Estimating Choice Models with Unobserved Expectations over Attributes

Agents often make choices based on expectations about uncertain attributes, yet the information used to form these expectations is typically unobserved by researchers. We develop two methods to estimate discrete choice models in which agents form expectations using heterogeneous and unobserved information sets. Preferences are point-identified by approximating the latent information structure with a finite mixture of information types, or set-identified with partial information about expectations. Both approaches apply to individual- and market-level data with minimal assumptions on expectation formation. Applications to firms' exporting decisions and consumers' automobile purchases show that accounting for unobserved information substantially affects estimated preferences.