Where the field is heading: adaptive trials and cumulative method

Unit one set the bound on a deployment claim: a result describes a system, a setting and a method together, so it licenses less than it reports. Two developments try to escape it: an experiment that randomises inside itself, and a method that accumulates pooled evidence and reusable measurement.

Both are real, and both relocate the generalisation problem rather than dissolving it. This last of ten units asks whether the field has escaped the bound the first one set.

The micro-randomised trial makes the deployment its own experiment

A micro-randomised trial randomises at each decision point within each person. Klasnja and colleagues (2015) set out the design and its analysis: at every occasion on which the system may act, it randomises whether to act, so a study can estimate the causal effect of that component on the proximal outcome and model its moderation.

Bidargaddi and colleagues (2018) show what that does at scale. They randomised 1255 users of a workplace well-being app at one of six times a day, for 89 days. Each needed a smartphone and a data plan of their own. A tailored message raised self-monitoring in the next 24 hours by 3.9 per cent, risk ratio 1.039. That effect attenuated over the study, though not significantly.

HeartSteps is sharper. Klasnja and colleagues (2019) ran a six-week trial with 44 adults. A suggestion raised the following 30-minute step count by 14 per cent at p = .06. The early effect was 66 per cent and diminished.

What the design optimises, and the theory it needs

A just-in-time adaptive intervention is defined by its parts. Nahum-Shani and colleagues (2018) set them out. They are the states that matter, the decision points at which the system may act, the tailoring variables, and the decision rules that map the two onto support. They report that many such interventions were built with minimal use of evidence, theory or treatment guidelines. Randomisation makes a rule testable, not right: the adaptive trial turns the generalisation problem into a specification problem and hands it to theory.

Suh and colleagues show what specifying the parts looks like on a real deployment. Their sensing-powered system sent stress-reduction microinterventions to 43 employees over four weeks, delivering 1,585 interventions, and they fitted logistic models over 27 candidate factors to find which predicted momentary engagement (Suh et al., 2024). The authors treat the factors as candidates for a decision rule, not as one already fixed. What the case does not show is an effect: engagement with an intervention is not a change in stress.

The theory the designs need is dynamic, and the measurement under them is not validated

Riley and colleagues (2011) found the application of behaviour theories in mobile interventions uneven: most smoking and weight-loss studies reported a theoretical basis, while most studies of adherence and disease management did not. They judge current theories inadequate for adaptive interventions and propose dynamic feedback system theories built from longitudinal device data. Kwasnicka and colleagues (2016) identified 117 behaviour theories and found maintenance rarely attained. A deployment that measures adoption has measured the easier half.

An adaptive system acts on states inferred from sensors, so the validity of the sensor sits inside the argument. Trifan and colleagues (2019) screened 7404 publications and included 118 studies of passive smartphone sensing. They conclude not that sensing fails, but that the outcomes it produces lack correlation with clinical knowledge. A rule that fires on an unvalidated signal has moved the claim from the intervention to the measurement.

Cumulative method: pooling gives a number, and maintenance has a different signature

Pooling is the field's other answer. Weisel and colleagues (2019) meta-analysed 19 randomised trials with 3681 participants, restricted to standalone apps, and found a significant effect for depression at Hedges g = 0.33; they conclude that standalone apps cannot be recommended on that evidence. Linardon and colleagues (2019) meta-analysed 66 trials of app-supported interventions, finding effects for depression at g = 0.28 and stress at g = 0.35, larger where a cognitive behavioural app came with professional guidance and reminders. The same group has since superseded both figures, with 176 trials returning g = 0.28 for depression and g = 0.26 for generalised anxiety (Linardon et al., 2024).

The two agree that apps help, and disagree about how much and for what. The disagreement follows from scope. One asks about a product used alone, and the other about a product embedded in support. The number a reader is handed is therefore a function of an inclusion rule. Neither reports the effect over a period long enough for maintenance to fail.

Hekler and colleagues (2013) make the time scale concrete. They analysed daily activity reports from 34 participants in a 16-week peer-delivered intervention for older adults and followed them to 18 months, when 10 were maintaining 150 minutes a week of moderate to vigorous activity and 24 were not. Maintainers took longer to reach a steady state and nonmaintainers began to drop at week 10. A trial that ends when the intervention ends measures initiation.

Contested question: two escapes from the setting, and neither has closed it?

On the first dispute, Klasnja and colleagues (2015) and Nahum-Shani and colleagues (2018) hold that the design is the cure: it replaces the single average a one-off deployment estimates with causal estimates of proximal effects and their moderation, shown working at scale and in context (Bidargaddi et al., 2018; Klasnja et al., 2019). The opposing position is that it relocates, because the estimate is conditional on the tailoring variables and the decision rule, the theory able to specify those is not yet in hand (Riley et al., 2011; Kwasnicka et al., 2016), and the effect decays with exposure, so a proximal effect is not the effect of the intervention over time.

