The Gold Standard of Experimental Design | Platform Trials, Non-Concurrent Controls and Temporal Drift

1. Define each comparison inside the platform

Back to contents

The platform is an infrastructure, not a single treatment contrast. For every arm, specify the eligible population, control definition, randomisation window, outcome, follow-up and estimand. Note which other arms were open at the same time and whether allocation ratios changed. A control participant may be concurrent for one comparison and non-concurrent for another.

2. See why non-concurrent controls break randomised comparability

Back to contents

Concurrent controls are randomised from the same eligible stream while an experimental arm is available. Earlier controls could not have been assigned to the later arm. If prognosis, background care, recruitment sites or outcome recording changes, treatment group and calendar period become associated. Randomisation within each active period does not automatically balance a comparison across periods.

3. Design calendar time into data collection

Back to contents

Predefine periods around arm openings, closures and major external changes. Preserve exact enrolment dates, site activation, eligibility versions, standard-of-care changes, outcome definitions and allocation probabilities. Keep screening logs so shifts in who enters the platform can be examined. Time adjustment cannot recover variables that were never measured or definitions that changed without trace.

4. Protect the concurrent comparison

Back to contents

Size and maintain a sufficient concurrent control group for each primary contrast whenever feasible. Sharing controls among simultaneously active arms can improve efficiency without abandoning contemporaneous randomisation. Adaptive allocation should not starve the control needed for interpretation. If non-concurrent data are planned, specify whether they affect the primary analysis, only precision, or sensitivity analyses.

5. Choose a time model before seeing the treatment result

Back to contents

Methods include categorical period effects, fixed-length calendar intervals, smooth trends, dynamic models and borrowing schemes that discount incompatible historical data. Each assumes something about how prognosis changes. A coarse step function can miss gradual drift; a flexible spline can overfit sparse periods. Use simulation under plausible abrupt, gradual and arm-specific trends to evaluate bias, error rates and precision.

6. Worked example: a respiratory platform after care improves

Back to contents

Arm C enters a hospital platform in July. Mortality among controls fell from 18% in winter to 11% in summer after vaccination and treatment practice changed. Arm C records 10% mortality. Pooling all earlier controls makes the arm appear strongly beneficial; comparing only concurrent controls, at 12%, yields a smaller and uncertain contrast. A prespecified period-adjusted model lies between them but depends on a common time-trend assumption. The report makes concurrent-only analysis prominent and shows how each borrowing rule changes the estimate.

7. Distinguish common drift from arm-specific changes

Back to contents

A common calendar effect assumes control outcomes and treatment potential shift in a shared way. But new sites, eligibility amendments or diagnostic changes may affect only participants available to a late arm. Interactions between period, site or subgroup may matter, yet sparse data limit estimation. Inspect covariate balance and control outcomes over time, but do not use a non-significant trend test as proof of exchangeability.

8. Changed-context transfer check: an education platform trial

Back to contents

A continuing school platform adds a new tutoring programme after an assessment redesign. The transferable principle is that earlier controls are not automatically comparable. The context adds academic cohorts, teacher learning and term effects. A score transformation may link tests, but it cannot by itself remove cohort or implementation drift. Concurrent schools remain the clearest anchor, with earlier cohorts used only under transparent sensitivity assumptions.

9. Know the common misconceptions

Back to contents

More controls do not guarantee less error; biased precision can make a wrong answer more convincing. Statistical adjustment does not recreate randomisation across time. Stable measured covariates do not prove stable unmeasured prognosis. Borrowing is not all-or-nothing, but data-adaptive borrowing also needs operating-characteristic evaluation. Platform flexibility does not excuse changing the estimand or analysis after results emerge.

10. Audit the evolving trial and every borrowing decision

Back to contents

Publish the master protocol history, arm-specific appendices, randomisation probabilities, eligibility changes, site calendar, concurrent windows, outcome versions and intervention availability. Plot control outcomes and key covariates over time without relying on those plots alone. Report concurrent-only, prespecified adjusted and alternative drift analyses, plus simulations showing performance under misspecification. Explain exactly which earlier participants influence each estimate.

11. Frequently asked questions

Back to contents

Are non-concurrent controls historical controls? They are internal earlier controls, but they share the central problem of lacking contemporaneous randomisation for a later arm. Must they be excluded? Not always, though assumptions and sensitivity analyses must match their role. Can period adjustment remove all drift? No; it only addresses patterns represented by the model and recorded data. Why run a platform? Shared infrastructure, controls and adaptive arm management can be efficient when each comparison remains interpretable.

12. The gold standard anchors efficiency to contemporaneous evidence

Back to contents

A strong platform trial treats every arm as a dated randomised comparison. It preserves concurrent controls, records how the platform changes and tests any borrowing method against credible drift. The gold standard welcomes efficiency only when the source of information remains visible: earlier controls may add evidence, but they cannot quietly inherit comparability they never had.