Decoding the Charitable Brain: Empathy, Perspective Taking, and Attention Shifts Differentially Predict Altruistic Giving

Donation task.

Altruistic behavior was measured using a charitable donation task (Böckler et al., 2016) (see Fig. 1A). Before the fMRI session, participants were given an endowment of €50. In each trial, they were first presented with a short description of a charitable organization, including its name, main goal, beneficiaries, and examples of the charity's activities (reading phase; up to 30 s). Participants indicated via a button press when they had finished reading. A continuous rating scale was then presented at the bottom of the screen (range of €0 to €50), and participants had to decide how much to donate to that charity (decision phase; up to 10 s). Participants responded by moving a slider (randomized initial position) using the index and middle finger of the right hand and confirmed their donation choice with a button press of the right ring finger. Intertrial intervals varied from 4 to 8 s, during which a black fixation cross was presented against a white background. Before scanning, participants were informed that one donation trial would be randomly selected at the end of the experiment and that their donation would be given to the charity (full amount donated in that trial). Notably, donations were costly to the subjects because they could keep 20% of the €50 that was not given to the randomly chosen charity. These rules ensured that participants treated each trial as if it was the only one (instead of dividing their endowment among different charities) and made donations consistent with their actual preferences. Partial payout was implemented to provide a moderate incentive for donations (Tusche et al., 2013). The 60 charities used in the task were selected based on pretests using independent samples. The charities were presented in 5 blocks, with the presentation order of charities randomized across participants, within and across donation blocks. Task blocks were divided by breaks of ∼30 s in which participants were allowed to rest.

Behavioral post-test.

Subsequent to scanning, participants completed a behavioral task designed to measure the extent to which participants deployed several social processes that have previously been associated with altruistic behavior (Eisenberg and Miller, 1987; Eckel and Grossman, 1996; Moll et al., 2006; Hare et al., 2010; Hein et al., 2010; Mathur et al., 2010; Batson, 2011; Ma et al., 2011; Waytz et al., 2012). During this task, participants were again presented with the mission statements of each charitable organization and were asked to provide the following ratings for each charity: (1) empathic responses (“how much empathy did you experience for the beneficiaries of the charity”); (2) perspective taking (“how much did you try to take the perspective of the beneficiaries of the charity”); (3) similarity (“how similar to you are the beneficiaries”); (4) closeness (“how likely is it that you or someone close to you would benefit from the organization”); (5) deservingness (“how much do the beneficiaries deserve to be supported”); and (6) opposition (“how much are you opposed to the general cause of the organization”). Importantly, participants were instructed to reconstruct and report their engagement in these processes during the previous donation decisions in the scanner. For each charity, the ratings were collected in randomized order using a continuous rating scale that ranged from 0 (not at all) to 5 (very much). Self-report ratings aimed to investigate the role of empathy and perspective taking in altruistic giving (on the within- and between-subject level). Additional ratings were included to examine the mediating role of social processes, such as judgments of similarity, closeness, or deservingness, on prosocial behavior and allowed investigating the specificity of neural responses in, for example, AI and TPJ.

EmpaToM task.

An additional fMRI task measured participants' empathy- and perspective taking-related responses independent of the donation task (Kanske et al., 2015b). These data were used to identify subject-specific neural markers of empathy and perspective taking, respectively. We also compared its behavioral data with self-reported empathy and perspective taking in the donation task as measured subsequent to scanning in the behavioral post-test. The EmpaToM task refers to cognitive perspective taking as “Theory of Mind” (ToM).

In each of the 48 trials, participants saw a short video (∼15 s) of an actor recounting an autobiographical episode that was either highly emotionally negative (high emotionality) or neutral (low emotionality). The assignment of videos to both conditions was defined a priori based on extensive validations of the EmpaToM paradigm (Kanske et al., 2015b). After each video, participants rated their own affect to measure spontaneous empathic responding (“How do you feel”; “very negative” to “very positive,” scale from −3 to 3) and their compassion for the person in the video (“How much compassion do you feel”; “none” to “very much,” scale from 0 to 6) (4 s per rating, fixed order). Participants responded by moving a slider using the index and middle finger of the right hand. After a variable delay of 1–3 s, participants were asked to answer a multiple-choice question concerning the content of the previous video clip (up to 15 s). Questions required either perspective taking (ToM) (e.g., “Anna thinks that …”) or factual reasoning (noToM) (e.g., “It is correct that …”). Participants responded by pressing one of three buttons, assigned to the three choice options. Trials were separated by variable delays of 1–5 s (plus a 2 s screen that displayed the name of the person in the following video). For a detailed description of task validation and example stories and questions for each experimental condition, see Kanske et al. (2015b).

