Does everyone receive the allocated intervention?

No. People may decline or discontinue it, and the main comparison follows original assignment.

Does single consent mean nobody consents?

No. It describes an arrangement in which consent to the new intervention is sought in that arm after allocation.

Can this design bypass ethics review?

No. Departures from ordinary consent and use of records need appropriate independent review and applicable authorization.

AI Glossary / Statistics and probability / Advanced

Zelen's design

Definition

Zelen's design is a randomized study design in which allocation occurs before consent to the assigned study intervention is sought. It changes the usual ordering of consent and randomization. Different versions use consent procedures in one or both study arms, with appropriate independent authorization for the departures from conventional research consent.

What it means

The design is associated with Marvin Zelen's work on randomized consent. Its practical appeal is that a study can evaluate an offer of an intervention among an eligible population, including people who decline that offer. This differs from enrolling only people already willing to accept random allocation between interventions.

The distinction matters for AI-supported services: a useful tool may have little population benefit if few people accept it. However, the design is not permission to experiment secretly on customers or patients. Consent arrangements, access to records and participant protections require a study-specific ethical and legal assessment. This entry explains the statistical idea rather than providing a compliance procedure.

How it works

Researchers define eligibility and the intervention comparison before allocation. In a single-consent version, those assigned the new intervention are approached about accepting it, while the comparison group continues usual care under the approved arrangements. In a double-consent version, people in both arms are approached after allocation about their assigned intervention.

The main analysis retains participants in their original assigned groups, whether they accept the offer or not. Outcome collection must cover those groups comparably; selective follow-up can undermine the comparison. Investigators separately document assignment, offers delivered, acceptance, actual receipt and outcomes. The design therefore distinguishes an allocated offer from the intervention a person ultimately receives.

Three practical examples

1. Service evaluation: Researchers consider whether offering an additional navigation service improves outcomes across all eligible users, including those who do not take up the offer. A suitably reviewed design preserves that population-level question.

2. Digital engagement: A research team evaluates an invitation to a support platform. Acceptance rates are part of the program's real-world effect rather than grounds for deleting nonusers from the assigned group.

3. Evidence review: An AI product manager reads a trial using post-randomization consent. The manager checks whether its results concern the policy of offering the service or the effect of actually receiving it before applying them to a rollout plan.

Think of it this way

Think of evaluating an invitation policy rather than reviewing only the experiences of people who accepted the invitation. Refusal is meaningful information about how the policy works. The analogy concerns the analysis; it does not remove the need to justify how people were included, approached and followed in a research study.

Worked example

An original hypothetical study assigns 100 eligible people to an offer of an optional program and 100 to usual service. Assume complete, comparable outcome records and an approved protocol. Sixty people in the offer group accept. Of those, 36 meet the predefined success outcome; another 10 successes occur among the 40 who decline. The usual-service group has 38 successes.

The assigned offer group's success rate is (36 + 10) / 100 = 46%. The comparison rate is 38 / 100 = 38%, giving an observed assignment-group difference of 8 percentage points. Uptake is 60 / 100 = 60%.

Comparing only acceptors gives 36 / 60 = 60%, apparently 22 points above the comparison group. That comparison no longer preserves randomization: acceptors may differ in motivation or prognosis. The 8-point estimate still needs an uncertainty interval and the usual validity checks. Dividing it by uptake does not automatically establish an effect for everyone who might use the program.

Illustrative case study

Fictional case: Alder Services considers a research evaluation of an AI-assisted appointment navigation offer. The proposed question is whether routinely offering the service improves attendance among an eligible population, rather than whether enthusiastic volunteers can use the interface.

Before implementation, the research team obtains the necessary independent review of allocation, consent and record use. It defines the attendance window and records unsuccessful contacts as well as accepted invitations. Its final report presents outcomes by assigned group and separately describes uptake. The team does not label the result a pure effect of the AI: the offer includes contact, explanation and access to staff support, all of which may contribute to the observed difference.

Common misconceptions and mistakes

Zelen's design is not an ordinary trial with a consent form accidentally obtained late. The ordering is intentional and must be justified in the protocol. Nor is it synonymous with cluster randomization; the allocation unit is a separate design choice.

Excluding refusers can create selection bias. Intention-to-treat analysis does not itself solve missing outcomes, contamination or poor measurement. An offer effect also cannot be assumed to equal the effect of compulsory participation or universal uptake. Finally, a low-risk intervention does not by itself establish that altered consent or use of identifiable records is permissible.

Business use

Organizations commissioning evidence should ask whether their decision concerns offering a service to a population or delivering it to willing users. The answer shapes which study can support the decision. A Zelen-style comparison may inform uptake-sensitive service policies when its ethical and operational requirements can be satisfied.

For procurement, request the allocation diagram, consent arrangements, participation flow and analysis population. Budget projections should use the actual offer and acceptance process tested. A different invitation, staffing model or eligible population may produce different uptake and outcomes even with the same underlying software.

Check your understanding

What comes first in the defining sequence?

Answer: Random allocation precedes the relevant intervention-consent process.

What is the observed assigned-group difference in the example?

Answer: Eight percentage points.

Why not compare only acceptors?

Answer: Their decision to accept can select a systematically different group.

Related terms

Selection bias
Estimand
randomized trial
informed consent
intention-to-treat analysis
pragmatic trial
treatment uptake

Learn next

Learn next how to distinguish assignment effects from treatment-receipt effects, handle missing outcomes and specify a study's estimand. Any proposed application also needs current, jurisdiction-specific research governance review.

Estimand

Sources

Zelen, Randomized consent designs for clinical trials: an update (1990), primary abstract: https://pubmed.ncbi.nlm.nih.gov/2218168/ . Simon, Shortreed and DeBar, Zelen design clinical trials: why, when, and how (2021), full methodological discussion: https://link.springer.com/article/10.1186/s13063-021-05517-w . Original Zelen full text was not consulted. Legal and ethics requirements need current specialist review; examples are fictional teaching material.

Previous: Z-test Next: Zero anaphor