Abstract:
|
Whether it is multiple phenotypes or multiple SNPs within a gene that are analyzed jointly, the problem of simultaneously evaluating multiple marginal test statistics is a persistent challenge throughout genetic epidemiology. There are many statistical tests that have been developed for this scenario such as variance component tests, max tests, and the higher criticism. When comparing the power of these methods, there is no universally most powerful method, and the best choice of method is highly dependent on the strength, direction, and sparsity of the signal. In order to gain an intuition behind the strengths and weaknesses of these popular approaches, we evaluate the acceptance/rejection regions of each test to see, geometrically, when they are similar or different from one another. We also explore how different alternatives effect these similarities and differences. We find that what drives the relative test performances is not the sparsity of the signal vector, but rather its direction.
|