Time: 23 January 2019, 1 - 2 pm Place: B705


One of the main aims in epidemiology and public health is to evaluate the disease burden due to some specific risk factor. In particular, one is often interested in the proportion of disease cases that could have been prevented had all subjects been unexposed to the specific risk factor. This quantity is formally called the attributable fraction (AF) or attributable risk.  The AF is a causal parameter and should ideally be estimated from a Randomized Clinical Trial (RCT) where the exposure is randomized in the population. However, for ethical or practical reasons this not always feasible and the AF is often estimated from observational data. The crucial problem of using observational data to estimate causal effects is confounding, i.e. common causes of the exposure and outcome, which create a spurious (i.e. non-causal) association between these. If all confounders (sufficient for confounding control) are observed, then the AF can be estimated by model-based adjustment.

However, a major concern in observational studies is that some confounders are typically unobserved. In these situations Instrumental Variable (IV) analysis could be used as a remedy.  The IV analysis was developed as a tool to mimic a RCT setting, by using an IV that can be considered randomized in the population. There are three main assumptions for an IV to be valid, 1) the IV should have a (preferably strong) association with the exposure, 2) the effect of the IV on the outcome should only go through the exposure and 3) the IV-outcome association should be unconfounded. In practice, finding IVs that simultaneously fulfil these requirements has proven to be challenging. For this reason Mendelian Randomization (MR), a special case of IV analysis, has gained popularity. In MR, genetic markers, usually Single Nucleotide Polymorphisms (SNPs), are used as instruments. This is motivated by Mendel's first and second law of inheritance which imply that SNPs are randomized at conception, and thus makes assumption 3) reasonable in many settings.

Since the AF is defined for binary exposures and outcomes we review the IV estimators for binary outcomes; the Two Stage estimator and G-estimator, and show how to estimate the AF in IV analysis. As an illustration of the methods, MR analysis is used to estimate the AF and investigate a potential causal effect between educational qualifications (attained a university degree) and CHD risk using the UK biobank data.