 1 (i.e., to punish independently) or Option 2 (i.e., to punish conditionally). Therefore, the per subject cost to deduct 4 points from the second mover was 3 points, 1.5 points, 1 point, and 0.75 points in Case 1, 2, 3, and 4 subjects participated in the punishment act, respectively (see Figure 1b, left and middle). Importantly, if subjects chose Option 2 (i.e., to punish conditionally), but no one chose Option 1 (i.e., to punish independently), punishment was not implemented (Figure 1b, right). The potential punishers' decision to deduct points was implemented only if the first mover had decided to transfer points and second mover had refused to transfer points. The first and second mover in the PD were aware that a group of four potential punishers could deduct 4 points from the second mover. The potential punishers were aware that the two other subjects knew that only the second mover in the PD could be punished.

2.3. Ratings of unfairness and motives for the punishment choice

After deciding on the punishment of the defecting second mover (when the first mover cooperated), the potential punishers had to indicate the perceived unfairness of the second mover's behavior (on a 5‐point Likert‐scale, 1 = very unfair, 5 = very fair). Additionally, the potential punishers had to rate their agreement with eight motive statements. The statements were taken partially from Balafoutas, Grechenig, & Nikiforakis (2014) and self‐created and aim to elicit potential motives underlying the punishment choice. For example, participants had to answer the following motive statements (for all statements, please see Table S1): (1) My choice was the morally right thing to do. (2) The second mover violated a social norm. (3) Deducting points from the second mover does not help anyone. Participants had to indicate how much they agreed with each motive statement (5‐point Likert‐scale, 1 = no agreement, 5 = full agreement). Finally, only conditional punishers had to rate their agreement with the following three motive statements thought to play a role for this type (5‐point Likert‐scale, 1 = no agreement, 5 = full agreement): (1) I wanted to help people deducting points by reducing the costs for them. (2) I did not want to be the only person who deducted points. (3) It was morally correct to support someone who chose to deduct points.

2.4. Acquisition of anatomical brain data

Anatomical brain data was acquired on a Siemens MAGNETOM Prisma 3.0 Tesla whole‐body scanner using a 64‐channel head coil. T1‐weighted 3D‐modified driven equilibrium Fourier transformation (MDEFT) images were acquired from each subject (176 slices, field of view: 256 × 256 × 176, slice thickness: 1 mm, no gap, repetition time: 7.93 ms, echo time: 2.49 ms, flip angle: 16°).

2.5. Preprocessing of anatomical brain data

Anatomical brain data was preprocessed with the computational anatomy toolbox (CAT 12, version 1450, Dahnke, Yotter, & Gaser, 2013) implemented in the statistical parametric mapping software (SPM 12, version 7487). CAT 12 is documented and freely available online (http://www.neuro.uni-jena.de/cat/) and covers diverse morphometric methods. Here we focused on voxel‐based morphometry (VBM)—a well‐established whole‐brain technique capable of discovering subtle, regionally specific changes in gray matter volume by averaging across subjects. This method is based on high‐resolution structural three‐dimensional magnetic resonance images, registered in standard space, and is designed to find significant regional gray matter differences throughout the whole brain by applying voxel‐wise statistics within the context of Gaussian random fields (Ashburner & Friston, 2000). Preprocessing of the data involved spatial normalization (to a MNI template), segmentation into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), modulation, and spatial smoothing with a Gaussian kernel (full width at half maximum = 8 mm; Ashburner, 2007; Ashburner & Friston, 2000, 2005). In detail, the segmentation approach is based on an Adaptive Maximum A Posterior technique without the need for a priori information of tissue probabilities, uses a Partial Volume Estimation with a simplified mixed model of at most two tissue types (Tohka, Zijdenbos, & Evans, 2004), and applies a classical Markov Random Field approach, which incorporates spatial prior information of adjacent voxels into the segmentation estimation (Rajapakse, Giedd, & Rapoport, 1997). Finally, gray matter segmentations are modulated by scaling with the amount of volume changes due to spatial normalization, so that the total amount of gray matter in the modulated image remains the same, as it would be in the original image.

