Can formative latent variables represent psychological concepts? Replication and extension of MacKenzie et al. (2005)

We revisit MacKenzie et al.'s (2005) influential work on bias in path estimates arising from the misspecification of formative latent variable models as reflective ones. Using simulations, we demonstrate that the bias in unstandardized path coefficients reported by MacKenzie et al. is due to scaling differences rather than model misspecification. When these differences are taken into account, the bias nearly disappears. Building on their work, we also demonstrate that the formative indicators in their model are unnecessary. This is because the formative latent variables derive their identity from the reflective indicators added for model identification. Removing one or even all formative indicators is inconsequential for estimating the paths between latent variables. Our study reveals a fundamental problem with this approach: Formative latent variables (whether first or higher order) cannot represent psychological concepts as intended because their identity derives from downstream reflective indicators rather than the formative indicators. We conclude by discussing the implications for measurement and index construction. (PsycInfo Database Record (c) 2026 APA, all rights reserved).