Department of Sociology, Columbia University, New York, NY 10027;
Department of Economics, Vrije Universiteit Amsterdam, 1081 HV Amsterdam, The Netherlands.
Proc Natl Acad Sci U S A. 2018 May 29;115(22):E4970-E4979. doi: 10.1073/pnas.1707388115. Epub 2018 Apr 23.
Identifying causal effects in nonexperimental data is an enduring challenge. One proposed solution that recently gained popularity is the idea to use genes as instrumental variables [i.e., Mendelian randomization (MR)]. However, this approach is problematic because many variables of interest are genetically correlated, which implies the possibility that many genes could affect both the exposure and the outcome directly or via unobserved confounding factors. Thus, pleiotropic effects of genes are themselves a source of bias in nonexperimental data that would also undermine the ability of MR to correct for endogeneity bias from nongenetic sources. Here, we propose an alternative approach, genetic instrumental variable (GIV) regression, that provides estimates for the effect of an exposure on an outcome in the presence of pleiotropy. As a valuable byproduct, GIV regression also provides accurate estimates of the chip heritability of the outcome variable. GIV regression uses polygenic scores (PGSs) for the outcome of interest which can be constructed from genome-wide association study (GWAS) results. By splitting the GWAS sample for the outcome into nonoverlapping subsamples, we obtain multiple indicators of the outcome PGSs that can be used as instruments for each other and, in combination with other methods such as sibling fixed effects, can address endogeneity bias from both pleiotropy and the environment. In two empirical applications, we demonstrate that our approach produces reasonable estimates of the chip heritability of educational attainment (EA) and show that standard regression and MR provide upwardly biased estimates of the effect of body height on EA.
在非实验数据中识别因果效应是一个长期存在的挑战。最近一种流行的解决方案是利用基因作为工具变量[即孟德尔随机化（MR）]的想法。然而，这种方法存在问题，因为许多感兴趣的变量在遗传上是相关的，这意味着许多基因可能直接或通过未观察到的混杂因素影响暴露和结果。因此，基因的多效性效应本身就是非实验数据中产生偏差的一个来源，这也会破坏 MR 纠正非遗传来源内生性偏差的能力。在这里，我们提出了一种替代方法，即遗传工具变量（GIV）回归，它在存在多效性的情况下提供了暴露对结果的影响的估计。作为一个有价值的副产品，GIV 回归还提供了结果变量的芯片遗传力的准确估计。GIV 回归使用与感兴趣的结果相关的多基因评分（PGS），这些评分可以从全基因组关联研究（GWAS）的结果中构建。通过将 GWAS 样本分为不重叠的子样本，我们得到了多个结果 PGS 的指标，可以相互作为工具变量使用，并且与兄弟姐妹固定效应等其他方法结合使用，可以解决多效性和环境引起的内生性偏差。在两个实证应用中，我们证明了我们的方法可以对教育程度（EA）的芯片遗传力进行合理估计，并表明标准回归和 MR 提供了身高对 EA 影响的向上偏估计。
