Partial Least Squares (PLS) Structural Equation Modeling (SEM) for Building and Testing Behavioral Causal Theory: When to Choose It and How to Use It

TLDR: First-generation second-generation techniques are superior for the complex causal modeling that dominates recent communication and behavioral research, and either covariance-based SEM or PLS should be selected.

Abstract: Problem: Partial least squares (PLS), a form of structural equation modeling (SEM), can provide much value for causal inquiry in communication-related and behavioral research fields. Despite the wide availability of technical information on PLS, many behavioral and communication researchers often do not use PLS in situations in which it could provide unique theoretical insights. Moreover, complex models comprising formative (causal) and reflective (consequent) constructs are now common in behavioral research, but they are often misspecified in statistical models, resulting in erroneous tests. Key concepts: First-generation (1G) techniques, such as correlations, regressions, or difference of means tests (such as ANOVA or ${\rm t}$ -tests), offer limited modeling capabilities, particularly in terms of causal modeling. In contrast, second-generation techniques (such as covariance-based SEM or PLS) offer extensive, scalable, and flexible causal-modeling capabilities. Second-generation (2G) techniques do not invalidate the need for 1G techniques however. The key point of 2G techniques is that they are superior for the complex causal modeling that dominates recent communication and behavioral research. Key lessons: For exploratory work, or for studies that include formative constructs, PLS should be selected. For confirmatory work, either covariance-based SEM or PLS may be used. Despite claims that lower sampling requirements exist for PLS, inadequate sample sizes result in the same problems for either technique. Implications: SEM's strength is in modeling. In particular, SEM allows for complex models that include latent (unobserved) variables, formative variables, chains of effects (mediation), and multiple group comparisons of these more complex relationships.