Parietal cortex and insula relate to evidence seeking relevant to reward-related decisions

Abstract

Decisions are most effective after collecting sufficient evidence to accurately predict rewarding outcomes. We investigated whether human participants optimally seek evidence and we characterized the brain areas associated with their evidence seeking. Participants viewed sequences of bead colors drawn from hidden urns and attempted to infer the majority bead color in each urn. When viewing each bead color, participants chose either to seek more evidence about the urn by drawing another bead (draw choices) or to infer the urn contents (urn choices). We then compared their evidence seeking against that predicted by a Bayesian ideal observer model. By this standard, participants sampled less evidence than optimal. Also, when faced with urns that had bead color splits closer to chance (60/40 versus 80/20) or potential monetary losses, participants increased their evidence seeking, but they showed less increase than predicted by the ideal observer model. Functional magnetic resonance imaging showed that urn choices evoked larger hemodynamic responses than draw choices in the insula, striatum, anterior cingulate, and parietal cortex. These parietal responses were greater for participants who sought more evidence on average and for participants who increased more their evidence seeking when draws came from 60/40 urns. The parietal cortex and insula were associated with potential monetary loss. Insula responses also showed modulation with estimates of the expected gains of urn choices. Our findings show that participants sought less evidence than predicted by an ideal observer model and their evidence-seeking behavior may relate to responses in the insula and parietal cortex.

Figure 1.

Beads task design. Each sequence of bead presentations was preceded by an instruction screen informing the participant of the color probability (0.8 or 0.6 majority bead color) and the potential monetary loss for incorrect urn choices ($10 or $0). After the presentation of each color, participants either chose to draw another bead or they chose an urn. Draw choices led to the presentation of another bead color and cost $0.25, while urn choices led to a feedback screen which displayed either a win of $10 for correct urn choices (blue is correct in this example) or a loss of $0 or $10 for incorrect urn choices.

Figure 2.

Behavioral performance as a function of color probability and loss. a, Mean and SE of the number of draws chosen by participants. b, Optimal number of draws derived from ideal observer. Although the same ideal observer model was always used to compute optimal performance for all sequences, we organized our results so that ideal observer performance scores could be separately matched to each participant's performance. These participant-matched scores were used to compute SEs (error bars) and statistical contrasts with participants. c, Mean and SE of the participants' accuracy. d, Optimal accuracy for ideal observer (SEs computed over participant-matched scores. e, The cost of incorrect urn choices C_w, estimated from participant behavior using parameterized Bayesian model.

Figure 3.

Urn choices versus draw choices contrast. a, Results of contrast urn choices > draw choices thresholded at p < 0.001 uncorrected, as observed in anterior cingulate, insula and striatum (left), and bilateral parietal cortex (right). b, Mean beta values and 90% confidence intervals at peak coordinates of areas in a, as a function of color probability and loss.

Figure 4.

Parietal cortex is associated with number of draw choices. a, Areas in right dorsolateral prefrontal cortex (left) and right parietal cortex (right) where beta values corresponding to urn choices showed positive linear relationships with average number of draws, thresholded at p < 0.001 uncorrected. b, Scatter plot of beta values for draw choices at the peak voxel in right parietal cortex, showing no positive associations with the number of draws. c, Scatter plot of beta values for urn choices for the same peak voxel, showing a positive association with the number of draws separately for all four conditions.

Figure 5.

Parietal cortex was associated with the majority color probability behavioral effect. a, Area in right parietal cortex where individual differences in the beta value contrast 0.6 > 0.8 (effect of color probability) predicted corresponding differences in the number of draw choices (0.6 > 0.8), thresholded at p < 0.001. b, Scatter plot showing no relationship between the color probability fMRI contrast (beta values at the peak voxel in right parietal cortex) for draw choices and the size of the probability effect on the number of draws. c, Scatter plot showing a positive linear relationship between the probability fMRI contrast (beta values at the peak voxel in right parietal cortex) for urn choices and the size of the color probability effect on the number of draws.

Figure 6.

Main effect of loss. a, The beta value contrast $10 > $0 yielded two areas: left posterior insula (left) and right inferior parietal cortex (right), thresholded at p < 0.001 uncorrected. b, Beta values for draw choices are shown for left posterior insula (left) and right inferior parietal cortex (right). Bars show the mean beta values and error bars show the 90% confidence intervals. c, Mean beta values and 90% confidence intervals for urn choices are greater for $10 losses than $0 at peak voxels in left posterior insula (left) and right inferior parietal cortex (right).

Figure 7.

Parametric modulation of urn choice responses by model-based estimate of value of the chosen action. We fit a parameterized version of our Bayesian model to the choices of our participants and thereby derived model estimates of the value Q associated with each chosen action. These “action values” were then related statistically to the fMRI data. a, Urn choice responses in right posterior insula modulated by action values Q, thresholded at p < 0.001 uncorrected. These results derived from testing whether the mean across all four conditions differed from zero. b, Means and 90% confidence intervals for action value modulation (beta values) as a function of color probability and loss for bead events leading to draw choices (left) and urn choices (right).