Interrupted mediation: A cautionary note on using derived metrics as intervening variables in path models

Abstract

Mediation - or models where an intervening variable is thought to propagate an affect from a cause to an outcome of interest - is a ubiquitous statistical tool in the behavioral and medical sciences. Despite a long history of thoughtful critiques of its use in applied research, mediation remains inferentially attractive and easy to implement in widely-available software. Here we highlight a challenge of mediation that has not yet received appropriate consideration - namely the potential for improper causal inferences when the mediator is a derived, or composite, score of a set of items or measures (e.g., sums, means, differences, etc.). We show that composites have the potential to confound or mask different causal processes that occur at the individual item level, leading to spurious mediation effects. As composite measures are near-universal in many fields from the psychological to biomedical sciences, this limitation of inferences potentially impacts a very broad set of research questions. In this paper we demonstrate this issue in a diversity of mediators, using both empirical examples of grey matter volume and depression questionnaires and a wide array of simulated examples. Additionally, we provide several potential analytical approaches with may help diagnose spurious mediation with composite measures and provide greater insight into the causal structure of our data. As part of this set of approaches, we advocate a transition from traditional, regression-based approaches to mediation towards a structural equation modeling (SEM) framework which provides many advantages. We also note additional conditions under which interrupted mediation can occur, providing directions for future developments in this area.

Examining the indirect role of perceived normativeness in the association between collectivistic values and psychological control of Turkish mothers

“…Although we do not anticipate using any other control variables, if the correlational results or t ‐tests on other demographic factors suggest significant effects, these will be added as control variables too. In order to avoid the possibility of interrupted mediation (see McCormick et al, 2022), the SEM model will be used for the normativeness of psychological control variable. Instead of using a composite mean score for our mediating variable (i.e., normativeness of psychological control), the normativeness items will be added to the model separately to predict the composite perceived normativeness variable (i.e., the latent variable model).…”

How should we model the effect of "change" - or should we?

“…However, unlike some composites, like product terms used to estimate interactions, purely additive composites (like difference scores) can not explain additional variance net their constituent parts. This means that we could not include x t , x t−1 , and ∆x in the same model and still obtain unique estimates (McCormick, Borgeest, et al, 2022), which we can do in the case of product composites. As we saw in Table 2, we also cannot avoid this by only including the ∆x predictor because of the highly unlikely constraint that places on the model -where it is equivalent to x t and x t−1 having regression coefficients of equal magnitude but opposite sign.…”