Fig. 4: Information value grows with a conserved form of uncertainty.

a, Families of uncertainty measures hypothesized to influence behavior using reward magnitudes and probabilities, magnitudes only or probabilities only (for example, SD, range and entropy). b, We presented individuals with three offer types (safe, 25/50/25 and 50/50) to dissociate these measures. Plots show their hypothesized effects on the subjective value of information (left) and fit quality (right); a.u., arbitrary units. c–e, Human information seeking is motivated by an SD-like form of uncertainty. c, Mean difference in percentage choice of Info versus Noinfo offers separately for each offer type. Error bars represent ± s.e.; *P < 0.05; **P < 0.01; ***P < 0.001; signed-rank tests. d, Mean fitted GLM weights for the effect of each uncertainty type on the subjective value of information. Insets show similar results for the subsets of participants classified as risk averse, risk non-significant or risk seeking. To ensure that this plot is not biased for any specific uncertainty measure, participants were classified as risk averse/seeking if they had a significant negative/positive Unc[r] effect according to any of the three models (SD, range or entropy; Supplementary Table 2). Also, to illustrate the willingness to pay for information, fitted parameters were scaled based on the fitted effect of E[r] to convert them from units of log odds to units of reward (money). e, Model comparison with shuffle-corrected log likelihoods relative to the SD model. Error bars represent ±bootstrap s.e. (n = 2,000 bootstraps); three asterisks (***) indicate the 99.9% bootstrap confidence interval excluding 0. f–h, Same as c–e but for animals, showing similar results. Data in f and g are from animal R; insets in g show each additional animal. i–k, LHb neuron information-related activity scales with an SD-like form of uncertainty. i, The same LHb neuron from Fig. 1 but separating its responses by offer type, revealing stronger information-related activity for 50/50 offers. The shaded area represents ±s.e. j, Left, SD-like form of uncertainty revealed by mean cross-validated GLM weights from fits to attribute-responsive LHb neurons with significant Info × Unc[r] effects. Error bars represent ±s.e.; *P < 0.05; ***P < 0.001; signed-rank tests. Right, mean time course of the cross-validated Info × Uncertainty Type effect on neuronal activity, measuring the enhancement of information-related activity by 50/50 relative to 25/50/25 offers (Methods). Data are shown in the same format as in Fig. 3b,e. k, Model comparison favors an SD-like form of uncertainty. Data are shown in the same format as in e. l–n, Same as i–k for the Pal. Model comparisons were performed using all neurons where the model converged to a stable fit in all bootstrap samples (n = 373 LHb, n = 293 Pal). See Supplementary Table 6 for details of all tests and P values.