The neural basis of belief updating and rational decision making

Abstract

Rational decision making under uncertainty requires forming beliefs that integrate prior and new information through Bayes' rule. Human decision makers typically deviate from Bayesian updating by either overweighting the prior (conservatism) or overweighting new information (e.g. the representativeness heuristic). We investigated these deviations through measurements of electrocortical activity in the human brain during incentivized probability-updating tasks and found evidence of extremely early commitment to boundedly rational heuristics. Participants who overweight new information display a lower sensibility to conflict detection, captured by an event-related potential (the N2) observed around 260 ms after the presentation of new information. Conservative decision makers (who overweight prior probabilities) make up their mind before new information is presented, as indicated by the lateralized readiness potential in the brain. That is, they do not inhibit the processing of new information but rather immediately rely on the prior for making a decision.

Keywords: Bayesian updating; LRP; N2; conservatism; representativeness heuristic.

Fig. 1

Experimental design. (A) Sequence of events during a single trial: fixation square (500 ms), prior probabilities (2000 ms), blank screen (500 ms), fixation square (500 ms) and sample of drawn balls. The latter vanished 200 ms after the participant’s response, followed by an intertrial interval (blank screen, 1500 ms). Balls shown here as black and white were green and blue in the screen, respectively. (B) k determines the prior for urn A (p_A), and m is the observed number of blue balls. For each (k,m) combination, the upper left-hand entry gives the posterior odds for urn A, and the remaining three entries are the prescriptions for Bayesian updating (upper right hand), conservativeness (lower left hand) and the representativeness heuristic (lower right hand). Light- and dark-shadowed entries indicate where Bayesian updating conflicts with the representativeness heuristic and conservativeness, respectively.

Fig. 2

The LRP. (A) Grand-average stimulus-locked LRP waveforms (k = 3 minus k = 1). Time zero corresponds to sample presentation; time −3000 to the presentation of prior probabilities. The more positive the value of the LRP amplitude, the stronger is the hand-specific activation of the response with the higher prior probability. Light gray bar indicates time window used for analyses. (B) Grand-average stimulus-locked LRP waveforms for participants with low vs high rates of conservative errors (median split).

Fig. 3

The N2 component at electrode FCz. (A) Grand-average N2 across all participants depending on situation type. Time zero corresponds to sample presentation. Light gray bar indicates time window used for analyses. To ensure that the increased negativity for conflict situations was indeed due to an increased N2, we considered the spatial distribution of the difference between neutral and conflict situations (for the representativeness heuristic) in this and subsequent time ranges. The picture shows that the conflict effect has the typical frontocentral distribution of an N2. Activity represents the difference between neutral and conflict situations in the indicated time ranges. (B) Grand-average N2 waveforms for participants with low vs high rate of errors in situations with a conflict between Bayesian updating and the representativeness heuristic (median split).