Although GWAS have been successful in identifying variants that influence a number of traits, there are still many exposures for which we do not yet have INK 128 order suitable instruments. In addition, genetic variants may be population-specific and not suitable for use in all ancestral groups. For example, a variant in the ALDH2 gene, which strongly influences alcohol consumption, is used in MR studies in East Asian populations, but occurs at too low a frequency for use in MR studies in European populations [30]. Crucially, genetic variants in MR studies must be associated with

the exposure of interest within the analysis sample and must show robust evidence for association with the same exposure in independent samples. Performing MR analyses using genetic instruments that have been discovered within the analysis sample but have not been independently replicated can lead to causal inference in the absence of true causal effects, because associations between genetic variants and exposures may just be chance findings. In addition, as effect sizes between genetic variants and phenotypes are often inflated in discovery samples (also known as the Beavis effect or Winner’s Curse), performing MR analyses within

discovery samples can result in biased causal effect sizes [31]. Biased estimates of effect sizes may also be obtained if the measured exposure does not fully capture the causal exposure through which the genetic variant operates [31]. For example, a variant in the nicotinic receptor alpha-5 subunit protein, rs16969968, influences lifetime tobacco find more exposure, but this is not well captured by self-report measures of smoking (e.g., cigarettes per day). MR of lung cancer data using cigarettes per day as the intermediate variable indicates a causal odds ratio for lung cancer of 2180 per pack of cigarettes smoked per day, compared to only 2.6 from observational analysis [32]. By

contrast, using cotinine, a metabolite of nicotine and a more precise objective measure of tobacco exposure, produces effect sizes during which are more consistent with observational findings [33]. In the absence of appropriate intermediate exposure measures, MR can still be used to infer causality, but it may not be possible to accurately estimate causal magnitudes of effect. Furthermore, MR studies can be informative about the effects of lifelong exposure to a risk factor, but are usually not appropriate for investigating the impact of short-term changes in risk factors on health outcomes. MR studies will also rarely provide information about the mechanisms underlying a causal relationship (although two-step MR can provide this). Although MR can minimise many of the biases associated with conventional epidemiological studies, there are ways in which MR can still be confounded.