Improved Direct Estimators for Small Areas

Unbiased direct estimators for small area quantities are usually considered too variable to be of any practical use. In this paper we propose a class of model-based direct estimators for small area quantities that appears to overcome this objection, in the sense that these estimators are comparable in efficiency to the indirect model-based small area estimators (e.g. empirical best linear unbiased predictors, or EBLUPs) that are now widely used. There are many practical advantages associated with such model-based direct (MBD) estimators, arising from the fact that they are computed as weighted linear combinations of the actual sample data from the small areas of interest. Note that in this case the weights ‘borrow strength’ via a model that explicitly allows for small area effects. One particular advantage that we explore in this paper is that estimation of mean squared error (MSE) is then straightforward, using well-known methods that are in common use for population level estimates. Empirical results reported in this paper show that the MBD estimator represents a real alternative to the EBLUP, with the simple MSE estimator associated with the MBD estimator providing good coverage performance. We also report results that indicate that the MBD estimator may be more robust than the EBLUP when the small area model is incorrectly specified. Furthermore, the MBD approach is easily extended to provide multi-purpose weights that are efficient across a range of variables, including variables that are unsuitable for EBLUP, e.g. variables that contain a significant proportion of zeros.