http://dx.doi.org/10.6129/CJP.2004.4604.02
Chinese Journal of Psychology, 46(4), 2004，283-292

鄭中平（國立台灣大學心理學系）；翁麗禎（國立台灣大學心理學系）

摘要

實徵研究經常遭遇資料遺漏，本研究即在探討資料遺漏機制為潛在變項組型混合模型時，結構方程模型的最大概似估計，並利用模擬資料為範例比較其與常用遺漏值處理法的差異。潛在變項組型混合模型為組型混合模型的延伸，此模型假設觀察變項的遺漏組型反映潛在變項之類別，而非外顯類別，且各類別可有相異的結構方程模型。本研究以隨機EM算則估計當結構方程模型的資料遺漏機制符合此模式時之參數，並以模擬資料瞭解其與不同遺漏值處理法表現之差異。結果顯示，本研究建議的方法對因素負載量與潛在類別比率等參數之估計良好。

關鍵詞：非隨機遺漏，組型混合模型，結構方程模型，最大概似法，潛在變項

Estimation of Structural Equation Models with Latent Variable Pattern Mixture Model via Stochastic EM Algorithm

Chung-Ping Cheng（Department of Psychology, National Taiwan University）；Li-Jen Weng（Department of Psychology, National Taiwan University）

Abstract

The maximum likelihood estimation method using the stochastic EM algorithm was developed for structural equation models (SEM) with latent variable pattern mixture model. Latent variable pattern mixture model is an extension of pattern mixture models with measurement errors and theoretical constructs considered. The patterns of missing were assumed to reflect latent classes rather than categories of manifest variables. Each latent class was allowed to have distinct structural equation model. The results of this simulation study indicated that the proposed estimation method via stochastic EM algorithm performed well compared to other missing data treatment methods and yielded satisfactory parameter estimates.

Keywords：nonignorable missingness, pattern mixture model, structural equation models, maximum likelihood method, latent variable