Abstract
Expression quantitative trait loci (eQTL) studies are performed to identify single-nucleotide polymorphisms that modify average expression values of genes, proteins, or metabolites, depending on the genotype. As expression values are often not normally distributed, statistical methods for eQTL studies should be valid and powerful in these situations. Adaptive tests are promising alternatives to standard approaches, such as the analysis of variance or the Kruskal-Wallis test. In a two-stage procedure, skewness and tail length of the distributions are estimated and used to select one of several linear rank tests. In this study, we compare two adaptive tests that were proposed in the literature using extensive Monte Carlo simulations of a wide range of different symmetric and skewed distributions. We derive a new adaptive test that combines the advantages of both literature-based approaches. The new test does not require the user to specify a distribution. It is slightly less powerful than the locally most powerful rank test for the correct distribution and at least as powerful as the maximin efficiency robust rank test. We illustrate the application of all tests using two examples from different eQTL studies.