The EMICS Tool to Design Mixed-methods Studies in Epidemiology.
The EMICS Tool to Design Mixed-methods Studies in Epidemiology.
复制标题
用于设计流行病学混合方法研究的 EMICS 工具。
DOI：
10.1097/ede.0000000000001718
复制
发表时间：
2024
期刊：
影响因子：
--
通讯作者：
Qi,Hanfei
中科院分区：
文献类型：
--
作者：
Houghton,LaurenC;Qi,Hanfei
我们的论文的关键信息是，当使用观察数据来估计妊娠期间干预措施的效果时，目标试验的规范和仿真可以防止设计偏差不朽的时间和选择。[1]为了传达这一信息，我们使用了一种简单的干预措施（疫苗接种），并描述了一种简单的分析方法（模拟与胎龄匹配的序贯试验）。本文没有详细讨论的另一个偏倚是混淆。为了关注设计问题，我们仅考虑在无残留时变混杂因素的简化假设下对基线协变量（包括胎龄）进行调整。这个假设在我们的例子中可能近似正确。我们提出了不同的方法来调整基线协变量，如逆概率加权，标准化（g公式），配对删失匹配。Latour等人2提出了在每个序贯试验中存在时变混杂时应该做些什么。答案是测量时变混杂因素，并使用g方法（例如逆概率加权和g公式）对其进行调整，即使在存在治疗混杂因素反馈的情况下，该方法也可以适当处理时变混杂因素。Latour et al.还指出，效应的大小可能因胎龄而异，因此效应估计值取决于人群中接种疫苗时孕周的分布。明确的靶向试验模拟促进了关于治疗效果异质性的讨论，这在许多以前的研究中没有提出，这些研究只是比较了接种疫苗和未接种疫苗之间的自然流产风险。通过回避对时变混杂的深入讨论，我们并不打算将混杂视为对观察性研究有效性的严重威胁，而是专注于确保不存在设计偏差的方法学方法，以便我们可以将注意力集中在混杂上。3
The key message of our paper was that specification and emulation of a target trial prevent design biases—immortal time and selection—when using observational data to estimate the effects of interventions during pregnancy. 1 To deliver this message, we used a simple intervention (vaccination) and described a simple analytic approach (emulation of sequential trials matched on gestational age). An additional bias that the paper does not cover at length is confounding. To focus on design issues, we only considered adjustment for baseline covariates (including gestational age) under the simplifying assumption of no residual time-varying confounding. This assumption is likely approximately true in our example. We proposed different methods to adjust for baseline covariates, such as inverse probability weighting, standardization (the g-formula), and matching with pairwise censoring. Latour et al. 2 ask what should be done in the presence of time-varying confounding within each sequential trial. The answer is to measure the time-varying confounders and adjust for them using g-methods—such as inverse probability weighting and the g-formula—that can appropriately handle time-varying confounders even in the presence of treatment-confounder feedback. Latour et al. also point out that the magnitude of the effect may vary across gestational age and thus the effect estimates depend on the distribution of gestational week at vaccination in the population. An explicit target trial emulation facilitates this conversation about treatment effect heterogeneity, which was not raised in many previous studies that simply compared the risk of spontaneous abortions between vaccinated and not vaccinated. By sidestepping an in-depth discussion about time-varying confounding, we did not intend to dismiss confounding as a grave threat to the validity of observational studies, but rather to focus on methodologic approaches that ensure the absence of design biases so that we can then focus our attention on confounding. 3