Distribution free testing of grouped Bernoulli trials

Abstract

Recently Khmaladze has shown how to `rotate' one empirical process to another. This paper is the first to apply this transform when successive data points are generated by a single distributional family, but with covariates varying over the sample. The application is to Bernoulli trials, and new results show how group sizes rotated are related to the number of parameters, and explore the impact of different types of data generating processes. The utility of the rotation is clear: goodness of fit tests after rotation to a distribution free process are easily computed, show excellent convergence properties, and exhibit high power to reject incorrect null hypotheses.