Overview

Not all adaptation means learning, and more variation does not prove more benefit. This chapter compares fixed state-sensitive rules, stochastic personalizing JITAIs, and human or social layers by what they can learn, what they ask of people, and what uncertainty or risk they add. Fixed rules favor transparency and auditability; stochastic rules create variation for learning but require safe exploration; human and social support address slower mechanisms with greater resource and coordination demands. The choice depends on fit, evidence, agency, safety, and meaningful outcomes.

Learning outcomes

A fixed JITAI can be state-sensitive without learning; its transparency helps when the signal, mechanism, and option are sufficiently understood.
Stochastic delivery creates variation needed to estimate responsivity, but exploration can withhold helpful support or deliver a less suitable option.
Human or social layers address slower mechanisms and coordination needs, but add time, cost, agency, privacy, and safety constraints.

Keywords

JITAIsPersonalizationFixed rulesStochasticityExplorationAgency

When More Adaptation Is Not More Personal

A JITAI can adapt its delivery without learning, and it can learn without becoming more helpful. This chapter compares fixed decision rules, stochastic personalizing rules, and human or social additions by the criteria that matter to the person receiving support: what the design can learn, what it asks the person to do, how much uncertainty it introduces, and whether a meaningful outcome improves. The goal is a conditional judgment about added adaptation, not a ranking in which the most automated system automatically wins.

A fixed rule can still be state-sensitive

It is useful to separate two meanings of adaptation. In a standard JITAI, a decision rule is fixed, but its output can change as the person’s state changes. If a rule says “deliver a prompt when no brushing has been detected in the previous 12 hours,” it may deliver a prompt at one decision point and withhold it at another. The state-dependent option changes; the rule does not learn from those decisions.

This arrangement has a virtue that is easy to overlook: transparency. Once the tailoring variable, threshold, and options are specified, the same conditions should produce the same decision. A researcher can ask whether the rule was followed. A participant can understand, at least in principle, why support appeared or did not. The design may be entirely appropriate when the mechanism is well understood, the relevant signal is available, and the intervention option has been tested under the conditions in which it will be used.

A fixed rule also places a limit on what the intervention can accommodate. It does not automatically discover that a person responds differently after repeated exposure, that a particular option works only in one context, or that the same threshold has different meaning for different people. If the environment or the person changes in a way the rule does not represent, the JITAI continues to apply its original logic. That can be a strength when stability and auditability matter; it can be a weakness when responsivity is genuinely uncertain.

The limitation is not solved merely by adding more sensors. More data can improve a tailoring variable, but a fixed rule still needs a prespecified relationship between the data and the option. A system can know more about the person and still make the same poorly matched decision. The question is not whether the rule is simple or sophisticated. It is whether its assumptions about state, mechanism, option, and receptivity are good enough for the purpose.

Why variation matters for learning

A personalizing JITAI, or pJITAI, adds another layer of adaptation. Artificial-intelligence methods or control-systems approaches can use data accumulated as people interact with the intervention to update the decision rules themselves. Instead of always mapping a given state to one fixed option, a stochastic rule assigns probabilities to different options. The resulting variation makes it possible to compare what happens under the alternatives in the same kinds of states.

The logic is easiest to see with the oral-care example. Suppose a fixed rule delivers a prompt every time the person has not brushed in the previous 12 hours. Across a group, every eligible decision point may lead to a prompt, and no comparable decision point under the same condition receives no prompt. The system can describe what happened after prompting, but it has little within-condition variation with which to estimate whether prompting was better than withholding.

A stochastic rule might instead deliver the prompt with probability 0.7 when no brushing was detected and with probability 0.4 when brushing was detected. The probabilities do not mean that the system is uncertain in a vague sense; they create a planned distribution of options. Over repeated decision points, some otherwise similar states receive a prompt and some do not.

Published by DeepMox

FAQ

How does a pJITAI differ from a standard JITAI?

A pJITAI differs from a standard JITAI by updating its decision rules from accumulated interaction data, whereas a standard rule remains fixed while its state-dependent outputs change. Stochastic options can create the variation needed to estimate responsivity.

Does individualized output prove personalization?

No. Individualized output describes what a person receives, while pJITAI personalization concerns improving the decision rules through continuing adaptation; neither claim proves better outcomes.

What costs does stochastic exploration add?

Stochastic exploration adds uncertainty, participant burden, and safety costs because the system may withhold an option that might help or deliver a less suitable one while learning. The acceptable scope depends on the stakes and reversibility of the decision.

Previous13 minJITAIs: What the Evidence Shows

Next10 minJITAIs: Match the Trial to the Decision

JITAIs: Fixed Rules or Personalization

When More Adaptation Is Not More Personal

A fixed rule can still be state-sensitive

Why variation matters for learning

FAQ