Meta-analysis and Mendelian randomization: A review.

Bowden J, Holmes MV. · 2019

Read Meta-analysis and Mendelian randomization: A review. by Bowden J, Holmes MV. (2019) 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-pmc6973275.txt.

Introduction

Bowden and Holmes review how meta-analysis tools strengthen Mendelian randomization studies. Their article appeared in Research Synthesis Methods. Mendelian randomization uses gene variants as instruments to test whether a risk factor truly causes a health outcome. Meta-analysis, long used to pool separate studies, now pools genome-wide data for many variants at once.

When causal estimates from different variants disagree, the method's core assumptions may be broken. The authors first explain the approach and the theory behind it. They then show how random effects models, meta-regression, and robust regression test for that disagreement. Such checks can adjust estimates and harden the conclusions. Their aim is rigour: causal claims from genetic data should survive the toughest statistical scrutiny available.

Catalog record

CREATORBowden J, Holmes MV.
YEAR2019
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC6973275

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Meta‐analysis and Mendelian randomization: A review By Jack Bowden, Michael V Holmes Published in: Research Synthesis Methods DOI: 10.1002/jrsm.1346 ABSTRACT Abstract Mendelian randomization (MR) uses genetic variants as instrumental variables to infer whether a risk factor causally affects a health outcome. Meta‐analysis has been used historically in MR to combine results from separate epidemiological studies, with each study using a small but select group of genetic variants. In recent years, it has been used to combine genome‐wide association study (GWAS) summary data for large numbers of genetic variants. Heterogeneity among the causal estimates obtained from multiple genetic variants points to a possible violation of the necessary instrumental variable assumptions. In this article, we provide a basic introduction to MR and the instrumental variable theory that it relies upon. We then describe how random effects models, meta‐regression, and robust regression are being used to test and adjust for heterogeneity in order to improve the rigor of the MR approach. Keywords: Mendelian randomization, meta‐analysis, pleiotropy, two‐sample summary data MR INTRODUCTION The primary aim of observational epidemiology is to determine the root causes of illness, with the focus of many epidemiological analyses being to examine whether exposure to a particular risk factor modifies the severity, or the likelihood of developing, a disease. Causal conclusions are rarely justified following a traditional analysis, even when strong statistical associations are measured between an exposure and outcome, because it is never certain that all confounders of the association have been identified, measured, and appropriately adjusted for. Mendelian randomization (MR)1 offers an alternative way to probe the issue of causality in epidemiological research, by using additional genetic variants that are hypothesized to satisfy the instrumental variable (IV) assumptions. Directed acyclic graphs (DAGs) are a useful tool, both to explain the rationale for an MR study and to clarify the IV assumptions that its validity rests on. Figure 1 shows a DAG relating the simplest single unit of genetic variation—a single nucleotide polymorphism (SNP) G—to an exposure, X, and outcome, Y, in the presence of unmeasured confounding, represented by U. [Figure: Causal directed acyclic graph (DAG) representing the hypothetical relationship between genetic variant G, exposure X, and outcome Y, in the presence of unobserved confounding, U. Solid arrows represent allowed relationships between the variables. Dashed lines represent relationships that are forbidden for G to qualify as a valid instrumental variable (IV). The G‐X and X‐Y arrows are parameterized by γ and β, with the latter denoting the causal effect of X on Y] The true causal effect of the exposure on the outcome is denoted by the arrow from X to Y in Figure 1 and the parameter β. The “associational” estimate obtained from a simple regression of the outcome on the exposure could be systematically different from this causal effect, because confounding may be responsible for all, or part of, its magnitude. From the DAG in Figure 1, this can be understood by noting that the association between X and Y is contributed to by the direct effect path X → Y and the “back door” path X ← U → Y.2 Suppose, however, that a SNP G exists, which robustly predicts a proportion of the exposure that is unrelated to any confounders of the exposure‐outcome relationship. This is represented by the path G → X and the absence of a path between G and U. If, in addition, G can only influence the outcome through the exposure, as represented by the absence of a direct path from G to Y, then this SNP is said to be a “valid IV.” G is usually coded as 0, 1, or 2 to reflect the number of exposure‐increasing alleles of a SNP an individual carries. This assumes that G exerts a linear per‐allele effect on X. The exposure itself is typically a continuous measure, for example, a person's blood pressure, body mass index, or cholesterol level. It will sometimes represent a binary health behavior, for example, whether an individual is a current smoker. The outcome of interest can be continuous but is often a binary variable, usually representing the presence or absence of a disease. Testing for causality 1.1. If G is a valid IV, then any nonzero association between this SNP and the outcome provides evidence to support the hypothesis that the exposure causally effects the outcome and β≠ 0. This can be understood by noting that G is “d‐separated” from Y (or independent of Y) in Figure 1 upon removal of the path X → Y.2 Genetic variants have successfully been used to test for causality by looking at their association with the outcome in a variety of settings. However, most MR studies go further, by attempting to estimate the magnitude of the causal effect, β. Specifically, this parameter quantifies the effect on the outcome when the exposure is intervened on and changed by one unit, with all other factors remaining fixed. QUANTIFYING THE CAUSAL EFFECT IN MR USING THE RATIO ESTIMATE AND TSLS If SNP G is a valid IV, the exposure can be assumed to causally affect the outcome in a linear fashion with no effect modification, then the underlying SNP‐outcome association (denoted by Γ) should be the product of the underlying SNP‐exposure association (denoted by γ) and the causal effect of the exposure on the outcome, β. That is, [formula: (1)Γ=βγ.] From Equation 1, the simplest estimate for β ( β^R, where R stands for “ratio”) is obtained by dividing the SNP‐outcome association estimate by the SNP‐exposure association estimate to give: [formula: (2)β^R=Γ^γ^.] The standard error of the ratio estimate can be approximated via a Taylor series expansion of β^R using the delta method.3 The ratio estimate in (2) is calculated from two summary estimates, but it is also equivalent to the estimate obtained by the following two‐step procedure applied to individual level data on Y, X, and G: - Step 1Regress the exposure on the SNP via the model: [formula: (3)X∣G=γ0+γG+εX.] - Step 2Regress the outcome on the fitted values of the regression in step 1, X^, via the model: [formula: (4)Y∣X^=β0+βX^+εY] and report its estimated regression coefficient, β^. This is referred to as “two‐stage least squares” (TSLS). When multiple SNPs are available as IVs, they can be easily incorporated into a TSLS analysis to yield a single causal estimate, by calculating fitted values based on a multivariable regression of the exposure on all SNPs together in model (3). In that case, γ and G would represent vectors of association parameters and SNP values for each individual. This automatically allows for any potential correlation between the SNPs, for example, due to linkage disequilibrium (LD). Standard errors for the TSLS estimate in (4) must take into account the uncertainty in the first stage model (3). This correction is performed as standard in most software packages. Equation 1, the ratio estimate in (2) and the TSLS procedure in (3) and (4) are only strictly correct when Y is itself continuous. When Y is binary, logistic regression is typically used to quantify the G‐Y association in (2) or the association between the genetically predicted exposure and outcome in (4). In this case, the causal effect of a unit change in the exposure on the risk of Y

Read the complete text