“…Such tests must rely either on variation in the agent's menu or on direct payoff information, and cannot be based solely on the agent's behavior in a single menu. Consistent with this observation, several studies consider the problem of testing whether a dataset consisting of choices and utilities over outcomes (where the latter is either given or inferred) from multiple decision problems is consistent with costly information acquisition, and if so, whether one can use these data to make inferences about the agent's cost of information (e.g., Caplin and Dean, 2015;de Oliveira et al, 2017;Caplin et al, 2020;Chambers, Liu, and Rehbeck, 2020;Dewan and Neligh, 2020;Dillenberger, Krishna, and S 2020;Lin, 2022). By studying the opposite question, namely, how one can use cost information to predict the agent's behavior, our paper shows one cannot hope for a more parsimonious class of tests than those considered by this literature.…”