結構方程模型估計方法的比較研究—大學生運動品牌形象認知模型的一個實例

目的：在於比較SEM與PLS在統計原理上的差異，方法：以大學生為母群體，共抽取15所大學，每所學校20位學生，共計施測300份，於2019年4月15日～5月14日完成施測。分析方法SEM為AMOS 20.0，PLS則為Smart PLS 3.0。並應用大學生運動品牌形象認知模型進行測量模型、模型評價、結構模型以及不同樣本數之間的比較示範。結果：（1）PLS反映型模型與形成型模型特性的差異，造成了同一資料兩者內模型參數估計結果的顯著差異。（2）SEM不適合小樣本量的研究而適合大樣本量的研究，而PLS無論是形成型模型還是反映型模型均適合小樣本量的研究。（3）PLS反映型模型強度最弱的路徑係數最容易受樣本量變化的影響。結論：目前，根據估計技術來區分，主要有兩大類建構模型。一種是基於最大概似估計（ML）的共變異數結構分析方法，該方法被稱為硬模型（Hard Modeling），以共變異數結構方程模型（Covariance-based SEM，CBSEM）為代表，運用的統計軟體有Amos、LISREL、EQS以及Mplus等。另一種則是基於偏最小平方法（PLS）的分析方法，被稱為軟模型（Soft Modeling），以主成份結構方程模型（component-based SEM）為代表，運用的統計軟體有Smart PLS、PLS-Graph、Visual PLS、PLS-GUI以及SPAD PLS等。SEM與PLS兩種方法學在實際運用中各有千秋，自稱典範。

關鍵字

偏最小平方法；測量模型；結構模型；模型評價

並列摘要

Purpose: To compare the differences between SEM and PLS in statistical principles. Methods: The research sample uses college students as the population. A total of 15 universities and 20 students from each school were selected, and 300 participants were tested. The test from April 15th to May 14th, 2019. The analysis method is AMOS 20.0 for SEM and Smart PLS3.0 for PLS. The cognition model of college students on sports brand image was adopted to compare and demonstrate the measurement model, model evaluation, structural model, and different sample numbers. Results: (1) The differences in the PLS reflective model and formative model characteristics resulted in significant differences in the estimation results of model parameters within the same data. (2) SEM was suitable for the study of large sample size instead of small sample size, while PLS was suitable for the study of small sample size regardless of the formative or reflective model. (3) The path coefficient with the weakest strength in PLS reflective model was most easily influenced by the change of sample size. Conclusion: There were two main types of construction models in the aspect of estimation techniques now. One was the covariance structure analysis method based on the maximum likelihood estimation (ML), which was called hard modeling (Hard Modeling) and represented by Covariance-based SEM (CBSEM). The statistical software used covered Amos, LISREL, EQS and Mplus. The other one was the analysis method based on the partial least square method (PLS), which was called soft modeling (Soft Modeling) and represented by the component-based SEM. The statistical software used covered Smart PLS, PLS-Graph, Visual PLS, PLS-GUI and SPAD PLS. SEM and PLS were provided with their unique advantages in practical application and could be honored as exemplary models.

並列關鍵字

partial least squares (PLS) ； measurement modeling ； structural modeling ； model evaluation

參考文獻

李茂能 (2008)。SEM 適配度指標的潛藏問題：最佳模式難求。測驗統計年刊，16，17-29。doi:10.6773/JRMS.200812.0007

徐茂洲 (2016)。體育領域應用偏最小平方法之現況。大專體育，136，55-65。doi:10.6162/SRR.2016.136.06

張國忠、陳文英 (2007)。高屏地區大學生運動用品品牌形象認知之研究。嘉大體育健康休閒期刊，6(1)，72-82。doi:10.6169/NCYUJPEHR.6.1.10

李曉鴻 (2012)。LISREL 與 PLS 建模方法的分析與比較。科技管理研究，32(20)，230-233。doi:10.3969/j.issn.1000-7695.2012.20.052

陳學濤、易東 (2009)。基於 PLS 結構方程模型的細分演算法。中國衛生統計，26(5)，527-528+538。doi:10.3969/j.issn.1002-3674.2009.05.027

主題瀏覽