國內教育、心理後設分析研究出版偏誤檢定之實徵分析

Empirical Assessment of Effect of Publication Bias on Meta-Analysis in Domestic Education and Psychology

後設分析技術自1976年起就逐漸成為社會科學重要的研究方法之一，但國內多數研究均是以應用導向為首要，其中，對於可能導致錯誤結論之出版偏誤議題，則較少以系統性的方式進行探討。據此，本研究經調查國內103篇教育、心理後設分析研究之出版偏誤檢定現況後，發現其檢定比率確實不高，只有近四成研究曾利用出版偏誤方法進行檢定，其中，幾乎有39篇研究都是以傳統Fail-safe N指標為考量，但此方法存在的實務與假設限制，是值得研究者重視。接續，本研究經引用30篇後設分析研究之實徵數據，採用等級相關與線型迴歸分析檢定funnel plot之不對稱性，並搭配trim and fill方法後，發現各方法檢定結果雖不一致，但卻顯示出版偏誤是可能存在於國內後設分析研究，並影響著分析結果。最後，經檢視各方法特色及實徵分析數據後，建議未來研究者在傳統Fail-safe N指標外，可考量同時運用線型迴歸分析、trim and fill方法產生的訊息，以較保守的態度來看待與呈現出版偏誤可能對其研究結果產生的影響。

Since 1976， meta-analysis has become an important method in social science. It has been observed that many meta-analysis have been used for application， but only few evaluated publication biases systematically. After surveying， a total of 103 educational and psychological meta-analysis were assessed， but only 39 (37.86%) evaluated publication bias. Also almost every meta-analysis only reported Fail-safe N， which should be treated with considerable caution due to some practical limitations. Therefore， researchers adopted rank correlation test and linear regression approach to measure funnel plot asymmetry， and assessed 30 published meta-analysis by trim and fill method. Although results are of discordance， researcher observed publication biases were common within meta-analysis and affected the results. Finally， some suggestions were proposed for future usage.