Identification of Panel Data Models with Endogenous Censoring

Abstract

This paper analyzes the identification question in censored panel data models, where the censoring can depend on both observable and unobservable variables in arbitrary ways. Under some general conditions, we derive the tightest sets on the parameter of interest. These sets (which can be singletons) represent the limit of what one can learn about the parameter of interest given the model and the data in that every parameter that belongs to these sets is observationally equivalent to the true parameter. We consider two separate sets of assumptions, motivated by the previous literature, each controlling for unobserved heterogeneity with an individual specific (fixed) effect. The first imposes a stationarity assumption on the unobserved disturbance terms, along the lines of Manski (1987), and Honore (1993). The second is a nonstationary model that imposes a conditional independence assumption. For both models, we provide sufficient conditions for these models to point identify the parameters. Since our identified sets are defined through parameters that obey first order dominance, we outline easily implementable approaches to build confidence regions based on recent advances in Linton et.al.(2010) on bootstrapping tests of stochastic dominance. We also extend our results to dynamic versions of the censored panel models in which we consider lagged observed, latent dependent variables and lagged censoring indicator variables as regressors.

Keywords: panel data, partial identification, endogenous censoring

JEL Classification: C23, C24, C32