2.6. Statistical analyses of psychometric data

Statistical analyses were run with the SPSS (version 25). See Section 3 for details about the statistical tests conducted, including independent t‐tests as well as univariate ANOVAs with between‐subject factor behavioral types (independent punishers, conditional punishers, and nonpunishers). Results were considered significant at the level of p <.05 (two‐tailed), except for the eight motive statements all participants had to answer. Here, we applied a correction for multiple comparisons according to Bonferroni, resulting in a corrected p‐value of 0.00625 (0.05/8). Thus, only findings reaching this p‐value survive Bonferroni correction. We use the abbreviation SD for standard deviation. As effect size measure eta² is reported, which is a measure of explained variance.

2.7. Statistical analyses of anatomical brain data

In order to explore whether the behavioral types can be characterized by anatomical brain characteristics, we performed univariate analysis of variance (ANOVA) with between‐subject factor behavioral types (independent punishers, conditional punishers, and nonpunishers) on the smoothed gray matter volume images in SPM 12. As is required for volumetric brain analyses, age, gender, and individual brain size were included in the design matrix as covariates of no interest to model and thus regress out any effects correlated with these covariates.

Given our a priori hypotheses (as outlined in the introduction), we focused in our analyses of anatomical group differences on brain areas playing a role in third‐party punishment, including areas involved in mentalizing (TPJ), emotion processing (amygdala, insula, and anterior cingulate cortex), behavioral control and strategic decision‐making (middle frontal gyrus), and reward processing (caudate nucleus). To this end, a bilateral mask comprising the amygdala, insula, anterior cingulate cortex, the middle frontal gyrus, the caudate, and TPJ (see below) was created, defined using the Automated Anatomical Labeling Atlas (Tzourio‐Mazoyer et al., 2002), and implemented using the WFU Pickatlas toolbox in SPM 12 (Maldjian, Laurienti, Kraft, & Burdette, 2003). For the anatomical definition of the bilateral TPJ, we included the bilateral angular and superior temporal gyrus (posterior to y = −40), because several recent meta‐analyses have consistently reported activation peaks from studies on mentalizing in this part of the brain (Carter & Huettel, 2013; van Overwalle, 2009). The same anatomical mask of the TPJ was used for small volume correction in previous studies on social decision‐making (e.g., Hutcherson, Bushong, & Rangel, 2015).

In order to control for multiple comparisons, we used nonparametric permutation statistics based on threshold‐free cluster enhancement (TFCE, Smith & Nichols, 2009) as implemented in the SPM 12 toolbox TFCE (version 201, http://dbm.neuro.uni-jena.de/tfce/). The idea of the TFCE approach is to combine focal effects with large voxel heights as well as broad effects (clusters). In contrast to common approaches that use cluster‐based thresholding no initial (and arbitrary) cluster‐forming threshold is necessary. TFCE takes a raw statistic image (e.g., t or f maps) and estimates a voxel‐wise metrics (the TFCE values) by combining spatially distributed cluster size and height information. As recommended for volumetric data, we applied a cluster‐size weighting of E = 0.5 and a height weighting of H = 2.0, and the Smith permutation method with 5,000 permutations (Smith & Nichols, 2009). Using voxel‐level nonparametric permutation testing (based on the estimated TFCE values), voxel‐wise p values are computed, which are family‐wise error (FWE) corrected. All analyses in the manuscript report FWE‐corrected (at p <.05) effects across the whole brain (whole‐brain FWE corrected) or across the small volume mask defined above (SV‐FWE corrected). In case of significant (FWE‐corrected) effects in the univariate ANOVA (see above), we extracted gray matter volume values from regions demonstrating anatomical differences, regressed out age, gender and individual brain size (the covariates of no interest), and conducted post‐hoc pairwise comparisons in SPSS (version 25) in order to explore the direction and specificity of the discovered anatomical differences between the three behavioral types.

