Weir Christopher J, Lees Kennedy R
Department of Medicine and Therapeutics, University of Glasgow, Gardiner Institute, Western Infirmary, Glasgow G11 6NT, UK.
Stat Med. 2003 Mar 15;22(5):705-26. doi: 10.1002/sim.1366.
Achieving balance on prognostic factors between treatment groups in a clinical trial is important to ensure that any observed treatment effect may be attributed to the treatment itself. Improving the balance on prognostic factors also potentially increases the statistical power attained in a trial. Substantial imbalances may occur by chance if simple randomization is used. Allocation of the treatment according to stratified random blocks based on clinical features is the conventional approach to obtain treatment groups that are as similar as possible. An alternative approach, known as minimization (or more generally as adaptive stratification), has also been proposed. We assessed the feasibility of adaptive stratification in the context of a clinical trial of insulin to control plasma glucose level following acute stroke. We determined suitable settings for the parameters in the adaptive stratification procedure by simulation studies. Specifically, we assessed: the optimal probability for allocating a patient to the preferred (leading to least imbalance on prognostic factors) treatment group; the number of variables that could be incorporated in the adaptive stratification algorithm; the weighting that should be given to each variable; and whether interactions between variables should be included. We then compared the statistical power, across a range of simulated treatment effects, between trials where treatments were allocated by stratified random blocks and by adaptive stratification. Finally, we considered the importance of the method of analysis in realizing the gain in power which may potentially be achieved by allocating treatments using stratified random blocks or adaptive stratification.
在临床试验中，使各治疗组间的预后因素达到平衡对于确保所观察到的任何治疗效果可归因于治疗本身至关重要。改善预后因素的平衡还可能提高试验所达到的统计效能。如果采用简单随机化，可能会偶然出现显著的不平衡。根据基于临床特征的分层随机区组进行治疗分配是获得尽可能相似治疗组的传统方法。还提出了另一种方法，即最小化法（或更一般地称为适应性分层法）。我们在一项关于胰岛素控制急性卒中后血糖水平的临床试验背景下评估了适应性分层的可行性。我们通过模拟研究确定了适应性分层程序中参数的合适设置。具体而言，我们评估了：将患者分配至首选治疗组（导致预后因素不平衡最小）的最佳概率；可纳入适应性分层算法的变量数量；应赋予每个变量的权重；以及是否应纳入变量之间的相互作用。然后，我们在一系列模拟治疗效果中比较了通过分层随机区组和适应性分层分配治疗的试验之间的统计效能。最后，我们考虑了分析方法对于实现通过使用分层随机区组或适应性分层分配治疗可能潜在获得的效能增益的重要性。
