临床试验中纵向缺失数据不同处理策略统计性能的比较

The statistical performance of different strategies for dealing with missing values in clinical trials

目的:评估不同缺失值处理方法在不同缺失模式、缺失机制及缺失比例下双臂优效临床试验中的统计性能,筛选出相应的最优缺失值处理方式。方法:使用SAS 9.4软件分别生成不同缺失模式(单调缺失或任意缺失)、缺失机制(完全随机缺失或随机缺失)以及不同缺失比例(0%~5%、5%~10%、10%~15%)的纵向模拟数据集,基于各模拟数据集分别使用不同缺失值处理策略进行疗效估计,并分别与完整数据集比较,以评价不同缺失值处理方式的统计性能。结果:当缺失比例<5%时,任意缺失模式下不同缺失值处理方式所得效应估计值均较为接近,单调缺失模式下只有重复测量混合效应模型(MMRM)以及多重填补后协方差分析误差较小。当缺失比例>5%时,不论缺失机制以及缺失模式,不同协方差矩阵结构的MMRM以及多重填补不同次数后协方差分析所得效应估计值与真实值之间仍最接近,且前者较后者更为稳定。而单一填补方式以及模式混合模型(PMM)随着缺失比例增加,其效应估计的误差也增加,尤其是单调缺失,当缺失比例为10%~15%时其误差最大。结论:不同缺失比例(0%~5%、5%~10%、10%~15%),缺失模式(单调缺失或任意缺失)以及缺失机制(完全随机缺失或随机缺失)下,MMRM误差最小,提示MMRM是处理双臂优效性临床试验中纵向定量数据缺失的首选方案。

Objective:To evaluate the statistical performance of various missing value processing methods in a two-arm superiority clinical trial with different missing patterns,mechanisms and proportions.Methods:We generated longitudinal simulation datasets containing monotonic and arbitrary missing patterns,missing mechanisms that occurred at random or completely at random,and missing proportions of 0%-5%,5%-10%and 10%-15%,respectively.Estimations of efficacy between the two groups after performing treatment in each simulated dataset by different missing value handling methods,and the statistical performance of the different methods were evaluated according to the magnitude of the difference in estimated efficacy from the full data set.Results:When the proportion of missingness was<5%,the effect estimates obtained from different missingness treatments in any missing patterns were close to each other,and the effect estimates from repeated measures mixed effects model(MMRM)and the analysis of covariance with multiple imputations in monotonic missing patterns were closest to the true values.When the proportion of missingness>5%,the effect estimates from MMRM with different covariance matrix structures and the analysis of covariance with multiple imputations were still closest to the true values regardless of the missing mechanisms and missing patterns,with the former more stable than the latter.In contrast,the error in the effect estimates of the single imputation approach and the mixed-mode model(PMM)increased with the proportion of missing,especially for the monotonic missing,where the error was greatest when the proportion of missing was 10%-15%.Conclusion:MMRM yielded the most accurate effect estimates for different missing proportions(0%-5%,5%-10%and 10%-15%),missing patterns(monotonic missing or arbitrary missing),and missing mechanisms(completely random missing or random missing),suggesting that MMRM is the preferred solution for dealing with missing longitudinal quantitative data in two-arm superiority clinical trials.