Tests for statistical significance of a treatment effect in the presence of hidden sub-populations

For testing the statistical significance of a treatment effect, we often compare between two parts of a population; one is exposed to the treatment, and the other is not exposed to it. Standard parametric or nonparametric two-sample tests are commonly used for this comparison. But direct applications of these tests can yield misleading results, especially when the population has some hidden sub-populations, and the effect of this sub-population difference on the response dominates the treatment effect. This problem becomes more evident if these sub-populations have widely different proportions of representatives in the samples obtained from these two parts. In this article, we propose some simple methods to overcome these limitations. These proposed methods first use a suitable clustering algorithm to find the hidden sub-populations, and then they eliminate the sub-population effect by using a suitable transformation of the data. Standard two-sample tests, when they are applied on the transformed data, usually yield better results. We analyze some simulated and real data sets to demonstrate the utility of these proposed methods.