Test sizes and the importance of finding global maxima of the likelihood function

This paper demonstrates the importance of taking care to find a global maximum when estimating nuisance parameters required for hypothesis tests. We find that sizes can be badly inflated when this is not done for a test of the general linear regression model.