Attention task.

Previous meta-analyses point to a substantial overlap of the neural networks involved in domain-general processes, such as reorienting of attention, and those involved in social processes, such as perspective taking and empathy (Decety and Lamm, 2007; Mitchell, 2008; Krall et al., 2015; but see Silani et al., 2013). Based on this evidence, it has been hypothesized that shifts of attention might be responsible for some of the neural activity in areas such as the TPJ during altruistic decision-making (Hare et al., 2010). If this is true and attentional shifts indeed contribute to altruistic giving (and hence account for, at least part of, the predictive information in the TPJ or other regions), neural responses that encode attentional reorienting should be reactivated during generous donations. To investigate this possibility, we asked a subset of participants to also perform a version of the Posner location-cueing task (Posner and Cohen, 1984; Corbetta and Shulman, 2002). The task was structured as follows: In each of the 240 trials, a centrally presented visual cue (white arrow pointing to the left or the right, 200 ms) indicated the spatial location of a following target stimulus that would appear at the left or the right side of the screen (ISI of 200–800 ms between cue and target). Targets were displayed for 200 ms and required participants to respond with a corresponding button press of their right index finger (target arrow upward) or middle finger (target arrow downward). Trials were separated by an average of 2925 ms. Compared with validly cued targets (80%, 192 trials), invalidly cued targets (20%, 48 trials) require reorienting of attention as reflected in increased reaction times (Muller and Rabbitt, 1989).

GLM 1 (high and low donations).

This GLM was used to obtain a measure of BOLD responses for generous and selfish donation decisions at the task-block level. To this end, for each participant, we defined high and low generosity trials based on participants' average donations (mean split). For each of the five donation blocks, a regressor for high (R1-R5) and low donation decision (R6-R10) was estimated. Each of these regressors lasted from the onset of the donation scale to the button press confirming the donation for the trial (decision phase). Similar block-wise estimates were created for the corresponding reading phases of high (R11-R15) and low (R16-R20) donations (with the modeled duration equal to the time between the presentation of the mission statement and the button press indicating that participants had completed reading). Movement parameters (R21-R26) were included as regressors of no interest. The estimated responses for the regressors of interest (R1-R10) were then used in the multivariate pattern analyses (MVPA) described below. GLM 1 estimated regressors for each of the five donation blocks separately, ensuring that the number of training and test data in the MVPA were perfectly balanced for participants' generous and selfish decisions.

GLM 2 (trial-wise donations).

This GLM was used to obtain a measure of BOLD responses during the donation task at the trial level. The model included a regressor for each decision phase (R1-R60), a separate regressor for each reading phase (R61-R120), and movement parameters (R121-R126). The duration of the regressors matched the durations modeled in GLM 1. The estimated responses for the donation decisions (R1-R60) were used as inputs for the multivariate support vector regressions (SVRs) described below.

GLM 3 (EmpaToM task).

