Instrumental Variable Estimation of Dynamic Linear Panel Data Models with Defactored Regressors and a Multifactor Error Structure

This paper develops an instrumental variable (IV) estimator for consistent estimation of dynamic panel data models with a multifactor error structure when both N and T, the cross-sectional and time series dimensions respectively, are large. Our approach projects out the common factors from observed variables, the exogenous regressors of the model, using principal components analysis and then uses the defactored regressors as instruments to estimate the unknown parameters, as in a standard 2SLS procedure. The approach requires estimating solely the common factors contained in the regressors, leaving those that only influence the dependent variable into the errors. Hence our approach is computationally attractive. Since our estimator is based on instrumental variables, it is not subject to the Nickell bias that arises with least squares type estimators in dynamic panel data models. The finite sample performance of the proposed estimator is investigated using simulated data. The results show that the estimator performs well in terms of bias, RMSE and size. The performance of an overidentifying restrictions test is also explored and the evidence suggests that it has high power when the key assumption, strong exogeneity of (a subset of) the regressors, is violated.