On the second, Weisel and colleagues (2019) and Linardon and colleagues (2024) hold that pooling is cumulative method. Their agreement on direction is the strongest quantitative result this unit can report: apps beat controls on depression in both. The estimate has narrowed as the evidence has grown, from g = 0.33 over 19 standalone-app trials to g = 0.28 over 176.

The objection to pooling is that a pooled trial estimate is the wrong shape of evidence for a deployment. Linardon and colleagues find larger effects where guidance and reminders accompanied the app, so the variable that may matter most in a service appears only as a moderator (Linardon et al., 2024).

The adaptive position rests on two micro-randomised trials at scale (Bidargaddi et al., 2018; Klasnja et al., 2019) and on a reinforcement learning analysis (Park et al., 2026). The pooling position rests on a series of meta-analyses from 19 trials to 176 (Weisel et al., 2019; Linardon et al., 2024). What would settle the first, and has not been run, is a micro-randomised trial with tailoring variables from a pre-specified dynamic model, a sensor validated against an external criterion, and a run long enough for maintenance patterns to appear (Hekler et al., 2013). What would settle the second is an aggregate reporting follow-up length, human support and engagement definition as moderators. The strongest objection is that a policy learned from one trial's logs is fitted to that trial, so it relocates the generalisation problem rather than solving it.

Where the area is heading

The adaptive strand has arrived at policy learning. Park and colleagues (2026) trained a two-stage offline reinforcement learning procedure on data from a six-week micro-randomised trial with 190 participants, choosing when and what to deliver; the learned policy beat the behaviour policy and a rule-based policy under fitted Q evaluation, across six off-policy estimators. They call the result hypothesis-generating and require prospective validation, and Trifan and colleagues (2019) find sensing outcomes uncorrelated with clinical knowledge. A field that generates hypotheses from its own logs faster than it can test them has not escaped the bound of unit one, because the checking still has to happen elsewhere.

The sources this unit rests on

Each note states what the cited document did and found. Titles are in British spelling, and identifiers were checked against registered metadata on 2 November 2026.

