Figures and data
Task timeline.
Subjects chose between two snack food lotteries on each trial. Subjects learned about the lotteries through random food draws. Every 4-8 seconds, subjects sampled a new draw from each lottery. They were allowed to sample as many times as they wanted but were incentivized to sample approximately 7 draws per trial. Sampled food draws were presented for 2 seconds, followed by a fixation cross appearing for 2-6 seconds with random jitter. The trial ended when the subject chose the left or right lottery, using the respective index finger. Upon making their choice, subjects were presented with a food drawn from their chosen lottery.
Example trial with the sampled value and accumulated value.
The sampled value (ΔSV; red) and accumulated value (ΔAV; black) are plotted for this example trial. For the first draw, ΔSV and ΔAV are identical. However, as the trial proceeds, the two signals diverge. In the model, a choice is made when the |ΔAV| reaches a pre-specified decision boundary.
Choice data.
(a) The probability of choosing left based on ΔSV and ΔAV. As the value difference becomes greater in favor of one option, the probability of choosing that option increases, for both ΔSV and ΔAV. (b) The effect of gaze on choice. The longer that subjects looked at one lottery over the other, over the course of the whole trial, the more likely they were to choose that lottery.
Regions responding to sampled value, accumulated value, and their gaze-weighted variants.
(a) vmPFC showed a significantly positive correlation with |ΔSV|. (b) Pre-SMA, IPS, and dlPFC showed significantly positive correlations with lagged |ΔAV|. (c) vmPFC and striatum showed significantly positive correlations with |ΔSVGaze|, as did pre-SMA, IPS, and dlPFC (not pictured). (d) Pre-SMA showed a significantly positive correlation with lagged |ΔAVGaze|.
GLM1 beta plots from the vmPFC, striatum, pre-SMA, IPS, and dlPFC.
Displayed are regression coefficients from each region for (a) non-gaze-weighted signals: absolute sampled value difference (|ΔSV|) and absolute lagged accumulated value difference (|ΔAV|), and (b) gaze-weighted signals: gaze-weighted sampled value (|ΔSVGaze|) and absolute lagged gaze-weighted accumulated value (|ΔAVGaze|). vmPFC shows a pattern of BOLD activity that scales positively with both |ΔSV| and |ΔSVGaze|, but does not respond to |ΔAV| or |ΔAVGaze|. The striatum shows sensitivity only to |ΔSVGaze|. In contrast, the pre-SMA, IPS, and dlPFC show strong positive correlations between BOLD activity and |ΔAV|, but minimal relation with |ΔSV|. The pre-SMA additionally shows sensitivity to |ΔAVGaze|, indicating that accumulated value representations in this region are modulated by the history of gaze allocation. Note: Bar heights reflect mean betas averaged across all voxels within each ROI. Statistical significance was determined using permutation tests with FWE correction, which identify spatially localized, reliable effects within each ROI. For the striatum, while |ΔSV| shows a numerically larger overall mean, only |ΔSVGaze| produced a spatially consistent cluster of voxels surviving FWE correction (118 voxels, p = 0.010).
Non-uniform temporal weighting.
In both gaze-weighted and non-gaze-weighted models, participants showed stronger recency than primacy effects, both in terms of (A) the model parameters, and (B) the resulting temporal weighting functions averaged across all trials. Error bars are standard errors clustered by participant.
Sample-level correlations between value signal regressors.
Pearson correlations computed across samples for different quantifications of sampled value (ΔSV) and accumulated value (ΔAV). Values represent correlations between absolute magnitudes of each measure. |ΔSV| = absolute sampled value difference; |ΔAV| = absolute accumulated value difference; lagged |ΔAV| = absolute accumulated value difference up to the previous sample; |ΔSVGaze| = gaze-weighted sampled value; |ΔAVGaze| = gaze-weighted accumulated value; lagged |ΔAVGaze| = gaze-weighted accumulated value from the previous sample.
