Bootstrap法与H-L法中位数差值的区间检验在非劣效试验中的模拟比较研究

Simulation comparison of interval test of median difference between Bootstrap method and H-L method in non-inferiority trial

毛昂; 曾子倩; 魏敏; 陈晓芳; 陈卫中

目的 通过计算机模拟研究，比较Bootstrap法和Hodges-Lehmann法(H-L法)在中位数差值非劣效性检验中的特点，为相关研究中统计学方法的选择提供依据。方法 以某临床试验中试验组与对照组咽痛消失时间的比较为基础，通过计算机模拟生成单组样本量分别为20、30、50、100、200各500个两独立样本，分别服从参数为90h（试验组）、100 h（对照组）的Poisson分布。针对每个样本采用基于正态近似和百分位数的Bootstrap法、H-L法求得中位数差值的置信区间，并通过置信区间下限与非劣性界值进行比较，得出三种方法的检验效能。结果 三种方法均表现为随着样本量增加，检验效能增加。在样本量为20时，H-L法与正态近似法检验效能相当（25% vs 24%），且都好与百分位数法(19%)。在样本量为30、50、100时，H-L法明显好于正态近似法与百分位数法，且正态近似法好于百分位数法。在样本量为200时，三种方法的检验效能相当,均在95%以上。结论 H-L法获得的区间最窄且最稳定，检验效能最高。正态近似法在样本较小时可以取得较好的检验效能，但区间较宽且波动范围大。在样本量较大时三种方法都可以取得较好的检验效能。

Objective Through computer simulation research, the performances of the Bootstrap method and Hodges-Lehmann method in the non-inferiority test of median difference were explored to provide a basis for the selection of statistical methods in related studies. Methods Based on the comparison of the disappearance time of pharyngeal pain between the experimental group and the control group in a clinical trial, the computer simulation was used to generate a single group of 500 samples with sample sizes of 20, 30, 50, 100 and 200, and the Poisson distribution with parameters of 90h (test group) and 100h (control group).For each sample, Bootstrap method and h-l method based on normal approximation and percentiles were used to obtain the confidence interval of the median difference, and the test effectiveness of the three methods was obtained by comparing the lower limit of the confidence interval with the non-inferior boundary value. Results When the sample size is 20, the H-L method is equivalent to the normal approximation method, and both are better than the percentile method. When the sample size is 30, 50, 100, the H-L method is significantly better than the normal approximation method and the percentile method, and the normal approximation method is better than the percentile method. When the sample size is 200, the three methods have the same performance. Conclusion The interval obtained by the H-L method is the narrowest and most stable, and has the highest test performance. The normal approximation method can achieve better test performance when the sample is small, but the interval is wide and the fluctuation range is large. When the sample size is large, all three methods can achieve better test performance.