Meta分析中各研究偏倚的量化分析  

Quantitative Analysis of Bias of Each Study in Meta-analysis

作　　者：付金玉[1] 秦超英[1] 

机构地区：[1]西北工业大学理学院应用数学系,西安710129

出　　处：《中国循证医学杂志》2016年第9期1112-1116,共5页Chinese Journal of Evidence-based Medicine

摘　　要：目的研究Meta分析中如何对各个研究的偏倚进行量化并进行估计。方法在随机效应模型中,假设纳入Meta分析中各个研究的效应量服从不同形状参数的偏正态分布,通过引入形状参数对偏倚进行量化,再利用马尔科夫估计和极大似然估计对总体效应量、表征偏倚的参数以及异质性方差进行估计。结果在模拟计算中,当效应量yi服从不同形状参数的偏正态分布时,总体效应量较效应量yi服从正态分布时更接近真实值;形状参数不同的偏态分布下研究间异质性对总体效应量的影响比形状参数相同的偏态分布和正态分布下研究间异质性对总体效应量的影响小;在实例分析中,总体效应量估计值的95%置信区间的长度比效应量yi服从正态分布时的短。结论由于将各个研究的偏倚纳入Meta分析随机效应模型中,通过将各个研究的偏倚进行量化,进而去除因偏倚引起的异质性对总体效应量造成的影响,使得总体效应量的估计值更接近真实值。Objective Study how to quantify the bias of each study and how to estimate them. Method In the random-effect model, it is commonly assumed that the effect size of each study in meta-analysis follows a skew normal distribution which has different shape parameter. Through introducing a shape parameter to quantify the bias and making use of Markov estimation as well as maximum likelihood estimation to estimate the overall effect size, bias of each study, heterogeneity variance. Result In simulation study, the result was closer to the real value when the effect size followed a skew normal distribution with different shape parameter and the impact of heterogeneity of random effects meta-analysis model based on the skew normal distribution with different shape parameter was smaller than it in a random effects meta- analysis model. Moreover, in this specific example, the length of the 95%CI of the overall effect size was shorter compared with the model based on the normal distribution. Conclusion Incorporate the bias of each study into the random effects meta-analysis model and by quantifying the bias of each study we can eliminate the influence of heterogeneity caused by bias on the pooled estimate, which further make the pooled estimate closer to its true value.

关 键 词：META分析 偏正态分布 极大似然估计 偏倚 异质性 合并效应量 

分 类 号：R-03[医药卫生]