Figure 2.

Behavioral Results

(A–C) Logistic regression results of experiment 1. Figures show how choices (0/1) of the option on the current trial are influenced by past rewards following choices of same option A (A); different options B or C (B); and again different options B or C, but depending on how often same option A had been chosen in the past 30 trials (C), depending on when the reward occurred relative to choice (bin 1, 0–0.5 s; 2, 0.5–1.5 s; 3, 1.5–2.5 s; 4, 2.5–3.5 s; 5, 3.5–4.5 s before reward). Values are mean ± SEM (across participants) of the regression coefficients obtained from the logistic regression.

(D) A separate linear regression shows that the average rate of responding in experiment 1 is, by definition, related to the rate of contingent rewards (CRs) but also to the rate of noncontingent rewards (NCRs), despite them being unrelated to behavior.

(E and F) Behavioral results of experiment 2. Using multiple logistic regression, we tested whether our instructions reliably induced contingent and noncontingent learning.

(E) Each box represents one condition, and each cell within a box represents a particular regressor. High parameter estimates are shown in white; low estimates in black. These regressors can be arranged into the lower quadrant of a square where the lead diagonal represents DIRECT learning (red), the next lower diagonal represents 1BACK learning (green), and the third diagonal 2BACK learning (yellow). For example, the first regressor in the top left box should receive loading if decisions under DIRECT instructions can be explained by a model in which any reward obtained on the previous trial (n – 1 column) is associated with the choice on that trial (n – 1 row). The plot shows that the DIRECT, 1BACK, and 2BACK conditions have predominantly yielded high parameter estimates in their respective red, green, and yellow regressors, while the FORWARD condition has led to loadings that are distributed across the different association types, as hypothesized.

(F) Averaging across the corresponding associations (red, green, and yellow diagonals, respectively) shows that the three different associations load differently depending on the instructed condition. See also Figures S1–S4.

All error bars represent SEM.