2.8. Discriminant analyses of motive statements and anatomical brain data

Finally, in order to get a more integrative view of all variables (structural brain characteristics and motive statements) characterizing the three types, we employed discriminant analyses. Discriminant analyses allow examining how well each individual subject can be classified into different types, based on motives, structural brain characteristics, or both. For that purpose, discriminant analyses determines the most parsimonious way to separate the types and discards predictors, which add little to the discrimination of the types. We use a stepwise estimation using Wilks' lambda for the inclusion of predictors (motives or structural brain characteristics or both), that is, in each step the predictor minimizing Wilk's lambda enters if it explains sufficient additional variance. The discriminant analyses involve deriving linear combinations of the included predictors, that is, discriminant functions, which yield coefficients that can be used to calculate a score for a respective discriminant function. The magnitude of the standardized coefficients indicate how strongly the predictors affect the score. Based on these estimated discriminant functions, individual discriminant scores for every subject and function can be calculated. Subsequently the centroids for each type can be determined and every subject can be classified based upon the distance to the centroids of each type. We use leave‐one‐out predictions by repeating the mentioned analyses n times with n − 1 subjects (i.e., the subject to be classified is not involved in the estimation of the discriminant functions) and we use equal priors, that is, we assume a 33.3% probability to be of either type. Discriminant analyses were calculated in SPSS (version 25).

3.1. Frequency of the behavioral types

The behavioral data consists of 102 participants in the role of the potential punisher. With our design, we can identify distinct types of third‐party punishment, that is, independent punishers, conditional punishers and nonpunishers. We find that 19.6% participants are independent punishers, 24.5% are conditional punishers, and 55.9% chose not to punish at all (Figure 2a). Note that, we use the subscripts _Indep for Independent Punishers, _Cond for Conditional Punishers, and _Non for Nonpunishers in the analysis section.

3.2. Characterization of the behavioral types with psychometric ratings

In order to characterize the behavioral types, we explored the relation between the types, the perception of unfairness, and the agreement with different statements about the underlying motives. We use univariate ANOVAs with between‐subject factor behavioral types for the analyses and subsequent post‐hoc independent t‐tests.

All participants first had to indicate how unfair they perceived the behavior of the defecting second mover in the PD (on a 5‐point Likert scale, 1 = very unfair, 5 = very fair). Even though the third‐parties behaved differently (and are of different types), the perception of unfairness did not differ between the three types (F _(2,99) = 2.05, p = .134). In fact, all types considered the behavior of a defecting second mover as very unfair (all means are below 1.9, see Figure 2b). Further, all participants also indicated their agreement on a variety of statements (eight statements, on a 5‐point Likert‐scale, 1 = no agreement, 5 = full agreement) targeted at specific motives underlying the punishment choice. In order to control for multiple comparisons, we applied a Bonferroni correction (0.05/8 = 0.00625, see Section 2) and report in the manuscript only motives that survived this correction procedure (Table S1 for the statistical analyses to all eight motive statements). We find that the three types differed in their agreement about whether “the second mover in the PD violated a social norm” (F _(2,99) = 10.46, p = .0001, eta² = 0.17). Independent punishers agreed almost entirely with the motive statement and nonpunishers only moderately, with conditional punishers being in‐between (mean_Indep = 4.3, mean_Cond = 3.52, and mean_Non = 2.96). Post‐hoc independent t‐tests confirmed the difference in pairwise comparisons (p _{Indep vs. Cond} = .016, p _{Indep vs. Non} <.0001, p _{Cond vs. Non} = .047, see Figure 2c). The behavioral types also differed in their opinion on whether “the deduction of points helps no one” (F _(2,99) = 11.85, p <.0001, eta² = 0.19). Here, nonpunishers agreed to a very large degree and more than both conditional and independent punishers (mean_Non = 4.26, mean_Cond = 3.24, mean_Indep = 3.05; p _{Cond vs. Non} = .0004, p _{Indep vs. Non} = .0001), while independent punishers and conditional punishers only agreed moderately with no significant differences between those two types (p _{Indep vs. Cond} = .610, see Figure 2c). Further, we also asked whether the “own choice was the morally correct thing to do.” Independent punishers agreed most (55% of which agreed entirely) followed by conditional punishers and nonpunishers (mean_Indep = 4.25, mean_Cond = 3.60, mean_Non = 2.84; F _(2,99) = 10.98, p < 0.0001, eta² = 0.18; p _{Indep vs. Cond} = .050, p _{Indep vs. Non} = .0001, p _{Cond vs. Non} = .012, see Figure 2c). Thus, the three behavioral types agreed to different degrees with the motives that we expected to be important based upon the description of the types in the introduction.

