Consistent Estimation in Mendelian Randomization with Some Invalid Instruments Using a Weighted Median Estimator.
Bowden J, Davey Smith G, Haycock PC, Burgess S. · 2016
Read Consistent Estimation in Mendelian Randomization with Some Invalid Instruments Using a Weighted Median Estimator. by Bowden J, Davey Smith G, Haycock PC, Burgess S. (2016) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4849733.txt.
Introduction
Jack Bowden, George Davey Smith, Philip C. Haycock, and Stephen Burgess proposed a weighted median method in 2016 for Mendelian randomization, a way to test cause using genes as natural levers. The standard method fails when some gene variants used as levers are invalid. Their estimator takes the middle of the variant estimates, weighted by precision.
The method stays sound even when up to half the data come from invalid variants. Simulations showed better error control than the standard method, and it pairs well with a second recent fix, MR-Egger regression. Tests on blood fats show how the tool works in practice. Genetic cause tests gain a sturdier option for messy real data.
Catalog record
| CREATOR | Bowden J, Davey Smith G, Haycock PC, Burgess S. |
|---|---|
| YEAR | 2016 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Consistent Estimation in Mendelian Randomization with Some Invalid Instruments Using a Weighted Median Estimator By Jack Bowden, George Davey Smith, Philip C Haycock, Stephen Burgess Published in: Genetic Epidemiology DOI: 10.1002/gepi.21965 ABSTRACT ABSTRACT Developments in genome‐wide association studies and the increasing availability of summary genetic association data have made application of Mendelian randomization relatively straightforward. However, obtaining reliable results from a Mendelian randomization investigation remains problematic, as the conventional inverse‐variance weighted method only gives consistent estimates if all of the genetic variants in the analysis are valid instrumental variables. We present a novel weighted median estimator for combining data on multiple genetic variants into a single causal estimate. This estimator is consistent even when up to 50% of the information comes from invalid instrumental variables. In a simulation analysis, it is shown to have better finite‐sample Type 1 error rates than the inverse‐variance weighted method, and is complementary to the recently proposed MR‐Egger (Mendelian randomization‐Egger) regression method. In analyses of the causal effects of low‐density lipoprotein cholesterol and high‐density lipoprotein cholesterol on coronary artery disease risk, the inverse‐variance weighted method suggests a causal effect of both lipid fractions, whereas the weighted median and MR‐Egger regression methods suggest a null effect of high‐density lipoprotein cholesterol that corresponds with the experimental evidence. Both median‐based and MR‐Egger regression methods should be considered as sensitivity analyses for Mendelian randomization investigations with multiple genetic variants. Keywords: Mendelian randomization, instrumental variables, robust statistics, Egger regression, pleiotropy INTRODUCTION Over the past decade, Mendelian randomization has become an established tool for probing questions of causality when characterizing the etiology of disease (Burgess and Thompson, 2015; Davey Smith and Ebrahim, 2003). The requirement for such approaches stems from a fundamental limitation of observational data, namely that causation cannot automatically be inferred from an association between an exposure and a disease. The association could be due to unobserved confounding between the exposure and the outcome, or reverse causation (the outcome affects the exposure) (Davey Smith and Ebrahim, 2004). These limitations are generally of no consequence when the aim is merely to predict the likelihood of future outcomes. However, if an exposure has a noncausal association with an outcome, then public health or pharmaceutical interventions targeted at the exposure will realize no material benefit and represent a waste of resources. The basic premise of Mendelian randomization relies on genetic variants that explain variation in the exposure, but do not affect the disease outcome except possibly through the exposure. Such genetic variants are known as instrumental variables (IVs) (Greenland, 2000). Subgroups of individuals with differing numbers of alleles of a genetic IV can be thought of as having been randomized to receive a different mean level of the exposure during their life course (Davey Smith and Ebrahim, 2005; Nitsch et al., 2006). If the randomization is indeed uncontaminated (in the sense that a person's genetic subgroup is independent of all factors, except the exposure and any causal consequence of the