As Hayes (2022) notes, the technical challenges associated with conducting moderation analyses using latent variables make such analyses impractical—if not impossible, in my experience. This has led to the common practice of estimating direct (or main) effects and mediation relationships using latent variables, while conducting moderation analyses at the observed-variable level.
Recently, however, I have encountered several critiques of this practice, primarily asserting that it is invalid. Yet, I find no strong methodological basis for outright rejecting this approach. Hayes (2022) has already challenged the previously held assumption that statistically significant direct and indirect effects are prerequisites for conducting mediation, moderation, or moderated mediation analyses. This suggests that the analyses of direct effects, mediation, and moderation involve conceptually and technically distinct processes.
Given this, I wonder why using different methods to address fundamentally different analytical questions would be problematic.
References:
Hayes, Andrew F. (2022): Introduction to mediation, moderation, and conditional process analysis. A regression-based approach. Third edition. New York NY: The Guilford Press (Methodology in the social sciences).
Although I do know the work from Andrew Hayes, I cannot remember what he meant with technical challenges to do moderation in a latent variable context. But while direct and indirect effects are pretty straight forward in SEM, moderation is not. There are several approaches already to construct latent interactions and none of them are as easy to do as mediation, for example. Maybe you can ask Holger Steinmetz here on RG directly (somehow I could not link his account....), who wrote an article about doing latent interaction.
Steinmetz, H., Davidov, E., & Schmidt, P. (2011). Three approaches to estimate latent interaction effects: Intention and perceived behavioral control in the theory of planned behavior. Methodological Innovations Online, 6(1), 95-110.
Addendum: I just found a new R package from 2024, which can conduct latent interaction with all relevant approaches. In their paper they review all approaches briefly and introduce the modsem package. I think this will be very helpful and was missing so far! Very interesting.
https://modsem.org/index.html
Slupphaug, K. S., Mehmetoglu, M., & Mittner, M. (2024). modsem: An R package for estimating latent interactions and quadratic effects. Structural Equation Modeling: A Multidisciplinary Journal, 1-13.
It seems, as if Holger deactivated the functions to be added in a discussion and to get mails via RG, I cant do anything either, sorry.
- Badges
- Science topic