Interestingreflections from the prof. Sarstedt and the prof. Ringle on the book ′′Structural Equation Models: From Paths to Networks ′′ (Westland 2019)
Sarstedt, M., Ringle, C.M. Structural Equation Models: From Paths to Networks (Westland 2019). Psychometrika (2020). https://doi.org/10.1007/s11336-020-09719-0
Structural equation modeling (SEM) is a statistical analytic framework that allows researchers to specify and test models with observed and latent (or unobservable) variables and their generally linear relationships. In the past decades, SEM has become a standard statistical analysis technique in behavioral, educational, psychological, and social science researchers』 repertoire.
From a technical perspective, SEM was developed as a mixture of two statistical fields—path analysis and data reduction. Path analysis is used to specify and examine directional relationships between observed variables, whereas data reduction is applied to uncover (unobserved) low-dimensional representations of observed variables, which are referred to as latent variables. Since two different data reduction techniques (i.e., factor analysis and principal component analysis) were available to the statistical community, SEM also evolved into two domains—factor-based and component-based (e.g., Jöreskog and Wold 1982). In factor-based SEM, in which the psychometric or psychological measurement tradition has strongly influenced, a (common) factor represents a latent variable under the assumption that each latent variable exists as an entity independent of observed variables, but also serves as the sole source of the associations between the observed variables. Conversely, in component-based SEM, which is more in line with traditional multivariate statistics, a weighted composite or a component of observed variables represents a latent variable under the assumption that the latter is an aggregation (or a direct consequence) of observed variables.
Future ImprovementsSimilar to any other textbook, there are also elements that we believe could be improved. While the structure is very clear, the chapters』 content distribution is sometimes uneven. For example, the author devotes an entire 22-page chapter (Chapter 2) and various sections (e.g., Chapter 4.4) throughout the 149-page book to PLS path modeling, only to conclude that 「it is an ideal tool for unscrupulous or lazy researchers interested in bogus theories with random data」 (Westland 2019, p. 38).
Apart from this structural concern, Westland’s (2019) conclusion underlines the general sentiment that component-based SEM and factor-based SEM are competing approaches. However, as suggested in other parts of the book, researchers』 functional background and adherence to a specific philosophy of science position contribute to the confusion over which method is 「right」 and which is 「wrong」 (Rigdon et al. 2017). These aspects have been extensively discussed in various social science fields (Henseler et al. 2014), including psychology (Rhemtulla et al. 2020; Rigdon et al. 2019). We believe a new edition would benefit from a more balanced description of the different viewpoints, given that recent advances in component-based SEM address the various concerns expressed in the book. For example, generalized structure component analysis offers a single optimization criterion as well as model fit metrics and allows researchers to impose model constraints (Hwang and Takane 2014). In addition, research has suggested extensions of component-based SEM methods that reliably estimate both factor- and component-model parameters (e.g., Dijkstra and Henseler 2015; Hwang et al. 2020).
Moreover, extensions of some of the simulations documented in the book would also be beneficial. For example, in Chapter 2, the author uses a simulation study to demonstrate the PLS path modeling method’s limitations. The simulation study assumes zero relationships between all the components, which is problematic, since the PLS path modeling algorithm—as also described in this textbook—requires a nomological net (i.e., at least some relationships that are not zero) to run adequately. The author’s general conclusion that 「PLS path estimates are biased and highly dispersed from small samples」 (Westland 2019, p. 36) should therefore only be considered in the context of this very specific model setup—see also Henseler et al. (2014).