The measurement error problem in dynamic panel data analysis: Modeling and GMM estimation

The Generalized Method of Moments (GMM) is discussed for handling the joint occurrence of fixed effects and random measurement errors in an autoregressive panel data model. Finite memory of disturbances, latent regressors and measurement errors is assumed. Two specializations of GMM are considered: (i) using instruments (IVs) in levels for a differenced version of the equation, (ii) using IVs in differences for an equation in levels. Index sets for lags and lags are convenient in examining how the potential IV set, satisfying orthogonality and rank conditions, changes when the memory pattern changes. The joint occurrence of measurement errors with long memory may sometimes give an IV-set too small to make estimation possible. On the other hand, problems of 'IV proliferation' and 'weak IVs' may arise unless the time-series length is small. An application based on data for (log-transformed) capital stock and output from Norwegian manufacturing firms is discussed. Finite sample biases and IV quality are illustrated by Monte Carlo simulations. Overall, with respect to bias and IV strength, GMM inference using the level version of the equation seems superior to inference based on the equation in differences.