The Gold Standard of Experimental Design | Umbrella Trials, Biomarker Stratification and Shared Screening

1. Define the master population and every substudy estimand

Back to contents

Specify disease stage, prior treatment, measurable disease, biomarker requirements and the population to which each result should apply. An estimand should identify treatment contrast, outcome, time window and handling of intercurrent events. “Effect in biomarker-positive patients” is incomplete when positivity depends on assay, threshold, tissue source and test success. Keep trial-wide screening objectives separate from substudy efficacy objectives.

2. Map the screening-to-allocation pathway

Back to contents

Draw the steps from consent and specimen collection through assay quality control, biomarker call, eligibility review, randomisation or assignment and treatment start. Record people who fail at each step. A central platform can reduce repeated testing, but tissue exhaustion, turnaround time and non-evaluable results can select a different population from all clinically eligible patients. Screening yield and treatment effect answer different questions.

3. Validate biomarkers in their context of use

Back to contents

Analytical validity concerns whether the assay measures the marker reliably; clinical validity concerns association with a clinical state or outcome; clinical utility concerns whether using it improves decisions in the intended setting. An assay suitable for exploratory stratification may not be ready for treatment assignment. Lock or version methods, thresholds and specimen rules, and bridge changes with paired samples where possible.

4. Resolve overlapping, nested and missing biomarker groups

Back to contents

Participants may carry several actionable markers, a broad pathway alteration and a more specific variant, or no usable result. Predefine hierarchy, randomisation among eligible options or another allocation rule. Otherwise investigator preference and slot availability can create hidden selection. Keep screen-negative and assay-failed categories distinct. A person whose specimen failed has an unknown marker status, not evidence of biomarker negativity.

5. Choose controls and multiplicity for the actual questions

Back to contents

Some umbrella substudies are randomised against a shared standard-of-care control; others are single-arm with an external benchmark. Shared controls can improve efficiency when eligibility, calendar time and outcome assessment align, but a control participant is not automatically valid for every biomarker stratum. Calibrate type I error, false-discovery or Bayesian decision thresholds across primary substudy questions and planned adaptations. The master protocol does not erase multiplicity.

6. Worked example: three markers and one shared screen

Back to contents

An invented trial screens 1,000 patients. Eighty have marker A, 120 marker B and 40 have both; 60 specimens fail. If A takes priority, the B substudy receives 80 B-only patients, not 120. Suppose the A assay is upgraded mid-trial and now detects twenty additional low-level positives, mostly later in calendar time. A crude response comparison mixes assay version, marker level and treatment era. The design instead records dual eligibility, randomises overlapping A/B patients where clinically appropriate, stratifies by assay version during the bridge period and reports screening denominators. The worked lesson is simple: allocation rules determine the population that each substudy actually estimates.

7. Protect trial integrity as substudies change

Back to contents

Adding or dropping an intervention can alter expectations, site behaviour and the availability of controls. Use prospective amendment rules, independent monitoring and controlled access to comparative results. Non-concurrent controls may differ because standards of care, diagnostics or supportive treatment changed. Version the master protocol and each substudy module, and state which participants and calendar periods enter every analysis. Operational flexibility needs stronger provenance, not less.

8. Changed-context transfer check: reading support by diagnostic profile

Back to contents

A school evaluates several reading supports within one year group and uses diagnostic tasks to route students to phonological, fluency or comprehension substudies. The transferable structure is a shared screen feeding prespecified strata. The context changes ethical and measurement details: profiles overlap, labels can change with learning and withholding support may be inappropriate. Randomisation might compare two acceptable supports within a profile, while all students receive a baseline programme. A diagnostic score is not a permanent identity.

9. Know the limits and common misconceptions

Back to contents

An umbrella trial is not any study with several treatment arms; biomarker-guided substudies within one disease are central to the usual design. A shared screening platform does not make substudy populations exchangeable. Biomarker positivity does not prove treatment benefit, and response in one stratum does not validate the assay for every use. Small rare strata remain uncertain. Complex operations can magnify missing tissue, delayed treatment and site selection even when the statistical model is correct.

10. Audit screening, assignment and analysis together

Back to contents

Publish the screening denominator, specimen flow, assay versions, failure rates, marker prevalence, overlap matrix, allocation rules, slot availability, substudy openings and closures, control eligibility and analysis populations. Retain code for decision thresholds and multiplicity simulations. Reanalyse after plausible assay misclassification, calendar drift and alternative handling of dual-positive participants. A clear diagram should connect every reported treatment effect back to the people who could have entered that comparison.

11. Frequently asked questions

Back to contents

How does an umbrella trial differ from a basket trial? Umbrella trials usually study several interventions within one disease; basket trials often study one intervention across several disease groups. Must assignment be randomised? No, but causal interpretation is stronger with an appropriate concurrent comparison. Can one participant enter several substudies? Sometimes sequentially, but rules, carryover and eligibility must be explicit. Is a failed assay biomarker-negative? No. Can controls be shared? Yes when eligibility, timing and outcome assessment support the intended contrasts.

12. The gold standard treats screening as part of the experiment

Back to contents

A strong umbrella trial begins before treatment assignment. It defines the master population, validates assays for their role, records every screening branch and resolves overlap without hidden preference. It then protects each treatment comparison with suitable controls, multiplicity rules and calendar-aware governance. The gold standard gains efficiency from shared infrastructure while keeping every substudy answer tied to a transparent biomarker and allocation pathway.