Abstract

Estimation of causal associations is fundamental in every area of science. From estimation of the effect of quantitative easing on the UK economy to the effect of alcohol intake on blood pressure, question of cause and effect are at the heart of all applied research. We shall explore the method of instrumental variables, which can be used to estimate causal associations from observational data free from bias due to confounding (correlation between the causal factor of interest and a competing risk factor) and reverse causation (true causal effect of the `outcome’ on the causal factor of interest. Of particular interest will be the case where the outcome is binary, as the non-linear association between causal factor and outcome leads to differences in the estimated causal effect from what would be estimated from a regression in some cases known as non-collapsibility. The talk will be illustrated throughout with data on the association between inflammation and coronary heart disease.