結構方程模型誤設程度量化方法之探討

Abstract

模擬研究在結構方程模型之發展與議題探討上扮演關鍵角色，其中模型誤設為一重要操弄變項，界定方式包括考量真實模型被忽略之參數的數目、誤設模型統計考驗之檢定力，以及誤設模型之RMSEA數值等取向。本研究採行圖示法同步呈現此三取向所得結果間之關係，考量三種模型類型，不同變項數目和參數數值，於圖中繪製忽略單一參數、二參數與三參數之模型所對應的檢定力與RMSEA數值。結果顯示三方法間之關係受模型大小、參數數值與模型類型影響，建議後續結構方程模型模擬研究者可參考三方法間之關係，選擇適合其研究議題的模型誤設量化方法。

Monte Carlo simulations play a critical role in furthering our understanding of results from structural equation modeling (SEM). Among the factors manipulated in SEM simulation studies, model misspecification is considered important and deserves attention from researchers. Three approaches have been suggested and used in quantifying severity of specification errors in SEM simulations, including the number of parameters in a true model omitted, power of a statistical test in rejecting a misspecified model, and the value of root mean square error of approximation (RMSEA) of a misspecified model. The present study illustrates the relations among these three approaches for quantifying degrees of model misspecification in SEM with a graphical representation. Graphs of power and RMSEA for models omitting one to three parameters were plotted against different numbers of variables and parameter values using three types of models. The relations among these three methods were shown to depend on model size, parameter values, and model type. Researchers interested in manipulating severity of model misspecification are encouraged to examine the relations among the quantities expressed by these three approaches to select an optimal method for quantifying specification errors for the research issue to be pursued.