Values encoded in orbitofrontal cortex are causally related to economic choices

Abstract

In the eighteenth century, Daniel Bernoulli, Adam Smith and Jeremy Bentham proposed that economic choices rely on the computation and comparison of subjective values¹. This hypothesis continues to inform modern economic theory² and research in behavioural economics³, but behavioural measures are ultimately not sufficient to verify the proposal⁴. Consistent with the hypothesis, when agents make choices, neurons in the orbitofrontal cortex (OFC) encode the subjective value of offered and chosen goods⁵. Value-encoding cells integrate multiple dimensions^6-9, variability in the activity of each cell group correlates with variability in choices^10,11 and the population dynamics suggests the formation of a decision¹². However, it is unclear whether these neural processes are causally related to choices. More generally, the evidence linking economic choices to value signals in the brain^13-15 remains correlational¹⁶. Here we show that neuronal activity in the OFC is causal to economic choices. We conducted two experiments using electrical stimulation in rhesus monkeys (Macaca mulatta). Low-current stimulation increased the subjective value of individual offers and thus predictably biased choices. Conversely, high-current stimulation disrupted both the computation and the comparison of subjective values, and thus increased choice variability. These results demonstrate a causal chain linking subjective values encoded in OFC to valuation and choice.

Extended Data Figure 1.

Exp.2, control for choice frequency. We noticed that across sessions the difference in value range (ΔV_A–ΔV_B) was correlated with the fraction of trials in which the animal chose juice A (% A choice) and with the relative value (ρ). In principle, these correlations could represent confounding factors. Indeed, 50 μA stimulation could partly disrupt the valuation process. As a result, the animal might respond by defaulting to the juice type most frequently chosen in that session, or to the preferred juice type. If so, the range-dependent bias would be akin to the order bias (Exp.1), in the sense that it would result from functional disruption as opposed to facilitation. To address this concern, we identified a subset of sessions for which choices between the two juices were split almost evenly. In this subset of sessions, the difference in value range and the fraction of A choices were not correlated. We reasoned that if the range-dependent bias observed for the whole data set was driven by a default to the most frequently chosen option, the bias should disappear when the analysis was restricted to this subset of sessions. However, this was not the case. In fact, the range-dependent bias measured for the selected subset was larger than that measured for the entire population. We concluded that range-dependent biases did not reflect simple heuristics. A. Correlation between the difference in value range and the fraction of A choices. Each data point represents one session. Considering the entire data set (black data points, N=96 sessions), the two measures were significantly correlated (r ≥ 0.71, p<10⁻¹⁵, Pearson and Spearman correlation tests). We defined a small ellipse centered on coordinates [0, 50] (axes = [9, 14]). The ellipse identified a subset of data (pink data points, N=31 sessions) for which the difference in value range and the fraction of A choices were not correlated (p ≥ 0.69, Pearson and Spearman correlation tests). B. Correlation between the difference in value range and the relative value. Considering the entire data set, the two measures were significantly correlated (r ≥ 0.33, p ≤ 0.001, Pearson and Spearman correlation tests). However, when the analysis was restricted to the subset of sessions identified in panel A (pink data points), the correlation changed sign. C. Range-dependent bias, same data as in Fig.4CD. Considering the entire data set, the change in relative value was significantly correlated with the difference in value range (r ≥ 0.34, p≤0.0007, Pearson and Spearman correlation tests). The correlation did not dissipate when the analysis was restricted to the subset of sessions identified in panel A (pink data points; r ≥ 0.45, p≤0.01, Pearson and Spearman correlation tests). In this figure, data from the two animals are combined. Black and pink lines in the three panels were obtained from Deming regressions.

Extended Data Figure 2.

Exp.2, results obtained in paired sessions. In N=33 instances, we ran two back-to-back sessions offering the same two juices and leaving the electrode in place, but changing the quantity ranges such that ΔV_A–ΔV_B would differ. A. Example of paired sessions. B. Population analysis. Each pair of sessions in the scatter plot is connected by a line, of which we computed the slope. Data points filled in green correspond to sessions in panel A. Data from the two monkeys are pooled. Across the population, slopes were typically >0 (p = 0.007, two-tailed Wilcoxon signed-rank test). Hence, range-dependent biases were not dictated by the juice pair or by the location of the electrode within OFC.

Extended Data Figure 3.

