Marginal stochastic choice

Models of stochastic choice typically use conditional choice probabilities given menus as the primitive for analysis, but in the field these are often hard to observe. Moreover, studying preferences over menus is not possible with this data. We assume that an analyst can observe marginal frequencies of choice and availability, but not conditional choice frequencies, and study the testable implications of some prominent models of stochastic choice for this dataset. We also analyze whether parameters of these models can be identified. Finally, we characterize the marginal distributions that can arise under two-stage models in the spirit of Gul and Pesendorfer [2001] and of kreps [1979] where agents select the menu before choosing an alternative.