Conditions initiales et estimation efficace dans les modèles dynamiques sur données de panel: Une application au comportement d'investissement des entreprises

Abstract

This paper provides a methodology for efficient estimation of dynamic panel data models under different assumptions concerning initial conditions and variance assumptions. In the analysis of panel data models that may involve lagged dependent variables, concerns about the initial conditions reflect the fact that most panel data sets have a relatively small time series dimension. Our model specification is a linear first-order dynamic model in which there are assumed to exist an identifying subset of strictly exogenous regressors. Under all but the most trivial assumption concerning the initial conditions the standard GLS procedure is inconsistent. We argue that our approach provides a natural procedure for testing assumptions concerning the process underlying the initial conditions. The methodology is applied to the estimation of a dynamic model of investment behaviour using panel data.