The Gold Standard of Experimental Design | Basket Trials, Borrowing Across Subgroups and Exchangeability

1. Define each basket and its estimand

Back to contents

Specify eligibility, biomarker assay, disease setting, prior treatment, outcome window and target effect for every subgroup. The shared intervention or target does not make populations identical. Decide whether the primary goal is a decision within each basket, an overall signal or both. A trial-wide response rate can conceal clinically important subgroup failure.

2. Distinguish pooling, stratification and borrowing

Back to contents

Complete pooling assumes one common effect; separate analyses assume no shared information. Hierarchical and model-averaging approaches allow partial or selective borrowing. These are not merely computational choices: they encode how similar basket effects are expected to be. Report what quantity shrinks toward what centre, and how much a basket with sparse data can be changed by the others.

3. Make exchangeability a scientific claim

Back to contents

Exchangeability may be supported by a shared target, pathway, dose, outcome definition and disease biology, but none is sufficient alone. Prognosis, co-mutations, tissue context and prior therapies can modify response. Predefine evidence for grouping baskets and reasons some should not borrow. Observed response similarity after unblinding is not a neutral basis for redesigning the model.

4. Control false-positive and false-negative decisions

Back to contents

Borrowing can raise power for small responsive baskets but may inflate erroneous positive decisions when effects differ. Conversely, a large null basket can over-shrink a genuinely active rare subgroup. Evaluate subgroup-wise and family-wise error, power, expected sample size and correct classification across homogeneous and heterogeneous scenarios. No design is uniformly best for every effect pattern.

5. Choose robust borrowing and stopping rules together

Back to contents

Bayesian hierarchical, exchangeability–non-exchangeability, commensurate-prior and model-averaging methods differ in how they discount conflict. Interim futility or efficacy stopping changes the information available for later borrowing. Calibrate thresholds through simulation with the actual accrual order and basket sizes. A method labelled robust can still borrow too much before heterogeneity becomes visible.

6. Worked example: one target across four tumour types

Back to contents

Four baskets test the same targeted therapy with planned sample sizes of 15 each. True response is strong in one type, modest in one, and absent in two. Complete pooling declares all four promising. Separate analyses miss the modest signal. A prespecified robust model borrows heavily between the two similar low-response baskets, lightly from the strong basket and leaves the modest basket uncertain. The report shows basket-specific posteriors, borrowing weights and error simulations rather than presenting the model as automatic truth.

7. Preserve randomised comparisons when they exist

Back to contents

Single-arm baskets rely on historical response benchmarks that may differ across diseases and sites. Randomised basket trials can estimate subgroup treatment effects against concurrent controls, though small strata remain difficult. Borrow treatment effects rather than raw outcomes when baseline risks vary, and model treatment-by-subgroup structure explicitly. A common control does not justify pooling untreated prognoses without adjustment.

8. Changed-context transfer check: one tutoring method across school groups

Back to contents

A programme is tested across four year levels. The transferable principle is partial borrowing only when subgroup effects are plausibly related. The context adds different curricula, baseline tests and teacher delivery. Shared branding is not a mechanism. Outcomes must be comparable and the model should allow a method that helps one level to be ineffective or harmful at another.

9. Know the common misconceptions

Back to contents

A common biomarker does not guarantee a common treatment effect. Dynamic borrowing does not eliminate exchangeability assumptions; it moves them into a model and prior. More power is not free when subgroup error rises. A small basket cannot be rescued by borrowing if its population or outcome is incomparable. Posterior probability thresholds are not self-justifying and must be calibrated to decisions and scenarios.

10. Audit subgroup definitions and borrowing influence

Back to contents

Publish basket eligibility, assay versions, accrual, outcome definitions, prior distributions, stopping boundaries and simulation code. Report separate and pooled summaries beside the chosen model. Show pairwise or basket-level influence where the method permits, and rerun under no borrowing, full pooling and plausible outlier patterns. Track protocol amendments that alter exchangeability during recruitment.

11. Frequently asked questions

Back to contents

Are basket trials only for cancer? No; the logic can apply to other subgrouped conditions. Must every basket use the same sample size? No, but operating characteristics must reflect the planned imbalance. Is Bayesian analysis required? No, though Bayesian hierarchical and model-averaging methods are common. Can baskets be added later? Sometimes, but eligibility, multiplicity, borrowing and calendar effects must be updated prospectively.

12. The gold standard makes borrowing answerable to biology

Back to contents

A strong basket trial defines subgroup decisions first, justifies exchangeability scientifically and calibrates borrowing across realistic heterogeneity. It shows how each basket influences the others and keeps separate evidence visible. The gold standard uses shared information as a measured resource, not a shortcut around small samples or biological difference.