Bayesian and frequentist models: legitimate choices for different purposes of clinical research

Abstract

Bayesian and frequentist approaches to statistical modelling in epidemiology are often pitted against each other as if they represented diametrically opposing philosophies. However, both approaches have a role to play in clinical epidemiology and the evaluation of clinical practice. Here I present an overview of the philosophical underpinnings of the Bayesian and frequentist approaches, showing that each model has its place depending on the philosophical and evaluative needs of the user. If the user's approach to a clinical problem places an emphasis on identifying causal relationships, a frequentist approach might be best suited. On the other hand, if the user takes an approach in which estimating a priori probabilities is appropriate, a Bayesian approach might be more appropriate. One could imagine both approaches used for the same study. Bayesian and frequentist approaches are complementary tools in the clinical evaluator's toolkit.