On the detection of influential subsets of categories in goodness of fit testing

In this paper we focus on the well-known test of goodness of fit for comparing observed counts to expected values under some null hypothesis. When the latter is rejected, we propose a simple method for detecting which subset(s) of category counts provoke(s) that rejection. The approach aims at building intervals iteratively and drawing appropriate conclusions on that basis. We discuss this method with respect to other classical approaches. We illustrate our purpose by treating some examples.