ideasIdea
Bayesian shrinkage for subgroup claims to stop false 'works in this group' stories
Trials look at dozens of patient subgroups and some will look good by chance. A statistical method that pulls extreme subgroup results toward the overall result would make these claims more honest.
Forest plots of subgroup effects are replaced or accompanied by hierarchical Bayesian shrinkage estimates, in which each subgroup effect is partially pooled toward the overall effect according to its precision. Regulators require shrunken estimates for any subgroup claim in a label and journals require them in publications.
Hypothesis
Shrinkage-based subgroup reporting will reduce the number of subgroup-driven label restrictions or expansions that are later reversed, and will improve replication of subgroup effects in subsequent trials.
Rationale
Naive subgroup estimates are noisy and biased toward extremes when selected; shrinkage is the standard remedy in other fields and would have avoided several well-known oncology subgroup reversals.
What would test it
Re-analyse subgroup claims from past label decisions with shrinkage and check which would have survived; compare with later replication data.
Maturity
speculative
Who has to act
regulator
Cost to try
Small (under $1M)
Years to first evidence
1
Bottlenecks it attacks
- Trial design, endpoints and cost · A phase 3 trial takes years and hundreds of millions of dollars, and often answers a question that has already moved on.
- Failures are hidden · Negative trials, failed drugs and abandoned programmes are rarely published, so the same mistakes are repeated.
Pages like this
not linked directly; found by shared links- IdeaPre-specified crossover-adjusted survival in every trial that allows crossover
- IdeaRequire a randomised phase 2 before any phase 3
Shares Failures are hidden, Trial design, endpoints and cost.
- IdeaStop failing trials earlier with pre-registered aggressive futility rules
Shares Failures are hidden, Trial design, endpoints and cost.
- TrialASPREE (ASPirin in Reducing Events in the Elderly)
Shares Hazard ratio (HR), Failures are hidden.
- IdeaBorrow from past control arms to shrink the control group in phase 3
- IdeaA public rulebook for when an external or synthetic control arm is acceptable
- IdeaA pre-registered standard for emulating trials with real-world data
- IdeaAgree in advance how to borrow evidence between similar rare cancers