Pembolehubah Pendam Teguh dalam Model Persamaan Berstruktur Kuasa Dua Terkecil Separa

Robust Latent Variables in The Partial Least Squares Structural Equation Model

Abstract

Model Persamaan Berstruktur (MPB) secara khususnya sesuai digunakan bagi memodelkan hubungkait antara konstruk bersandar dan tak bersandar berbilang secara serentak. MPB mampu menjawab set persoalan kajian yang berhubungkait secara komprehensif dan sistematik kebiasaannya dengan menggunakan dua pendekatan, iaitu, MPB berdasarkan anggaran kebolehjadian maksimum (MPB-AKM) dan kuasa dua terkecil separa (MPB-KTS). Bagi membangunkan MPB yang melibatkan konstruk pendam psikologi. Konstruk pendam psikologi perlu dianggarkan. Diketahui daripada beberapa kajian, data psikologi tidak bertaburan secara normal, dan wujud multikolinearan antara pembolehubah penunjuk yang membentuk pembolehubah konstruk tersebut. Dalam kajian ini, didapati bahawa wujud multikolinearan antara pembolehubah penunjuk dalam konstruk psikologi. Pendekatan MPB-AKM mengandaikan data bertaburan secara multivariat normal. Walau bagaimanapun pendekatan MPB-KTS adalah bebas-taburan, di samping tidak memerlukan andaian yang ketat berbanding pendekatan MPB-AKM. Justeru itu, MPB-KTS adalah lebih sesuai digunakan. Walau bagaimanapun, MPB-KTS sensitif terhadap kewujudan data pencilan. Kajian ini menambah baik kaedah penganggaran pembolehubah pendam apabila wujudnya data pencilan dengan menggunakan MPB-KTS melalui penggunaan varians-kovarians teguh.

____________________________________________________________________
Structural Equation Modelling (SEM) is particularly suitable to answer a set of interrelated research questions in a systematic and comprehensive analysis by modelling the relationships among multiple independent and dependent constructs simultaneously using two approaches, namely, SEM based on Maximum Likelihood Estimator (MLE) (SEM-MLE) and Partial Least Squares (PLS) (SEM-PLS). To develop SEM which involved the latent psychology constructs, the latent psychology constructs are needed to be estimated. It is known from various studies, that psychology data is not normally distributed, moreover there exists a multicollinearity among the manifest variables that are used in estimating the constructs. In this study, it is found that there are multicollinearity among the manifest variables. The SEM-MLE approach assumes that observations are governed by multivariate normal distribution. However, SEM-PLS is a distribution-free, thus, requires much less stringent assumptions than the SEM-MLE approach, and thus, SEM-PLS is preferably to be used. However, SEM-PLS is sensitive in the presence of outliers, therefore, this study seeks to improve the method of estimation of latent constructs with the presence of outliers by using SEM-PLS through robust variance-covariance.