The Effect of Serial Correlation in Estimating Dynamic Panel Data Models

Abstract

There are several methods of estimating dynamic panel data models in the context of both micro-economic and macro-economic data.  This paper investigates the performance of five different estimators of dynamic panel data models (the random effect model).  A  Monte Carlo experiment was conducted when individual, N is large and time dimension, T is finite and the error component model is assumed to be serially correlated. The bias and Root Mean Square Error criterion were used to access the performance of different estimators under consideration. We find that the Anderson-Hsiao using lagged differences as instrument (AH(d)) performs better when the time dimension is small (T=5), Anderson-Hsiao using lagged levels as instrument (AH(l)) performs better when T is moderate(T=10) and the first step Arellano-Bond estimator (ABGMM1) outperforms all other estimators when T increases to 20, this confirms the work of Kiviet (1995)  and Judson-Owen(1996) that no estimator has been found to be appropriate choice in all circumstances. For a dynamic panel data with large time dimension we suggest that the first step Arellano-Bond Estimator (ABGMM1) Estimator is appropriate.  The result shows that the bias of the first step Arellano-Bond estimator (ABGMM1) estimate is severe with small time dimension and the ordinary Least Square (OLS) and Least Square Dummy Variable (LSDV) are also bias when T is small.  It was discovered that the effect of serial correlation is negligible irrespective of the order. Keywords: Autocorrelation, Dynamic Panel data, Econometric models, Generalized Method of Moment (GMM), Moving Average.