配对设计率差及率比置信区间构建新方法

发布时间：2018-03-12 12:08

本文选题：非参数置信区间　切入点：配对设计　出处：《南方医科大学》2017年博士论文　论文类型：学位论文

【摘要】：背景:配对率数据在生物统计学应用领域中极为常见。对于配对率数据,不论在非劣、优效还是等效性验证中,置信区间都扮演着重要的角色。对于配对率差的置信区间构建,目前推荐的Tango score近似法的置信区间宽度相对较宽,推荐的非条件精确置信区间及对应的近似非条件精确置信区间运算量较大。Lange等人将灵敏度、特异度当成特殊的AUC,为配对率的置信区间构建提供了一条新思路。但现有非参数法计算复杂,限制了其应用。对于配对率比的置信区间构建,已有的模拟研究得出了不一致的结论,因此需要对已有方法重新进行模拟比较。本研究在已有方法的基础上增加了新的基于不同相关系数处理方法的MOVER法构建策略。目的:①利用AUC基于U统计量的非参数构建方法构建配对率差的非参数置信区间,并与已有方法进行模拟比较;②基于MOVER法提出两种新的配对率比置信区间构建策略,并与已有方法进行模拟比较。方法:①将率数据表示为U统计量的表达式,基于structural component方法构建配对率协方差阵;基于正态分布及t分布近似构建配对率差的非参数置信区间方法。②采用两种不同的相关系数调整方法构建配对率差MOVER Wilson score置信区间,并将所有方法进行模拟比较。③模拟比较评价指标采用置信区间覆盖率、平均/中位宽度以及尾侧不覆盖率比值指标。结果:基于U统计量,本研究构建了基于正态分布及自由度为n-1的t分布近似配对率差置信区间。基于配对率相关系数φ估计,本研究基于φ的连续性矫正法给出了新的MOVER置信区间构建策略。同时本研究基于将配对率四格表数据当做多项分布数据,得出了基于多项分布率比的MOVER置信区间构建策略。模拟研究得出:①对于配对率差的置信区间,本研究提出的基于自由度n-1的t分布非参数法与Tango score法的覆盖率在各参数设置下均高于94%,并且除了在小样本情况下,本研究提出的非参数方法置信区间宽度比Tango score法置信区间要窄,尾侧不覆盖率指标各参数设置下均有较好的表现。相比精确置信区间法本研究提出的非参数法计算更为简单。②对于配对率比置信区间法,本研究提出的φ连续性校正MOVER Wilson score覆盖率整体表现最优,置信区间宽度较窄。近似score法及本研究提出的基于多项分布MOVER法表现次之。就尾侧不覆盖率来讲,基于近似score法、基于多项分布对数变换MOVER法及φ连续性校正MOVER Wilson score法要优于其他方法。结论:本研究建立了基于U统计量的配对率差置信区间构建t分布近似非参数法。该方法计算简便,而且模拟研究显示不论在大样本还是小样本下其均有较好的表现,适合在应用中推广使用。将配对率数据当成多项分布数据及基于相关系数φ连续性校正所构建的MOVER Wilson score法的配对率比置信区间相比已有方法表现较好,其中以基于连续性校正φ估计的方法表现最优。综合考虑本研究推荐应用中使用连续性校正φ估计的MOVER Wilson score法。
[Abstract]:Background: matching rate data in biostatistics applications is very common. For paired data, regardless of the non inferior, superior efficiency or equivalenceverification, confidence intervals have played an important role. For the matching rate difference confidence interval construction, the width of confidence interval currently recommended Tango score approximation method is relatively wide approximate, non conditional non recommended conditions of exact confidence intervals and the corresponding confidence interval computation by.Lange sensitivity, specificity as a special AUC for paired rate confidence interval provides a new way to construct a non parametric method. But the existing computational complexity, which limits its application. For the construction of confidence interval rate matching ratio, simulation studies have yielded conflicting results, so we need to re simulate the comparison of existing methods. This research is based on the existing method adds a new base in With the method of MOVER correlation coefficient method to construct strategy. Objective: 1. Using AUC to construct the construction method of non parametric paired confidence interval rate difference based on nonparametric U statistics, and by comparing with existing methods; the MOVER method based on the proposed two new pairing rate than the confidence interval construction strategy, and with existing methods comparison of simulation. Methods: the data rate expressed as U statistic expression, structural component method of constructing covariance matrix based on the matching rate; normal distribution and t distribution approximation to construct non parametric paired confidence interval method. The rate difference based on correlation coefficients of two different adjustment methods to construct pairing rate difference MOVER score confidence interval Wilson and, all methods are compared. The simulation simulation and comparison of evaluation index by using the confidence interval coverage rate, average / median width and tail side coverage ratio index: Based on U statistics, this study constructed a normal distribution and N-1 degrees of freedom t distribution approximation matching rate difference confidence interval. Based on the correlation coefficient matching rate estimation based on this research, based on the continuity of the new correction method with MOVER confidence interval construction strategy. At the same time this study based on the matching rate of four case table data as multinomial distribution data obtained multinomial distribution ratio MOVER confidence interval based on the simulation of the confidence interval matching rate difference, the degree of freedom of n-1 t distribution and Tango non parametric method score method based on coverage in all parameters was higher than 94%. And except in the case of small samples, this study proposes a non parametric method of confidence interval width is narrower than the Tango score confidence interval, caudal not coverage parameters provided better performance were compared. The exact position The non parameter method to calculate the confidence interval method proposed is simple. For the matching rate than the confidence interval method, this study proposes a MOVER Wilson score with continuity correction coverage best overall performance, the confidence interval width is narrow. The approximate score method and the MOVER method based on the multinomial distribution of performance. The tail side does not cover rate, approximate score method based on the multinomial distribution logarithmic transformation method of MOVER and MOVER Wilson score with continuity correction method is better than other methods based on. Conclusion: This study established the U statistic matching rate difference confidence interval constructed t distribution approximation method based on non parameter. This method is simple in calculation, and simulation research shows that in large or small sample size which have good performance, suitable for use in the application. The paired data rate as multinomial distribution data and the correlation coefficient based on continuity correction The matching rate of the MOVER Wilson score method is better than that of the confidence interval method, and the method based on the continuous correction method is the best. The MOVER Wilson score method is recommended for the recommended application.

【学位授予单位】：南方医科大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：O212.1

【相似文献】