Chapter 7:

In this chapter, we consider robust methods, which do not require all genetic variants to be valid instrumental variables to give consistent estimates of a causal parameter. If a particular gene region (or regions) has a specific biological link to the exposure, then we would generally advocate basing the primary Mendelian randomization analysis on variants from that region. However, particularly for complex risk factors such as body mass index or blood pressure, there is no single gene region that encodes the risk factor, and so a polygenic Mendelian randomization analysis is necessary. If several genetic variants in different gene regions have similar associations with the outcome, then a polygenic analysis may even provide stronger evidence of a causal relationship, as the analysis is not dependent on the validity of the IV assumptions for a single gene region. However, in many cases, not all genetic variants associated with the exposure will tell the same story.

We consider robust methods for Mendelian randomization in three categories: consensus methods, outlier-robust methods, and modelling methods. Findings from a polygenic Mendelian randomization investigation are more reliable when several methods that make different assumptions give similar answers. However, each of these methods still relies on untestable assumptions, and appropriate caution in their interpretation is needed, particularly when results from methods differ.