节点文献
基于发表时间顺序的发表偏倚诊断方法研究
A Study of the Method to Diagnose Publication Bias Based on Publication Date
【作者】 郑辉烈;
【导师】 王增珍;
【作者基本信息】 华中科技大学 , 流行病与卫生统计学, 2010, 博士
【摘要】 目的本次研究拟验证发表时间顺序与发表偏倚的关系这一假设,将发表时间顺序引入发表偏倚的诊断过程中,利用Meta分析中各独立研究的发表时间顺序和效应尺度信息构建一种新的基于发表时间顺序的发表偏倚诊断方法(简称时序法),通过有偏和无偏Meta分析研究模拟数据及从文献中获得的Meta分析实际数据,对时序法和传统方法的诊断准确度和诊断一致性进行评价,并提出最佳的联合诊断方案。方法本次研究构建的时序法以效应尺度-发表时间顺序散点图、按发表时间顺序排列的效应尺度序列的游程总数和效应尺度-发表时间顺序回归模型来诊断发表偏倚是否存在。利用有偏和无偏Meta分析研究模拟数据,依据在保证一定的特异度水平的条件下,灵敏度处于可接受水平的原则,确定游程总数法诊断发表偏倚的参考界值范围;利用有偏和无偏Meta分析研究模拟数据计算时序法和传统方法诊断发表偏倚的灵敏度和特异度,评价时序法的诊断准确度。利用Meta分析研究模拟数据和实际发表的文献数据,采用Kappa指数及其95%可信区间评价时序法与传统方法以及两种时序法之间的诊断一致性。利用Meta分析研究模拟数据,观察时序法和传统方法以及两种时序法之间的联合诊断(并联)发表偏倚的准确度,评价联合诊断的效果并从中选择最佳的时序法和传统方法的联合诊断方案。使用SAS8.2软件编写时序法和传统方法的宏程序,并应用于Meta分析的实例研究。本次研究利用Matlab7.0软件产生Meta分析模拟数据,以Monte Carlo法估计灵敏度和特异度,经预模拟研究,确定模拟次数为1000次,不同参数设置条件下的随机种子数不同。假设检验水准设置为0.20。结果研究发现,时序法(游程总数法和回归模型法)诊断发表偏倚的灵敏度较低,特异度尚可;游程总数法的灵敏度和特异度低于回归模型法;未加权回归模型法的灵敏度高于加权回归模型法,特异度则低于加权回归模型法;时序法的灵敏度普遍低于传统方法(Egger回归法和Begg秩相关法),而两者间的特异度则是互有高低,比较相近;时序法与传统方法的诊断准确度可能与Meta分析研究的总体效应量大小有关,当总体效应量接近于0时,各方法的诊断发表偏倚的灵敏度较低;总体效应量从小到大变化时,各方法诊断发表偏倚的灵敏度呈现先增大后减小的趋势,而特异度基本相等;回归模型法与传统方法的诊断准确度可能与Meta分析研究中子研究的个数有关,子研究个数越多,各方法的灵敏度越大,特异度则越小;时序法和传统方法诊断一致性差,游程总数法和回归模型法对发表偏倚的诊断一致性相对较好,但其一致性程度(Kappa指数)依然处于较低水平;时序法与传统方法间联合诊断(并联)的灵敏度普遍高于传统方法间联合诊断的灵敏度,时序法和传统方法的联合诊断具有较理想的灵敏度和特异度,以加权回归模型法和Begg秩相关法的联合诊断效果最佳;游程总数法和回归模型法的联合诊断效果稍差于时序法和传统方法的联合诊断;应用SAS系统语言编写的包含了时序法和传统方法的发表偏倚诊断宏程序的运行结果与常用软件一致,方便了时序法的推广应用;Meta分析的实例研究发现,对于包含较多子研究个数的Meta分析,回归模型法与传统方法结论基本一致,而子研究个数较少时,回归模型法相对于传统方法更为灵敏。本次Meta分析实例研究的子研究个数较少,而其游程总数都比较大,游程总数法提示可能存在发表偏倚。结论时序法与传统方法具有相似的诊断准确度,因其具有与传统方法完全不同的理论基础,所以该方法有特殊的应用价值和前景,特别是与传统方法进行发表偏倚的联合诊断。时序法中的加权回归模型法与Begg秩相关法的并联联合诊断方案具有相对较好的灵敏度,且其特异度也处于较高水平,是较理想的一种联合诊断方案。创新提出了一种新的基于发表时间顺序的发表偏倚诊断方法——时序法;提出了一种产生包含发表时间顺序的无偏和有偏Meta分析模拟数据的方法;提出了时序法(游程总数法和回归模型法)和传统发表偏倚诊断方法(Egger回归法和Begg秩相关法)的联合诊断方案。
【Abstract】 Objectives To validate the hypothesis that the publication bias is associated with the sequence of publication date, we used the effect size of substudy in Meta-analysis and the sequence of publication date to construct a new method to detect publication bias, then evaluated the accuracy to detect the publication bias of the new method and the consistency of detecting the publication bias for new and traditional methods and put forward the combined diagnosis project of the new and traditional methods to detect the publication bias. Because the new method to detect publication bias is associated with publication date, so we named this new method "the method based on the sequence of publication date".Methods The new method to diagnose publication bias comprises the effect size-the sequence of publication date scatterplot, the total runs of the sequence of effect size sorted by publication date and the effect size-the sequence of publication date linear model.Using the simulated Meta-analysis data being of publication bias and the simulated Meta-analysis data being of no publication bias, the two-side reference interval of the total runs diagnosing publication bias for the method of the total runs was determined according to the principle that the sensitivity to detect publication bias comes to acceptable level and the specificity to detect publication bias reaches to a high level simultaneously; using the simulated Meta-analysis data being of publication bias and the simulated Meta-analysis data being of no publication bias, the sensitivity and specificity of new method and traditional method to detect publication bias were calculated. Afterward, we compared the accuracy of new method with traditional method; using the simulated and published Meta-analysis data, the consistency of detecting the publication bias for new and traditional methods were examined with the kappa index and its 95% confidence interval (CI); using the simulated meta-analysis data, the accuracy of combined diagnosis project (in parallel) with new methods and traditional methods was discussed. According to the accuracy of combined diagnosis project, then the optimal combined diagnosis project with new method and traditional method was put forward.The macro of new methods and traditional methods was written with Statistical Analysis System (SAS) language and carried out in one published Meta-analysis. The macro could output the results of new methods and traditional methods in report forms.The simulated Meta-analysis data were generated by means of Matlab7.0. This study employed the method of Monte Carlo to estimate the accuracy of the new methods and traditional methods to detect publication bias in simulated Meta-analysis for each combination of parameters which was set when generating the simulated Meta-analysis data. According to the pilot simulation study, simulation repeated 1000 times in each different combination of parameters. Simulation was based on different random seed in different combination of parameters. The significant level of hypothesis test was at 0.2.Results The sensitivity to detect