Abstract

This presentation first explores the motivation for conducting epidemiological research: such research ranks well below the ‘gold standard’ of randomised clinical trials in the conventional hierarchy of evidence, yet observational studies (sometimes on >100 million patients) provide valuable information that is rarely obtained from such trials. Hence the problems of confounding and bias (two sides of the same coin) that beset observational studies are well worth addressing, and in this presentation the statistical tools that epidemiologists use for the purpose – adjustment for covariates, case-control matching etc. – are reviewed. The statistics that epidemiologists typically present are incidence and prevalence rates, and their ratios between exposed and unexposed patients. These statistics are interpreted in terms of the Poisson and binomial distributions, rather than the Normal distribution used for continuous variables, so the statistician must be familiar with the methods for fitting models using these distributions – Poisson regression with overdispersion, logistic regression, etc. He or she also needs to be aware of the difficulty of specifying confidence intervals for parameter estimates based on these discontinuous, asymmetric distributions: there is no single ‘right answer’. In the author’s experience, epidemiologists have breathtakingly well developed intuition for the ways in which data can mislead, and what to do about it. But the majority do not habitually express their ideas in algebra and geometry, and when the tasks required of them take them beyond the edge of their statistical comfort zone they will welcome the support of a statistician who is a fast learner and a good communicator.