The Mundlak Estimator in a Panel Data Model with Serially Correlated Error Component Disturbances

This paper shows that the classic Mundlak (1978) result where the random effects estimator reduces to the fixed effects estimator when the regressors are all correlated with the individual effects, may not hold if the remainder disturbances have a general serial correlation variance-covariance matrix. This includes the popular AR(1), MA(1) and ARMA(p, q) processes for serial correlation. This is illustrated with an empirical example for the AR(1) case.