STUDY GUIDE: Building Health Behavior Models to Guide JITAIs
Building Health Behavior Models to Guide JITAIs
Nahum-Shani, Hekler & Spruijt-Metz · Health Psychology, 2015, Vol. 34, Suppl. · pp. 1209–1219
Existing health behavior models are too static to guide JITAI development. This paper provides a pragmatic framework — organized around three areas of emphasis — to translate static models into dynamic ones capable of informing when, where, and how to deliver just-in-time adaptive support.
What Is the Paper About?
The Technology–Science Gap
Mobile technology now makes it technically possible to deliver highly personalized, real-time health interventions. The problem is that the scientific models used to design behavioral interventions have not kept pace — they mostly capture stable risk factors (e.g., diagnosis, demographics) rather than the rapidly changing states (e.g., momentary stress, location, mood) that JITAIs need to target.
What the Authors Propose
A pragmatic framework built around three sequential areas of questions: (1) defining the problem, (2) defining what “just-in-time” means in that context, and (3) formulating the adaptation strategy. The framework organizes existing evidence into a useful scientific model and exposes the gaps that still need to be filled.
Illustrative Running Example
Throughout the paper the authors use “Joe,” a full-time employee aged 30–65 who engages in hazardous drinking (defined as >14 drinks/week or >4/occasion for men). The goal of the hypothetical JITAI is to transition Joe to non-hazardous drinking patterns within a year. All framework concepts are illustrated with reference to Joe’s case.
Two Pillars: JIT and Adaptation
Just-in-Time (JIT)
Provide the right type of support, precisely when needed, and only when needed — minimizing waste and fitting the real-life setting. Support should be offered when the person is (a) in a state of vulnerability or opportunity, and (b) receptive.
Adaptation (Dynamic Individualization)
Using time-varying information — not just stable baseline data — to decide what intervention option to offer at each moment. This distinguishes JITAIs from standard adaptive interventions that tailor only at baseline.
Vulnerability / Opportunity
A transient tendency to experience adverse outcomes (vulnerability) or a transient window for positive change (opportunity). Both are shaped by the interplay of stable factors (traits, diagnoses) and dynamic ones (momentary stress, social context).
Receptivity
The person’s transient tendency to receive, process, and use the support provided. Influenced by both stable (personality, working memory) and dynamic factors (current emotions, distractors, ethical considerations such as driving).
Key Terms
Click any card to reveal the definition. Cards are grouped by concept category.
Core Concepts
JITAI Design Elements
Methods & Research Design
The Pragmatic Framework
Three sequential areas of emphasis, each answered through a series of guiding questions.
Defining the Problem
- Who are you trying to help? Identify the target population and their key attributes (e.g., employed adults may be unable to use time-intensive interventions at work). Use formative methods such as user-centered design.
- What is the distal outcome? Articulate a clinically meaningful long-term goal (e.g., transition to non-hazardous drinking within one year).
- What is the temporal progression of key factors? Build a dynamic model using timescales to describe how the process leading to the distal outcome unfolds over time. This is the most novel and essential step.
- What are the contender proximal outcomes? Select malleable, intermediate targets — starting at the shortest timescale to prevent snowballing effects. Multiple proximal outcomes may be targeted simultaneously.
Timescale Hierarchy (Joe’s Drinking Example)
Each level uses the unit below it as its variable. Factors accumulate upward; chronic outcomes can feed back downward.
Defining JIT in Context
- What factors mark a state of vulnerability/opportunity? For each candidate proximal outcome, identify the conditions that predict adverse outcomes or enable positive change. These inform the selection of decision points.
- What intervention options can affect proximal outcomes JIT? Identify options feasible at the relevant timescale. If the timescale is too fast (e.g., minute-level physiological reactivity), JIT delivery may not be feasible — consider other options or a different proximal outcome.
- What factors mark a state of unreceptivity? Identify conditions under which the person should not receive support — e.g., driving a car (safety), being in a meeting (intrusiveness), high intervention fatigue. Always include “provide nothing” as a valid option.
