Many clinical trials involve partially clustered data, where some observations belong to a cluster and others can be considered independent. For example, neonatal trials may include infants from single or multiple births. Sample size and analysis methods for these trials have received limited attention. A simulation study was conducted to (1) assess whether existing power formulas based on generalized estimating equations (GEEs) provide an adequate approximation to the power achieved by mixed effects models, and (2) compare the performance of mixed models vs GEEs in estimating the effect of treatment on a continuous outcome. We considered clusters that exist prior to randomization with a maximum cluster size of 2, three methods of randomizing the clustered observations, and simulated datasets with uninformative cluster size and the sample size required to achieve 80% power according to GEE‐based formulas with an independence or exchangeable working correlation structure. The empirical power of the mixed model approach was close to the nominal level when sample size was calculated using the exchangeable GEE formula, but was often too high when the sample size was based on the independence GEE formula. The independence GEE always converged and performed well in all scenarios. Performance of the exchangeable GEE and mixed model was also acceptable under cluster randomization, though under‐coverage and inflated type I error rates could occur with other methods of randomization. Analysis of partially clustered trials using GEEs with an independence working correlation structure may be preferred to avoid the limitations of mixed models and exchangeable GEEs.

中文翻译：

---

部分聚类试验中连续结果的混合效应模型和广义估计方程的性能，包括独立数据和配对数据


许多临床试验涉及部分聚类数据，其中一些观察结果属于一个聚类，而其他观察结果可以被认为是独立的。例如，新生儿试验可能包括单胎或多胎婴儿。这些试验的样本量和分析方法受到的关注有限。进行模拟研究的目的是 (1) 评估基于广义估计方程 (GEE) 的现有功效公式是否能够充分近似混合效应模型所实现的功效，以及 (2) 比较混合模型与 GEE 在估计治疗对持续结果的影响。我们考虑了随机化之前存在的聚类，最大聚类大小为 2，三种随机化聚类观察的方法，以及具有无信息聚类大小的模拟数据集以及根据具有独立性的基于 GEE 的公式实现 80% 功效所需的样本大小或可交换的工作关联结构。当使用可交换 GEE 公式计算样本量时，混合模型方法的经验功效接近名义水平，但当样本量基于独立 GEE 公式时，混合模型方法的经验功效往往过高。独立GEE在所有场景下始终保持收敛且表现良好。可交换 GEE 和混合模型的性能在聚类随机化下也是可以接受的，尽管其他随机化方法可能会出现覆盖不足和夸大的 I 类错误率。使用具有独立工作相关结构的 GEE 进行部分聚类试验的分析可能是首选，以避免混合模型和可交换 GEE 的限制。