(A) Subjective weights of influence of recent rewards on object choice, as derived from logistic regression. Filled symbols indicate significance (p<0.005, t-test; pooled across animals). (B) Subjective weights of influence of recent choice on object choice, as derived from logistic regression. (C) Approach for deriving subjective risk from the variance of recent reward history. Upper panel: vertical black bars represent rewarded choices for object A. Middle/lower panels: trial-by-trial estimates of value (middle) and risk (lower) calculated, respectively, as mean and variance of reward history over last 10 trials (using weights shown in A). The dashed magenta line indicates the subjective risk estimate used for neuronal and behavioral analysis. Heavy lines: running average of weighted estimates. Thin solid lines: unweighted, objective value/risk. Value from reward history was highest in the high-probability block, whereas risk was highest in medium-probability block (inverted U-shaped relationship, see Figure 1A). All units were normalized to allow for visual comparisons. (D) Positive influence of subjective risk on choice (risk-seeking attitude) and separate value and risk influences on choice. Object-choice probability increased with risk difference between objects (ΔRisk, sorted by median split; ‘high’ indicates higher risk with object A compared to object B; p<0.002 for all pair-wise choice probability comparisons between adjacent relative-risk levels, χ2-tests; N = 16,346 trials). The risk effect added monotonically and consistently to the increase of choice probability with object value difference (ΔValue). (E) Logistic regression. Coefficients (β) for relative value (ΔValue, p=4.4 × 10−39), relative risk (ΔRisk. p=2.8 × 10−4) and left-right bias (Bias, p=0.698) across sessions (t-tests, random-effects analysis). The constant (bias) was not significant, suggesting negligible side bias. The inset shows coefficients for a subset of trials where value difference was minimized (10% of trials); only risk difference was significantly related to choice (p=0.0072) but not value difference (p=0.182), data pooled across animals. (F) Psychometric functions relating relative value and risk to choice probability (across animals and sessions).