Department of Methodology and Statistics, Utrecht University, Utrecht, The Netherlands.
Clin Trials. 2020 Aug;17(4):420-429. doi: 10.1177/1740774520913042. Epub 2020 Mar 19.
BACKGROUND/AIMS: This article studies the effect of attrition in the cluster randomized crossover trial. The focus is on the two-treatment two-period AB/BA design where attrition occurs during the washout period. Attrition may occur at either the subject level or the cluster level. In the latter case, clusters drop out entirely and provide no measurements in the second period. Subject attrition can only occur in the cohort design, where each subject receives both treatments. Cluster attrition can also occur in the cross-sectional design, where different subjects are measured in the two time periods. Furthermore, this article explores two different strategies to account for potential levels of attrition: increasing sample size and replacing those subjects who drop out by others.
The statistical model that takes into account the nesting of subjects within clusters, and the nesting of repeated measurements within subjects is presented. The effect of attrition is evaluated on the basis of the efficiency of the treatment effect estimator. Matrix algebra is used to derive the relation between efficiency, the degree of attrition, cluster size and the intraclass correlations: the within-cluster within-period correlation, the within-cluster between-period correlation and (in the case of a cohort design) the within-subject correlation. The methodology is implemented in two Shiny Apps.
Attrition in a cluster randomized crossover trial implies a loss of efficiency. Efficiency decreases with an increase of the attrition rate. The loss of efficiency due to attrition of subjects in a cohort design is largest for small number of subjects per cluster-period, but it may be repaired to a large degree by increasing the number of subjects per cluster-period or by replacing those subjects who drop out by others. Attrition of clusters results in a larger loss of efficiency, but this loss does not depend on the number of subjects per cluster-period. Repairing for this loss requires a large increase in the number of subjects per cluster-period. The methodology of this article is illustrated by an example on the effect of lavender scent on dental patients' anxiety.
This article provides the methodology of exploring the effect of attrition in cluster randomized crossover trials, and to repair for attrition. As such, it helps researchers plan their trial in an appropriate way and avoid underpowered trials. To use the methodology, prior estimates of the degree of attrition and intraclass correlation coefficients are needed. It is advocated that researchers clearly report the estimates of these quantities to help facilitate planning future trials.
背景/目的：本文研究了在整群随机交叉试验中失访的影响。重点是在两处理两周期 AB/BA 设计中，在洗脱期发生失访的情况。失访可能发生在个体水平或群体水平。在后一种情况下，整个群体完全退出，在第二阶段不提供任何测量值。个体失访只能发生在队列设计中，每个个体都接受两种处理。群体失访也可能发生在横断面设计中，其中不同的个体在两个时间点接受测量。此外，本文还探讨了两种不同的策略来考虑潜在的失访水平：增加样本量和用其他人替代退出的个体。
本文提出了一种考虑个体在群体内嵌套和重复测量在个体内嵌套的统计模型。根据处理效果估计量的效率来评估失访的影响。矩阵代数用于推导出效率、失访程度、群体大小和组内相关系数之间的关系：同一群体内同一时期的相关系数、同一群体内不同时期的相关系数以及（在队列设计的情况下）个体内的相关系数。该方法在两个 Shiny 应用程序中实现。
在整群随机交叉试验中失访会导致效率降低。效率随失访率的增加而降低。在队列设计中，个体失访导致的效率损失在每个群体-时期的个体数量较少时最大，但通过增加每个群体-时期的个体数量或用其他人替代退出的个体，可以在很大程度上修复这种损失。群体失访会导致更大的效率损失，但这种损失不依赖于每个群体-时期的个体数量。修复这种损失需要大量增加每个群体-时期的个体数量。本文通过一个关于薰衣草气味对牙科患者焦虑影响的例子来说明该方法。
本文提供了探索整群随机交叉试验中失访影响并修复失访的方法。因此，它可以帮助研究人员以适当的方式规划试验，避免试验效能不足。要使用该方法，需要事先估计失访程度和组内相关系数。本文提倡研究人员明确报告这些数量的估计值，以帮助促进未来试验的规划。
