La falta de reproducibilitat de la investigació: L'estadística com a legitimació del resultat

Abstract

Scientific research is validated by reproduction of the results, but efforts to reproduce spurious claims drain resources. We focus on one cause of such failure: false positive statistical test results caused by random variability. Classical statistical methods rely on p-values to measure the evidence against null hypotheses, but Bayesian hypothesis testing produces more easily understood results, provided one can specify prior distributions under the alternative hypothesis. We describe new tests, UMPBTs, which are Bayesian tests that provide default specification of alternative priors, and show that these tests also maximize statistical power.