Multivariate Refutation of Aetiological Hypotheses in Non-Experimental Epidemiology

Extension of Karl Popper's logic of refutation from the realm of contingency tables to multivariate modelling leads to the conclusion that rigorously scientific multivariate analysis in non-experimental epidemiology differs from the traditional quasi-scientific approach. Instead of aiming for high sensitivity in detecting aetiological agents, the goal in refutation is high specificity--to give the best defence of the 'innocence' of every exposure hypothesized as being a cause. Instead of 'forward selection' or 'backward elimination', multivariate refutation uses the method of 'forward elimination'. This entails a likelihood approach (which may be complemented by, but should be demarcated from, Bayesian methods) not only for statistical inference but also, by analogy, for study design and conduct: one starts with the conclusion (the estimate or hypothesis) and works backwards to the observations (the likelihood of the data or the design of the study). Differences in practice can sometimes be large, as illustrated by a study of hypothesized triggers of myocardial infarction. Multivariate refutation should replace the concept of multivariate modelling in non-experimental epidemiology.