A Practical Causal Mediation Workflow for Psychologists: From Concepts to Implementation
Abstract
Mediation analysis is widely used across psychology to support claims about psychological mechanisms. Such claims often carry causal interpretations that depend on assumptions that are frequently left unstated, and may not be justified under traditional mediation approaches. Despite substantial advances in mediation best practices emerging from the field of causal inference, uptake of these approaches has remained low, possibly because the literature is highly technical and lacks comprehensive, integrated practical implementation guidance. This paper addresses that gap by integrating developments in causal mediation into an accessible, fully demonstrated workflow for applied researchers without prior causal inference knowledge. We identify two major limitations of traditional mediation approaches: vulnerability to mediator–outcome confounding and reliance on a product-of-coefficients approach that cannot account for predictor–mediator interactions and certain nonlinear models more generally. We then present concrete solutions that can be applied throughout the research workflow with limited additional work: before data collection (design-based solutions), during mediation effect estimation, and after analysis to assess robustness (sensitivity analysis). In doing so, we introduce key causal inference frameworks (directed acyclic graphs and potential outcomes) to facilitate understanding of these problems, bridge conceptual reasoning and practical solutions, and equip readers to apply these concepts in their own study contexts. Finally, we provide a fully worked example, including R code that readers can adapt for their own use. Overall, the proposed workflow requires only modest extensions to standard practice but can substantially improve the validity of causal claims drawn from mediation analysis.
Related
- On some concepts of convergence and their connections
- Quantiles for multivariate distribution: a short survey of population concepts
- Shared Cause Adjusted Mediation (SCAM): Introduction of method with empirical applications
- A novel Arctan-based model: Bayesian analysis, progressive censoring, and practical applications
- PDLP: a practical first-order method for large-scale linear programming
BibTeX
@misc{wenger2026practical,
title = {{A Practical Causal Mediation Workflow for Psychologists: From Concepts to Implementation}},
author = {Joshua D Wenger and Zoe Baker and Vasant G Honavar},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/sz7vx_v2},
url = {https://osf.io/preprints/psyarxiv/sz7vx_v2/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A Practical Causal Mediation Workflow for Psychologists: From Concepts to Implementation
AU - Joshua D Wenger
AU - Zoe Baker
AU - Vasant G Honavar
PY - 2026
DA - 2026/09/04
PB - PsyArXiv
DO - 10.31234/osf.io/sz7vx_v2
UR - https://osf.io/preprints/psyarxiv/sz7vx_v2/
AB - Mediation analysis is widely used across psychology to support claims about psychological mechanisms. Such claims often carry causal interpretations that depend on assumptions that are frequently left unstated, and may not be justified under traditional mediation approaches. Despite substantial advances in mediation best practices emerging from the field of causal inference, uptake of these approaches has remained low, possibly because the literature is highly technical and lacks comprehensive, integrated practical implementation guidance. This paper addresses that gap by integrating developments in causal mediation into an accessible, fully demonstrated workflow for applied researchers without prior causal inference knowledge. We identify two major limitations of traditional mediation approaches: vulnerability to mediator–outcome confounding and reliance on a product-of-coefficients approach that cannot account for predictor–mediator interactions and certain nonlinear models more generally. We then present concrete solutions that can be applied throughout the research workflow with limited additional work: before data collection (design-based solutions), during mediation effect estimation, and after analysis to assess robustness (sensitivity analysis). In doing so, we introduce key causal inference frameworks (directed acyclic graphs and potential outcomes) to facilitate understanding of these problems, bridge conceptual reasoning and practical solutions, and equip readers to apply these concepts in their own study contexts. Finally, we provide a fully worked example, including R code that readers can adapt for their own use. Overall, the proposed workflow requires only modest extensions to standard practice but can substantially improve the validity of causal claims drawn from mediation analysis.
ER -