Identification, Estimation and Testing in Panel Data Models with Attrition: The Role of the Missing at Random Assumption

This paper discusses identification, estimation and testing in panel data models with attrition. We focus on a situation which often occurs in the analysis of firms: Attrition (exit) is endogenous and depends on the outcomes of an observed stochastic process and the interest-parameters characterizing this process. Thus attrition is non-ignorable even if selection is based only on observed variables - that is, even if the missing items are missing at random (MAR). The likelihood function obtained by ignoring the attrition mechanism is a pseudo likelihood function. Assuming that the MAR condition holds, this paper establishes conditions for identification and consistent estimation based on the pseudo likelihood function. It is also shown that the MAR hypothesis has testable implications in many situations that are encountered in practice. Simulations suggest that in the case of the autoregressive model with random effects, the efficiency of the pseudo likelihood estimator (based on normality) is not much affected even by strong departures from normality. In a variety of simulation models, the pseudo likelihood estimator clearly outperforms the moment estimators - even when the latter are consistent.