Power-up: A reanalysis of ‘power failure’ in neuroscience using mixture modeling.

Abstract

Recently, evidence for endemically low statistical power has cast neuroscience findings into doubt. If low statistical power plagues neuroscience, then this reduces confidence in the reported effects. However, if statistical power is not uniformly low, then such blanket mistrust might not be warranted. Here, we provide a different perspective on this issue, analyzing data from an influential study reporting a median power of 21% across 49 meta-analyses (Button et al., 2013). We demonstrate, using Gaussian mixture modeling, that the sample of 730 studies included in that analysis comprises several subcomponents so the use of a single summary statistic is insufficient to characterize the nature of the distribution. We find that statistical power is extremely low for studies included in meta-analyses that reported a null result and that it varies substantially across subfields of neuroscience, with particularly low power in candidate gene association studies. Therefore, whereas power in neuroscience remains a critical issue, the notion that studies are systematically underpowered is not the full story: low power is far from a universal problem.