This GLM was used to investigate whether the subject-specific influence of empathy and perspective taking in the donation task is related to participants' empathy and perspective taking responses in an independent task. In particular, we tested whether empathy- and perspective taking-related neural signatures obtained in the EmpaToM task can predict the degree to which these processes contribute to donation behavior. To this end, GLM 3 estimated four regressors of interest for (R1) high emotional video phases, (R2) low emotional video phases, (R3) question phases with perspective taking demands (ToM), and (R4) question phases without perspective taking demands (noToM). The two rating phases (R5, R6) and six movement parameters (R7-R12) were included as regressors of no interest. As in previous studies (Kanske et al., 2015b), two contrast images were estimated for each participant: To identify empathy-related brain responses, we contrasted “high emotional” minus “low emotional” video phases (R1-R2). To obtain perspective taking-related brain responses, we contrasted “ToM” − “noToM” question phases (R3-R4). Individual contrast images were entered into two separate random-effects group analyses (simple t test in SPM8). Only voxels that were significant at a statistical threshold of p < 0.05 (FWE-corrected, small volume corrected for the respective main effect in a sample of n = 178) (Kanske et al., 2015b) were considered relevant for the neural encoding of empathy and perspective taking and were used as input (i.e., features) for the MVPA. All voxels identified in the respective contrast were used together as input for the MVPA described below. The selection of voxels (i.e., features) used for the subsequent MVPA based on statistical criteria, such as t tests, is a standard approach in MVPA (Mitchell et al., 2004; Polyn et al., 2005) to identify “relevant” voxels (i.e., potentially high level of signal and low level of noise) for the neural encoding of a condition/variable of interest and to reduce the dimensionality of the MVPA (for detailed introductions to this topic see, e.g., Haynes and Rees, 2006; Misaki et al., 2010). The selection of voxels based on comparisons of task-related experimental conditions of interest is conceptually well grounded, data-driven, and completely independent from the donation task.

GLM 4 (attention task).

This GLM was used to investigate whether activity related to the generosity of donations is related to stimulus-driven shifts of attention. GLM 4 estimated the average responses in each target period of the Posner cueing task (R1-R240). Cueing periods and movement parameters were modeled as regressors of no interest. Estimated responses for the 48 invalidly cued target trials and for 48 validly cued target trials (randomly drawn from the total of 192 valid trials) were used as input in the MVPA described below. Only correct trials were considered as input for the MVPA and the number of validly and invalidly cued trials was matched.

Neural decoding of generous donation decisions.

This decoding analysis was designed to identify brain regions that decode generous versus selfish giving decisions as measured by a mean split of donations at the subject level. To this end, we used GLM 1's parameter estimates for brain responses during high (R1-R5) and low (R6-R10) donations for each participant. We applied a whole-brain “searchlight” decoding approach that does not depend on a priori assumptions about informative brain regions and ensures unbiased information mapping throughout the whole brain (Kriegeskorte et al., 2006; Haynes et al., 2007). The searchlight decoding was done as follows: For each participant, we defined a sphere (radius = 4 voxels) around a given voxel v_i of the measured brain volume (Bode and Haynes, 2009; Kahnt et al., 2011a; Tusche et al., 2014; Wisniewski et al., 2015). For each of the N voxels within this sphere, we extracted parameter estimates and transformed them into an N-dimensional pattern vector. Pattern vectors were created separately for high and low donations and for each donation block. Pattern vectors of all blocks but one (“training data”) were used to train a linear kernel support vector machine (SVM) classifier (http://www.csie.ntu.edu.tw/∼cjlin/libsvm) using a fixed cost parameter c = 1. This provided the basis of the following classification of the pattern vectors of the remaining donation block (“test data”) as representing either high or low donations. The procedure was repeated five times, always using a different donation block as test dataset (leave-one-pair-out cross-validation). The amount of predictive information on generosity was represented by the average percentage of correct classifications across the fivefold cross-validation and was assigned to the central voxel v_i of the searchlight sphere. This support vector classification (SVC) was successively performed for all spherical clusters created around every voxel of the measured brain volume, resulting in a 3D map of average classification accuracies for each participant. These accuracy maps were then smoothed (8 mm FWHM) and entered into a random-effects group analysis (simple t test in SPM8) to identifying brain regions that encoded generosity across participants. Chance level of the binary classification (high vs low donations) was 50%. In this and later searchlight decoding analyses, brain regions are reported as significant only if they yielded decoding accuracies significantly above chance at a stringent statistical threshold of p < 0.05 (FWE-corrected for multiple comparisons at the voxel level).

Post hoc permutation tests were performed for each identified cluster (using all voxels within a cluster) to illustrate how likely cluster-wise decoding accuracies were achieved by chance, compared with data-driven permutation-based null distributions. For each participant, permutation distributions were created by breaking up the mapping of labels (high or low donation) and response patterns (1000-fold). To allow for permutation tests at the group level, we then compared average “real” decoding accuracies (cluster-wise mean across participants) to sampled permutation null distribution (averaged across participants for each of the 1000-fold). Cluster-wise permutation-based null distributions were also created for all other whole-brain searchlight decoding analyses, verifying theoretical chance levels (e.g., of 50% for binary classifications) without exception. Complementary permutation tests, which further illustrated that cluster-wise decoding accuracies are unlikely achieved under the null hypothesis, are only displayed for the whole-brain decoding of high and low donations.

