Abstract

Inferring causality from observational data is difficult as it is not always clear which of two associated variables is the cause, which the effect, or whether both are common effects of a third unobserved variable, or confounder. Instrumental variable (IV) methods are widely used in econometrics and provide a means to test for or estimate a causal effect when confounding is believed to be present but not fully understood. Mendelian randomisation refers to the situation when the instrument is a genetic variant and has received a lot of attention in the epidemiological literature recently.

Testing for the presence of a causal effect is generally straightforward but point estimates are only obtainable under additional parametric and distributional assumptions. Moreover, there are several IV estimators to choose from, all requiring different assumptions and targeting different causal effects. As is usual in causal inference, such assumptions cannot always be verified from the data and have to be supported by background knowledge. Problems particularly arise when the outcome of interest is a binary indicator of disease status, for example, although there are special cases where these can be satisfactorily addressed.

I will introduce a formal causal framework and discuss Mendelian randomisation in the context of binary epidemiological outcomes. I will explore the different assumptions underpinning the various IV estimators and show how violations of these assumptions may bias the results of an analysis.