Trial design picker

Every trial design in the glossary as a card: when to use it, a trial in OnCo that used it, and the risk that most often undoes it. Below the cards is a plain-English list that starts from the question you want answered and names the design that answers it, then the statistics you will meet when the result comes out and the lifecycle every trial passes through.

29 designs6 familiesAll trial terms All endpoint terms Trials in plain words Failure museum

Comparative designs6 Single-arm and externally controlled designs4 Master protocols3 Adaptive designs7 Dose-finding designs3 Pragmatic, cluster and real-world designs6 Which design fits my question Reading the result Lifecycle

Comparative designs

6 designs

Two or more arms, assigned by chance. The only designs that can prove a treatment causes a benefit.

Randomised trial

A trial in which a coin toss (done by computer) decides which treatment each patient gets, so the groups are alike in every way except the treatment. It is the only reliable way to prove a treatment causes a benefit.

When to use it: You want to know whether a new treatment is better than the current standard, and enough patients exist to compare them fairly.
Main risk: Slow and expensive; the standard of care can move on before the trial reads out.
Worked example: CheckMate 067Positive945 patients randomised equally across three arms, followed for ten years.

Clinic basics Clinical trial Placebo Blinded trial

Non-inferiority trial

A trial designed to show a new treatment is not meaningfully worse than the standard, rather than better: used when the new option is shorter, cheaper, less toxic or easier (five radiotherapy fractions instead of 25, six months of trastuzumab instead of twelve).

When to use it: The new option is shorter, cheaper, less toxic or easier, and you need to show it gives up little or nothing.
Main risk: A sloppy trial drifts towards no difference, which here looks like success; the margin decides everything.
Worked example: PERSEPHONEPositiveSix months of trastuzumab against twelve, judged against a pre-specified margin.

Trials De-escalation, escalation and response-adapted therapy Hypofractionation Intention-to-treat (ITT) and per-protocol analysis

Non-inferiority margin and equivalence trials

The margin is the amount of benefit a trial is allowed to lose and still call the new treatment good enough; it is chosen before the trial starts, and where it is set decides what the result means.

When to use it: You are reading or writing a non-inferiority protocol and need to justify how much loss is acceptable.
Main risk: A wide margin makes almost anything look good enough; a superiority miss is not a non-inferiority win.
Worked example: FAST-ForwardPositiveOne-week breast radiotherapy met its margin on five-year local relapse.

Trials Non-inferiority trial Randomised trial Intention-to-treat (ITT) and per-protocol analysis

Stratified randomisation, allocation concealment and minimisation

Randomisation is done by computer and often within groups (by stage, by biomarker, by region) so that each arm gets a fair share of the patients who matter most, and nobody can steer a particular patient to a particular arm.

When to use it: The trial is small or a few factors (stage, biomarker, region) strongly predict outcome and must be balanced by construction.
Main risk: Too many strata leave cells empty; the factors chosen become the subgroups everyone will later argue over.
Worked example: IMvigor011PositiveOnly patients with a positive blood test for leftover cancer were randomised.

Trials Randomised trial Blinding Blinded trial

Blinding (double-blind, open-label, placebo-controlled)

Whether patients and doctors know which treatment is being given. Double-blind: neither knows (a placebo hides it). Open-label: both do, unavoidable for surgery or radiotherapy but a source of bias when judging progression and symptoms.

When to use it: The outcome involves judgement (symptoms, scan reads, when to stop treatment) and expectations could colour it.
Main risk: Many cancer treatments cannot be disguised; blinding then shifts to the people reading the scans.
Worked example: ACT IVNegativeA vaccine against a control vaccine, both with chemotherapy, in 745 patients.

Trials Blinded independent central review Control arm and comparator Intention-to-treat (ITT) and per-protocol analysis

Crossover in trials

When patients in a trial's control arm are allowed to switch to the experimental drug after their cancer progresses. It is fair to patients but blurs the survival comparison, because the control group has now had the drug too.

When to use it: Control patients will be offered the experimental drug at progression, for ethical or practical reasons.
Main risk: Dilutes the survival comparison; a real benefit can vanish, and its absence can be excused.
Worked example: PSMAforePositiveMost control patients crossed over to the radioligand, blurring overall survival.

Trials Overall survival Progression-free survival Surrogate endpoint

Single-arm and externally controlled designs

4 designs

Everyone gets the new treatment. Fast and small, and unable on their own to show anyone lives longer.

