Influence of vmPFC on dmPFC Predicts Valence-Guided Belief Formation

Abstract

When updating beliefs about their future prospects, people tend to disregard bad news. By combining fMRI with computational and dynamic causal modeling, we identified neurocircuitry mechanisms underlying this optimism bias to test for valence-guided belief formation. In each trial of the fMRI task, participants (n = 24, 10 male) estimated the base rate (eBR) and their risks of experiencing negative future events, were confronted with the actual BR, and finally had the opportunity to update their initial self-related risk estimate. We demonstrated an optimism bias by revealing greater belief updates in response to good over bad news (i.e., learning that the actual BR is lower or higher than expected) while controlling for confounds (estimation error and personal relevance of the new information). Updating was favorable when the final belief about risks improved (or at least did not worsen) relative to the initial risk estimate. This valence of updating was encoded by the ventromedial prefrontal cortex (vmPFC) associated with the valuation of rewards. Within the updating circuit, the vmPFC filtered the incoming signal in a valence-dependent manner and influenced the dorsomedial prefrontal cortex (dmPFC). Both the valence-encoding activity in the vmPFC and its influence on the dmPFC predicted individual magnitudes of the optimism bias. Our results indicate that updating was biased by the motivation to maximize desirable beliefs, mediated by the influence of the valuation system on further cognitive processing. Therefore, although it provides the very basis for human reasoning, belief formation is essentially distorted to promote desired conclusions.SIGNIFICANCE STATEMENT The question of whether human reasoning is biased by desires and goals is crucial for everyday social, professional, and economic decisions. How much our belief formation is influenced by what we want to believe is, however, still debated. Our study confirms that belief updates are indeed optimistically biased. Critically, the bias depends on the recruitment of the brain valuation system and the influence of this system on neural regions involved in reasoning. These neurocircuit interactions support the notion that the motivation to maximize pleasant beliefs reinforces those cognitive processes that are most likely to yield the desired conclusion.

Keywords: DCM; belief update; computational modeling; optimism bias; value; vmPFC.

Figure 1.

Outline and examples of experimental trials. Each experimental trial consisted of four succeeding events. With respect to a specific adverse life event (e.g., suffering from cancer), subjects had to estimate the BR (eBR) and their own risk (E1). They were then presented with the actual BR and had the opportunity to estimate their own risk again (E2). After identical eBR and E1, the upper progression of the hypothetical trial example shows a BR lower than expected indicating good news, whereas the lower progression shows a BR higher than expected indicating bad news. EEs corresponded to the difference between the eBR and the actual BR and the update corresponded to the difference between the first and the second self-risk estimate. Note that, in both trial examples, the EE is 10 and the update is 8. For eBR, E1, and E2, subjects were instructed to use response buttons to adjust the displayed number to match their estimate as soon as the number font changed to green (after 2 s). Interstimulus intervals between eBR, E1, and E2, as well as intertrial intervals after E2, were jittered and consisted of a fixation cross (not shown here).

Figure 2.

Task performance and computational modeling. A, Bars show subjects' updates, that were significantly larger after good news (GOOD) than after bad news (BAD). White dots represent simulations of updates by the “biased” computational model that assumes asymmetric learning rates for good and bad news (αA, two free parameters, α and A). Gray dots indicate simulated updates resulting from the “unbiased” model that assumes identical learning rates for good and bad news (α, one free parameter, α). The simulated unbiased updates provide a normative benchmark for rational updating with learning rates estimated for each subject under consideration of her or his exact trial history. Error bars indicate SEs. B, Bayesian model comparison confirmed that the biased model αA best predicted subjects' updates. Model frequencies show that the majority of subjects were best described by the αA model above and beyond chance (red dashed line). Error bars indicate the posterior variance. C, Learning rates extracted from the winning model αA were significantly higher after good than bad news. Error bars indicate SEs. D, Optimism bias (update_GOOD − update_BAD) and A (estimated for each subject by the model αA) were significantly correlated (dots represent single subjects). E, F, Correlations between task variables separately for trials with good news (E) and those with bad news (F). *p < 0.05, **p < 0.01.

Figure 3.

Brain regions encoding errors and the valence of belief updating. A, When being confronted with the actual BR, errors in BR estimation (weighted by the PR) were tracked by the ACC, the IFG, the anterior insula, the middle orbital gyrus and the dlPFC. The line chart shows that the activity in the dlPFC (representative of all clusters) increased with decreasing error size (parametric modulation by error, negative correlation). Of all the involved regions, only in the dlPFC did the magnitude of the error tracking correlate with the general learning rate component α (see scatter plot). Therefore, subjects with a stronger error tracking in the dlPFC also more strongly adjusted their initial beliefs in response to errors. B, During the second risk estimation, the activity in the vmPFC encoded the valence of updating, adjusted for EE and PR. The gray box schematically illustrates the opposed valences of increasing updates after good and bad news (in this example, eBR = E1). After good news, large updates are favorable because they ultimately change beliefs toward lower risk estimates and small updates are unfavorable because they let the opportunity to improve risk estimates pass by. In contrast, after bad news, large updates are unfavorable because they ultimately change beliefs toward higher risk estimates and small updates are favorable because they prevent worsening of risk estimates. Resulting valences are summarized in the table below: unfavorable (U), mid (M), and favorable (F) updates. The line chart shows that the activity in the vmPFC tracked the positive valence because it increased with increasing update sizes after good news but decreased with increasing update sizes after bad news. The scatter plot shows that subjects with a stronger optimism bias also demonstrated a greater tracking of favorable updating in the vmPFC. In A and B, the line charts and the scatter plots were not used for statistical inference (which was performed in parametric modulation and covariate analyses within the SPM framework); they are shown solely for illustrative purposes. C, After demonstrating the valence effect with the more precise parametric modulation analysis presented in B, a simplified analysis of updating was conducted as a basis for DCM. Here, all trials were assigned to three valence categories: those with unfavorable (U), mid (M), and favorable (F) updates (adjusted for EE). Conjunction across these three categories revealed a distributed network involved in general updating, overlapping with the error tracking effect in the dlPFC. Comparing trials with favorable and unfavorable updates revealed the differential recruitment of the vmPFC and the dmPFC during updating. The line charts show contrast estimates in the dlPFC, vmPFC, and dmPFC, respectively.

Figure 4.

Neurocircuitry mechanisms underlying optimistic belief updating. A, Ten different dynamic causal models varying in intrinsic connectivity and contextual modulation (unfavorable and favorable updating, U and F) were specified. The model space encompassed three brain regions involved in updating: dlPFC, vmPFC, and dmPFC. B, Bayesian model selection revealed that the model m1 best explained subjects' BOLD signal above and beyond chance (red dashed line). In this model, the coupling between dlPFC and vmPFC was differentially modulated by unfavorable and favorable updating. Therefore, the vmPFC filtered the incoming information in a valence-dependent manner and furthermore influenced the dmPFC. C, Connectivity parameters derived from m1 show that the coupling between dlPFC and vmPFC tended to be weaker in the context of unfavorable relative to favorable updating. D, Optimism bias correlated with two parameters of m1 (highlighted in red): differential modulation of the dlPFC-vmPFC connection by favorable versus unfavorable updating (F-U) and the strength of the vmPFC-dmPFC connection (vmPFC::dmPFC). Therefore, subjects with a stronger optimism bias also demonstrated a greater valence-dependent filtering of incoming information by vmPFC and a greater transmission of this differential signal further to dmPFC.