To compare these decoding results with classic univariate effects, we also estimated a GLM 5 that was identical to GLM 1 (including parameter estimates for “high” and “low” donations) with the exception that spatially smoothed functional data were used (Gaussian kernel of 8 mm FWHM). Individual contrast images were then used to compute random-effects group analyses using paired t tests as implemented in SPM8. Similar to the multivariate decoding approach, we applied a stringent statistical threshold of p < 0.05 (FWE-corrected).

In addition to binary classifications of high and low donations, we also examined whether trial-by-trial variations in donation size were encoded in multivoxel response patterns, using a support vector regression (SVR) analysis in combination with a searchlight approach (as outlined above). This analysis used trial-wise donations as labels and trial-wise response patterns (GLM 2) as input features of the prediction (for a detailed description of SVR, see below).

Neural decoding of social processes at the time of donation decisions.

Next, we examined whether predictive neural information on generous donations can be linked to the recruitment of social processes (measured after scanning) that, as reported below, were found to impact donations at the behavioral level. In particular, we tested whether the amount of reported empathy and perspective taking for the beneficiaries of each charity were encoded in brain regions that encoded the generosity of donation decisions. This was done as follows: For each rating obtained in the behavioral post-test, we applied a separate SVR analysis in combination with a searchlight approach as described above. To take full advantage of stimulus-specific ratings (e.g., the degree of perspective taking), we used trial-wise parameter estimates (R1-R60, GLM 2) as input for all linear kernel SVRs (ν-SVR; LIBSVM as used for binary classifications; http://www.csie.ntu.edu.tw/∼cjlin/libsvm) together with participants' ratings as labels. A slightly reduced cost parameter of c = 0.01 was used (preselected based on previous studies using a similar whole-brain searchlight SVR approach) (Kahnt et al., 2011a, 2014) to account for substantially increased computational costs compared with block-wise, binary SVC. Similar to the decoding of high and low donations, pattern vectors of four of the five donation blocks were used to train the regression model. This provided the basis for the prediction of the ratings of the 12 trials of the remaining donation block (independent test data) based on their neural response patterns. This procedure was repeated several times, always using trials of a different task block as test data (fivefold cross-validation). Information about the mental process of interest (e.g., perspective taking) was defined as the average Fisher's z-transformed correlation coefficient between the rating predicted by the SVR model and participants' actual ratings in these trials (Kahnt et al., 2011a, b, 2014; Gross et al., 2014). Similar to the SVC, this accuracy value was assigned to the central voxel of the searchlight cluster, and the procedure was repeated for every voxel of the measured brain volume. For each rating, resulting decoding maps of all participants were smoothed (8 mm FWHM) and submitted to random-effects group analyses. Statistical tests were restricted to brain regions predictive of the generosity of donations (see Table 2) using the inclusive masking option as implemented in SPM8. Only regions that passed the statistical threshold of p < 0.05 (FWE-corrected at voxel level) are reported.

As before, we compared these results with those of a univariate analysis by estimating a GLM 6 (based on smoothed data, 8 mm FWHM) with the following regressors: reading phases (R1), decision phases (R2), and the decision phase interacted with empathic responses (R3), perspective taking (R4), deservingness (R5), similarity (R6), closeness (R7), and opposition (R8). Movement parameters were included as regressors of no interest (R9-R14). Contrast images were calculated for each of the parametric regressors (R3-R8) and used for separate random-effects group analyses using single t tests as implemented in SPM8 (p < 0.05, FWE-corrected). We also repeated this GLM without the serial orthogonalization of R3-R8 as implemented in SPM.

Neural decoding of the subject-specific influence of empathy and perspective taking in donation decisions.

