統合分析中非常態異質性之評估與處理

Abstract

隨機效應模型統合分析是目前主流的研究統整方法。在此方法中，統合分析中各研究的異質性常被假設為常態分布，但這其實是一個很強的假設，且常常未被注意或是驗證。雖然已有統計方法可評估常態分布假設，但在統合分析中由於個別研究的研究本身標準誤不同，目前的方法並無法直接應用。我們首先提出了標準化方法，以統計檢定與分位-分位圖來評估統合分析中的常態分布假設。模擬研究顯示這樣的方式有著控制良好的型一誤差率且也有一定的統計檢定力。若常態分布假設不成立，則呈現未來研究分布的預測區間估計也會受到影響。因此，我們也提出了無母數的方法以估計預測區間與預測分布。模擬研究顯示這樣的方法與現行應用常態分布假設的方法相比，可以得到更為不偏的估計。我們系統性地回顧了高影響因子的期刊上刊登的統合分析，發現在真實研究中，常態分布假設的確不能一體適用。我們並提供了真實的統合分析例子來呈現分析非常態異質性的重要。

Random-effects meta-analysis is one of the mainstream methods for research synthesis. The heterogeneity in meta-analyses is usually assumed to follow a normal distribution. This is a strong assumption which often receives little attention and is used without justification. Although methods for assessing the normality assumption are readily available, they cannot be used directly because the included studies have different within-study standard errors. We first present a standardization framework for evaluation of the normality assumption. We use both a formal statistical test and a quantile–quantile plot for visualization. Simulation studies show that our normality test has well-controlled type I error rates and reasonable power. Prediction intervals show the range of true effects in future studies and have been advocated to be regularly presented. We provide a simple method to estimate prediction intervals and predictive distributions nonparametrically when the normality assumption is implausible. Simulation studies show that this new method can provide approximately unbiased estimates compared with the conventional method. We then examine the normality assumption in real-world meta-analyses with a meta-epidemiological study. Systematically reviewing meta-analyses in high-impact journals, we find that the normality assumption is not universally applicable in meta-analyses. Real examples are also provided to illustrate the significance of analyzing non-normal heterogeneity in meta-analyses.