Single-arm trial

A trial where everyone gets the new drug and there is no comparison group; success is judged by the share of tumours that shrink and for how long. Fast and small, it underpinned accelerated approvals for larotrectinib, sotorasib, tarlatamab and most CAR-T products, but it cannot show that patients live longer.

When to use it: A rare or heavily pre-treated population with no good comparator, and a drug whose effect on tumour shrinkage should be large and obvious.
Main risk: Patient selection and response-evaluable definitions flatter the result; confirmatory trials sometimes fail.
Worked example: L-MINDPositive81 patients, a 60 percent response rate and a matched real-world comparison behind an approval.

Trials Objective response rate Duration of response External and synthetic control arms

External and synthetic control arms

Instead of randomising patients to a control group, comparing a single-arm trial against patients treated in the past or recorded in registries, matched by propensity scores. Regulators accepted this for eflornithine in neuroblastoma and other rare cancers, but unmeasured differences between the groups, including changes in supportive care between eras, can masquerade as drug effects.

When to use it: Randomisation is impossible or unethical and good historical or registry data on similar patients exist.
Main risk: Differences between eras and populations masquerade as drug effects; both arms of ACT IV beat historical expectations.
Worked example: NMTRC003/003B (DFMO maintenance)Positive105 children on eflornithine compared with a propensity-matched external control.

Trials Single-arm trial Real-world evidence Control arm and comparator

Window-of-opportunity trial

Giving a new drug for a few weeks in the gap between diagnosis and scheduled surgery, then examining the removed tumour to see what the drug did to it. Patients lose nothing (surgery proceeds as planned) and researchers get a direct look at the drug's biological effect.

When to use it: You need to see what a drug does inside a tumour and surgery is already scheduled a few weeks away.
Main risk: Cannot measure survival; a short exposure can miss slow effects; delaying surgery must be justified.
Worked example: Single-injection depot progesterone before breast surgery (Tata Memorial)MixedA single pre-operative injection, with survival followed for years afterwards.

Trials Neoadjuvant / adjuvant / perioperative Pathologic complete response First-in-human (FIH) trial

Real-world evidence

Data from routine care rather than clinical trials, used to check whether results hold outside the trial population.

When to use it: You want to know whether a trial result holds in older, sicker and more diverse patients treated in routine care.
Main risk: Confounding by indication: patients who received the drug differ from those who did not in ways records do not capture.
Worked example: PATHFINDER 2CompletedAbout 35,000 adults given a blood test alongside standard screening in a single-arm study.

Trials External and synthetic control arms Pragmatic trial Registry-based randomised trial

Master protocols

3 designs

One infrastructure, many questions: one drug across cancers, many drugs within a cancer, or arms that come and go for years.

Basket trial

A basket trial tests one drug in patients whose cancers share a mutation or marker, whatever organ the cancer started in; each organ type is a basket, and the trial asks whether the drug works across them.

When to use it: A molecular alteration appears in a few percent of many cancers and a drug is built for it.
Main risk: Small baskets give wide confidence intervals, and a target can drive one cancer and be a passenger in another.
Worked example: NAVIGATEPositiveLarotrectinib across 17 tumour types, the first tumour-agnostic approval.

Trials Basket, umbrella, and platform trials Umbrella trial Seamless, adaptive and Bayesian trial designs

Umbrella trial

An umbrella trial takes one cancer, tests every patient's tumour for a panel of markers, and routes each patient to the sub-study whose drug matches their marker, so several targeted drugs are tested at once under one roof.

When to use it: One cancer splits into many small molecular subgroups and patients should be tested once and considered for many arms.
Main risk: Screening hundreds to fill sub-studies of thirty is slow, and many patients screen into no arm at all.
Worked example: myeloMATCHRecruitingEvery new leukaemia sequenced within days and routed to a matched sub-study.

Trials Basket, umbrella, and platform trials Basket trial Seamless, adaptive and Bayesian trial designs

Basket, umbrella, and platform trials

Trial designs that test one drug across several cancers sharing a mutation (basket, such as NCI-MATCH), several drugs each matched to a biomarker within one cancer (umbrella, such as Lung-MAP), or keep adding and dropping arms over time (platform, such as I-SPY 2). Basket results underpin tumour-agnostic approvals.

When to use it: The disease will have new candidate treatments for years and a shared control arm can serve them all.
Main risk: Control arm drift over a long trial; governance and statistics are heavier than for a single trial.
Worked example: STAMPEDEPositiveArms added and dropped against one hormone-therapy control since 2005, in more than 12,000 men.

