遺漏值處理法與模型設定對結構方程模型適合度指標之影響

Effects of Missing Data Treatments and Model Specification on Fit Indices in Structural Equation Modeling

遺漏值處理是社會科學研究經常面對的問題，本研究目的即探討結構方程模型不同遺漏值處理法與模型適合度指標的關係。經由模擬研究，討論不同遺漏值處理法在模型設定正確與錯誤下的過度適配情形，及不同適合度指標在具遺漏資料時之表現。結果顯示，當模型設定錯誤時，結構化最大概似法在部分指標上的確有過度適配情形，且隨遺漏值比率增加而越形嚴重；採用無結構最大概似法，再進行結構方程模型分析的兩階段方法則無過度適配情形，然模型設定正確時，其第一類型錯誤偏高。本研究結果並未發現在所有情形下表現優良之遺漏值處理法或適合度指標，使用者宜考量不同情形，選取適當的遺漏值處理法與適合度指標的組合。本研究亦發現Hu與Bentler（1998）所建議部分指標的檢定力過低，使用上宜加注意。

This Monte Carlo study explored effects of missing data treatment and model specification on 8 recommended fit indices in structural equation modeling. The results indicated that the structured maximum likelihood method tended to overestimate the degree of model-data fit, and the degree of overfitting increased as the percentage of missing data increased. Overfitting was not observed with unstructured maximum likelihood method, although this method tended to reject the model too often when the model was correctly specified. None of the fit index or missing data treatment was found to be superior across all conditions. The power of Gamma hat and Mc was found to be low. A careful selection of missing data treatment and fit indices was called for.