• Klasnja, P., Hekler, E. B., Shiffman, S., Boruvka, A., Almirall, D., Tewari, A. and Murphy, S. A. (2015). Microrandomised trials: an experimental design for developing just-in-time adaptive interventions. Health Psychology. doi:10.1037/hea0000305 Introduces the micro-randomised trial, which randomises each decision point within each person, and the analyses that estimate proximal effects and their time-varying moderation. It is this unit's method for making a deployment an experiment.
• Bidargaddi, N., Almirall, D., Murphy, S., Nahum-Shani, I., Kovalcik, M., Pituch, T., Maaieh, H. and Strecher, V. (2018). To prompt or not to prompt? A microrandomised trial of time-varying push notifications to increase proximal engagement with a mobile health app. JMIR mHealth and uHealth. doi:10.2196/10123 Ran a micro-randomised trial inside a workplace well-being app for 89 days with 1255 users, randomising at one of six daily times whether to send a tailored message. The message raised self-monitoring in the next 24 hours by 3.9 per cent, risk ratio 1.039, and the effect attenuated over the study without reaching significance.
• Klasnja, P., Smith, S., Seewald, N. J., Lee, A., Hall, K., Luers, B., Hekler, E. B. and Murphy, S. A. (2019). Efficacy of contextually tailored suggestions for physical activity: a micro-randomised optimisation trial of HeartSteps. Annals of Behavioural Medicine. doi:10.1093/abm/kay067 A six-week micro-randomised trial with 44 adults in which a suggestion raised the following 30-minute step count by 14 per cent at p = .06, with a 66 per cent early effect that diminished and antisedentary suggestions undetected. It shows that the estimate depends on when in the study it is taken.
• Nahum-Shani, I., Smith, S. N., Spring, B. J., Collins, L. M., Witkiewitz, K., Tewari, A. and Murphy, S. A. (2018). Just-in-time adaptive interventions in mobile health: key components and design principles for ongoing behaviour support. Annals of Behavioural Medicine. doi:10.1007/s12160-016-9830-8 Defines the intervention through its states, decision points, tailoring variables and decision rules, and reports that many such systems were built with minimal use of evidence, theory or treatment guidelines. It states the specification problem that the adaptive trial hands to theory.
• Riley, W. T., Rivera, D. E., Atienza, A. A., Nilsen, W., Allison, S. M. and Mermelstein, R. (2011). Health behaviour models in the age of mobile interventions: are our theories up to the task? Translational Behavioural Medicine. doi:10.1007/s13142-011-0021-7 Reviews how behaviour theories were applied in mobile interventions and finds a basis reported by most smoking and weight-loss studies but missing from most adherence and disease-management studies. It judges current theories inadequate for adaptive interventions and proposes dynamic models built from longitudinal device data.
• Kwasnicka, D., Dombrowski, S. U., White, M. and Sniehotta, F. (2016). Theoretical explanations for maintenance of behaviour change: a systematic review of behaviour theories. Health Psychology Review. doi:10.1080/17437199.2016.1151372 Identified 117 behaviour theories, of which 100 met inclusion criteria, and synthesised five themes: motives, self-regulation, resources, habits, and environmental and social influences. It finds maintenance rarely attained and explained differently from initiation.
• Trifan, A., Oliveira, M. and Oliveira, J. L. (2019). Passive sensing of health outcomes through smartphones: systematic review of current solutions and possible limitations. JMIR mHealth and uHealth. doi:10.2196/12649 Screened 7404 publications and included 118 studies of passive smartphone sensing, dominated by physical activity and mental health. It finds the outcomes produced lack correlation with clinical knowledge, so the measurement layer of an adaptive system is not yet validated.
• Weisel, K. K., Fuhrmann, L. M., Berking, M., Baumeister, H., Cuijpers, P. and Ebert, D. D. (2019). Standalone smartphone apps for mental health: a systematic review and meta-analysis. npj Digital Medicine. doi:10.1038/s41746-019-0188-8 Meta-analysed 19 randomised trials with 3681 participants restricted to standalone apps, finding significant effects for depression at g = 0.33 and smoking at g = 0.39 and none for anxiety, suicidal ideation, self-injury or alcohol use. The authors conclude standalone apps cannot be recommended.
• Linardon, J., Cuijpers, P., Carlbring, P., Messer, M. and Fuller-Tyszkiewicz, M. (2019). The efficacy of app-supported smartphone interventions for mental health problems: a meta-analysis of randomised controlled trials. World Psychiatry. doi:10.1002/wps.20673 Meta-analysed 66 randomised trials of app-supported interventions, finding effects for depression at g = 0.28, generalised anxiety at g = 0.30 and stress at g = 0.35. Effects were larger where a cognitive behavioural app came with professional guidance, making support part of the effect.
• Linardon, J., Torous, J., Firth, J., Cuijpers, P., Messer, M. and Fuller-Tyszkiewicz, M. (2024). Current evidence on the efficacy of mental health smartphone apps for symptoms of depression and anxiety: a meta-analysis of 176 randomised controlled trials. World Psychiatry. doi:10.1002/wps.21183 Meta-analysed 176 randomised trials, the successor to the 2019 pair. Apps held a small effect on depression at g = 0.28 (number needed to treat 11.5) and generalised anxiety at g = 0.26.
• Hekler, E. B., Buman, M. P., Poothakandiyil, N., Rivera, D. E., Dzierzewski, J. M., Aiken Morgan, A., McCrae, C. S., Roberts, B. L., Marsiske, M. and Giacobbi, P. R. (2013). Exploring behavioural markers of long-term physical activity maintenance. Health Education and Behaviour. doi:10.1177/1090198113496787 Analysed daily activity reports from 34 participants in a 16-week peer-delivered intervention for older adults, followed to 18 months, when 10 were maintaining at least 150 minutes a week of moderate to vigorous activity and 24 were not. Maintainers took longer to reach a steady state and nonmaintainers began to drop at week 10.
• Park, C., Heo, J. i., Park, J. Y., Jeon, M., Do, G., Park, S., Jung, D. and Lim, M. H. (2026). Data-driven personalisation of just-in-time adaptive mental health intervention via two-stage reinforcement learning approach. npj Digital Medicine. doi:10.1038/s41746-026-03086-3 Trains a two-stage offline reinforcement learning procedure on data from a six-week micro-randomised trial with 190 participants to choose when and what to deliver, and the learned policy beat the behaviour and rule-based policies across six estimators. The authors frame the result as hypothesis-generating and require prospective validation.
• Suh, J. et al. (2024). Toward tailoring just-in-time adaptive intervention systems for workplace stress reduction: exploratory analysis of intervention implementation. doi:10.2196/48974 Deployed a sensing-powered just-in-time adaptive intervention to 43 employees over four weeks, delivering 1,585 microinterventions. Logistic models over 27 candidate factors identified which trait, state and intervention factors predicted momentary engagement, as an exploratory basis for a tailoring rule.