How to Enhance the Power to Detect Brain-Behavior Correlations With Limited Resources

Abstract

Neuroscience has been diagnosed with a pervasive lack of statistical power and, in turn, reliability. One remedy proposed is a massive increase of typical sample sizes. Parts of the neuroimaging community have embraced this recommendation and actively push for a reallocation of resources towards fewer but larger studies. This is especially true for neuroimaging studies focusing on individual differences to test brain-behavior correlations. Here, I argue for a more efficient solution. Ad-hoc simulations show that statistical power crucially depends on the choice of behavioral and neural measures, as well as on sampling strategy. Specifically, behavioral prescreening and the selection of extreme groups can ascertain a high degree of robust in-sample variance. Due to the low cost of behavioral testing compared to neuroimaging, this is a more efficient way of increasing power. For example, prescreening can achieve the power boost afforded by an increase of sample sizes from n=30 to n=100 at ~5% of the cost. This perspective article briefly presents simulations yielding these results, discusses the strengths and limitations of prescreening and addresses some potential counter-arguments. Researchers can use the accompanying online code to simulate the expected power boost of prescreening for their own studies.