Confounding and reverse causality have prevented us from drawing meaningful clinical interpretation even in well-powered observational studies. Confounding may be attributed to our inability to randomize the exposure variable in observational studies. Mendelian randomization (MR) is one approach to overcome confounding. It utilizes one or more genetic polymorphisms as a proxy for the exposure variable of interest. Polymorphisms are randomly distributed in a population, they are static throughout an individual's lifetime, and may thus help in inferring directionality in exposure-outcome associations. Genome-wide association studies (GWAS) or meta-analyses of GWAS are characterized by large sample sizes and the availability of many single nucleotide polymorphisms (SNPs), making GWAS-based MR an attractive approach. GWAS-based MR comes with specific challenges, including multiple causality. Despite shortcomings, it still remains one of the most powerful techniques for inferring causality.With MR still an evolving concept with complex statistical challenges, the literature is relatively scarce in terms of providing working examples incorporating real datasets. In this chapter, we provide a step-by-step guide for causal inference based on the principles of MR with a real dataset using both individual and summary data from unrelated individuals. We suggest best possible practices and give recommendations based on the current literature.
Keywords: Causal inference; Genome-wide association study; Individual data; Instrumental variable; Mendelian randomization; Observational epidemiology; Pleiotropy; Reverse causation; Summary data; Unobserved confounding.