Causal Inference

This unit covers modern statistical methods for assessing the causal effect of a treatment or exposure from randomised or observational studies. The unit begins by explaining the fundamental concept of counterfactual or potential outcomes and introduces causal diagrams and directed acyclic graphs (DAGs) to identify visually confounding, selection and other biases that prevent unbiased estimation of causal effects. Key issues in defining causal effects that can be estimated in a range of contexts are presented using the concept of the “target trial” to clarify exactly what the analysis seeks to estimate. A range of statistical methods for analysing data to produce estimates of causal effects are then introduced. Propensity score and related methods for estimating the causal effect of a single time point exposure are presented, together with extensions to longitudinal data with multiple exposure measurements, and methods to assess whether the effect of an exposure on an outcome is mediated by one or more intermediate variables. Comparisons will be made throughout with “conventional” statistical methods. Emphasis will be placed on interpretation of results and understanding the assumptions required to allow causal conclusions. Stata and R software will be used to apply the methods to real study datasets.