Trials Basket trial Umbrella trial Seamless, adaptive and Bayesian trial designs

Adaptive designs

7 designs

Rules fixed in advance let the trial change as data come in: merge phases, resize, drop arms, tilt randomisation.

Seamless, adaptive and Bayesian trial designs

Trial designs that change as data come in: merging phases so successful drugs move forward without pause, dropping arms or doses that are not working, adding new ones, and shifting randomisation toward what seems to help. Faster and more efficient, but they need careful statistics to stay honest.

When to use it: A successful phase 2 cohort should roll straight into a registrational trial without a pause between phases.
Main risk: Alpha control and a firewall around the interim data are essential; the phase 2 patients may not be quite the phase 3 population.
Worked example: NRG-HN002 & NRG-HN005 (HPV+ de-escalation)MixedA phase 2/3 de-escalation trial stopped when the reduced-dose arms did worse.

Trials Basket, umbrella, and platform trials Interim analysis, readout and data cut-off First-in-human (FIH) trial

Response-adaptive randomisation

In a response-adaptive trial the computer tilts the odds as results come in, so later patients are more likely to be assigned to the arm that seems to be working and arms that are failing are dropped.

When to use it: Several arms compete and it is ethical and practical to steer later patients towards what is working.
Main risk: Less efficient than 1:1, vulnerable to time trends, and can leak which arm is winning.
Worked example: STAMPEDEPositiveMulti-arm multi-stage: arms that fail early stages are retired, the rest carry on.

Trials Seamless, adaptive and Bayesian trial designs Bayesian trial design Basket, umbrella, and platform trials

Bayesian trial design

A Bayesian trial states what was believed before the trial, updates that belief with each patient's result, and reports the probability that the treatment works, rather than a yes-or-no verdict against a p-value.

When to use it: You have real prior information (adult data, earlier trials) and a small population, or you need continuous updating for dose finding or graduation rules.
Main risk: The prior is a judgement; regulators want it justified in advance and the false-positive rate simulated.
Worked example: NMTRC003/003B (DFMO maintenance)PositiveAn externally controlled approval of the kind Bayesian borrowing formalises.

Trials Seamless, adaptive and Bayesian trial designs Response-adaptive randomisation Dose-escalation designs

Statistical power, sample size and re-estimation

Before a trial starts, statisticians work out how many patients, or how many deaths or relapses, are needed to detect the benefit they hope for; an adaptive trial can check that guess part way through and enlarge itself if the guess was wrong.

When to use it: The effect size used to plan the trial is uncertain and you want the option to enlarge the trial at an interim without inflating false positives.
Main risk: Unblinded re-estimation must be paired with methods that hold alpha; sizing on inflated early data is the classic way to fail.
Worked example: DREAM3RNegativeSized on encouraging early data and stopped early without meeting its endpoint.

Trials Statistical significance Seamless, adaptive and Bayesian trial designs Interim analysis, readout and data cut-off

Group sequential design, stopping rules and alpha spending

A group sequential trial plans in advance how many times it will peek at the data and how strong the evidence must be at each peek to stop early, so that looking several times does not inflate the chance of a false positive.

When to use it: You will look at the data more than once and want the right to stop early for benefit, harm or futility.
Main risk: Early stopping overestimates the effect and truncates secondary endpoints and long-term safety.
Worked example: ADAURAPositiveUnblinded early on the monitoring committee's recommendation; survival benefit confirmed later.

Trials Interim analysis, readout and data cut-off Futility analysis Statistical significance

Enrichment and biomarker-stratified designs

An enrichment design enrols only patients whose tumours carry the marker the drug needs; a stratified design enrols everyone but tests marker-positive and marker-negative patients separately, to learn whether the marker predicts benefit.

When to use it: You need to learn whether a marker predicts benefit, not just prognosis, or to co-develop a drug with its companion test.
Main risk: Enriching on an unvalidated marker excludes patients who would have benefited; separate p-values are not an interaction test.
Worked example: MAGNITUDEMixedMarker-positive and marker-negative cohorts run side by side; one stopped for futility, one won.

Trials Biomarker Seamless, adaptive and Bayesian trial designs Stratified randomisation, allocation concealment and minimisation

SMART design (sequential multiple assignment randomised trial)

A SMART trial randomises patients once at the start and again at a decision point, such as after the first scan or a residual disease test, to compare whole treatment strategies rather than single drugs.

When to use it: The real question is a strategy (start with A, switch to B on poor response) rather than a single drug.
Main risk: Needs more patients because they are split at each stage, and the decision rule must work identically at every site.
Worked example: CAPTIVATEPositivePatients with undetectable residual disease re-randomised to stop or continue.

