Latent dimensions in neural representations predict choice context effects

Choices are affected by the context of available alternatives, a phenomenon termed choice context effects. Current models of context effects require options to be described by two explicit numerical attributes. However, decision-makers might represent these options by additional latent attributes, which are hard to define a-priori. We propose to use participants neural representations to access the full attribute set they consider and predict context effects without modelling any explicit attributes. We first estimated the context effects elicited by lotteries using a behavioral sample. Then we recruited two fMRI samples with preregistered design to estimate the neural representations of each lottery without the context of choice. We predicted the context effects using only the similarity in neural representations between the individual lotteries, improving both out-of-sample and in-sample predictions compared to traditional methods. These neural representations encoded a mixture of explicit and latent attributes, previously inaccessible to researchers using only behavioral methods.