Combining information on multiple instrumental variables in Mendelian randomization: comparison of allele score and summarized data methods.
Burgess S, Dudbridge F, Thompson SG. · 2016
Read Combining information on multiple instrumental variables in Mendelian randomization: comparison of allele score and summarized data methods. by Burgess S, Dudbridge F, Thompson SG. (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-pmc4832315.txt.
Introduction
Stephen Burgess, Frank Dudbridge, and Simon G. Thompson compare two ways to combine genetic instruments in Statistics in Medicine in 2016. Mendelian randomization uses gene variants as natural experiment levers to test whether a risk factor causes a disease. Modern studies hold many levers at once, and two schools combine them: fold each person's variants into one allele score, or compute an estimate per variant from summary data and pool them.
The paper shows the two roads meet: score estimates can be rebuilt from summary data alone. It also extends both to variants that travel together. Simulations warn that the summary method leans toward zero when the weights come from shaky estimates, while score estimates stay clean. An example on cholesterol genes and heart disease shows the gain from using several variants in one gene region.
Catalog record
| CREATOR | Burgess S, Dudbridge F, Thompson SG. |
|---|---|
| YEAR | 2016 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 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:
Combining information on multiple instrumental variables in Mendelian randomization: comparison of allele score and summarized data methods By Stephen Burgess, Frank Dudbridge, Simon G. Thompson Published in: Statistics in Medicine DOI: 10.1002/sim.6835 ABSTRACT Mendelian randomization is the use of genetic instrumental variables to obtain causal inferences from observational data. Two recent developments for combining information on multiple uncorrelated instrumental variables (IVs) into a single causal estimate are as follows: (i) allele scores, in which individual‐level data on the IVs are aggregated into a univariate score, which is used as a single IV, and (ii) a summary statistic method, in which causal estimates calculated from each IV using summarized data are combined in an inverse‐variance weighted meta‐analysis. To avoid bias from weak instruments, unweighted and externally weighted allele scores have been recommended. Here, we propose equivalent approaches using summarized data and also provide extensions of the methods for use with correlated IVs. We investigate the impact of different choices of weights on the bias and precision of estimates in simulation studies. We show that allele score estimates can be reproduced using summarized data on genetic associations with the risk factor and the outcome. Estimates from the summary statistic method using external weights are biased towards the null when the weights are imprecisely estimated; in contrast, allele score estimates are unbiased. With equal or external weights, both methods provide appropriate tests of the null hypothesis of no causal effect even with large numbers of potentially weak instruments. We illustrate these methods using summarized data on the causal effect of low‐density lipoprotein cholesterol on coronary heart disease risk. It is shown that a more precise causal estimate can be obtained using multiple genetic variants from a single gene region, even if the variants are correlated. © 2015 The Authors. Statistics in Medicine published by John Wiley & Sons Ltd. INTRODUCTION An instrumental variable (IV) can be used to estimate the causal effect of a risk factor on an outcome from observational data 1, 2. A valid IV must be associated with the risk factor of interest but not associated with other factors on alternative causal pathways. This implies that it is not associated with any confounder of the risk factor–outcome association and that any causal pathway from the IV to the outcome passes through the risk factor 3. Much recent attention has been devoted to IV analysis in the context of Mendelian randomization, defined as the use of genetic variants as IVs 4, 5. The causal effect of the risk factor on the outcome with a single IV can be estimated by dividing the coefficient from the regression of the outcome on the IV by the coefficient from the regression of the risk factor on the IV 6. This is known as the ratio of coefficients method. Alternatively, the same estimate can be obtained by first regressing the risk factor on the IV and then regressing the outcome on the fitted values of the risk factor from the first‐stage regression 7. This is known as the two‐stage least squares (2SLS) method. The 2SLS method can be extended for use with multiple IVs 8. As the number of IVs increases, overfitting in the first‐stage regression model leads to systematic finite‐sample bias in the causal estimate 9. This bias, known as weak instrument bias, acts in the direction of the confounded observational association between the risk factor and outcome 10. When there is a single IV, the median bias of the ratio (or 2SLS) method estimator is negligible for all but the weakest of IVs 11. A recent methodological development to exploit this fact is to aggregate multiple IVs into a single univariate score, and to use this score as a single IV rather than to use multiple IVs 12. In Mendelian randomization, this is known as an allele score, genetic risk score or gene score. An alternative approach to combine information on multiple IVs is to use summarized data on the associations of genetic variants with risk factors and disease outcomes. These data are increasingly becoming available from large consortia, such as the Global Lipids Genetics Consortium (GLGC) for lipid fractions 13 and DIAGRAM for type 2 diabetes 14. Causal estimates can be obtained from these associations for a single genetic variant using the ratio method without the need for individual‐level data. Two methods for obtaining causal estimates from summarized data for multiple IVs have been proposed: a summary statistic method, in which the ratio estimates from each IV are combined in an inverse‐variance weighted meta‐analysis 15, 16, and a likelihood‐based method, in which the summarized data are modelled directly using a likelihood function 17, 18. The summary statistic method requires that the IVs are uncorrelated in their distributions (for genetic IVs, the variants are in linkage equilibrium). In this paper, we review and extend the literature on IV estimation methods with summarized data, currently described in disparate sources. In Section 2, we lay out the assumptions made in this paper for the identification of causal effects. In Section 3, we demonstrate how an allele score estimate with a pre‐specified choice of weights can be reproduced using summarized data. We rederive the known result for uncorrelated IVs that the allele score and summary statistic methods using an internally derived choice of weights give the same estimates as a (multivariable) 2SLS method; the estimates differ for other choices of weights. We investigate the bias and coverage properties of the allele score and summary statistic methods in simulation studies for different choices of weights, in particular with weak instruments. In Section 4, we derive extensions to the previously described methods that can be used when the IVs are correlated and similarly investigate their statistical properties. In Section 5, the methods are illustrated using summarized data on the causal effect of low‐density lipoprotein cholesterol (LDL‐c) on coronary heart disease (CHD) risk, comparing causal estimates obtained using a single genetic variant with those obtained using multiple genetic variants from the same gene region. Finally, we discuss the relevance of these methodological developments to applied practice (Section 6). For reference, a summary of methods for IV estimation discussed in this paper is given in Table 1. Sample code for implementing the methods is given in Appendix A.1. We clarify that the individual‐level data methods require individual participant data on the genetic variants used as IVs, risk factor and outcome. The summarized data methods only require data on the associations of the IVs with the risk factor and with the outcome. If limited individual‐level data are available (for example, on the IV–risk factor relationship but not the IV–outcome relationship), then summarized associations can be obtained from the individual‐level data, and the analysis can proceed using summarized data only. [Table: Summary of instrumental variable (IV) estimation methods discussed in this paper.] MODELLING ASSUMPTIONS In this paper, the situation of a continuous risk factor and a continuous outcome will be assumed, although the binary outcome case can be handled in a similar way. We assume that the causal effect of the risk factor on the outcome is linear and homogeneous in the population without effect modification. We also assume that the associations of the IVs with the risk factor are linear and