Hosted by the MRC Integrative Epidemiology Unit
Summary: There is continuing debate over how to measure statistical evidence. Classical testing with P-values is prone to misunderstanding, while Bayesian testing requires prior distributions that can substantially influence results. For a more objective approach, I propose Empirical Bayes factors in which the prior is estimated from the data, giving a measure of evidence that can be interpreted from both frequentist and Bayesian perspectives. To provide a qualitative sense of stronger and weaker evidence, I propose a scale of measurement based on modelling the effect of evidence on beliefs. Together these proposals offer steps towards more objective reporting of statistical evidence.
Biography: Frank Dudbridge is Professor of Statistical Genetics at the University of Leicester, and a Visiting Professor at the University of Bristol. He has worked on statistical methods for genetic epidemiology, including family-based association studies, genome-wide association studies, Mendelian randomisation and polygenic risk prediction. His applied collaborations span cardiovascular, respiratory, psychiatric and cancer genomics