Mendelian Randomization

Abstract Causal claims from observational epidemiological studies are influenced by reverse causation and confounding. Mendelian randomization assesses whether genetic predictors of a risk factor (such as alcohol consumption) are associated with risk of a disease outcome (such as stroke). Due to Mendel's laws and the random nature of genetic inheritance, genetic variants are typically only associated with traits that they cause. This means genetic associations are less susceptible to confounding. Furthermore, as genetic variants are fixed from birth, the potential for reverse causation is diminished. Therefore, if the genetic variants affect the risk factor in a specific way (formally, if they satisfy the assumptions of an instrumental variable), then any genetic association with the outcome is indicative of a causal effect of the risk factor on the outcome. The approach has been used to strengthen (and refute) the causal nature of relationships for a wide range of risk factors and outcomes. Key Concepts Genetic variants that affect a risk factor in a specific way can be treated analogously to randomization of that risk factor in a randomized trial. In particular, an association between such a genetic variant and a disease outcome provides evidence that the risk factor has a causal effect on the outcome. Mendelian randomization is the use of genetic variants as instrumental variables to make causal inferences. The instrumental variable assumptions cannot be formally proved – their validity can be probed, but ideally they should be justified on the basis of our scientific knowledge. ‘Plausible’ Mendelian randomization analyses are those where the genetic variant(s) are located in a gene region with known biological relevance to the risk factor, and, hence, pleiotropy and population stratification are unlikely. ‘Questionable’ Mendelian randomization analyses are those where the validity of the instrumental variable assumptions is unclear, and, hence, a cautious causal conclusion is only justified if there is consistency of evidence. Two‐sample Mendelian randomization analyses can be performed quickly and routinely using publicly available data from genome‐wide association studies. Such analyses can provide valuable insights into causal mechanisms. However, they can also be performed in an unthinking way, without critical thought or judgement; such analyses can be of limited value. Many extensions to the Mendelian randomization paradigm are available, which provide causal inferences under weaker assumptions, or else answer translational questions about the impact of interventions in a risk factor.