# Cross-trial comparison

Source: https://onco.cc/terms/cross-trial-comparison/  
OnCo record `cross-trial-comparison` (Term). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

Results from separate studies can differ because the people, care or measurements differ. Comparing their headline percentages does not by itself show which treatment works better.

## Summary

A comparison between separate studies must account for differences in participants, treatment, background care, endpoints, clocks and follow-up. Randomisation protects the comparison within a trial; it does not randomise participants between unrelated trials. Formal indirect comparisons can retain within-trial contrasts through a common comparator, but depend on compatible effect modifiers and methods. A higher headline percentage in one study is insufficient to rank treatments.

## Fields

- Kind: Term
- Last checked: 2026-11-10
- Also known as: cross-trial comparison; indirect treatment comparison

## Notes

- A higher percentage in one paper is a reason to examine its study, rather than proof that its treatment is better. Randomisation protects comparisons within a trial, not between separately conducted trials. Comparing their experimental arms discards that protection. Disease burden, biomarker assay and threshold, prior treatment, background care, dose, endpoint rules, analysis population and follow-up can differ. Formal indirect comparison through a common comparator preserves each trial's treatment-versus-control contrast but still requires compatible questions and methods. Pooling active arms across trials is prone to bias, and even an adjusted indirect estimate remains less secure than a suitable direct comparison (Bucher and colleagues, 1997). Network meta-analysis adds a transitivity assumption: important effect modifiers must be sufficiently comparable across the network. A shared drug name or statistical agreement cannot establish that unmeasured compatibility. Similar endpoint labels can hide different definitions: a clinical benefit rate may include different durations of stable disease. Adjustment for measured differences does not remove every bias, and lower hazard ratios against different comparators cannot automatically rank treatments.
- Two fictional single-arm studies each enrol 100 people with the same response definition, assessment period and complete denominator. In the favourable baseline group, A has 40 responses among 80 people (50%) and B has 12 among 20 (60%). In the other group, A has 4/20 (20%) and B 24/80 (30%). Overall, A has 44% response and B 36%, despite B's higher rate in both groups. A enrolled an 80/20 mixture of the two groups, while B enrolled 20/80. Composition reverses the headline ranking. This arithmetic does not establish that B is superior: people were not randomised between studies, and one measured feature cannot resolve every difference.
- A fictional vaccine report of 90% recurrence-free at two years in selected postoperative patients cannot be ranked against 30% objective response in a different advanced-disease drug study. Immune response, pathological response, imaging response and survival are different outcomes in different settings. A comparative effect estimate needs a shared clinical question and defensible design. For time-to-event results, align the event definition, origin, assessment schedule, cutoff and follow-up as well as population and care. FDA's December 2018 endpoint guidance warns that external survival comparisons can reflect selection, imaging and supportive-care differences. Neither fictional example certifies an actual indirect comparison.

## Sources

- The results of direct and indirect treatment comparisons in meta-analysis of randomized controlled trials.: https://pubmed.ncbi.nlm.nih.gov/9250266/
- Cochrane: network comparisons and transitivity: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-11
- FDA: oncology clinical-trial endpoints (2018): https://www.fda.gov/media/71195/download

## Connected records

- terms: [Absolute versus relative benefit (number needed to treat)](https://onco.cc/terms/absolute-benefit/), [Confidence interval](https://onco.cc/terms/confidence-interval/), [Hazard ratio (HR)](https://onco.cc/terms/hazard-ratio/), [Kaplan-Meier curve, censoring and proportional hazards](https://onco.cc/terms/kaplan-meier-curve/)
- fronts: [Drug Discovery Platforms](https://onco.cc/fronts/drug-discovery/)

---
JSON: https://onco.cc/api/v1/entities/cross-trial-comparison.json