Exp.2, analysis of response times (RTs). A. Example session 1. Each data point represents one trial type and the two lines were obtained from linear regressions. Under normal conditions (stimOFF, black), RTs decreased as a function of the chosen value (x-axis). Electrical stimulation (stimON, red) generally reduced RTs. Linear fits reveal that lower RTs were due to a lower intercept, as opposed to a steeper (i.e., more negative) slope. BC. Population analysis, monkey D (N=35). For each session, we regressed RTs onto the chosen value, separately for stimOFF and stimON trials. We then compared the intercepts and the slopes at the population level. The picture emerging from panel A was confirmed for the population. In panel B (intercept), each data point represents one session. The population is significantly displaced below the identity line (p=0.018, two-tailed Wilcoxon test). In panel C (slope), it can be noticed that the slope under stimulation was shallower (less negative), probably due to a floor effect. Filled data points correspond to the session shown in panel A. D. Example session 2. Same format as in panel A. EF. Population analysis, monkey G (N=61). Same format as in panels BC. Electrical stimulation significantly lowered the intercept but did not significantly alter the slope. Filled data points correspond to the session shown in panel D. In panels BCEF, values indicated in the insert refer to the difference between the stimON measure and the stimOFF measure, averaged across the population. All p values are from two-tailed Wilcoxon tests, and t tests provided very similar results.

Extended Data Figure 4.

Exp.1, range-dependent choice biases. ABC. Results obtained when electric current was delivered at 25 μA, 50 μA and ≥100 μA. In each panel, x- and y-axes represent the difference in value range (in uB) and the difference in relative value, respectively. Each data point represents one session. Sessions from the two animals and with different stimulation times (offer1 or offer2) were pooled. Gray lines were obtained from linear regressions. Each panel indicates the p values obtained from Pearson and Spearman correlation tests. In essence, the choice bias imposed by the stimulation (δρ) was correlated with the difference in value ranges (ΔV_A–ΔV_B) at low current (25 μA; weakly) and intermediate current (50 μA), but not at high current (≥100 μA).

Extended Data Figure 5.

Stimulation in Exp.2 did not systematically alter the sigmoid steepness. For this analysis, the two groups of trials (stimOFF, stimON) were examined separately (see Methods). The two axes represent the sigmoid steepness in the two conditions. Sessions from the two animals were pooled (N=95, 2 outliers removed), and each data point represents one session. The gray ellipse represents the 90% confidence interval. The p value is from a Wilcoxon test and similar results were obtained with a t test.

Extended Data Figure 6.

Exp.1, interpretation of the order bias. A. Decelerating response function. The black line represents an ideal response function, which relates the number of spikes emitted by a cell in a given time window (y-axis) to the synaptic current entering the cell (x-axis). In the condition highlighted in yellow, I_O is the synaptic current due to the offer on the monitor, r is the corresponding response, I_S is the synaptic current due to the stimulation, and δr is the corresponding increase in the number of spikes. The condition highlighted in blue is similar, except that I_O is larger (I_O,blue > I_O,yellow). Because the response function is decelerating, δr in the blue condition is smaller (δr_blue < δr_yellow). In Exp.1, only one good was presented at the time. Neurons associated with that good were naturally more active (higher I_O) than neurons associated with the other good. Thus deceleration in the response function induced a bias favoring the good not offered during the stimulation (order bias). For given I_O,yellow and I_O,blue, the difference δr_yellow – δr_blue increases with I_S. Hence, higher stimulation currents induced larger order biases. B. Concurrent presence of order bias and range-dependent bias. The cartoon illustrates an ideal session in Exp.1. We assume that under normal conditions there is no order bias (stimOFF, continuous lines). Thus the two sigmoids for AB trials and BA trials coincide. We also assume that stimulation is delivered during offer1, and that ΔV_A–ΔV_B > 0. The order bias separates the two sigmoids such that under stimulation the sigmoid for AB trials is on the left of that for BA trials (stimON, dashed lines). The range-dependent bias imposes a shift on the total sigmoid, including both AB and BA trials (not shown), which moves to the right compared to normal conditions. The two choice biases are complementary and independent.

Figure 1.