Trials Seamless, adaptive and Bayesian trial designs De-escalation, escalation and response-adapted therapy Minimal / molecular residual disease

Dose-finding designs

3 designs

How a first-in-human trial climbs to a useful dose, and why the highest tolerable dose is no longer the goal.

Dose-escalation designs (3+3, BOIN, dose-expansion)

How a phase 1 trial climbs from a tiny starting dose to a useful one: the old 3+3 design treats three patients at a time and moves up if none has serious toxicity; newer statistical designs (BOIN, CRM) use all the data to pick doses more accurately with fewer patients on ineffective levels.

When to use it: A new agent is entering people for the first time and the dose-toxicity relationship is unknown.
Main risk: The 3+3 design treats many patients at ineffective doses and finds the maximum tolerated dose poorly; model-based designs need a statistician.
Worked example: REJOICE-Ovarian01RecruitingEscalation, then a randomised dose-optimisation part, then phase 3 at the chosen dose.

Trials Dose-limiting toxicity Maximum tolerated dose Recommended phase 2 dose

Project Optimus

An FDA programme pushing companies to find the best dose of a cancer drug, not just the highest tolerable one.

When to use it: A targeted drug or antibody-drug conjugate saturates its target below the maximum tolerated dose and the pivotal dose has not been compared with a lower one.
Main risk: Adds a randomised dose cohort and months to development; skipping it can mean an approved dose nobody can stay on.
Worked example: Low-dose nivolumab plus metronomic chemotherapy (Tata Memorial)PositiveOne twentieth of the standard dose tested in a randomised trial, for affordability.

Regulation & policy Dose-escalation designs Maximum tolerated dose Recommended phase 2 dose

First-in-human (FIH) trial

The first time a new drug is given to people, after animal and laboratory testing and clearance of an Investigational New Drug application. In oncology these are phase 1 trials in patients with advanced cancer who have exhausted standard options, starting at a fraction of the dose predicted to be safe; modern designs run straight through to registrational cohorts.

When to use it: Preclinical work is complete and the question is whether the drug is safe enough, at any dose, to study further.
Main risk: Late toxicities are missed in short observation windows; a small early signal is easily over-read.
Worked example: AUGMENT-101PositiveA phase 1/2 that carried a new drug class from first dose to approval.

Trials Dose-escalation designs Maximum tolerated dose Seamless, adaptive and Bayesian trial designs

Pragmatic, cluster and real-world designs

6 designs

Testing treatments where and how they will actually be used, from a single patient to whole districts.

Pragmatic trial

A pragmatic trial tests a treatment the way it would actually be used: ordinary patients, ordinary clinics, usual care as the comparison and an outcome that matters to patients, so the answer applies in the real world and not only in the trial.

When to use it: The intervention is already in use (a generic drug, an exercise programme, a screening schedule) and the question is whether it works in ordinary care.
Main risk: Usual care varies and drifts, adherence is lower, and record-based outcomes are less complete, all pulling towards no difference.
Worked example: CHALLENGE (CCTG CO.21)PositiveA coached exercise programme after chemotherapy, delivered across five countries over three years.

Trials Randomised trial Registry-based randomised trial Decentralised trial

Cluster-randomised trial

A cluster-randomised trial randomises whole groups (villages, clinics, hospitals) rather than individual people, which is the only fair way to test something delivered to a community, such as a screening programme.

When to use it: The intervention is delivered to a community, clinic or hospital and cannot be given to one person and withheld from the next.
Main risk: Needs more participants for the same power, and few clusters cannot be rescued by many individuals.
Worked example: Osmanabad cervical screening trial (HPV testing vs cytology vs VIA)Positive52 villages randomised to four screening strategies; only HPV testing cut deaths.

Trials Randomised trial Stepped-wedge design Pragmatic trial

Stepped-wedge design

In a stepped-wedge trial every hospital or region eventually switches to the new approach, but the order in which they switch is randomised, so the periods before and after each switch can be compared fairly.

When to use it: Every site will adopt the new approach eventually and it cannot be rolled out everywhere at once.
Main risk: Intervention periods are later than control periods, so anything else that changes over time is partly confounded.
Worked example: SANOPositiveDutch hospitals switched to active surveillance in randomised order.

Trials Cluster-randomised trial Non-inferiority trial Non-inferiority margin and equivalence trials

Registry-based randomised trial

A registry-based trial randomises patients who are already being tracked by a national or disease registry and uses the registry, not trial visits, to record what happens to them, making very large trials cheap enough to run.

