Meta和pooled分析方法在癌症流行病学研究中的应用和比较

Application and comparison of Meta and pooled analyses in the study of cancer epidemiology

目的 介绍癌症流行病学研究中经常采用的Ｍｅｔａ 和ｐｏｏｌｅｄ 两类综合分析方法，结合应用实例，进行分析对比。方法 以非吸烟女性肺癌危险因素的多中心病例对照研究中的两个危险因素（肺癌家族史和父母吸烟史） 为例，首先进行各研究单位的独立分析，然后采用Ｍｅｔａ 和ｐｏｏｌｅｄ 分析法进行综合分析，并将两类综合分析的结果进行比较。结果 对两个变量各中心资料独立分析时，其参数估计值（ 比数比ＯＲ） 及统计学检验结果均不相同。“肺癌家族史”变量以上海的ＯＲ有统计学意义，而“父母吸烟史”变量则是哈尔滨的ＯＲ有统计学意义。在固定效应模型假设下，各种综合分析方法的结果相当接近，包括未调整、调整（ＭａｎｔｅｌＨａｅｎｓｚｅｌ 法） 和多因素分析（ｌｏｇｉｓｔｉｃ 模型） 情形。当固定效应模型假设不符，而采用随机效应模型时，ｐｏｏｌｅｄ 分析的结果与Ｍｅｔａ 分析的结果稍有不同，但差异不大。在ｐｏｏｌｅｄ 分析中，分层ｌｏｇｉｓｔｉｃ 模型似乎比基本ｌｏｇｉｓｔｉｃ 模型要好。结论 大样本多中心病例对照研究资料的Ｍｅｔａ 和ｐｏｏｌｅｄ

Objective To introduce two comprehensive methods of analysis, Meta and pooled, frequently used in cancer epidemiology, and to compare them with authentic examples. Methods From data of a multicenter case control study of lung cancer risk factors among non smoking females, “family history of lung cancer” and “smoking history of parents” were the two variables selected for analysis. Based on the data from individual centers, Meta and pooled methods were employed to perform comprehensive analysis and the results compared. In the former, both fixed and random effect models were considered. Results When the data from individual study centers were analyzed, odds ratios (OR) differed. The OR of variable “family history of lung history” was statistically significant in Shanghai, while that of variable “smoking history of parents” was statistically significant in Harbin. Given the fixed effect model, the results of analysis were identical under different conditions of analysis (non adjusted, adjusted with Mantel Haenszel methods and multivariate logistic analysis). The results were similar when random effect model was used. In case the fixed effect model was not valid and there were some differences obtained by the two methods, combined parameters calculated from random effect model would be used. Stratified logistic model seemed better than basic logistic model in the pooled analysis. Conclusion The results from Meta and pooled comprehensive analyses are essentially consistent.