Department of Biostatistics, University of Alabama at Birmingham, Birmingham, Alabama, USA.
Department of Medicine, Hospital for Special Surgery and Weill Cornell Medicine Center, New York, New York, USA.
Genet Epidemiol. 2021 Jul;45(5):549-560. doi: 10.1002/gepi.22387. Epub 2021 May 16.
Mendelian randomization (MR) applies instrumental variable (IV) methods to observational data using a genetic variant as an IV. Several Monte-Carlo studies investigate the performance of MR methods with binary outcomes, but few consider them in conjunction with binary risk factors.
To develop a novel MR estimator for scenarios with a binary risk factor and outcome; and compare to existing MR estimators via simulations and real data analysis.
A bivariate Bernoulli distribution is adapted to the IV setting. Empirical bias and asymptotic coverage probabilities are estimated via simulations. The proposed method is compared to the Wald method, two-stage predictor substitution (2SPS), two-stage residual inclusion (2SRI), and the generalized method of moments (GMM). An analysis is performed using existing data from the CLEAR study to estimate the potential causal effect of smoking on rheumatoid arthritis risk in African Americans.
Bias was low for the proposed method and comparable to 2SPS. The Wald method was often biased towards the null. Coverage was adequate for the proposed method, 2SPS, and 2SRI. Coverage for the Wald and GMM methods was poor in several scenarios. The causal effect of ever smoking on rheumatoid arthritis risk was not statistically significant using a variety of genetic instruments.
Simulations suggest the proposed MR method is sound with binary risk factors and outcomes, and comparable to 2SPS and 2SRI in terms of bias. The proposed method also provides more natural framework for hypothesis testing compared to 2SPS or 2SRI, which require ad-hoc variance adjustments.
孟德尔随机化（MR）将工具变量（IV）方法应用于观察数据，使用遗传变异作为 IV。几项蒙特卡罗研究调查了具有二项结局的 MR 方法的性能，但很少将其与二项风险因素结合考虑。
为具有二项风险因素和结局的情况开发一种新的 MR 估计量；并通过模拟和真实数据分析与现有的 MR 估计量进行比较。
将双变量伯努利分布应用于 IV 环境。通过模拟估计经验偏差和渐近覆盖概率。通过模拟比较了拟议的方法与 Wald 方法、两阶段预测器替换（2SPS）、两阶段残差纳入（2SRI）和广义矩方法（GMM）。使用 CLEAR 研究中的现有数据进行分析，以估计吸烟对非裔美国人类风湿关节炎风险的潜在因果效应。
拟议方法的偏差较低，与 2SPS 相当。Wald 方法通常偏向于零假设。拟议方法、2SPS 和 2SRI 的覆盖范围是足够的。Wald 和 GMM 方法的覆盖范围在几种情况下较差。使用各种遗传工具，吸烟与类风湿关节炎风险之间的因果关系没有统计学意义。
模拟表明，拟议的 MR 方法对于具有二项风险因素和结局的情况是可靠的，与 2SPS 和 2SRI 相比，在偏差方面相当。与 2SPS 或 2SRI 相比，拟议的方法还为假设检验提供了更自然的框架，后者需要特定的方差调整。
