Ask Mallard
Study design

Sequential multiple assignment randomized trials (SMART)

5 min read

A SMART randomizes patients at successive treatment decisions to learn which sequence of care works best for different responses.

NHMRC EVIDENCE LEVEL · 2009Level II · InterventionProvisional design classification. Study quality and certainty of findings require separate appraisal.
Why this level?Participants or clusters are randomly allocated to concurrent comparison groups. NHMRC 2009 source

Patients are randomized at each specified decision. Comparisons of complete treatment strategies preserve those assignments and account for shared histories. Randomization does not remove bias from missing outcomes or departures from the planned analysis.

Fit

When to use it, and when not to

Use it when

  • You need to choose what to do next when initial treatment works or fails.
  • You can measure response at fixed times and randomize between acceptable next steps.

Not when

Before you start

What you need

Decide these first
  1. The decision treeassignmentDraw initial treatments, response groups and later choices. State who is randomized again.
  2. The response rule and decision timesDefine response before enrollment. Use a rule clinicians can apply at the planned visit.
  3. Each allocation probability and the target contrasttargetsummaryMeasureRecord probabilities at every decision. Choose which complete treatment strategies to compare.
It rests on these
  1. Each treatment decision follows its randomized allocation.How to check: Keep the known probabilities for each eligible treatment, given the history available at that decision.
  2. Every treatment in the target strategy has a positive chance of assignment.How to check: Check every branch. A strategy cannot be evaluated where its required treatment was never offered.
  3. The analysis allows for shared participant histories.How to check: Responders may support several strategies. Weight by the relevant allocation probabilities and account for correlated strategy estimates.
  4. Missing responses and outcomes need an explicit plan.How to check: Record why values are absent. Justify the missingness assumption and assess departures from it.

Plain language

How it works

Chance assigns initial care. At a later visit, a rule classifies response. Eligible patients are randomized again between next steps. A complete rule such as start A, continue if responding, otherwise switch is an embedded strategy. Compare strategies on an outcome measured at a common horizon.

The design structure
Sequential randomization and response-guided care. This example randomizes only nonresponders again. A strategy includes initial treatment, the response rule and a next step. Responders can contribute to several strategies; they remain the same people.Sequential randomization and response-guided careInitial randomizationStart AStart BAssess response by a prespecified ruleResponders: continue initial treatmentNonresponders: randomize againSwitch treatmentAdd treatmentThis example randomizes only nonresponders again. A strategy includes initial treatment, the response ruleand a next step. Responders can contribute to several strategies; they remain the same people.
Conceptual diagram. It shows the structure of the idea, not data from any study.
Why the whole path matters

Patients can be consistent with several strategies. An analysis can weight each observed path by the inverse of its allocation probability. Use a method designed for these shared histories, not an ordinary comparison of final treatment groups.

Worked example

A worked example

Invented for illustration
Setting
Patients start treatment A or B with equal probability.
Question
How are patients represented in a response-guided strategy?
Approach
Responders continue their first treatment. Nonresponders are randomized equally to switch or add treatment.
Result
The allocation probability product for A then switch is 25 in 100, or 0.25. The inverse-probability weight is 4.
What it means
An A responder has weight 2. The weights use allocation chances, not the probability of response or raw subgroup success rates.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

A SMART compares response-guided treatment sequences, not just the last treatments received. Preserve each randomization probability and account for shared participants when comparing embedded strategies.

  • Comparing only the last treatments as if they were baseline groups.Instead: Compare prespecified complete strategies. Later treatment depends on earlier treatment and response.
  • Confusing adaptive care with changing the trial after interim results.Instead: A SMART adapts care to individual response. Trial-level adaptation changes trial rules and needs separate safeguards.
  • Counting a responder shared by two strategies as two independent people.Instead: Keep participant identifiers and use variance estimates that allow for the shared histories.

Inside Mallard

How Mallard uses this

Plans do not link to this design yet. Mallard has no sample-size form for sequential multiple assignment randomized trials (smart). A plan sized by simulation gets no headline sample size; Mallard does not run that simulation.

No Mallard code example points to this page.
Atlas node design.smart, the id a plan cue records.

Related Learn pages

Related designs and methods

Concepts

Evidence

Sources & further reading

  1. An experimental design for the development of adaptive treatment strategies (opens in a new tab) · Stat Med · doi 10.1002/sim.2022
  2. Experimental design and primary data analysis methods for comparing adaptive interventions (opens in a new tab) · Psychol Methods · doi 10.1037/a0029372
Prepared with AI assistance. Source titles, authors, publication years and available DOI metadata were checked against PubMed on 2026-11-06. No independent methodological approval is claimed. Murphy (2005) introduces the sequential randomized design. Nahum-Shani et al. (2012) give the embedded-strategy analysis, inverse-allocation weights and variance methods for shared participant histories. Reporting guidance: use CONSORT for randomized comparisons, with extensions suited to the actual design. A reporting checklist is not a design manual.
Last source review: 2026-11-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.