the publication bias of the new method based on the sequence of publication date (the method of the total runs and the method of effect size-the sequence of publication date linear model) was low and the specificity to detect the publication bias of the new method was acceptable; the accuracy of the method of the total runs was less than the accuracy of the method of effect size-the sequence of publication date linear model; the sensitivity of the method of unweighted linear model was more than the sensitivity of the methods of weighted linear model, however, the specificity of the method of unweighted linear model was less than the specificity of the method of weighted linear model; the sensitivity of the new method based on the sequence of publication date was less than the sensitivity of the traditional method (Egger’s method and Begg’s method), but the specificity of the new methods based on the sequence of publication date was not less than the sensitivity of the traditional method; when the population effect size (SMD or ln(OR)) approximated to zero, then the sensitivity of all methods was low, but when the population effect size increased, the sensitivity of all methods increased early and decreased latter and the specificity of all methods immobilized. That is, the accuracy of all methods might relate to the magnitude of the population effect size; when the number of the substudy of Meta-analysis increased, the sensitivity of all methods, except for the method of the total runs, increased and the specificity decreased. That is, the accuracy of all methods might be associated with the number of substudy of Meta-analysis; the consistency for new method and traditional method based on different theories was low; the consistency for the two new methods was still low (the Kappa index was small), but the consistency for the two new methods was better than the consistency for new method and traditional method; the sensitivity of combined diagnosis with new method and traditional method in parallel was more than the sensitivity of combined diagnosis with two traditional methods (Egger’s method and Begg’s method) in parallel; when the new method and the traditional method were combined in parallel, the sensitivity of the project of combined diagnosis increased, however, the specificity did not decrease obviously; among all projects of combined diagnosis of new method and traditional method, the project of the method of weighted linear model and the Begg’s method was optimal; the sensitivity of the project of combined diagnosis of the two new methods was acceptable, but the specificity was small compared to the project of combined diagnosis of the new method and the traditional method; the macro of the new method and the traditional method was written with SAS language and its output was identical to the output of the software in common use (STATA), therefore, the macro made the new method and the traditional method applied conveniently. When the new method and traditional method to detect publication bias were applied to the published Meta-analysis, we found that the results of the method of the linear model were similar to the results of the traditional method for the Meta-analysis with large number of substudy, however, the sensitivity of the method of the linear model was better than the sensitivity of the traditional method for the Meta-analysis with small number of substudy. Because of the magnitude of the total runs was large when the number of substudy in published Meta-analysis was small, the method of the total runs indicated that there might be the publication bias in this published Meta-analysis.Conclusions The accuracy of the new method based on the sequence of the publication date was similar to the traditional method. Because the theories behind the new method and the traditional method were different, therefore, the new method had the special application value and prospects for planning out the project of combined diagnosis with the traditional method in parallel. In all projects of combined diagnosis of the new method and the traditional method, the sensitivity of the project of combined diagnosis of the method of weighted linear model and the Begg’s method was acceptable and the specificity reached to a high level, therefore, this project was regarded as an optimal project.Innovations A new method based on publication date to diagnose publication bias was advanced; a method to generate simulated Meta-analysis data with the sequence of publication date was put forward; the optimal project of combined diagnosis with the new method (the method of the total runs and the method of effect size-the sequence of publication date linear model) and the traditional method (the Egger’s method and Begg’s method) was put forward.
【Key words】 Meta-analysis; Publication bias; Sequence of publication date; Total runs; Linear model; Combined diagnosis; Sensitivity; Specificity;