Finally, we explored to what extent the motives for participating in collective actions apply to conditional punishers. Note that, only conditional punishers answered these motive statements because they target specific aspects of conditional behavior. Here, we find that conditional punishers demonstrated a moderate to high agreement on the three motives, providing further evidence for the correct classification of this type. More specifically, conditional punishers agreed moderately with the motives of helping the deducting person (“I wanted to help the people deducting points by lowering the costs for them,” mean = 3.2, SD = 1.22) and not wanting to punish all by themselves (“I did not want to be the only person to deduct points,” mean = 3.2, SD = 1.32). Further, conditional punishers agreed most with the motive “it was morally correct to support someone who decided to deduct points” (mean = 3.72, SD = 0.98, see Figure 2d).

So far, we characterized the types based on each motive independently. In order to provide a more integrative view, we will now explore whether and to what extent we can classify a participant based upon the motives using a discriminant analysis and leave‐one‐out predictions (see Figure 4a). Note that only the eight motives all participants answered could be used in this analysis. In a first step, the analysis kept only those motives that explain sufficient variance by minimizing Wilk's Lambda. Three out of 8 motives survive this first step and these motives explain 37.7% of the variance in the three types. The three motives are the same as discussed above (also displayed in Figure 2c) and are used to create two discriminant functions, which separate the types maximally. Figure 4a shows both function scores for each subject and the centroids for each type. The first and second function combined are significant (χ ² = 46.4, p <.001) while the second function itself is not (χ ² = 1.96, p = .375). The first function yields negative (standardized) coefficients for the motive statements concerning the moral correctness of the own choice (−.53) and the perceived norm violation of the second mover (−0.52), and yields a positive (standardized) coefficient for agreeing that deducting points does not help (.58). As can be inferred from Figure 4a, the first function is especially good at discriminating the nonpunishers and independent punishers. We classify each participant using leave‐one‐out predictions. In total, 60.8% of subjects are classified correctly, which is considerably better than the 33.3% chance level. More specifically, 65% of independent punishers, 28% of conditional punishers and 73.7% of nonpunishers are classified correctly (see Figure 4a). This suggests that the motives are particularly good at classifying nonpunishers and independent punishers.

3.3. Characterization of the behavioral types with structural brain characteristics

To explore whether the three types can be characterized by distinct neural signatures, we performed quantitative morphometric analyses of T1‐weighted anatomical images using VBM implemented in the computational anatomy toolbox (CAT 12). VBM is a whole brain technique capable of discovering subtle, regionally specific changes in gray matter volume (see material and methods for details).

We applied univariate analysis of variance (ANOVA) in SPM 12 with gray matter volume as dependent variable and behavioral types (independent punishers, conditional punishers, and nonpunishers) as between‐subject factor. We controlled in the analysis for individual brain size, gender, and age, as is required in volumetric analyses. In order to control for multiple comparisons, we used nonparametric permutation statistics based on threshold‐free cluster enhancement (see Section 2 for details). Findings revealed that three brain areas showed volumetric differences (at p <.05, SV‐FWE corrected) between the three behavioral types (Figure 3a–d), including the right TPJ (TPJ: x = 47, y = −66, z = 35, peak F‐value: 17.71, peak TFCE‐value: 21099), the right DLPFC (DLPFC: x = 45, y = 11, z = 44, peak F‐value = 14.31, peak TFCE‐value: 10966), and the bilateral dorsal caudate (left caudate: x = −14, y = −6, z = 23, peak F‐value: 19.48, peak TFCE‐value: 29133; right caudate: x = 20, y = −3, z = 23, peak F‐value: 14.84, peak TFCE‐value: 11233). In order to explore the direction and specificity of the observed anatomical differences, we extracted the gray matter volume values from these significant regions (for the extraction thresholded at p <.001, uncorrected), regressed out age, gender and individual brain size (the covariates of no interest), and conducted post‐hoc pairwise comparison in SPSS. Below we report the findings of parametric statistics. See Table S2 for nonparametric statistics. Notably, all findings hold, irrespective of whether we conduct parametric or nonparametric tests.

Pairwise comparisons between the three behavioral types revealed a distinctive pattern in the right TPJ, right DLPFC and bilateral dorsal caudate. As shown in Figure 3a, independent punishers demonstrated larger gray matter volume in the right TPJ (in the area of the angular gyrus) than both conditional punishers (F _(1,43) = 12.48, p = .001, eta² = 0.225) and nonpunishers (F _(1,75) = 11.66, p = .001, eta² = 0.135). Conditional punishers and nonpunishers did not differ significantly with respect to gray matter volume in the right TPJ (F _(1,80) = 0.531, p = .468, eta² = 0.007).

