PLS-based SEM Algorithms: The Good Neighbor Assumption, Collinearity, and Nonlinearity

Abstract

The partial least squares (PLS) method has been extensively used in information systems research, particularly in the context of PLS-based structural equation modeling (SEM). Nevertheless, our understanding of PLS algorithms and their properties is still progressing. With the goal of improving that understanding, we provide a discussion on the treatment of reflective and formative latent variables in the context of three main algorithms used in PLS-based SEM analyses –PLS regression, PLS Mode A, and PLS Mode B. Two illustrative examples based on actual data are presented. It is shown that the “good neighbor†assumption underlying modes A and B has several consequences, including the following: the inner model influences the outer model in a way that increases inner model coefficients of association and collinearity levels in tandem, and makes measurement model analysis tests dependent on structural model links; instances of Simpson’s paradox tend to occur with Mode B at the latent variable level; and nonlinearity is improperly captured. In spite of these mostly detrimental outcomes, it is argued that modes A and B may have important and yet unexplored roles to play in PLS-based structural equation modeling analyses.