Formulating the Adaptation Strategy
- What are the tailoring variables? Select variables that mark both vulnerability and receptivity. Consider the feasibility of measuring them at the required timescale using sensors, EMA, or digital footprints.
- For each level of the tailoring variable, which option is likely best? Think through the expected effect of each option at different levels. Existing evidence often only establishes associations, not the cutpoints needed to distinguish when one option beats another.
- What decision rules can operationalize effective adaptation? Synthesize the above into a coherent rule set linking tailoring variable levels to intervention options. Accept that initial rules will be imperfect — plan for iterative improvement using real-time data (machine learning, control systems approaches).
Four Key Limitations Identified
1. Insufficient Temporal Evidence
- Existing evidence rarely specifies how factors are ordered and related over time.
- Intensive longitudinal study designs (EMA, wearable sensors) are needed.
- Modern analytic methods — machine learning, control systems engineering — are required to handle the resulting high-volume data.
2. Lack of Cutpoint Precision
- Even well-known associations (distress predicts poor coping) do not tell us the cutpoint on the tailoring variable that differentiates when intervention A beats intervention B.
- Micro-randomized trials (MRTs) are a promising design to generate this causal, cutpoint-level evidence.
3. Limited Research on Receptivity Markers
- Behavioral science has devoted little attention to identifying what marks a state of receptivity.
- Human-computer interaction research offers some insights but much more work is needed.
- Ethical considerations (privacy, safety, welfare) must be reviewed for every intervention option and population.
4. Decision Rules Cannot Be Fully Specified A Priori
- Between-person and within-person variability in vulnerability/receptivity means initial rules will always be incomplete.
- Adaptive, real-time methods are needed to “train” decision rules as the person progresses through the JITAI.
Quiz
5 questions covering definitions, framework logic, and application.
Current health behavior models are too static to guide JITAI development. A pragmatic framework — organized around defining the problem, defining what “just-in-time” means, and formulating an adaptation strategy — can bridge the gap by translating static models into dynamic ones and surfacing the empirical questions still to be answered.
-
Time Is the Missing Dimension in Existing Models
Most behavioral science models focus on stable risk and protective factors. JITAIs require models that specify the temporal nature of each factor — when does vulnerability arise, how quickly does it pass, at what timescale does an intervention need to fire? Without this, developers cannot specify decision points or decision rules.
-
Two Conditions Must Be Met Simultaneously: Vulnerability AND Receptivity
JIT support should only be delivered when a person is both (a) in a state of vulnerability/opportunity and (b) receptive to support. Delivering support when only one condition is met wastes resources, creates burden, and may cause harm — hence “provide nothing” is always a legitimate intervention option.
-
Short Timescales Feed into Long Ones — and Vice Versa
The fluid analogy makes explicit that minute-level hassles accumulate into hourly distress, then daily stressors, then chronic stress — which can in turn erode stable traits like coping capacity. JITAIs targeting micro-moments can therefore have cascading effects at longer timescales, and this dynamic logic must be modelled explicitly.
-
The Field Needs Intensive Longitudinal Data and New Analytic Methods
Standard between-subjects RCTs were designed for stable, slowly-changing outcomes. JITAI science requires intensive longitudinal designs (EMA, sensors), and analytic tools from data science, machine learning, and control systems engineering to extract decision-relevant knowledge from high-frequency data streams.
-
Micro-Randomized Trials (MRTs) Are the Right Causal Tool
To estimate which intervention option produces the best proximal outcome under which level of a tailoring variable — and at which cutpoint — the MRT design randomizes participants to options at each decision point, generating the causal evidence standard RCTs cannot provide for dynamic decision rules.
-
Decision Rules Must Be Adaptive, Not Fixed
Because vulnerability and receptivity vary both across people and within a person over time, pre-specified decision rules will always be imperfect. Methods that can update decision rules in real time — learning from the person’s own data as they use the JITAI — are essential for achieving optimal personalization at scale.