Figure 2.
Choice history is systematically associated with the monkeys' current behavior. a, We use a GLM to analyze the statistical effect of different covariates on the monkeys' choices for n = 75 sessions (n = 41 for Monkey 1; n = 34 for Monkey 2). b, The choice prediction performance (aROC; see Materials and Methods) evaluated only for no-signal trials for different covariates: the choice on the previous trial (Cn-1, mean CPP 0.609, p < 10−8); target on the previous trial (Tn-1, mean CPP 0.525, p < 0.001); whether the preceding trial was rewarded (win, Wn-1, mean CPP 0.522, p < 0.01), the spike count on the current trial (Sn, mean CPP 0.564, p < 10−7); the spike count on the previous trial (Sn-1, mean CPP 0.504, p = 0.378). For the full model, CPP is 0.666, p < 10−10. The p values indicate whether CPPs were significantly different from chance performance (0.5) using Wilcoxon signed rank tests. c, The weights for the target on the previous trial (Tn-1) are plotted against those of the previous choice (Cn-1) for all n = 75 sessions (circles: Monkey 1, n = 41; squares: Monkey 2, n = 34). Green symbols represent means across all sessions for each monkey. d, The mean choice prediction performance converted to correlation coefficients (choice correlation; see Materials and Methods) are plotted for models incorporating the variables on the x-axis cumulatively. From left to right the variables used are as follows: (Cn-1); (Cn-1, Tn-1); (Cn-1, Tn-1, Wn-1); (Cn-1, Tn-1, Wn-1, Sn-1); (Cn-1, Tn-1, Wn-1, Sn-1, Sn). The height of each bar reflects the incremental improvement of the model prediction caused by the variable plotted on the x-axis. Significant increments are caused by Cn-1 (p < 10−8), Tn-1 (p < 10−5), Wn-1 (p = 0.046), Sn (p < 10−4), but not additionally by Sn-1 (p = 0.954), Wilcoxon signed rank tests for all. The horizontal dashed line marks the value for a model incorporating only Sn. Error bars are ±1 SE; Colors as in b.