Robust Mediation Analysis
Also known as: robust indirect effects, outlier-resistant mediation, robust causal mediation
Robust mediation analysis estimates the indirect effect of an independent variable on an outcome through one or more mediators using estimators that resist the influence of outliers and non-normal error distributions. By combining robust regression (such as median or M-estimation) with percentile or bias-corrected bootstrap confidence intervals, it yields trustworthy conclusions when standard ordinary-least-squares mediation would be distorted by extreme observations.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use robust mediation analysis when the data contain outliers or display heavy-tailed, skewed, or otherwise non-normal residuals that could distort standard mediation estimates. It is particularly warranted when diagnostic checks reveal influential observations in either regression equation, or when theoretical knowledge of the domain suggests extreme values are plausible (e.g., reaction-time data, income, clinical biomarkers). Do not use it as a routine replacement for standard mediation when residuals are approximately normal and no influential points are present, as the efficiency loss relative to OLS may be unnecessary. It is also not the appropriate tool when the mediator or outcome is binary or categorical — dedicated methods such as logistic-path mediation or causal mediation with non-linear links should be used instead.
Strengths & limitations
- Produces unbiased or near-unbiased indirect effect estimates even when outliers or heavy-tailed errors would substantially distort OLS-based mediation.
- Bootstrap confidence intervals make no normality assumption about the distribution of the product ab, yielding more accurate Type I error control than the Sobel test.
- Yields interpretable path coefficients (median shifts or robust conditional means) that correspond to the centre of the data rather than its extremes.
- Compatible with the familiar two-equation mediation framework, so results can be compared directly with standard mediation analyses reported in prior literature.
- Available in widely used software including R (robmed, WRS2) and can be approximated in SPSS via Hayes's PROCESS macro with bootstrap options.
- Robust estimators are less efficient (have larger standard errors) than OLS when the data genuinely follow a normal distribution, so power may be lower in clean data.
- Median regression assumes a location-shift model; if the effect of X on M changes shape across the distribution, quantile regression or other distributional methods may be more informative.
- Bootstrap inference requires larger samples for stable interval estimates; with very small samples (n < 50) bootstrap CIs can be unreliable regardless of the estimator.
- Multiple mediators in parallel or serial chains increase model complexity and require care in specifying which paths are estimated robustly.
- Interpretation of median-regression coefficients differs subtly from OLS coefficients and should be stated explicitly when reporting.
Frequently asked
How does robust mediation differ from simply using bootstrap with standard OLS mediation?
Standard OLS with bootstrap addresses the non-normality of the sampling distribution of the product ab but leaves the path coefficients themselves sensitive to outliers. Robust mediation replaces the OLS estimator with a resistant one (e.g., median regression) so that extreme observations have limited leverage on a and b before the bootstrap resampling is even applied. The two approaches can be combined: robust coefficient estimation followed by bootstrap inference.
Which robust estimator should I choose — median regression or Huber M-estimation?
Median regression (least absolute deviations) is the simplest and most interpretable robust estimator but can be less efficient for moderately non-normal data. Huber M-estimation iteratively re-weights observations, providing a smooth trade-off between OLS efficiency and outlier resistance. Yuan and MacKinnon (2014) found median regression performed well across many simulation conditions; the choice should be guided by software availability and the specific nature of the non-normality.
Do I need a large sample for robust mediation analysis?
Robust regression estimators generally require at least moderate sample sizes (roughly n = 100 or more) to be stable, and bootstrap confidence intervals for indirect effects need adequate n to converge. With very small samples, neither standard nor robust mediation provides reliable inference, and alternative designs or meta-analytic approaches should be considered.
Can robust mediation handle categorical or binary mediators?
Standard robust regression based on median or M-estimation is designed for continuous outcomes. When the mediator or outcome is binary, count-based, or ordinal, appropriate non-linear extensions (logistic, Poisson, ordinal regression) or causal mediation frameworks are needed instead.
Is robust mediation available in SPSS or R?
In R, the robmed package implements robust mediation directly, and the WRS2 package offers robust regression utilities that can be combined manually. In SPSS, Hayes's PROCESS macro (model 4) performs OLS mediation with bootstrap CIs, which addresses the inferential step but not outlier-resistant estimation of the paths themselves; analysts needing full robustness should use R or Mplus with robust estimators.
Sources
- Yuan, Y., & MacKinnon, D. P. (2014). Robust mediation analysis based on median regression. Psychological Methods, 19(1), 1–20. DOI: 10.1037/a0033820 ↗
- Hayes, A. F. (2013). Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-Based Approach. Guilford Press. ISBN: 978-1609182304
How to cite this page
ScholarGate. (2026, June 3). Robust Mediation Analysis. ScholarGate. https://scholargate.app/en/statistics/robust-mediation-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Mediation AnalysisStatistics↔ compare
- Moderated MediationStatistics↔ compare
- Robust Path AnalysisStatistics↔ compare
- Robust Structural Equation ModelingStatistics↔ compare
- Structural Equation ModelingResearch Statistics↔ compare