High-current stimulation of OFC disrupts valuation. A. Experiment 1, design. Offers, represented by sets of squares, appeared centrally and sequentially. In this trial, the animal chose between 2 drops of grape juice and 6 drops of peppermint tea. B. Example session 1. In half of the trials, we delivered 125 μA current during offer1. The panel illustrates the choice pattern for AB trials (red) and BA trials (blue), separately for stimOFF trials (light) and stimON trials (dark). Data points are behavioral measures and lines are from probit regressions (Eq.1). In each condition (stimOFF, stimON), the order bias (ε) quantified the distance between the two flex points. In stimOFF trials, a small order bias favored offer2 (ε_stimOFF =0.02). In stimON trials, the order bias increased (ε_stimON =0.07). Hence, stimulation biased choices in favor of offer2. CDE. Population results for stimulation during offer1 (N=29 sessions, ≥100 μA). Stimulation did not affect relative values (C); it did not consistently affect the sigmoid steepness (D); and it biased choices in favor of offer2 (E). F. Example session 2. Here 125 μA current was delivered during offer2. Stimulation induced a bias in favor of offer1 (ε_stimON<ε_stimOFF) and increased choice variability (shallower sigmoids in stimON trials; η_stimON<η_stimOFF). GHI. Population results for stimulation during offer2 (N=25 sessions, ≥100 μA). Stimulation did not affect relative values (G); it reduced the sigmoid steepness (H); and it biased choices in favor of offer1 (I). In panels CDEGHI, green symbols are from sessions shown in B and F; ellipses indicate 90% confidence intervals. All p values are from two-tailed Wilcoxon tests, and very similar results were obtained using t tests.

Figure 2.

Effects of electrical stimulation at different current levels. The whole data set includes N=29/22/29 sessions in which 25/50/≥100 μA were delivered during offer1, N=17/22/25 sessions in which 25/50/≥100 μA were delivered during offer2, and N=50 control sessions (0 μA; 194 sessions total). A. Relative value. B. Sigmoid steepness. C. Order bias. In each panel, blue and yellow refer to stimulation during offer1 and offer2, respectively. Data points are averages across sessions and error bars indicate SEM. Asterisks highlight measures that differed significantly from zero (all p<0.005, two-tailed Wilcoxon test). All other measures were statistically indistinguishable from zero (all p>0.05, two-tailed Wilcoxon test). Extended Data Table 1 provides the exact p values. Statistical analyses based on t tests provided very similar results.

Figure 3.

Prediction of range-dependent choice bias induced by electrical stimulation (facilitation). A. Experiment 2, design. Two offers are presented simultaneously. After a brief delay, the animal indicates its choice with a saccade. Electrical stimulation (50 μA) is delivered throughout offer presentation. BCD. Predictions for one example session. In OFC, the encoding of offer values is predominantly positive (higher activity for higher values). Panels B and C represent the (mean) tuning curves for pools of offer value A cells and offer value B cells under adapted conditions. Firing rates (y-axis) are plotted as a function of the offer values (x-axis) expressed in units of juice B (uB). Red horizontal lines represent the two value ranges, with ΔV_A>ΔV_B. The same firing rate interval δr corresponds to different value intervals, with δV_A>δV_B. Panel D represents choice patterns. Electrical stimulation increases both offer values, but the net effect is a choice bias in favor of juice A (δρ>0). Conversely, in sessions where ΔV_A<ΔV_B, δr induces δV_A<δV_B, and electrical stimulation biases choices in favor of juice B (δρ<0, not shown). See Methods.

Figure 4.

Range-dependent choice bias induced by neuronal facilitation of OFC. A. Example session 1. In this session, we set ΔV_A<ΔV_B. Consistent with the prediction, electrical stimulation biased choices in favor of juice B (δρ<0). B. Example session 2. In this case, we set ΔV_A>ΔV_B. Electrical stimulation biased choices in favor of juice A (δρ>0). CD. Population analysis. The two panels refer to the two animals. In each panel, the choice bias (δρ, y-axis) is plotted against the difference in value range (ΔV_A–ΔV_B, x-axis). Each data point represents one session, and the gray line is from a linear regression. Value ranges are expressed in units of juice B (uB). The two measures are significantly correlated in both monkey D (r=0.53, p=0.001, Pearson correlation test; r=0.49, p=0.003, Spearman correlation test) and monkey G (r=0.29, p=0.024, Pearson correlation test; r=0.36, p=0.005, Spearman correlation test). Green data points are from sessions illustrated in panels A and B.