Micro-randomized trials
A micro-randomized trial repeatedly assigns brief treatment options at decision points to learn their near-term effects when people are available.
Why this level?
Participants or clusters are randomly allocated to concurrent comparison groups. NHMRC 2009 sourceTreatment is randomized at eligible decision points. This supports a causal comparison of brief treatment options under the trial schedule. It does not automatically identify a long-term policy effect, and repeated decisions are not independent participants.
Fit
When to use it, and when not to
Use it when
- You need to decide when a prompt or other brief treatment component helps.
- You can record availability, treatment chances and a near-term outcome at each decision point.
Not when
- You compare whole courses of care with later treatment stages. Consider a SMART. Read: Sequential multiple assignment randomized trials (SMART)
- You need one person to compare reversible treatments over longer periods. Consider an N-of-1 trial. Read: N-of-1 trial
Before you start
What you need
- Decision times and availabilitySpecify when randomization occurs and when an intervention is safe and feasible.
- The near-term outcome windowoutcomeTypeDefine the response after each decision. Separate it from the final clinical outcome.
- Randomization chances and participant historiesassignmentrepeatedMeasuresRecord the probability used at each available decision, including any dependence on past treatment.
- Availability is determined before current treatment assignment.How to check: Record why a decision point is unavailable. Do not code unavailable moments as randomized no-treatment moments.
- Both compared options have positive known allocation chances.How to check: Check the probabilities given the recorded history. Assignment must follow those probabilities.
- Repeated outcomes within a person can be dependent.How to check: Use participant-level variance estimates. Allow for overlapping windows and carryover in the target effect.
- A missing outcome is distinct from being unavailable for treatment.How to check: State the missingness assumption and method. Check sensitivity when outcomes may be missing because of unobserved activity.
Plain language
How it works
At each available decision, chance assigns a brief treatment option. Measure the outcome over a prespecified near-term window. Pooling these decisions can reveal when a component helps. Weighted and centered least squares is one approach for suitable outcomes and target effects. Keep the known treatment chances and account for dependence within people.
A proximal causal excursion effect compares a brief departure from the trial treatment schedule among available decision points. Its meaning depends on that schedule and the outcome window. Unlike a conventional crossover, the target is repeated brief decisions; unlike an N-of-1 trial, the study can pool many people.
Worked example
A worked example
- Setting
- A walking app randomizes a prompt or no prompt when sending one is safe.
- Question
- Does a prompt increase walking during the next hour?
- Approach
- For illustration, estimate a constant near-term contrast across available decisions, allowing for repeated observations within each person.
- Result
- Walking occurs after 60 of 200 prompt decisions and 40 of 200 no-prompt decisions. The difference is 10 per 100; the risk ratio is 1.5.
- What it means
- These counts illustrate a contrast, not its uncertainty or a long-term policy effect.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Many decision points are not many independent people. A near-term effect among available decision points does not establish the long-term effect of a treatment policy.
- Treating each phone notification as an independent participant.Instead: Keep repeated decisions grouped by person for estimation and uncertainty.
- Claiming a prompt effect proves the best long-term treatment policy.Instead: A causal excursion effect changes treatment briefly against the trial history. A long-term policy effect needs a separate target and analysis.
- Ignoring prompt burden and later availability.Instead: Track fatigue, carryover and engagement. Earlier treatment may change whether later treatment can be given.
Inside Mallard
How Mallard uses this
Plans do not link to this design yet. Mallard has no sample-size form for micro-randomized trials. A plan sized by simulation gets no headline sample size; Mallard does not run that simulation.
design.micro_randomized, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Microrandomized trials: An experimental design for developing just-in-time adaptive interventions (opens in a new tab) · Health Psychol · doi 10.1037/hea0000305
- Assessing Time-Varying Causal Effect Moderation in Mobile Health (opens in a new tab) · J Am Stat Assoc · doi 10.1080/01621459.2017.1305274
- Estimating time-varying causal excursion effect in mobile health with binary outcomes (opens in a new tab) · Biometrika · doi 10.1093/biomet/asaa070