The Gold Standard of Experimental Design | Micro-Randomised Trials, Proximal Outcomes and Time-Varying Treatment Effects

1. Begin with the decision point

Back to contents

Identify moments when an intervention could reasonably be delivered and a choice remains unresolved. A decision point may occur five times a day, after a sensor event or when an app is opened. Repetition alone does not make a trial micro-randomised. State the options, randomisation probabilities, information available before assignment and the action the future just-in-time rule might take.

2. Define availability before randomisation

Back to contents

A participant may be unavailable while driving, sleeping, in a meeting, recently contacted or at risk of overload. Availability must be measured before treatment assignment using prespecified rules. Randomising only when available targets an effect among available decision points. Changing the availability rule during analysis changes the estimand and can confuse safety constraints with lack of treatment effect.

3. Choose a proximal outcome and window

Back to contents

The proximal outcome should be affected soon enough to guide component optimisation: steps in the next 30 minutes, engagement by evening or craving over two hours. The window begins after randomisation and should avoid contamination from the next decision when possible. A distal goal such as monthly wellbeing can motivate the intervention, but repeated proximal effects do not mechanically add up to that distal outcome.

4. Keep randomisation probabilities interpretable

Back to contents

Probabilities may be constant or vary by time and prespecified context. They must remain known and bounded away from zero where comparisons are required. Very frequent treatment can create burden, while a low probability yields little information. Use design constraints such as maximum daily prompts transparently; do not let a hidden recommendation algorithm determine assignment and then call the result randomised evidence.

5. Plan moderation and time variation in advance

Back to contents

Effects may differ by study day, recent activity, location, stress or prior treatment. Prespecify a small set of moderators measured before the decision point. Smooth trends or basis functions can represent gradual change, while categorical periods can reveal nonlinearity. Searching hundreds of sensor features after the trial risks discovering unstable rules rather than dependable personalisation.

6. Worked example: a walking suggestion after inactivity

Back to contents

For six weeks, adults are randomised at up to five available times each day to receive or not receive a short walking suggestion. The proximal outcome is step count in the next 30 minutes. Suggestions increase steps early in the study, but the effect declines by week four and is near zero after another suggestion in the previous hour. The analysis reports an average excursion effect and prespecified moderation by study week and recent treatment. It recommends spacing rules for a later intervention, not a claim that the programme improved long-term health.

7. Use analysis that respects repeated decisions

Back to contents

Outcomes within a person are correlated, treatments can influence later availability and proximal windows may overlap. Weighted and centred estimating methods target causal excursion effects under sequential randomisation assumptions. Standard errors must account for repeated observations. Treating thousands of decision points as independent people produces false precision, while a simple participant-level treatment average discards the timing question.

8. Changed-context transfer check: revision prompts

Back to contents

A learning app randomises a retrieval-practice prompt when a student opens the app after a long gap. The proximal outcome is completion of a five-minute practice set. The transferable logic is repeated decisions with a near-term outcome. The context adds school schedules, parental device rules and learning quality: opening the set is not mastery, and repeated prompts may displace self-directed study. Longer assessments are needed before claiming educational benefit.

9. Know the common misconceptions

Back to contents

An MRT is not the same as a SMART: the former usually targets frequent component decisions, while the latter studies a small sequence of treatment stages. More randomisations do not eliminate missing sensor data or disengagement. A null average may hide helpful and harmful contexts, yet subgroup searches can also overfit. Proximal improvement can coexist with burden, habituation or no distal change. The design optimises a component; it does not validate a finished intervention by itself.

10. Audit every decision opportunity

Back to contents

Archive the decision schedule, availability state, randomisation probability, assigned option, delivery confirmation, engagement, proximal window, sensor provenance and reasons for suppression. Plot treatment and missingness by time of day and study week. Test alternate outcome windows, overlap rules and plausible missing-data assumptions. Report prompt burden, adverse events and how often the proposed rule would act in practice.

11. Frequently asked questions

Back to contents

Must every participant be randomised hundreds of times? No; frequency follows the decision problem. Can probabilities change? Yes, if the scheme is prespecified or otherwise known and analysis accounts for it. Is engagement a valid outcome? It can be a proximal target, but it is not the same as health or learning. Can an MRT prove a personalised rule works? It informs rule construction; the completed intervention usually needs separate evaluation.

12. The gold standard links each randomisation to one decision

Back to contents

Micro-randomisation is powerful because it places experiments inside the moments when support might be offered. The gold standard defines availability, outcome windows, probabilities and moderators before analysis, models dependence and monitors burden. A time-varying proximal effect becomes useful when it leads to a testable delivery rule while keeping long-term claims modest.