Response-adaptive randomisation

In a fixed randomised trial every patient has the same chance of each arm from the first enrolment to the last. Response-adaptive randomisation updates those chances at planned intervals using the outcomes seen so far, within a biomarker subgroup if the design has them. Arms that look better get a larger share of new patients; arms that look worse get fewer and may be dropped for futility; an arm whose predicted probability of success in a future phase 3 crosses a threshold can graduate. The attraction is ethical and practical: more patients receive the better treatment during the trial itself, and drugs that are going nowhere stop consuming patients early.

I-SPY 2 in neoadjuvant breast cancer is the design's best-known user, randomising adaptively within biomarker signatures and graduating drugs on predicted pathological complete response, and GBM AGILE applies the same machinery in glioblastoma. In the corpus, STAMPEDE shows the multi-arm multi-stage version of the idea: since 2005 it has added and dropped arms against a common hormone-therapy control in more than 12,000 men, retiring arms that fail early stages and carrying forward the ones that pass, which is how docetaxel and abiraterone were shown to extend life at first diagnosis of metastatic disease. myeloMATCH re-routes leukaemia patients between sub-studies as measurable residual disease results arrive, an adaptation at the level of the individual patient rather than the randomisation ratio.

The costs are real. Adaptive randomisation is less statistically efficient than a fixed 1:1 split and can be fooled by time trends: if the type of patient enrolled changes during the trial, an arm that recruited mostly later patients is being compared with earlier controls. The analysis has to account for the adaptation, the operating characteristics have to be simulated before the trial starts, and the emerging allocation ratios can leak information about which arm is winning to investigators who should be blind to it. Regulators accept these designs in early phases readily and in confirmatory trials only with pre-specified rules and a firewall between the adapting statisticians and everyone else.