When to use it: A complete, linkable registry already records the outcome you care about and the intervention is simple.
Main risk: Data are only as good as the registry; scan-defined and patient-reported outcomes are out of reach.
Worked example: CIRCULATE-Japan (GALAXY / VEGA / ALTAIR)ActiveAn observational cohort tested for circulating tumour DNA feeds randomised sub-studies.

Trials Pragmatic trial Real-world evidence Randomised trial

Decentralised trial

A decentralised trial brings the trial to the patient: consent by video, drug delivered to the home, blood drawn at a local clinic, symptoms reported on a phone, so people far from a cancer centre can take part without travelling to it every few weeks.

When to use it: Eligible patients live far from centres, the drug is oral or the intervention is behavioural, and enrolment or diversity is the bottleneck.
Main risk: Infusions, biopsies and central imaging still need sites; local scans add noise; digital tools exclude some patients.
Worked example: BWEL (Breast Cancer Weight Loss, Alliance A011401)MixedA two-year weight-loss programme delivered by telephone to more than 3,000 women.

Trials Pragmatic trial Accrual and enrolment Informed consent

N-of-1 trial

An N-of-1 trial is a randomised experiment in a single patient: the person alternates between treatment and comparison in random order, often blinded, to find out what works for them rather than for the average patient.

When to use it: A fast-acting, reversible treatment for a stable symptom, and the question is what works for this patient.
Main risk: Useless for treatments with carry-over or for outcomes that cannot be measured repeatedly, which rules out anticancer drugs.
Worked example: Low-dose olanzapine for cancer anorexia (Tata Memorial)PositiveA group-level answer to an appetite question that a series of single-patient trials could personalise.

Trials Randomised trial Crossover in trials Placebo

Which design fits my question

Start from the question, not the design. Each answer links to the term that explains the trade-offs.

1.
Is the new treatment better than what we do now?
A randomised controlled trial, blinded if the outcome involves judgement.
Randomised trial
2.
Can we give less (shorter, fewer fractions, no chemotherapy) without losing much?
A non-inferiority trial with a justified margin, analysed both intention-to-treat and per protocol.
Non-inferiority margin and equivalence trials
3.
Does this drug work in a rare mutation that appears across many cancers?
A basket trial, with response rate per basket and a confirmatory commitment.
Basket trial
4.
Which of many targeted drugs helps which molecular subgroup of one cancer?
An umbrella trial with a shared screening panel and matched sub-studies.
Umbrella trial
5.
Will there be new candidate treatments in this disease for years to come?
A platform trial with a shared control arm and pre-specified rules for adding and dropping arms.
Basket, umbrella, and platform trials
6.
Does the marker predict benefit, or just prognosis?
A biomarker-stratified design that randomises within marker-positive and marker-negative groups and tests the interaction.
Enrichment and biomarker-stratified designs
7.
What dose should go into the pivotal trial?
Model-based dose escalation followed by a randomised comparison of two or more doses.
Project Optimus
8.
Should treatment change depending on early response or residual disease?
A SMART design that re-randomises at the decision point and compares whole strategies.
SMART design (sequential multiple assignment randomised trial)
9.
Does the intervention work in ordinary clinics and ordinary patients?
A pragmatic trial with broad eligibility, usual care as comparator and a hard outcome.
Pragmatic trial
10.
Is the intervention delivered to a community or a whole clinic?
A cluster-randomised trial, or a stepped-wedge design if every site will adopt it eventually.
Cluster-randomised trial
11.
Can we randomise at all?
If not, a single-arm trial with an external control, and honesty about what it cannot show.
External and synthetic control arms
12.
Which treatment works for this one patient's symptom?
An N-of-1 trial, if the treatment acts and wears off quickly.
N-of-1 trial

Reading the result

The statistics a trial report throws at you, each with a worked example from the corpus. The forest plot puts every hazard ratio in OnCo on one axis and Trials in plain words turns results into people out of 100.

Kaplan-Meier curves and censoring Hazard ratio Absolute versus relative benefit Confidence interval P values and multiplicity Estimands and intercurrent events Which surrogates have earned trust Quality-adjusted survival Patient-reported outcomes Why trials fail

The lifecycle every trial passes through

Protocol, ethics review, registration, enrolment, monitoring, readout, publication, label. The regulatory timeline dates the last step for every product in OnCo and the catalyst calendar lists the readouts still to come.

From protocol to label Protocol and analysis plan Ethics review Informed consent Registration and results reporting Equipoise Data monitoring committee Interim analysis and readout Futility Phases 1, 2 and 3