As shown in Figure 3b, conditional punishers demonstrated larger gray matter volume in the right DLPFC than independent punishers (F _(1,43) = 13.31, p = .001, eta² = 0.236) and nonpunishers (F _(1,80) = 7.62, p = .007, eta² = 0.087). Independent punishers and nonpunishers did not differ significantly with respect to gray matter volume in the right DLPFC (F _(1,75) = 1.42, p = .237, eta² = 0.019).

Finally, as shown in Figure 3c/d, conditional punishers demonstrated lower gray matter volume in the bilateral dorsal caudate than both independent punishers (left caudate: F _(1,43) = 16.57, p <.001, eta² = 0.278; right caudate: F _(1,43) = 8.72, p = .005, eta² = 0.169) and nonpunishers (left caudate: F _(1,80) = 7.86, p = .006, eta² = 0.090; right caudate: F _(1,80) = 6.973, p = .010, eta² = 0.080). Independent punishers and nonpunishers did not differ significantly with respect to gray matter volume in the bilateral caudate (left caudate: F _(1,75) = 2.52, p = .116, eta² = 0.033; right caudate: F _(1,75) = 1.29, p = .259, eta² = 0.017).

As for the motive statements, we conducted a discriminant analysis with the extracted gray matter volume values of the four discovered brain areas (adjusted for brain size, age, and gender) in order to provide a more integrative view and to examine how good each individual participant can be classified. When we employ the discriminant analysis, the right TPJ, right DLPFC, and the left caudate add sufficient explanatory power (based on Wilks' lambda), resulting in the exclusion of the right caudate. The three included areas explain 30.6% of the variance in the three behavioral types and are used to create two discriminant functions, which separate the types maximally. Figure 4b shows both function scores for each subject and the centroids for each type. The first function and the second function combined are significant (χ ² = 35.73, p <.001) while the second function itself is not (χ ² = 4.311, p = .116). The first function yields positive (standardized) coefficients for the right TPJ (.58) and the left caudate (.59) and a negative (standardized) coefficient for the right DLPFC (−.56). As can be inferred from Figure 4b, the first discriminant function is especially good at discriminating conditional punishers and independent punishers. The leave‐one‐out predictions yield a total of 47.1% of subjects being classified correctly, which is better than the 33.3% chance level. More specifically, 65% of independent punishers, 60% of conditional punishers and 35.1% of nonpunishers are correctly classified (see Figure 4b). In contrast to the motives, structural brain characteristics seem to be better at classifying conditional punishers but less good at classifying nonpunishers.

Finally, to complete the picture, we employed a discriminate analysis with all eight motive statements and four brain areas (using the extracted and adjusted gray matter volume measures). Notably, all the motives and brain areas included in the other two discriminant analyses also yield sufficient discriminating power to be included in the combined motive and brain analysis (using Wilks' lambda for the inclusion of predictors). These six predictors explain 56.5% of the variance and are used to create two discriminant functions. Figure 4c shows both function scores for each subject and the centroids for each type. The first and second function combined are significant (χ ² = 80.438, p <.001) and the second function itself is also significant (χ ² = 27.629, p <.001). The first function is mainly affected by the negative (standardized) coefficient for the motive of “deduction does help no one” (−.417) and positive (standardized) coefficients for the right TPJ (.462), the perceived norm violation of the second mover (.444), and the moral correctness of the own choice (.567). As can be inferred from Figure 4c, the first function distinguishes independent punishers and nonpunishers well. The second function is mainly affected by the positive (standardized) coefficient for the right DLPFC (.524) and negative (standardized) coefficient for the left caudate (−.582). As can be inferred from Figure 4c, the second function discriminates between conditional punishers and the other two behavioral types. The leave‐one‐out predictions show that at least 60% of every type are correctly classified, with a total accuracy of 63.7% (which is better than the 33.3% chance level). Here, 70% of independent punishers, 60% of conditional punishers and 63.2% of nonpunishers are correctly classified (see Figure 4c). Thus, it seems that the combination of both structural brain characteristics and motives is particularly well suited to classify the three behavioral types.