【作者】 周军；

【导师】 陈平雁；

【作者基本信息】 南方医科大学 ， 流行病与卫生统计学（生物统计）， 2018， 硕士

【摘要】 背景近年来循证医学(Evidence Based Medicine,EBM)得到了迅速发展,基于meta分析的循证医学研究结果也作为较高级别的证据来指导临床实践。当前对于均值、率以及生存结局数据的meta分析方法得到了全面应用及发展,其方法学也较为成熟,但是对于等级资料的分析目前并无较好的处理方法。对于等级结局数据,目前常采用将等级资料近似为连续型数据或将其拆分成二分类结局后采用均值或率数据的meta分析方法进行分析。此方式无疑会损失有用的数据信息甚至可能得出错误的结论。鉴于此,本研究拟就等级资料的meta分析方法进行研究。Alan Agresti 1980年曾针对等级资料提出两组间比较的Generalized Odds Ratios(GenOR)值,其当等级只有两类时正好为率数据比较的OR值,因此可以理解为是常用率数据组间比较OR值就等级资料的推广,可将其称为等级资料OR值。该OR值良好的统计性能为本研究解决等级资料meta分析方法提供了良好的基础。目的本研究拟针对两独立样本和两配对样本等级资料组间比较,构建基于等级资料OR值——GenOR的meta分析方法,研究将同时构建固定效应模型及随机效应模型,满足不同应用需求。方法研究首先计算每个研究GenOR统计量,然后基于逆方差法进行加权,分别在固定效应模型及随机效应模型假设下得到合并后的效应量,并构建其95%置信区间。所有方法均采用MonteCarlo统计模拟以评价其统计性能。点估计准确性评价指标包括点估计的平均偏差(Bias)、平均相对偏差及均方误差(Mean Square Error,MSE),95%。置信区间评价指标包括置信区间覆盖率、单侧不覆盖率。结果本研究基于逆方差加权法构建了两独立样本及两配对样本等级资料meta分析固定效应模型及随机效应模型。针对两独立样本,模拟研究显示固定效应下,本研究所提方法不论在大样本还是小样本情况下,其点估计值的平均偏差,平均相对偏差,均方误差均较小,95%置信区间覆盖率接均近于95%,说明点估计及区间估计均较准确;随机效应模型下各参数设置下置信区间覆盖率不理想,有待进一步改进。针对两配对样本,模拟研究显示本研究所提方法不论是固定效应模型还是随机效应模型的点估计的偏差,平均相对偏差,均方误差均较小。除随机效应模型小样本下置信区间覆盖率较低外,其他参数下95%置信区间覆盖率均接近95%。结论本研究构建的等级资料meta分析方法除两独立样本随机效应模型有待改进外,其他模型统计性能优良,可满足应用需求,建议推广应用。更多还原

【Abstract】 BackgroundIn recent years,evidence based medicine(EBM)has grown rapidly.Evidence-based medical research based on meta-analysis also serves as a high-level evidence to guide clinical practice.At present,the meta-analysis methods for the mean,rate,and survival outcome data has been fully applied and developed,and its methodology is also relatively mature.However,there is currently no good meta-analysis method for the analysis of ordinal data.For ordinal outcome data,meta-analysis methods are currently used to approximate the ordinal data as continuous data or split it into dichotomous outcomes and then use the meta-analysis methods of mean or rate data for analysis.This operation will undoubtedly lose data information and may even lead to erroneous conclusions.In view of this,this study intends to conduct a meta-analysis of ordinal data.Alan Agresti proposed the Generalized Odds Ratios(GenOR)values between the two groups for ordinal data in 1980.When there are only two types of grades,it is exactly the OR value for the proportion data comparisons,so it can be treated as a generalization of the OR for the proportion data.The GenOR provides a good basis for the study of the meta-analysis of ordinal data because of the good statistical properties.ObjectiveThis study intends to construct a meta-analysis method based on the OR of ordinal data-GenOR for the comparison of two sets of ordinal data sets under parallel design and paired design,and to construct a fixed-effects model and a random-effects model simultaneously to meet the needs of different applications.MethodThe study first calculates the GenOR statistics for each study and then weights them based on the inverse variance weighting method,the combined effects were obtained under the fixed effect model and the random effect model respectively,and their 95%confidence intervals were constructed.All methods use Monte Carlo statistical simulations to evaluate their statistical performance.The evaluation indicators include 95%confidence interval coverage,one-sided non-coverage ratio,point estimation bias,mean difference percentage,and mean square error(MSE).ResultIn this study,a meta-analysis of parallel design and matching design was constructed based on inverse variance weighting method,including fixed effects model and random effects model.For the parallel design,the simulation study shows that in the case of large samples and small samples,the point estimation error,the percentage difference in average,the mean square error are small,and the 95%confidence interval coverage is close to 95%,indicating that point estimates and interval estimates are quite accurate.For the paired design,the simulation study shows that,the point estimation error,the percentage difference in average,the mean square error are small,and the 95%confidence interval coverage is close to 95%,showing this method is also suitable for matching design data.ConclusionThe meta-analysis method of ordinal data constructed in this study has excellent statistical performance except the random effect model under parallel design and can meet application requirements,and it is recommended for application.更多还原