Right Temporoparietal Junction Underlies ... - Journal Of Neuroscience

Computational modeling

To examine how participants evaluated payoffs of each party and integrated them into a subjective value (SV), we compared the following eight models with different utility functions characterizing participants' choices.

Model 1 was adapted from a recent study on moral decision-making by Crockett et al. (2014, 2017), which could be formally represented as follows: where SV denotes the SV of the given trial if the participant chooses to accept. For rejection trials, SV is always 0 given the rule of the task (i.e., neither beneficiaries would gain the money; same for all models). Ms and Mo represent the payoff (gain or loss) for oneself and payoffs donated to the corresponding association. α (0 < α < 1) is the unknown parameter of social preference that arbitrates the relative weight on the payoff for the participant in the decision. θ (0 < θ < 1) is the unknown parameter characterizing the audience effect, which is modulated by an indicator function q (0 for private, 1 for public; same below). This model assumes that the subjective value was computed as a weighted summation of personal payoffs and payoffs donated to the association, and that people cared less about their own payoffs but increased the weights on the benefits donated to the association in public (vs private). Model 2 was similar to Model 1 except that it adopted two separate α values depending on the moral context in that trial.

Model 3 has a logic similar to that of Model 1 and was built on studies adopting a donation task (Lopez-Persem et al., 2017; Qu et al., 2020), as follows: where α and β are unknown parameters that capture the weight of the payoff for either the participant or the association involved in the trial (−20 < α, β < 20). Again, θ (0 < θ < 10) describes the audience effect, which is represented by the indicator function q. Model 4 was similar to Model 3 except that it adopted two separate pairs of α and β according to the association involved in that trial (i.e., good cause or the bad cause).

Models 5–8 were established on the basis of the Fehr–Schmidt model (Fehr and Schmidt, 1999), as follows: where α and β measure the degree of aversion to payoff inequality in disadvantageous and advantageous situations respectively (i.e., how participants dislike that they themselves gained less/more than the association; 0 < α, β < 5). Among them, Model 5 adopted a fixed pair of α and β values in all four conditions. Model 6 and Model 7 took different pairs of α and β values either in terms of the audience or the moral context. Model 8 assumed that people showed distinct advantageous and disadvantageous inequality aversion that changed in each of the four conditions.

Given the softmax rule, we could estimate the probability of making a moral choice (i.e., accept in the Good context or reject in the Bad context) as below: where τ refers to the inverse softmax temperature (0 < τ < 10), which denotes the sensitivity of an individual's behavior to the difference in SV between moral and immoral choices.

We leveraged a hierarchical Bayesian analysis (HBA) approach (Gelman et al., 2014) to fit all the above candidate models via the “hBayesDM” package (Ahn et al., 2017). In general, HBA has several advantages over the traditional maximal likelihood estimation approach such that it could provide more stable and accurate estimates, and estimate the posterior distribution of both the group-level and individual-level parameters simultaneously (Ahn et al., 2011). The hBayesDM package performs a full Bayesian inference and provides actual posterior distribution using a Markov chain Monte Carlo (MCMC) sampling manner through the Stan language (Gelman et al., 2015). Conforming to the default setting in this package, we assumed that the individual-level parameters were drawn from a group-level normal distribution: individual-level parameters ∼ normal (μ, σ). We fit each candidate model with four independent MCMC chains using 1000 iterations after 2000 iterations for the initial algorithm warmup per chain that results in 4000 valid posterior samples. The convergence of the MCMC chains was assessed through Gelman–Rubin R-hat Statistics (Gelman and Rubin, 1992).

For model comparisons, we computed the leave-one-out information criterion (LOOIC) score for each candidate model (Bault et al., 2015). LOOIC score provides the estimate of out-of-sample predictive accuracy in a fully Bayesian way, which makes it more reliable than the point estimate information criterion [e.g., Akaike information criterion (AIC)]. By convention, the lower LOOIC score indicates better out-of-sample prediction accuracy of the candidate model. A difference score of 10 on the information criterion scale is considered decisive (Burnham and Anderson, 2004). We selected the model with the lowest LOOIC as the winning model for subsequent analysis of key parameters. A posterior predictive check was additionally implemented to examine the absolute performance of the winning model. In other words, we tested whether the prediction of the winning model could capture the actual behaviors. In terms of the actual trial-wise stimuli sequences, we used each individual's joint posterior MCMC samples (i.e., 4000 times) to generate new choice datasets correspondingly (i.e., 4000 choices per trial per participant). Then we calculated the mean proportion of moral choices of each experimental condition in these new datasets for each subject, respectively. We performed a Pearson correlation to examine to what degree the predicted proportion of moral choice correlated with the actual proportion across individuals in each condition, respectively.