Behavioral regressions in the donation task (described below) revealed that the relative influence of empathy and perspective taking in donation decisions varied across individuals. Following up on this evidence, we examined whether these subject-specific contributions of empathy and perspective taking for donation decisions are related to participants' more general empathy and perspective taking responses, independent of altruistic giving. In particular, we tested whether task-elicited neural signatures of empathy and perspective taking can predict the degree to which participants rely on these processes in a separate donation task. Contrary to the previously described decoding analyses that tested for predictive neural information within each participant (across trials or conditions), this analysis focused on interindividual differences and hence used a cross-subject decoding approach. Moreover, instead of focusing on information contained in local spatially distributed response patterns (as identified in searchlight decoding approaches), this analysis used multivoxel representations across brain regions to characterize affective (empathy) and cognitive (perspective taking) mental states in individual subjects.

For each participant, based on functional data in the EmpaToM task, we created two independent contrast images that represented neural signatures for empathy and perspective taking (GLM 3). Two separate group-level analyses were implemented to identify voxels relevant for the encoding of both processes (i.e., feature selection independent of the donation task). For the 119 identified empathy-relevant voxels (see Table 5), we extracted parameter estimates from the subject-specific contrast image for empathy (yielding a 119-dimensional pattern vector for each participant). Subject-specific empathy-patterns were then used to train a linear support vector regression (ν-SVR, c = 1) together with participants' measures of the relative influence of empathy in the donation task as labels (β_empathy obtained in the behavioral regression of donations). Cross-validation was achieved by a split-half of the sample (i.e., n_train = 15, n_test = 15), and repeated 1000 times. Predictive accuracies were defined as the average Fisher's z-transformed correlation coefficient between the regression weights predicted by the SVR model and actual regression weights of participants in the test sample. To assess the statistical significance of the predictions, accuracies were contrasted against permutation-based null distributions (1000-fold) (paired-sample Wilcoxon signed rank test as implemented in MATLAB, The MathWorks). The decoding procedure was then repeated for perspective taking and the other post-test measures (separately), based on both the empathy-related and perspective taking-related responses in the EmpaToM task. Bonferroni correction was used to adjust p values for multiple comparisons across the total of 12 independent tests.

Neural decoding of stimulus-driven reorientation of attention.

To investigate the role of domain-general attentional shifts in altruistic decisions, we performed two different types of decoding analyses for attention data: within-attention task and cross-task decoding.

The within-attention task MVPA identified brain regions that encode stimulus-driven reorienting of attention using a whole-brain searchlight approach similar to the binary classification of generous and selfish donations described above. Parameter estimates of validly and invalidly cued target trials from GLM 4 were used to create trial-specific pattern vectors. Pattern vectors for 48 valid and 48 invalid trials were then assigned to four sets according to their presentation order. Matching the decoding of high and low donations, a linear SVC (c = 1) was trained on data of three sets and tested on independent data of the remaining fold. Average correct classification across the fourfold cross-validation was assigned to the central voxel of the searchlight. Participants' decoding maps were smoothed (8 mm FWHM) and tested at the group level against the chance level of 50% (p < 0.05, FWE-corrected).

The cross-task MVPA tested whether overlaps in areas predicting attention shifts and generous donations are indicative of shared neural code across both tasks. Thus, assuming that attentional shifts indeed contribute to altruistic giving, we hypothesized that neural responses that encode attentional reorienting might be reactivated during generous donations. Comparisons of neural activation patterns across tasks (Tusche et al., 2014; Corradi-Dell'Acqua et al., 2016) allowed us to explicitly test for common neural codes across both tasks. To this end, we performed ROI-based cross-task decoding using clusters in right TPJ and right pSTS as described in Table 2. Decoding parameters and neural response patterns were identical to the within-task decoding of donations and attention shifts, with the exception that we now trained a classifier on all response patterns of the attention task (validly vs invalidly cued targets) and tested it on response patterns of one donation block (low vs high donations). Decoding accuracies obtained in this fivefold out-of-sample prediction (i.e., of independent data of one particular donation block) were then averaged and reflected the amount of shared predictive information in the ROI.

For each ROI, we then tested whether participants' decoding accuracies were statistically significant above the chance level of 50% using permutation tests (1000-fold). Significant results (corrected for the number of ROIs) would point to shared information across both tasks and support the notion that predictive information on generosity in this brain region is, at least partly, due to reinstatement of neural codes recruited for domain-general processes, such as reorienting of attention.