A demonstration of construct validity based on neurometric information.

(A) A support vector machine algorithm was used to train the trust model. (B) An independent trust dataset was used to validate the trust model’s generalizability. (C) We tested the model on independent datasets such as the Balloon Analog Risk Task to assess the convergent and discriminant validity of the trust model. (D) We hypothesized that trust was associated with beliefs of safety, feelings of anticipated reward, and affect, but not other processes. The distances and thickness of edges represent the generalizability of the trust model to other domains, and the size of nodes represents sample size of the dataset from each domain.

The trust model and its performance in the training and validation dataset.

(A) The trust model is a whole brain pattern of voxel weights that can be linearly combined with new data to predict psychological levels of trust. We visualize the voxels that most reliably contribute to the classification using a bootstrap procedure (thresholded p < 0.005 uncorrected for visualization). (B) The receiver-operating-characteristic (ROC) plot highlights the sensitivity and specificity of the model in cross-validation and in an independent holdout dataset. (C) We plot the pattern expression, which is the spatial correlation between the cross-validated model and each participant’s average activity in the trust and distrust conditions based on their actual decisions in the game. Each participant (N=40) is depicted as a separate row with a line connecting the correlation values from each condition. The model correctly predicted the decisions of 36 out of 40 participants (90% accuracy).

Construct validity of the trust model and model generalizability.

(A) (1) Network plot illustrates that the trust model significantly generalizes to safety and affect datasets, but not to reward and other processing datasets. The distances and thickness of edges are weighted based on the rank of classification accuracy, and the size of nodes represents sample size of each dataset. (2) The forced-choice classification accuracy for each dataset within the four domains was shown in the bar plot. Only the safety and affect domains demonstrated above chance accuracy across datasets. (B) Trust model pattern expression differences between the two conditions in the: (1) trust testing datasets, (2) two safety datasets, (3) two affect datasets, (4) four reward datasets, as well as (5) five datasets involving cognitive control and social cognition.

Spatial pattern similarity across all brain data from fifteen datasets.

(A) Visualization of spatial similarity of whole brain maps from 1,484 participants from 15 datasets using UMAP nonlinear manifold learning. Each dot represents a beta map from each participant. Parameters were arbitrarily selected to aid in visualization (n-neighbors=50, minimum distance=0.001). Trust was more similar to safety, no-reward, neutral and positive affect; whereas distrust was more similar to risk, reward, and negative affect. (B) Hierarchical clustered heatmap of correlation across the mean spatial pattern from each condition (31 conditions from 15 datasets) also revealed similar findings as above. The dendrogram colors indicate how the tasks cluster together. (C) Independent component analysis (ICA) showed that the two safety, the two trust, three no-reward/loss conditions, and one neutral affect condition loaded positively on the first component, while the two risk, the two distrust, three reward conditions, and two negative affect conditions loaded negatively on this component. See Figure S1 for the other two components

Basic information of each dataset as well as forced-choice classification accuracy and p values for each generalization testing.

Condition loadings on each of the three ICA components.

(A) The two trust conditions, safe conditions, no-reward/loss conditions, and one neutral affect condition loaded positively on the first component. By contrast, the two distrust conditions, risk conditions, reward conditions, and one negative affect condition loaded negatively on the first component. (B) Spatial patterns for each of the three components.