Fall 2002 Seminar Series - October 17, 2002
University of Minnesota
School of Statistics
College of Liberal Arts

Goodness-of-fit tests for normal mixed model

Christian Ritz
Department of Mathematics and Physics
The Royal Veterinary and Agricultural University
Copenhagen, Denmark

Thursday, October 17, 2002
4:00 PM, 115 Ford Hall
Minneapolis, East Bank Campus
Social at 3:30 PM, 300 Ford Hall

Abstract

     To assess the normality assumption of the random effects in a normal mixed linear model, the so-called estimated best linear unbiased predictors (EBLUPs) could be used. One such approach, taken by Lange and Ryan (Ann. Statist. 17 624--642, 1989), is to construct a weighted normal plot based on a pointwise limit result for a weighted empirical distribution function of the standardized EBLUPs. Using a weighted normal plot it is possible to adjust for design imposed imbalances. Looking at the corresponding (weighted) empirical process, it is possible to obtain goodness-of-fit statistics for testing the null hypothesis that the random effects follow a normal distribution in a mixed linear model. The test is performed by applying a classical goodness-of-fit statistic, eg Kolmogorov-Smirnov, to the difference between the empirical process based on standardized EBLUPs and some suitable reference process. The limiting distributions of these statistics are governed by the limiting distribution of the underlying processes, for which weak convergence is established. The power of the tests against several alternatives is explored in a small simulation study. An example illustrates the use of the tests.