Will PLS have to become factor-based to survive and thrive?

The article below provides an overview of various SEM approaches. It argues that minimization of type I and II errors, or false positives and negatives respectively in hypothesis testing, can only happen if latent variables are implemented as factors (and not as composites). It is argued that this requires the use of modern, factor-based PLS methods (known as PLSF methods), which have some advantages not only over classic PLS implementations, but also over covariance-based SEM approaches. We discussed a PLSF type in the article; namely type CFM3.

Kock, N. (2024). Will PLS have to become factor-based to survive and thrive? European Journal of Information Systems, 33(6), 882-902.

Link to full-text file for this article:

Will PLS have to become factor-based to survive and thrive?

Abstract:

Structural equation modelling (SEM) is a general method that aims at estimating models with latent variables (LVs), where the LVs are measured indirectly and with some imprecision via questionnaires. This is done usually employing question-statements answered on Likert-type scales. In this paper we discuss various forms of SEM, and demonstrate that composite-based models, common in classic partial least squares (PLS) implementations, are poorly aligned with the very idea of SEM. We argue that minimisation of type I and II errors, or false positives and negatives respectively in hypothesis testing, can only happen if LVs are implemented as factors (and not as composites). This requires the use of modern, factor-based PLS methods, which have some advantages not only over classic PLS implementations, but also over covariance-based SEM approaches. Our main goal with this paper is to stimulate debate, whether pro or against our views. If we are generally correct in our thinking, the impact on how quantitative research is conducted in the field of information systems, as well as many other fields, could be quite dramatic. The reason for this is the widespread use of SEM in information systems, business, and the behavioural sciences.

Note: Some readers of this blog have brought to our attention that a critique of the article above is already out, and with a number of mistakes and incorrect statements, such as that: they (the critics) used CFM1 because this is the only PLSF type documented in the WarpPLS User Manual (untrue and very easy to check); and that the algorithm that they analyzed (PLSF-CFM1) is a slow version of Dijkstra’s PLSc technique (CFM1 does not use PLSc at all); among other easy-to-avoid mistakes and incorrect statements. We are aware of this critique.

Best regards to all!