To pool or not to pool: Do we really need to make a choice?

This paper revisits an old econometric issue “to pool or not to pool a panel” from a new perspective and proposes a pooling averaging estimator. To implement the pooling averaging technique, we provide a preliminary screening procedure (prior to averaging) to rule out the poorly specified models, which not only facilitates the computation and also improves the estimation accuracy. We examine the asymptotic property of a pooling averaging estimator based on the Mallows criterion and derive the upper bound of its risk. A variety of panel data estimators are compared in extensive simulation designs. We find that there does not exist a uniformly best estimator, and the performance of an estimator depends on the degree of heterogeneity, the sample size, and noises in the model. This explains, to some extent, the discrepancy between the theoretical justification of non-pooling methods and empirical evidence that advocates pooling. The method is also applied to examine the determinants of sovereign credit default swap spreads, and the empirical results echo our theoretical arguments.