When Will Mendelian Randomization Become Relevant for Clinical Practice and Public Health?

Obtaining reliable evidence regarding factors that cause disease or influence disease progression is clearly central to progress in science. There have been many high-profile failures in this regard, which range from health-related behaviors (eg, dietary components that observational epidemiology strongly suggested protected against chronic disease that failed when tested in large-scale randomized clinical trials [RCTs]), through vitamin supplement use and menopausal hormone therapy, to a large number of drugs that failed at phase 3 trial stage.1,2 Observational epidemiologic studies are prone to confounding, reverse causation (ie, when the disease process influences the exposure, rather than vice versa), and a variety of other biases.1 Furthermore, laboratory studies at the cellular through to the whole-animal level have also led to many erroneous conclusions being drawn.3 Many strategies need to be adopted to ameliorate this situation, applying the overarching philosophy that only very rarely does any single source of evidence adequately establish the veracity of causal claims. Several methods should be applied to any question, each of which would be expected to have biases, but the origins of—and predicted influence of—these biases should be unrelated to each other and therefore not bias the findings to the same extent or even in the same direction. Findings that are similar when these different methods are applied are more likely to be robust. With such “triangulation” of evidence, more reliable causal inference should be achievable.4 A powerful component of this evidence base can be provided by the application of a methodology that incorporates the natural randomization inherent in the generation of genetic individuality—the process of mendelian randomization.5,6 The basic principle of mendelian randomization is straightforward: it is that genetic variation generates differences between individuals that influence health outcomes that are not subject to the confounding or reverse-causation bias that can distort observational findings.5,6 This process can be considered analogous to randomization in an RCT,5 and there are now many examples of its application.7 These range from proofs of principle (eg, that low-density lipoprotein cholesterol, high blood pressure, obesity, and smoking increase the risk of coronary heart disease [CHD]), demonstration of factors unlikely to be causal (eg, C-reactive protein in relation to CHD, diabetes, and several cancers), dispelling claims of health protection (eg, moderate alcohol intake is not beneficial with respect to CHD risk), and the prediction of the findings from both successful and unsuccessful pharmacological RCTs.6-8 Mendelian randomization studies have benefited from analytical methods adapted from “instrumental variables” approaches in econometrics, allowing estimation of effect sizes and their precision.6 These estimates require careful translation, because they can differ in magnitude from reliable estimates obtained from other sources. For example, because genetic variants generally relate to lifetime differences in the exposure (eg, low-density lipoprotein cholesterol) they relate more strongly to disease outcome (eg, CHD) than would be seen in an observational study or a relatively short-term (a few years) RCT of cholesterol reduction.5,8 More importantly, an exposure that influences disease risk at a critical period of the life course—as has been suggested for vitamin D levels during the preadult stage in relation to multiple sclerosis risk—will be uncovered by a mendelian randomization study (because the genetic variants influence vitamin D levels across life), but an intervention (or observational study) outside of that period will miss the critical window of exposure and would not be expected to recapitulate this finding.9 Early mendelian randomization studies tended to use single genetic variants and focus on a specific risk factor– disease association within a single study population. More recently, the rapid growth in established genotype-phenotype associations derived from genome-wide association studies (GWASs) has led to large numbers of genetic variants being identified for many exposures. These data can be combined with data from the large number of disease outcome GWASs to allow what is known as 2-sample mendelian randomization analysis.6 In this issue of JAMA, Emdin and colleagues10 report an extensive mendelian randomization analysis of whether—and if so, how—abdominal obesity (indexed by waist-to-hip ratio [WHR]) influences risk of CHD and type 2 diabetes. The authors used publicly available GWAS results from more than 322 154 participants in 4 studies to estimate the relationship of abdominal adiposity with disease outcomes in 2-sample mendelian randomization analysis.6 This approach has additional assumptions to single-sample studies6 but has the very considerable advantage of allowing estimation of the causal effects of a large array of potential exposures on many disease outcomes. A platform, MR-Base, is now available that facilitates rapid 2-sample mendelian randomization interrogation of myriad such exposure-outcome associations.11 In addition to data from the 4 GWASs, Emdin et al also used individual data from 111 986 participants from the UK Biobank to replicate the estimated causal association of WHR adjusted for body mass index (BMI) with CHD and type 2 diabetes. In further analysis, the Related article page 626 Opinion