fMRI data preprocessing

Functional imaging data were analyzed using SPM12 (Wellcome Trust Center for Neuroimaging, University College London). The preprocessing procedure followed the pipeline recommended by SPM12. In particular, functional images (EPI) were first realigned to the first volume to correct motion artifacts, unwarped, and corrected for slice timing. Next, the structural T1 image was segmented into white matter, gray matter, and CSF with the skull removed, and coregistered to the mean functional images. Then all functional images were normalized to the Montreal Neurologic Institute (MNI) space, resampled with a 2 × 2 × 2 mm3 resolution, in terms of parameters generated in the previous step. Last, the normalized functional images were smoothed using an 8 mm isotropic full-width at half-maximum based on a Gaussian kernel.

Within-subject representational similarity analyses

To clarify what information rTPJ exactly represents during the decision period that distinguished ASD participants from HC participants, we conducted a within-subject RSA in Python 3.6.8 using the nltools package (version 0.3.14; https://github.com/cosanlab/nltools). Some preparation was performed before implementing RSA. In particular, we established a trial-wise general linear model (GLM) for each participant, which included the onsets of the decision screen with the duration of decision time of each valid trial. Here, valid trials were those that conformed to neither the exclusion criterion for the behavioral data (trials with extremely fast or slow responses; see above for details) nor the fMRI data (trials in runs with excessive head motion). The onsets of button press and invalid trials were also modeled as separate regressors of no interest. In addition, six movement parameters were added to this GLM as covariates to account for artifacts of head motion. The canonical hemodynamic response function was used and a high-pass temporal filtering was performed with a default cutoff value of 128 s to remove low-frequency drifts. After the parameter estimation, we built up the trial-wise contrasts that were used for subsequent RSA.

Our analyses concentrated on rTPJ given our hypotheses. Notably, we took two different ways to define the cluster of rTPJ to circumvent the potential effect of ROI selection on results. These included defining it via a whole-brain parcellation based on meta-analytic functional coactivation of the Neurosynth database (i.e., the parcellation-based ROI; https://neurovault.org/collections/2099/; including a total of 1750 voxels, with a volume of 2 × 2 × 2 mm3 per voxel; same below) or via a coordinate-based manner given a recent meta-analysis on neural correlates of ToM (Schurz et al., 2014; i.e., the coordinate-based ROI; a sphere with a radius of 10 mm centering on the MNI coordinates of 56/−56/18; 515 voxels in total).

We first extracted the parameter estimates (i.e., contrast value in arbitrary units) of rTPJ from these first-level contrast images of valid trials for each participant, respectively. Next, we constructed the individual-level neural representation distance matrix (RDM) by computing the pairwise correlation dissimilarity of activation patterns within this mask between each pair of valid trials. We also built up the same neural RDM for left TPJ (lTPJ) as a control region (i.e., the parcellation-based ROI, 1626 voxels in total; the coordinate-based ROI (Schurz et al., 2014), a sphere with a radius of 10 mm centering on the MNI coordinates of −53/−59/20; 515 voxels in total). In line with our research goal, we constructed two main cognitive RDMs in light of the trial-wise information of reputation (i.e., arbitrary code: 0 = Private, 1 = Public), and Moral Context (i.e., 0 = Bad, 1 = Good) by calculating the Euclidean distance between each pair of trials. We also built up two additional cognitive RDMs using the trial-wise information of payoffs for the participant (i.e., from 1 to 8 in step of 1), and payoffs for associations (i.e., from 4 to 32 in step of 4) as control subjects. These cognitive RDMs measured the dissimilarity between trials given corresponding information. Notably, we sorted all trials according to the order of Audience, Moral Context, payoff for the participant, and payoff for associations (the charity or the bad cause) to guarantee the information contained by both the neural and cognitive RDMs was matched with each other. To make these cognitive RDMs comparable, we rescaled them within the range from 0 (i.e., the most similar) to 1 (i.e., the most dissimilar). Then we performed a Spearman's rank-order correlation between the neural RDM and the cognitive RDM for each participant.

For the group-level statistical tests, we first implemented the Fisher r-to-z transformation on the Spearman's ρ, and then performed the permutation-based two-sample t test (i.e., the number of permutations was 5000) on these statistics between the two groups for each cognitive RDM separately. To further examine the robustness of these findings, we applied the above analyses using all 256 trials. To this end, a new GLM was established that modeled the onset of the decision screen of all trials to further construct the neural RDM. The remaining details and procedures were the same as mentioned above.