On Estimation of the Logarithm of the Mean Squared Prediction Error of A Mixed-effect Predictor

Abstract

: The mean squared prediction error (MSPE) is an important measure of uncertainty in small-area estimation. It is desirable to produce a second-order unbiased MSPE estimator, that is, the bias of the estimator is o ( m − 1 ), where m is the total number of small areas for which data are available. However, this is difficult, especially if the estimator needs to be positive, or at least nonnegative. In fact, very few MSPE estimators are both second-order unbiased and guaranteed to be positive. We consider an alternative, easier approach of estimating the logarithm of the MSPE (log-MSPE), thus avoiding the positivity problem. We derive a second-order unbiased estimator of the log-MSPE using the Prasad–Rao linearization method. The results of empirical studies demonstrate the superiority of the proposed log-MSPE estimator over a naive log-MSPE estimator and an existing method, known as McJack. Lastly, we demonstrate the proposed method by applying it to real data.