A solution to the feature binding problem for risky choice

Sequential sampling models describe the cognitive mechanisms at play in preferential decision making. These models can predict attention, choice, and response time in simple choices, but are currently unable to specify how decision-makers deliberate in more complex settings, such as those involving multi-branch gambles. To make rational decisions for such gambles, decision-makers need to compute interactions between payoffs and probabilities. Current sequential sampling models, which model information sampling as sequentially independent, are unable to capture within-branch attribute interactions, and thus make absurd predictions for multi-branch gambles. This is analogous to the feature binding problem in cognitive science, which involves the integration of the perceptual properties of objects. In this paper, we propose a solution to the feature binding problem for risky choice. Specifically, we propose that attribute sampling in multi-branch gambles is sequentially non-independent and that decision-makers are more likely to sample the probability of a branch if they observe a highly desirable payoff in that branch (and vice versa). We show that such a non-independent sampling process allows sequential sampling models to make utility-maximizing predictions. We test our model on data from four existing Mouselab and eye-tracking experiments, and two novel Mouselab experiments, and find that most participants display the non-independent attribute sampling proposed by our model. Additionally, we show that participants who display stronger non-independent sampling are also less likely to deviate from expected value/utility maximization. Overall, our results show how feature binding implemented in existing sequential sampling models can be used to predict sophisticated risky choice behavior.