SMART Designs: Sequential Multiple Assignment Randomized Trials

A Sequential Multiple Assignment Randomized Trial (SMART) does not test one treatment against another. It tests a sequence of decisions — which treatment to start with, and how to adjust based on how a participant responds — against other plausible sequences. The output isn’t “drug A beat drug B.” It’s a decision rule: start with X; if the person hasn’t responded within a defined window, switch to Y; if they have, continue X at a reduced intensity. That decision rule is called a dynamic treatment regime (DTR), and building one well is the entire point of the design.

SMARTs matter specifically where a single fixed treatment rarely works for everyone and clinicians already adapt in practice — behavioral intervention research, mental health and substance-use treatment, chronic disease management, and increasingly implementation science and mHealth — but the adapting has historically happened by clinical judgment rather than by evidence. A SMART turns that adaptation itself into the thing being tested.

What a SMART Actually Randomizes

A SMART has two or more stages. At each stage, participants are randomized again — but not everyone is randomized to the same set of options at the second stage. Who gets re-randomized, and to what, depends on how they did at the first stage. This is the structural feature that separates a SMART from a standard multi-arm RCT: assignment at stage two is a function of an observed tailoring variable — typically early response or non-response, though it can be any measured characteristic a clinician would realistically use to adjust care.

A minimal two-stage structure looks like this:

• Stage 1: Participants are randomized to an initial treatment (e.g., Treatment A vs. Treatment B).
• Response assessment: After a pre-specified interval, each participant is classified as a responder or non-responder on a pre-specified criterion.
• Stage 2: Non-responders (and sometimes responders too, if the scientific question calls for it) are re-randomized to a second-stage option — continue, augment, or switch.

Every participant follows exactly one of the treatment sequences the design generates, and every sequence maps back to one of a small number of complete, replicable decision rules.

Embedded Adaptive Interventions

The complete set of decision rules a given SMART can produce — every legal path through its stages — are the adaptive interventions embedded in that trial. A two-stage SMART with two first-stage options and two second-stage options per response status typically embeds somewhere between four and eight complete adaptive interventions, depending on how the design branches. The trial isn’t just comparing treatments; it’s comparing these embedded regimes against one another, using the randomized data pooled correctly across the stages a given regime is consistent with (the “replicate the data” weighting logic from Murphy’s original SMART methodology, later extended into Q-learning approaches for building more deeply tailored rules than the embedded set alone provides).

This is also where the design underdelivers most often in practice: a 2022 systematic review of published SMART trials in the Journal of Clinical Epidemiology found that despite this being one of the design’s central advantages, 83.3% of trials rarely went on to actually analyze and compare their embedded adaptive interventions — running a SMART’s randomization scheme without doing the comparison the scheme exists to enable.

Primary Aims Determine the Sample Size — Not the Other Way Around

A SMART’s primary scientific aim has to be specified before the sample size calculation is meaningful, because different aims imply different comparisons and different effective sample sizes at each stage:

• Stage-specific comparisons (e.g., which first-stage treatment is better, ignoring stage two) use standard two-arm comparison logic on the stage-one randomization.
• Comparing embedded adaptive interventions against each other requires accounting for participants who are shared across multiple regimes (a non-responder who continued Treatment A is “in” every embedded regime that starts with A and continues on non-response) — this changes the effective variance and the sample size formula compared with a simple two-arm trial.
• Second-stage comparisons depend on the first-stage response rate: a lower response rate leaves more people eligible for the second randomization, which changes second-stage power independently of the total enrollment.

Reporting quality here is a documented, specific problem, not a hypothetical one. The same 2022 Journal of Clinical Epidemiology systematic review (Bigirumurame, Uwimpuhwe & Wason) found:

• 66.7% of published SMART trials did not report the parameters needed to replicate their own sample size calculation.
• 75% did not report whether or how they adjusted for multiple comparisons.
• 91.7% did not specify what software was used for the sample size calculation, and 58.3% omitted the analysis software entirely.

The review’s conclusion was blunt: SMART trials should report all key design components — primary aim, sample size parameters and assumptions, multiple-testing handling, and software — as a baseline, not an afterthought, precisely because the design’s added complexity over a standard RCT makes under-reporting harder to catch and easier to misinterpret downstream.

SMART Designs vs. Related Designs

SMARTs are frequently confused with, or lumped together under, “adaptive trial designs” generally — worth separating clearly:

• SMART vs. response-adaptive/group-sequential adaptive trials: CASRAI’s guide to adaptive trial designs covers designs that change randomization ratios, sample size, or stopping decisions within a single treatment comparison based on accumulating data. A SMART instead re-randomizes participants to entirely different, subsequent treatment decisions based on individual response — the adaptation is at the level of each participant’s care pathway, not the trial’s overall statistical parameters. The two are genuinely different tools and are sometimes combined (a SMART with a response-adaptive randomization rule at one stage), but “adaptive” in each case means something structurally different.
• SMART vs. crossover designs: a crossover trial gives every participant every treatment in sequence, primarily to control between-subject variability. A SMART gives each participant one path through a decision tree, chosen partly by design (randomization) and partly by their own observed response — it is not trying to give everyone everything.
• SMART vs. factorial designs: a factorial design tests multiple independent factors simultaneously against a fixed set of participants. A SMART’s second-stage assignment is explicitly not independent of the first stage — it is conditioned on it, which is exactly what makes DTR construction possible.
• SMART vs. stepped-wedge or cluster-randomized designs: CASRAI’s guide to stepped-wedge and cluster-randomized designs covers staggered rollout of a single intervention across clusters over time. That’s an implementation and logistics problem; a SMART is a treatment-sequencing problem, and the two can in principle be combined but answer different questions.

Where SMARTs Fit in Implementation and Practice Research

Because a SMART produces an actual decision rule rather than a single winning treatment, it pairs naturally with implementation science work on getting evidence into practice — CASRAI’s guides on choosing an evidence-based practice model and the ARCC model cover the organizational side of that translation. A SMART is how the underlying adaptive intervention gets built and evidenced in the first place, before an implementation team adapts it for rollout in a specific setting.

Frequently Asked Questions

What is a SMART design in research?

A Sequential Multiple Assignment Randomized Trial is a multi-stage trial design in which participants are randomized more than once, with later-stage randomization options and eligibility depending on how each participant responded to their earlier treatment. Its purpose is to build and compare complete, replicable treatment-sequencing rules (dynamic treatment regimes), not to declare a single winning treatment.

What is a tailoring variable?

A tailoring variable is the measured characteristic — most commonly response or non-response to an initial treatment — that determines which second-stage treatment options a participant is eligible for and randomized among. A well-chosen tailoring variable is one a clinician could realistically observe and act on outside a trial setting.

How is sample size calculated for a SMART trial?

It depends on the primary aim. A stage-specific comparison uses standard two-arm sample size logic. Comparing embedded adaptive interventions against each other requires accounting for participants shared across regimes, which changes the effective variance compared with a simple two-arm calculation. Second-stage comparisons are also constrained by the observed first-stage response rate, since only non-responders (typically) are eligible for re-randomization.

What’s the difference between a SMART design and an adaptive trial design?

“Adaptive trial design” usually refers to changing a trial’s statistical parameters — randomization ratios, sample size, or stopping rules — based on accumulating data within a single treatment comparison. A SMART instead re-randomizes individual participants to different subsequent treatments based on their own response, in order to build a treatment-sequencing rule. They can be combined, but they answer different design questions.