Statistische Methoden zur Replizierbarkeitsbewertung im Rahmen mehrstufiger Studien

Abstract

In replicability analyses, we are concerned with the analysis of partial conjunction null hypotheses. Valid combination p-values for the latter are often too large. Randomized p-values replace too large p-values with a random number between zero and one, whereas conditional p-values discard these entirely; both need to rescale the remaining p-values. In a multiple testing setup, where many partial conjunction null hypotheses are considered, using either can for example improve the estimation of the proportion of true null hypotheses or the power of a multiple test. Furthermore an analysis of different combination functions in different settings has been conducted, some turning out to be more advantageous for spread-out evidence and others for sparse evidence.