Title: Bootstrap methods for multiple comparison in small areas statistics
Authors: Stefan Sperlich - University of Geneva (Switzerland) [presenting]
Abstract: A huge effort has been made in order to estimate the mean squared error for estimators in small areas. An increasingly popular approach for approximating these errors is that of using resampling methods. In the more general literature on mixed effects or multilevel morels there exist already various proposals of different bootstrap procedures. These have also been proposed in order to directly construct prediction intervals for specific estimators in small area estimation. A generally known problem is that these are typically constructed in a way that they exhibit individual (or say average) coverage probability in the sense that for 95 percent PIs we have 5 percent of areas for which the PI does not include the parameter of interest. We do not know anything about joint converage probabilities and can not use them for comparison of two or more areas. In a first step we try to find out which of these bootstrap methods can easily be extended to exhibit a joint or uniform coverage probability.