http://dx.doi.org/10.6129/CJP.2007.4904.05
Chinese Journal of Psychology, 49(4), 2007， 383-405

邱皓政（國立中央大學企業管理學系）

摘要

潛在構念的估計是社會與行為科學研究的重要命題，結構方程模式取向是當代據以進行潛在構念估計與關係探究重要典範，但是，如果研究樣本涉及叢集抽樣而呈現階層結構，研究者即必須進行多層次結構方程模式分析（MSEM）。階層資料的重要特性是低階層次變數在高階會形成脈絡變數，因此MSEM 分析必需能夠處理潛在脈絡變數模型的萃取與模型評估。本文的目的，即在說明如何利用MSEM 程序來進行潛在脈絡變數的因素模型與結構模型分析，並以一項口試實證資料（76 位教師甄試口試委員針對841 位考生的口試評分資料），以Mplus4 進行單層次與多層次結構方程模式分析。分析結果證實了MSEM 程序可以有效分離考生共變矩陣與評審共變矩陣，並證實了個體層次與總體層次的潛在因素模型並不相同，為一非等值模型（non-equivalent model）。個體層次測量變數在評審層次是一個單一完整的潛在構念，亦即方法效應或月暈效果的概念，但在考生層次則仍保有原先的二因素架構。本文的
結論指出， MSEM 比傳統的MLM 取向在處理脈絡變數上更具有彈性，在方法學上具有重要的價值，但是在技術層次仍有諸多有待突破的瓶頸，值得進一步深究。

關鍵詞：脈絡變數、潛在脈絡變數、多層次資
料、多層次結構方程模式、評分者效應

Multilevel Latent Variable Modeling of Contextual Variables: Application of Multilevel Structural Equation Modeling on Rater Effects

Hawjeng Chiou（Department of Business Administration National Central University）

Abstract

The estimation of latent construct is the core issue of social and behavioral science. Structural equation  modeling is the most important paradigm for estimation and for exploring the relationships among constructs. If the data for analysis involving multilevel or clustered structure, however, the multilevel SEM have to be applied for the special form of dataset. One of the features of multilevel data is that the observed variables in the  individual level could be aggregated as a contextual variable into higher level. In terms of latent construct, contextual variable may behave as a latent form and should be estimated in a way of factorial analysis instead of manifest procedure. In a circumstance that the latent contextual variables have significant effects on dependent variable is called latent contextual effects. The purpose of present paper is to clarify the significance of the latent contextual effects in the social research. Rating data of oral examination examined in the present study contains a total of 76 raters and 841 interviewees. A series of single and multiple level SEM models was proposed for examining the latent contextual effects. Results of Mplus indicated that the MSEM modeling can effectively extract the latent contextual variables and furthermore used to explain the dependent variables. Different factorial structure on latent contextual variables in both macro and micro level reveals that the contextual model is not equivalent across levels. The present study proved that the MSEM approach to examine contextual variables is more flexibility than traditional MLM paradigm. However, technical as well as methodological issues have to be solved.

Keywords: contextual variables, latent contextual variables, multilevel data, multilevel structural equation modeling, rater effects