The Gold Standard of Experimental Design | Sequential Multiple Assignment Trials, Embedded Regimes and Dynamic Treatment Rules

1. Start with the sequence of decisions

Back to contents

Map the clinical, behavioural or educational decisions before drawing a trial diagram. Define the initial options, response assessment, tailoring variables, subsequent options, timing and final outcome. A dynamic treatment rule links observed history to a treatment recommendation at each decision point. The SMART should randomise the unresolved decisions that matter, not add stages merely because adaptation sounds personalised.

2. Distinguish paths from embedded regimes

Back to contents

A treatment path is the sequence one participant actually receives. An embedded dynamic treatment regime is a rule that specifies what would be recommended for every relevant response history. One observed path can be consistent with more than one regime when those regimes share early decisions. Count regimes from their full decision rules, not from visible branches alone, and state which contrasts are primary.

3. Define tailoring variables before observing outcomes

Back to contents

Response status, adherence, burden, side effects or early progress can guide later treatment, but each variable needs a measurement time, threshold and action. A data-dependent threshold chosen after looking at final outcomes weakens the causal question. Tailoring variables should be available when the decision is made, measured consistently and distinguished from outcomes that occur after the subsequent treatment begins.

4. Randomise only among available and ethical options

Back to contents

At each stage, specify who is eligible for re-randomisation and the probabilities for each option. Responders and non-responders may face different choices; some participants may continue without another randomisation. Conceal upcoming assignments and prevent clinicians or implementers from altering response classification to obtain a preferred option. Record refusals, unavailable treatments and departures from the designed path.

5. Choose primary aims and power the right contrast

Back to contents

A SMART can estimate first-stage effects, second-stage effects within response groups, comparisons of embedded regimes or components of an optimal rule. It is rarely powered equally for all. Sample size must account for response rates, re-randomisation probabilities, shared paths and expected outcomes. Prespecify the principal estimand and treat additional regime searches as exploratory unless the design supports them.

6. Worked example: two-stage support for depression

Back to contents

Participants begin with either guided digital therapy or brief problem-solving therapy. At eight weeks, non-responders are re-randomised to add coaching or switch treatment; responders are re-randomised to maintenance messages or monthly check-ins. One embedded regime starts digital therapy, adds coaching after non-response and uses messages after response. Estimating its mean outcome requires contributions from people whose observed assignments match the rule, with weights reflecting the randomisation probabilities. It is not the raw mean of one simple branch.

7. Analyse shared paths and repeated randomisation correctly

Back to contents

Regime estimators often use inverse-probability weights so participants representing rarer assignment sequences receive appropriate influence. Robust standard errors, weighted regression or generalised estimating equations can compare regime means and incorporate baseline covariates. The same participant may contribute to more than one compatible regime, creating dependence. Ordinary subgroup means or separate stage-wise tests do not reproduce the intended regime comparison.

8. Changed-context transfer check: adaptive mathematics support

Back to contents

A school trial begins students with worked-example practice or retrieval practice. After four weeks, students below a prespecified mastery threshold are re-randomised to tutoring or an intensified digital module; responders receive spacing schedules. The transferable logic is a sequence of decisions. The context adds classroom clustering, timetable constraints and measurement effects: mastery must be assessed before the second-stage assignment, and analysis must respect both cluster dependence and embedded regimes.

9. Know the common misconceptions

Back to contents

SMART does not mean the randomisation probabilities must adapt to accumulating trial results. It is not automatically a confirmatory test of a final personalised algorithm, and more stages do not ensure better care. Re-randomising only non-responders can still embed multiple regimes, while responders may or may not face another decision. Selecting the best-looking regime from many without correction exaggerates performance and should be validated in a separate trial.

10. Audit design, conduct and regime construction

Back to contents

Archive the decision map, eligibility rules, response thresholds, randomisation lists, probabilities, treatment availability, timing windows, deviations and analysis code. Show which participants are consistent with each embedded regime and how weights were formed. Test sensitivity to missing tailoring variables, threshold uncertainty, nonadherence and extreme weights. Report stage-specific evidence separately from full-regime comparisons.

11. Frequently asked questions

Back to contents

Is a SMART the same as an adaptive randomised trial? Not necessarily; SMART usually refers to planned sequential randomisations used to build adaptive interventions. Does every participant get re-randomised? Only those whose pathway includes another decision. Can the best regime be deployed immediately? It usually needs practical refinement and independent evaluation. Why can one participant represent several regimes? Several rules may prescribe the same treatment along that person’s observed history.

12. The gold standard aligns trial paths with decision rules

Back to contents

A credible SMART begins with real sequential decisions and uses randomisation to learn how later choices should depend on earlier response. The gold standard names embedded regimes, powers a clear estimand, analyses shared paths correctly and separates optimisation from final confirmation. The design is successful when it makes an adaptive strategy testable—not when a complicated flowchart merely looks personalised.