Center for Evidence Synthesis in Health, School of Public Health, Brown University, Providence, RI; Department of Health Services, Policy & Practice, School of Public Health, Brown University, Providence, RI.
Clinical Epidemiology Program, Ottawa Hospital Research Institute, Ottawa, Canada; School of Epidemiology and Public Health, University of Ottawa, Ottawa, Canada.
J Clin Epidemiol. 2021 Nov;139:307-318. doi: 10.1016/j.jclinepi.2021.06.011. Epub 2021 Jun 22.
Incorporating cluster randomized trials (CRTs) into meta-analyses is challenging because appropriate standard errors of study estimates accounting for clustering are not always reported. Systematic reviews of CRTs often use a single constant external estimate of the intraclass correlation coefficient (ICC) to adjust study estimate standard errors and facilitate meta-analyses; an approach that fails to account for possible variation of ICCs among studies and the imprecision with which they are estimated. Using a large systematic review of the effects of diabetes quality improvement interventions, we investigated whether we could better account for ICC variation and uncertainty in meta-analyzed effect estimates by imputing missing ICCs from a posterior predictive distribution constructed from a database of relevant ICCs.
We constructed a dataset of ICC estimates from applicable studies. For outcomes with two or more available ICC estimates, we constructed posterior predictive ICC distributions in a Bayesian framework. For a selected continuous outcome, glycosylated hemoglobin (HbA1c), we compared the impact of incorporating a single constant ICC versus imputing ICCs drawn from the posterior predictive distribution when estimating the effect of intervention components on post treatment mean in a case study of diabetes quality improvement trials.
Using internal and external ICC estimates, we were able to construct a database of 59 ICCs for 12 of the 13 review outcomes (range 1-10 per outcome) and estimate the posterior predictive ICC distribution for 11 review outcomes. Synthesized results were not markedly changed by our approach for HbA1c.
Building posterior predictive distributions to impute missing ICCs is a feasible approach to facilitate principled meta-analyses of cluster randomized trials using prior data. Further work is needed to establish whether the application of these methods leads to improved review inferences for different reviews based on different factors (e.g., proportion of CRTs and CRTs with missing ICCs, different outcomes, variation and precision of ICCs).
将整群随机试验 (cluster randomized trials, CRTs) 纳入荟萃分析具有挑战性，因为并非总是报告适当的研究估计量的聚类标准误差。CRTs 的系统评价通常使用单个常数外部估计值来调整研究估计量的标准误差，以促进荟萃分析；这种方法未能考虑到 ICC 在研究之间的可能变化，以及它们估计的不精确性。我们使用一项关于糖尿病质量改进干预效果的大型系统评价，研究是否可以通过从相关 ICC 数据库中构建的后验预测分布中插补缺失的 ICC 来更好地解释荟萃分析中效应估计的 ICC 变化和不确定性。
我们从适用的研究中构建了一个 ICC 估计数据集。对于具有两个或更多可用 ICC 估计值的结局，我们在贝叶斯框架中构建了后验预测 ICC 分布。对于一个选定的连续结局，糖化血红蛋白 (glycosylated hemoglobin, HbA1c)，我们比较了当估计干预成分对糖尿病质量改进试验的治疗后平均值的影响时，将单个常数 ICC 纳入与从后验预测分布中插补 ICC 对纳入案例研究的影响。
使用内部和外部 ICC 估计值，我们能够构建 13 个综述结局中的 12 个的 59 个 ICC 数据库（每个结局 1-10 个），并估计 11 个综述结局的后验预测 ICC 分布。我们的方法对 HbA1c 的综合结果没有明显改变。
构建后验预测分布以插补缺失的 ICC 是一种可行的方法，可以使用先前的数据促进对整群随机试验的有原则的荟萃分析。需要进一步研究以确定这些方法的应用是否会基于不同的因素（例如，CRTs 的比例和缺失 ICC 的 CRTs、不同的结局、ICC 的变化和精度）为不同的综述提供改进的综述推断。
