临床研究样本代表性评估方法的对比研究

A Comparative Study on Evaluation Methods of Sample Representativeness for Clinical Research

目的对现有样本代表性评估方法进行全面比较和探讨,为临床研究样本代表性评估方法选择提供参考。方法结合国内肺癌患者特征的分布以及国内临床研究样本筛选的实际情况,模拟肺癌患者目标人群,抽取不同样本量和不同偏离程度的样本,使用现有样本代表性评估方法计算样本代表性,同时计算疗效估计偏差(bias),通过建立各方法代表性测量值与bias之间的相关性模型,分析各方法评估代表性的准确性和稳定性。结果整体结构差异率(rate of overall struction variation,RV)RV1和RV2及基于倾向评分的C统计量、基尼集中比求和(sum Gini concentration ratio,SGCR)及K-S距离(kolmogorov-smirnov distance,KSD)均能较好地测量不同样本的偏离程度。在不同样本量下,RV2和RV1与bias相关模型的R^(2)值均大于0.90,C统计量、SGCR及K-S距离的R^(2)大于0.80。结论因考虑了特征权重,整体结构差异率更为准确、稳定,尤其是RV2能更好地测量不同偏离程度样本的代表性、准确反映估计偏差;在难以获得特征重要性信息时,SGCR及利用倾向评分的方法中的C统计量和K-S距离测量代表性的可靠性也可以接受。

Objective To compare the existing evaluation methods of sample′s representativeness and provide reference for selection of sample representativeness evaluation methods in clinical research.Methods Simulate the target population of lung cancer patients and select samples with different sample sizes and different degrees of deviation based on the distribution of traits of lung cancer patients in China and the actual situation of sample screening in domestic clinical studies.Calculate sample representativeness using the existing evaluation methods of sample′s representativeness,and calculate estimation deviation(bias).By constructing the correlation model between the measured value of each method and bias,analyze the accuracy and stability of each method.Results The overall structural variance rate RV1、RV2、C-statistic based on propensity score、SGCR and K-S distance could well measure the degrees of deviation of different samples.Under different sample sizes,the R^(2) of RV2 and RV1 are greater than 0.90,and R^(2) of C-Statistic、SGCR and K-S distance were greater than 0.80.Conclusion The overall structural variance rate is more accurate and stable because the traits weight is taken into account.In particular,RV2 can better measure the representativeness of samples with different degrees of deviation and accurately reflect the estimation deviation.However,when it is difficult to obtain the feature importance information,the reliability of the representative measurement of SGCR as well as C-statistic and K-S distance used the propensity score-based method are acceptable.