Simultaneous testing for mean and variance differences with nasty data
by
Chris Francis
NIWA
Coauthors: Bryan Manly (WEST Inc, Cheyenne, Wyoming)

Possibly the most commonly performed type of statistical test is for differences in means. It is often advisable to precede such a test by another test for differences in variances because the mean-tests commonly perform poorly in the presence of such differences. We describe an approach, based on randomization, that tests for both these differences, and evaluate its performance via a simulation experiment. The approach is shown to be useful as a robust, conservative method in cases when the samples come from very non-normal distributions. One possible outcome of the test is the conclusion that there are mean and/or variance differences, but it is not possible to say which.

Date received: August 31, 2001