exposure), then differences in the outcome between genetic subgroups can be causally attributed to the exposure (Didelez and Sheehan, 2007). The following three assumptions are necessary for a genetic variant to be a valid IV (Martens et al., 2006): - IV1: the variant is predictive of the exposure; - IV2: the variant is independent of any confounding factors of the exposure—outcome association; - IV3: the variant is conditionally independent of the outcome given the exposure and the confounding factors. The IV assumptions are illustrated in Figure 1. IV1 is the only assumption that can be fully empirically tested, because IV2 and IV3 depend on all possible confounders of the exposure—outcome association, both measured and unmeasured. Any statistical method for obtaining causal inferences must by necessity make an untestable assumption. The validity of a causal conclusion from a Mendelian randomization analysis depends on the plausibility of these assumptions. [Figure: Illustrative diagram representing the hypothesized relationships between genetic variant Gj, exposure X, disease Y, and confounders U when Gj is a valid instrumental variable (IV). Crosses indicate violations of assumptions IV2 and IV3 that potentially lead to invalid inferences from conventional methods.] Early implementations of the Mendelian randomization approach were largely constrained by limitations of power as investigations were undertaken in small sample sizes, and used only a handful of genetic variants (each explaining a small proportion of the variance in the exposure). However, a revolution in the field is under way led by the identification of increasing numbers of genetic variants robustly associated with particular traits, and the public release by many large consortia of summary association estimates for hundreds of thousands of genetic variants with exposures and disease outcomes (Burgess et al., 2015b), such as the Global Lipids Genetics Consortium (GLGC) for lipid fractions [GLGC, 2013] and the CARDIoGRAM consortium for coronary artery disease (CAD) risk (CARDIoGRAMplusC4D Consortium, 2013). The availability of such summary data has facilitated powerful Mendelian randomization investigations to be conducted in a two‐sample framework by providing genetic associations precisely estimated in large sample sizes (Burgess et al., 2013). The inclusion of multiple variants in a Mendelian randomization analysis typically leads to increased statistical power (Freeman et al., 2013), but presents new challenges (Glymour et al., 2012). First, if there is substantial overlap in the datasets from which the association estimates with the exposure and with the outcome were obtained, then the resulting analysis suffers from bias and inflated type 1 error rates when the included variants are “weak” (i.e., they do not explain a substantial proportion of variation in the exposure in the dataset under analysis) (Burgess et al., 2011; Pierce and Burgess, 2013). Second, it may not be the case that all included genetic variants are valid IVs. In this paper we propose a new method for Mendelian randomization using summary data that offers protection against invalid instruments: the weighted median estimator. This approach can provide a consistent estimate of the causal effect even when up to 50% of the information contributing to the analysis comes from genetic variants that are invalid IVs. We explore the statistical properties of the weighted median estimator in a realistic simulation study, and compare with an alternative summary data analysis method also robust to some violations of the IV assumptions, MR‐Egger regression (Bowden et al., 2015). We explain how the two approaches differ in their assumptions, and when they each work well or fail. We provide an illustrative estimate of two‐sample Mendelian randomization using summary data on the associations of 185 genetic variants with high‐density lipoprotein cholesterol (HDL‐c), low‐density lipoprotein cholesterol (LDL‐c), and triglycerides from the GLGC, and with CAD risk from the CARDIoGRAM consortium. We conclude with a discussion of the issues raised and the potential for future research. METHODS Consider data from a Mendelian randomization study on J genetic variants G1,...,GJ, a continuous exposure X and a continuous outcome Y. All confounding variables are subsumed into a single variable U. We initially assume that all genetic variants are valid IVs, and further assume that all the relationships between variables in Figure 1 are linear without heterogeneity or effect modification: [formula: X|Gj=γ0+γjGj+εXjY|Gj=Γ0+ΓjGj+εYj.] Assumption IV1 tells us that all variants are associated with the exposure, so γj≠0 for all j. Assumptions IV2 and IV3 tell us that the genetic associations with