Testing for Redundancy in Normal Mixture Analysis
Marco Bee · Communications in Statistics - Simulation and Computation · 2004
In this paper we deal with the problem of testing for redundancy of variables in a mixture of two p-variate normal populations. If the investigator is only interested in estimation of the parameters of the first q variables, the last p − q variables are said to be redundant when they can be safely discarded, i.e., when they carry no additional information about the parameters of interest. The main contribution of this work consists in finding the likelihood ratio test for this hypothesis. In general, the distribution of the test is nonstandard, so that bootstrap methods are employed to analyze its properties; as an illustration, the techniques are applied to a real-data problem in the field of quality control. However in a specific case, namely when the Mahalanobis distance is large, the usual χ2 asymptotic approximation is appropriate; thus we study, by means of a simulation experiment, the relationship between the Mahalanobis distance and the limiting distribution of the test.