多分类结局指标中两类别占比之比的统计推断方法

摘要

Objective Statistical inference methods for comparisons between two categories of the multinomial outcome are not available now.This study aims to develop hypothesis testing and interval estimation methods based on the percent ratio(PR).Methods Firstly,the variance of log transformed PR was constructed based on the delta method and the hypothesis testing method was established using normal approximation method.The confidence interval was estimated based on Koopman method,logarithm transformation method and adjusted logarithm transformation method.Type I error,statistical power and the coverage rate of confidence interval were assessed by Monte Carlo simulation methods.Results Type I error of the developed hypothesis testing method was well controlled.All coverage rates of constructed 95%confidence interval methods were around.Koopman method was superior to logarithm transformation method and adjusted logarithm transformation method,but all methods were unstablewhen the sample size was too small(for instance,less than 20).Conclusion The hypothesis testing method and confidence interval methods brought up in the paper can meet application requirements and the CI estimation method base on Koopman''s method is recommended for confidence interval estimation.%目的 针对多分类结局指标数据,就某两类占比之比构建相应的假设检验及置信区间估计方法.方法 先根据Delta法构建对数变换后比值的方差,然后用正态近似法构建其假设检验方法,分别基于Koopman法、对数变换法和校正的对数变换法构建其置信区间.通过模拟验证假设检验方法的一类错误、检验效能和置信区间覆盖率.最后以实例进行说明.结果 基于占比比值的假设检验方法可以较好的控制一类错误.三种置信区间方法的覆盖率均在95%左右,其中基于Koopman法更优.当样本量太小(如不足20例)时,所有方法均不够稳健.结论 本研究构建的多分类结局指标某两类占比之比的统计推断方法表现能满足应用需求,并推荐基于Koopman法的置信区间估计.