Pooling in dynamic panel-data models: An application to forecasting GDP growth sates

in this article, we analyze issues of pooling models for a given set of N i ndividual units observed over T periods of time. When the parameters of the models are different but exhibit some similarity, pooling may lead to a re duction of the mean squared error of the estimates and forecasts. We invest igate theoretically and through simulations the conditions that lead to imp roved performance of forecasts based on pooled estimates. We show that the superiority of pooled forecasts in small samples can deteriorate as the sam ple size grows. Empirical results for postwar international real gross dome stic product growth rates of 18 Organization for Economic Cooperation and D evelopment countries using a model put forward by Garcia-Ferrer, Highfield, Palm, and Zellner and Hong, among others illustrate these findings. When a llowing for contemporaneous residual correlation across countries, pooling restrictions and criteria have to be rejected when formally tested, but gen eralized least squares (GLS)-based pooled forecasts are found to outperform GLS-based individual and ordinary least squares-based pooled and individua l forecasts.