Borhan Sayem, Mallick Rizwana, Pillay Mershen, Kathard Harsha, Thabane Lehana
Department of Health Research Methods, Evidence and Impact, McMaster University, Hamilton, ON, Canada.
Biostatistics Unit, Research Institute of St Joseph's Healthcare, Hamilton, ON, Canada.
Contemp Clin Trials Commun. 2019 Jul 5;15:100405. doi: 10.1016/j.conctc.2019.100405. eCollection 2019 Sep.
The assessment of the sensitivity of statistical methods has received little attention in cluster randomized trials (CRTs), especially for stratified CRT when the outcome of interest is continuous. We empirically examined the sensitivity of five methods for analyzing the continuous outcome from a stratified CRT - aimed to investigate the efficacy of the Classroom Communication Resource (CCR) compared to usual care to improve the peer attitude towards children who stutter among grade 7 students. Schools - the clusters, were divided into quintile based on their socio-political resources, and then stratified by quintile. The schools were then randomized to CCR and usual care groups in each stratum. The primary outcome was Stuttering Resource Outcomes Measure. Five methods, including the primary method, were used in this study to examine the effect of CCR. The individual-level methods were: (i) linear regression; (ii) mixed-effects method; (iii) GEE with exchangeable correlation structure (primary method of analysis). And the cluster-level methods were: (iv) cluster-level linear regression; and (v) meta-regression. These methods were also compared with or without adjustment for stratification. Ten schools were stratified by quintile, and then randomized to CCR (223 students) and usual care (231 students) groups. The direction of the estimated differences was same for all the methods except meta-regression. The widths of the 95% confidence intervals were narrower when adjusted for stratification. The overall conclusion from all the methods was similar but slightly differed in terms of effect estimate and widths of confidence intervals.
Clinicaltrials.gov, NCT03111524. Registered on 9 March 2017.
在整群随机试验（CRT）中，统计方法敏感性的评估很少受到关注，尤其是当感兴趣的结局为连续性变量时的分层CRT。我们通过实证研究了五种分析分层CRT连续性结局的方法的敏感性——旨在研究课堂交流资源（CCR）相较于常规护理在改善七年级学生对口吃儿童同伴态度方面的疗效。学校——作为整群，根据其社会政治资源分为五分位数，然后再按五分位数分层。然后将各学校随机分配到每个分层中的CCR组和常规护理组。主要结局是口吃资源结局测量。本研究使用了包括主要方法在内的五种方法来检验CCR的效果。个体水平的方法有：（i）线性回归；（ii）混合效应法；（iii）具有可交换相关结构的广义估计方程（主要分析方法）。整群水平的方法有：（iv）整群水平线性回归；（v）meta回归。这些方法还比较了有无分层调整的情况。十所学校按五分位数分层，然后随机分配到CCR组（223名学生）和常规护理组（231名学生）。除meta回归外，所有方法估计差异的方向均相同。调整分层后，95%置信区间的宽度变窄。所有方法得出的总体结论相似，但在效应估计和置信区间宽度方面略有不同。
Clinicaltrials.gov，NCT03111524。于2017年3月9日注册。
