E3 PreliminaryModerate confidencePEM not requiredMethods-PaperPeer-reviewedMachine draft
Beyond total treatment effects in randomised controlled trials: Baseline measurement of intermediate outcomes needed to reduce confounding in mediation investigations.
Landau, Sabine, Emsley, Richard, Dunn, Graham · Clinical trials (London, England) · 2018 · DOI
Quick Summary
This study explains the best way to analyze how a treatment works by breaking down its effects into direct and indirect pathways. The researchers tested different statistical methods using a rehabilitation trial in ME/CFS patients and found that one approach (called analysis of covariance) gives the most reliable answers by measuring patients' symptoms before and after treatment.
Why It Matters
This study is important because understanding *how* rehabilitation and other treatments work in ME/CFS—through which mechanisms they produce benefit—requires rigorous mediation analysis. The recommendation to always measure baseline mediators and use appropriate statistical methods ensures that future trials can reliably identify which aspects of treatment truly drive clinical improvement, informing more targeted therapeutic development.
Observed Findings
- Post-treatment-only mediation analysis produced biased estimates when baseline measures of clinical or intermediate outcomes were correlated with post-treatment measures.
- Change-score mediation analysis provided unbiased estimates only when baseline and change scores of the intermediate variable were independent—an assumption often violated in practice.
- Analysis of covariance conditioning on baseline values of both mediator and outcome remained unbiased across all three tested confounding processes.
- When applied to the ME/CFS rehabilitation trial, the proportion of treatment effect mediated by activity limitation varied from 57% to 86% depending on the analytical method used.
Inferred Conclusions
- Baseline measurement of both putative mediators and clinical outcomes is essential for valid mediation analysis in randomized trials.
- Analysis of covariance should be the default approach for mediation analysis of continuous outcomes, not merely for improved precision but critically to avoid bias from baseline confounding.
- Trialists must carefully select mediation methods with consideration for the underlying confounding structure, as failure to do so can yield substantially different and potentially misleading conclusions about mechanism.
Remaining Questions
- How should mediation analysis be conducted for binary or time-to-event clinical outcomes and mediators?
What This Study Does Not Prove
This is a methodological study, not a clinical trial; it does not prove that rehabilitation is effective for ME/CFS or establish the true magnitude of any mediation effect. The findings apply only to continuous outcomes; categorical or time-to-event mediators may require different approaches. The causal conclusions rest on the assumption that the data generating model used in simulations accurately reflects real confounding structures.
Tags
Symptom:Fatigue
Method Flag:Exploratory Only
Metadata
- DOI
- 10.1177/1740774518760300
- PMID
- 29552919
- Review status
- Machine draft
- Evidence level
- Early hypothesis, preprint, editorial, or weak support
- Last updated
- 8 April 2026
About the PEM badge: “PEM required” means post-exertional malaise was an explicit required diagnostic criterion for participant inclusion in this study — not that PEM was studied, observed, or discussed. Studies using criteria that do not require PEM (e.g. Fukuda, Oxford) are tagged “PEM not required”. How the atlas works →
Spotted an error in this entry? Report it →