Response to: How to design and analyse cluster randomized trials with a small number of clusters? Comment on Leyrat et al.
Response to: How to design and analyse cluster randomized trials with a small number of clusters? Comment on Leyrat et al.
复制标题
回应：如何设计和分析少量聚类的整群随机试验？
DOI：
10.1093/ije/dyy062
复制
发表时间：
2018
影响因子：
7.7
通讯作者：
Leyrat C
中科院分区：
文献类型：
--
作者：
Leyrat C
我们要感谢Van Breukelen和Candel对我们手稿的评论。虽然我们大体上同意他们的观点，但我们想澄清几点。首先，他们认为，根据现有文献，我们的结果是可以理解的。我们同意，我们文章中的一些结果是众所周知的(例如，未加权的聚类级分析失去了效率)。然而，这些方法仍然普遍使用，2因此我们将它们包括在内，以便从经验上证明其他方法的好处。此外，我们不知道广义估计方程(GES)、混合效应模型和针对连续结果的群集级分析之间的任何经验比较。我们同意Van Breukelen和Candel的观点，即这些方法的一些理论结果是可用的；然而，这些结果通常基于不总是转化为现实场景的近似(特别是关于小样本修正)，因此使用模拟来评估这些方法在一系列现实场景中的属性是有用的。3其次，Van Breukelen和Candel对我们模拟研究中使用的样本量公式提出了异议。因为样本量公式取决于基本的分析模型，所以没有一个公式适用于所有被比较的分析方法。然而，我们的目标是在第一类错误率和功率方面对每种分析方法的相对性能进行基准测试。鉴于所使用的具体样本量公式不会影响哪种分析方法表现最好，我们不确定为什么
We would like to thank Van Breukelen and Candel for their comments on our manuscript. 1 Although we broadly agree with them, we would like to clarify several points. First, they argue that our results can be understood in light of the existing literature. We agree that some of the results in our article are well known (eg that unweighted cluster-level analyses lose efficiency). However, these approaches are still commonly used, 2 and so we included them in order to empirically demonstrate the benefit of other approaches. Furthermore, we are unaware of any empirical comparison between generalized estimating equations (GEEs), mixed-effect models and cluster-level analyses for continuous outcomes. We agree with Van Breukelen and Candel that some theoretical results are available for these approaches; however, these are often based on approximations which do not always translate to realistic scenarios (particularly regarding small-sample corrections), and so it is useful to assess the properties of these approaches across a range of realistic scenarios using simulation. 3 Second, Van Breukelen and Candel take issue with the sample size formula used in our simulation study. Because sample size formulas depend on the underlying analysis model, there is no single formula which is appropriate for all the analysis methods being compared. However, our aim was to benchmark the relative performance of each analysis method in terms of type-I error rate and power. Given that the specific sample size formula used will have no impact on which analysis approach performs best, we are unsure why