Fig. 4: Gaze sequences predict out-of-sample trial-by-trial cooperation.

a Gaze sequence: Differences in cooperation rates between trials in which an AOI is never looked at and trials during which it is sampled during the first, second, third or last fixation. Each dot represents the difference in cooperation rate for one participant, the error bars represent the average across participants and the s.e.m. Stars represent significant two-tailed paired t-tests (p < 0.00025 (***), p < 0.0025 (**), p < 0.0125 (*), using a Bonferroni correction) between cooperation rates for trials on which the AOI is never looked at and trials on which it is sampled at the specific time-point. b Quality of fit (mean square error) of different classifiers predicting the outcome (cooperate vs defect) on each trial using gaze patterns. The quality criterion (mean squared error) is computed using the average across trials for each participant (left) or the average across participants for each trial (right) and used to compare models capturing data the best. Models using the first four and last fixation combined with payoff values provide the best out-of-sample predictions. c Out-of-sample predictions of the winning classifier (payoffs; fixations 1 to 4 and last). The predictions are averaged across trials for each participant (left) or across participants for each trial (right) and plotted against the average of the true labels. The models are fitted using a leave-one-participant-out procedure, where data from all but one participant are used for training the classifier. The trained classifier is used to predict the cooperation of the left-out participant on each trial (out-of-sample predictions), which are displayed here, showing that the first four and last fixation combined with payoff values are good predictors of cooperation. The dotted lines represent chance levels, computed with permutation tests. All panels: N = 88 participants, except for (c), right N = 96 games across 88 participants.