Chapter 3: Partial Least Squares Path Modeling

At the conceptual level, structural models generally consist of relations among a number of idealized constructs (Bagozzi, 2011). Before such a model can be evaluated empirically, a researcher must find some way to represent those constructs in terms of data. In structural equation modeling (SEM), constructs are most often represented by the communality or shared variance among a set of observed variables—in other words, by a common factor.

But alternatives exist. One alternative approach is to craft representations of the theoretical variables as weighted composites of observed variables. Despite the dominance of factor-based approaches in SEM, decades of literature contrasts factor-based and composite-based approaches to data analysis, revealing strengths and weaknesses on both sides (e.g., Velicer & Jackson 1990). Factor-analytic method promise statistical advantages, but those benefits come at the cost of restrictive assumptions, which do not always hold. In some circumstances, researchers will have little hope of meeting the requisite conditions; in others, the researcher’s focus may be on something other than statistical optima. Researchers in possession of promising data but lacking either a well-developed theoretical framework or the luxury of multiple rounds of measure development may find in composite-based methods the opportunity for insight, using methods that are convenient, fast, and tolerant of less-than-ideal conditions. So, while researchers will continue to value the perfect world strengths of factor-based SEM, researchers should also appreciate the real world advantages of compositebased methods.