A review of oncology evidence found that none of eight eligible Cochrane meta-analyses analytically addressed treatment switching, even though 40 of 68 randomized trials, or 59%, explicitly allowed patients to move from one treatment to another. In 26 trials, or 38%, the review could not find a statement saying whether switching was permitted, and only two trials, or 3%, explicitly said it was not allowed.
An estimand is the precise clinical question a treatment effect is meant to answer. The authors warn that pooling estimates aimed at different clinical questions may obscure the meaning of the pooled effect, and note that none of the included reviews explicitly discussed estimands or other post-randomization events.
How the review was assembled
The search was restricted to 2021 onward and focused on peer-reviewed Cochrane reviews of pairwise meta-analyses involving randomized oncology trials and newer therapies. Eligible reviews had to assess progression-free survival or overall survival. The search identified 162 reviews; eight eligible meta-analyses contained 81 unique trials, and 68 randomized trials met the review’s final criteria.
Two reviewers independently screened the records and extracted study details, checking protocols, statistical analysis plans and trial registrations when available. The review followed PRISMA guidance and used a descriptive synthesis rather than conducting a new meta-analysis.
Switching was common, but the details were thin
Among the 68 trials, 46, or 68%, were Phase 3 and 52, or 76%, were open-label. Fifty-eight trials, or 85%, reported both progression-free survival, generally measured from randomization to first progression or death, and overall survival, generally measured from randomization to death from any cause. Seven reported overall survival only, while three reported progression-free survival only.
Primary progression-free and overall survival analyses used the intention-to-treat set, meaning the original assigned treatment groups, in 55 studies, or 81%. But 10 studies, or 15%, did not explicitly say which analysis set they used.
Among the 40 trials that allowed switching, disease progression was a reported reason in 35, or 85%; safety or tolerability in 16, or 40%; clinician discretion in 12, or 30%; and patient preference in seven, or 18%. The categories could overlap, and four trials, or 10%, gave no reason.
Reporting was much thinner on timing: just one study, or 3%, said when switching occurred. Of the 40 switching-allowed trials, 28, or 70%, allowed crossover between trial arms. Eleven permitted both directions, 12 allowed only control-to-intervention crossover, and five allowed only intervention-to-control crossover; 27 of those 28 studies, or 96%, reported the crossover rate.
Across the 40 switching-allowed trials, 23, or 59%, reported control-to-intervention crossover, 16, or 41%, reported intervention-to-control crossover, and 18, or 46%, reported switching to another therapy not studied in the trial. In five trials, or 13%, the direction was not reported.
The statistical gap
At the review level, none of the eight Cochrane reviews mentioned estimands, intercurrent events or post-randomization events. Six meta-analyses reported no analytic strategy for treatment switching; the other two described plans that varied and were not tied to a specific outcome. Only two of the eight considered intercurrent events in their risk-of-bias assessments.
One of the clearest gaps involved censoring, the rule for deciding when a patient’s later experience stops contributing to a particular survival analysis. Publicly available protocols or statistical analysis plans were found for only 29 of 68 trials, or 43%. Among 28 trials with assessable progression-free survival switching-censoring rules, 15, or 54%, censored for switching-related reasons, 10, or 36%, did not, and three, or 11%, had indeterminate rules.
Even where switching was used as a censoring point, only four of the 15 trials, or 27%, reported that rule in the primary study report. No included trial censored overall survival at switching. In one meta-analysis, trials used different primary progression-free survival censoring approaches, yet their treatment-effect estimates were pooled.
Secondary or exploratory switching analyses were planned in 14 of 28 trials assessing progression-free survival, or 48%, and in 10 of 29 assessing overall survival, or 35%.
Why the finding matters—and where it stops
The authors interpret the pattern as limited attention to treatment switching in the reviewed meta-analyses and poor trial reporting. They call for explicit estimand specification, improved reporting and meta-analytic methods that can accommodate multiple estimands.
The findings are bounded by the review’s scope: it examined treatment switching in oncology and published Cochrane meta-analyses. Protocols and analysis plans were publicly available for fewer than half of the trials, limiting what could be checked. Industry-sponsored or industry-performed reviews were not assessed and might have produced different findings.
The authors also note that the interval after release of the ICH E9(R1) addendum may not have been long enough for meta-analytic practice to adapt. This was a descriptive review, not a new meta-analysis; its conclusion concerns how oncology evidence reports and combines treatment switching.
Paper data and sources
Original title: Treatment switching in evidence synthesis in oncology: A systematic review of current meta-analytical practices
Authors: Metcalfe R, Vuong Q, Yan Y et al.
Journal/Repository: Research Synthesis Methods
Status: Peer-reviewed
First online: 2026-08-19
DOI: Not available
Original paper · Full text