Uddin Md Jamal, Groenwold Rolf H H, de Boer Anthonius, Belitser Svetlana V, Roes Kit C B, Hoes Arno W, Klungel Olaf H
Division of Pharmacoepidemiology and Clinical Pharmacology, Utrecht Institute for Pharmaceutical Sciences, University of Utrecht, Utrecht, the Netherlands.
Pharmacoepidemiol Drug Saf. 2014 Feb;23(2):165-77. doi: 10.1002/pds.3555. Epub 2013 Dec 5.
Instrumental variable (IV) analysis is becoming increasingly popular to adjust for confounding in observational pharmacoepidemiologic research. One of the prerequisites of an IV is that it is strongly associated with exposure; if it is weakly associated with exposure, IV estimates are reported to be biased. We aimed to assess the performance of IV estimates in various (pharmaco-)epidemiologic settings.
Data were simulated for continuous/binary exposure, outcome and IV in cohort and nested case-control (NCC) designs with different incidences of the outcome. Pearson's correlation, point bi-serial correlation, odds ratio (OR), and F-statistic were used to assess the IV-exposure association. Two-stage analysis was performed to estimate the exposure effect.
For all types of IV and exposure in the cohort and NCC designs, IV estimates were extremely unstable and biased when the IV was very weakly associated with exposure (e.g. Pearson's correlation < 0.15 for continuous or OR < 2.0 for binary IV and exposure; although specific cut-off values depend on simulation settings). For stronger IVs, estimates were unbiased and become less variable compared with weaker IVs in the case of continuous and binary (risk difference scale) outcomes. For a similar IV-exposure association (e.g. OR = 1.4 and 5% incidence of the outcome), the variability of the estimates was more pronounced in the NCC (standard deviation = 2.37, case : control = 1:5) compared with the cohort design (standard deviation = 1.14). The variability was even more pronounced for rare (≤1%) outcomes. However, IV estimates from the NCC design became less variable with an increasing number of controls per case. Moreover, estimates were biased when the IV was related to confounders even with strong IVs.
Instrumental variable analysis performs poorly when the IV-exposure association is extremely weak, especially in the NCC design. IV estimates in the NCC design become less variable when the number of control increases. As NCC does not use the entire cohort, in order to achieve stable estimates, this design requires a stronger IV-exposure association than the cohort design.
在观察性药物流行病学研究中，工具变量（IV）分析在调整混杂因素方面正变得越来越流行。IV的前提条件之一是它与暴露密切相关；如果它与暴露的关联较弱，据报道IV估计值会有偏差。我们旨在评估IV估计值在各种（药物）流行病学环境中的表现。
针对队列研究和巢式病例对照（NCC）设计中的连续/二元暴露、结局和IV，在结局发生率不同的情况下进行数据模拟。使用Pearson相关性、点二列相关性、比值比（OR）和F统计量来评估IV与暴露的关联。进行两阶段分析以估计暴露效应。
对于队列研究和NCC设计中的所有类型的IV和暴露，当IV与暴露的关联非常弱时（例如，连续变量的Pearson相关性<0.15或二元IV与暴露的OR<2.0；尽管具体的临界值取决于模拟设置），IV估计值极其不稳定且有偏差。对于较强的IV，在连续和二元（风险差尺度）结局的情况下，估计值无偏差，并且与较弱的IV相比变异性更小。对于相似的IV与暴露的关联（例如，OR = 1.4且结局发生率为5%），与队列设计（标准差 = 1.14）相比，NCC设计（标准差 = 2.37，病例：对照 = 1:5）中估计值的变异性更明显。对于罕见（≤1%）结局，变异性甚至更明显。然而，随着每个病例对照数量的增加，NCC设计的IV估计值变异性降低。此外，即使IV较强，但当IV与混杂因素相关时，估计值仍有偏差。
当IV与暴露的关联极其微弱时，工具变量分析表现不佳，尤其是在NCC设计中。当对照数量增加时，NCC设计中的IV估计值变异性降低。由于NCC不使用整个队列，为了获得稳定的估计值，该设计比队列设计需要更强的IV与暴露的关联。
