Decision bias and sampling asymmetry in reward-guided learning

Abstract

Human economic decisions are highly sensitive to contexts. Deciding between two competing alternatives can be notoriously biased by their overall value ( magnitude effect) or by a third decoy option ( distractor effect). Some prominent explanations appeal to diminishing value sensitivity and divisive normalization in value representations, i.e., representational bias, that feed into the choice stage. However, these explanations have recently come under scrutiny due to empirical inconsistencies and mounting alternative theories. Here, we posit that context-sensitive choices may not stem from representational biases but rather emerge as by-products of asymmetric sampling during value learning. In a reward-guided choice task, participants aimed to maximize cumulative rewards through trial and error. The task introduced alternating blocks with either a colored distractor or a neutral notional distractor. We observed decreased choice accuracy when higher-value distractors were present, a pattern that persisted even in the notional distractor blocks. Using computational modeling, we show that this phenomenon falls out naturally from a simple learning rule without relying on any additional mechanism such as divisive normalization or nonlinear utility. Furthermore, we found that, contrary to divisive normalization, choice accuracy was not influenced by distractor value but strongly depended on the magnitude of the targets values per se. This magnitude sensitivity was also found in the notional distractor conditions and could lawfully be reproduced by the learning model. Importantly, when counterfactual feedback eliminated sampling asymmetry, the observed decision bias vanished. Our results suggest that the genesis of context-sensitive choices may lie in the learning dynamics themselves, specifically sampling asymmetry, rather than in pre-decisional representational biases. This finding reframes the discourse on irrational decision-making, attributing it to acquired biases during the learning process, not necessarily computational intricacies at the choice stage.