How the triggering worked — and why it matters

The study tested four different “alarm systems” for deciding when to send the prompt. Understanding the difference between them is essential to understanding the results.

Conditions 1 and 2 are the distress-based triggers. They watch the participant’s EMA survey answers and try to infer that the person needs support without asking them directly. To understand what that actually means, it is worth looking at what the EMA survey consisted of in practice.

Six times a day, at a random moment within a two-hour window, the participant’s phone buzzed with a notification from the m-Path app. When opened, it presented a short series of single questions. The four distress questions were each answered on an explicitly 7-point Likert scale:

That was essentially the entire survey. Four slider questions and one three-option question, delivered six times a day — sometimes while the participant was walking, eating, or half-distracted. Each question answered in seconds. And yet those numbers were being fed directly into algorithms that decided whether to send a mental health intervention.

This is worth sitting with. The fixed cutoff in Condition 1 set its threshold at ≥ 5 on that 7-point scale — which is actually the upper third of the scale, not the midpoint. A participant had to be reporting fairly high distress before it fired. That makes the poor timing appropriateness ratings for Condition 1 even more interesting: participants were scoring 5 or above on distress and still reporting that the intervention arrived at the wrong moment. High distress and readiness to seek support, it turns out, are not the same thing.

Condition 2 uses personalized Shewhart control charts — one per distress variable, per person. During the first seven days, the algorithm learns each participant’s personal baseline and natural variability for each of the four distress variables independently. It then calculates a personal upper control limit (UCL) for each one. From day 4 onward, the JITAI fires whenever any distress score exceeds that person’s own UCL — not a group-level threshold, but a deviation from their own normal. A person who routinely scores 6 on stress will not be triggered by a 6; a person who normally scores 3 will be. The formulas underpinning this personalization come from Montgomery’s foundational statistical quality control textbook, adapted for psychological EMA data by Schat et al. (2023).

Condition 3 takes a completely different approach. Rather than watching scores and inferring need, the EMA survey simply asks the support need question described above. If the answer is yes, the JITAI fires. The participant themselves is the trigger.

Condition 4 is the control: no JITAI is sent, even if the criteria for one of the other three conditions would have been met. This is what makes causal comparison possible.

The randomization logic — and why it was necessary

At every single one of the 6 daily EMA surveys, the system randomly assigned that moment to one of the four conditions with equal probability — 25% each. This happened regardless of how distressed the participant was or what time of day it was.

This design choice is the methodological backbone of the study. If the researchers had simply let each condition fire whenever its criterion was met, the conditions would never be fairly comparable. Condition 1 fired 221 times; Condition 3 fired only 46 times. Without randomization, any difference in outcomes between them could simply reflect the fact that Condition 3 moments were rarer and therefore more personally meaningful — not that the triggering strategy itself was better.

By randomizing the condition assignment first, the researchers broke that link. And Condition 4 in particular is what enables causal inference: because the control condition is randomly assigned, you occasionally get moments where a participant is highly distressed and would have met the criteria for an active condition — but no JITAI was sent. Those moments become the counterfactual: what would have happened if nothing had been done? Without random assignment you can never answer that question cleanly.

The baseline learning problem

Condition 2 required a learning phase before it could function properly — seven days of EMA data to estimate each participant’s upper control limit. But the researchers started the intervention phase on day 4, not day 8, because asking depressed individuals awaiting therapy to go two full weeks without any intervention felt ethically uncomfortable and would likely have hurt compliance. This means the Shewhart control chart UCLs were still being estimated during days 4 through 7 and only stabilized after day 7. It was a deliberate tradeoff: personalization accuracy sacrificed for participant welfare and retention.

Why six surveys a day — and what that frequency costs

The paper presents six EMA surveys per day as a given rather than a decision requiring justification, but there is an established methodological rationale behind it. Six is a commonly used frequency in EMA research on mood and daily life, and the specific implementation here matters: surveys were not sent at fixed times but at one random moment within each of six consecutive two-hour windows spanning 8 AM to 10 PM. This stratified random sampling approach ensures coverage across the whole waking day while preventing participants from anticipating the next survey — which matters because predictable timing would allow participants to prepare answers mentally in advance, undermining the “momentary” quality of the assessment.

The frequency of six is a tradeoff between three competing pressures. More surveys per day gives finer-grained emotional data and more decision points for the JITAI — especially important for Condition 2, whose algorithm needs sufficient observations to estimate personal baselines reliably. But more surveys also increase burden, fatigue, and the risk of careless responding or dropout. Six has become a pragmatic consensus in the field: enough to capture meaningful within-day emotional variation without overwhelming participants.

What the paper does not address is whether six was the right number for this specific population. Individuals with depressive symptoms may experience survey fatigue differently from healthy participants, and 126 surveys over 21 days is not a trivial ask. The fact that 80% of participants met the ≥70% compliance threshold is reassuring — but 20% did not, and one participant dropped out specifically citing lifestyle incompatibility with the survey schedule. There is also a subtler risk: with six decision points per day over three weeks, participants may gradually learn to recognize the triggering logic — noticing patterns in when the app responds and when it does not. That kind of learning could alter how honestly they answered distress questions, which would undermine the entire measurement system the JITAI depends on. The study does not resolve this tension.

What the support need condition revealed

The support need condition (Condition 3) outperformed both distress-based conditions on every user-rated measure: timing appropriateness (4.41 vs. 2.82–2.98 out of 7), helpfulness, and behavior adoption (43% vs. 22–23%). The average effect size across these outcomes was Cohen d = −0.69, which is meaningful.

The reason appears to be that Condition 3 captured not just vulnerability (being distressed) but receptivity — whether the participant was actually ready and willing to act. A participant can be stressed or lonely and have no interest in calling anyone: they might be in a meeting, the source of their stress might be a relationship conflict, or they simply might not feel ready. Distress and readiness to seek support are not the same thing, and the distress-based triggers were conflating the two.

However, there is an important counterpoint. When looking at actual subsequent support-seeking behavior rather than user ratings, Condition 1 (fixed cutoff) showed the largest numerical effect — suggesting that distress detection may catch moments of implicit need that the participant has not consciously recognized or articulated. So the two approaches may not be competing so much as complementary: one captures readiness, the other captures need the person has not yet named.

Who was in the study — and why it matters

The study was conducted at the University of Zurich in Switzerland. Of the 25 participants, 88% self-identified as women, 12% as men, and 84% as White. Mean age was 35.1 years. Participants were recruited through psychotherapist referrals and social media posts.

The paper acknowledges the gender skew as a limitation, noting that gender differences in social support preferences and help-seeking behavior may affect engagement and effectiveness. However it does not address the ethnic homogeneity as a limitation at all — which is itself notable.

For a feasibility study conducted in Zurich, the sample is arguably representative of the population the intervention was designed for. Switzerland has specific structural features relevant to the study’s premise: long therapy waiting times, high smartphone penetration, and a particular cultural context around help-seeking. Within that context the demographics are not surprising.

But the skew becomes a problem the moment anyone considers scaling or generalizing the intervention. Social support norms vary enormously across cultures — who you turn to, how you ask, whether seeking help carries stigma, and what kinds of support are culturally legible. An intervention built on the assumption that participants have an accessible, willing social network and feel comfortable activating it may work very differently in populations where those assumptions do not hold.

The honest bottom line

On the outcome that actually matters clinically — whether participants felt better — the JITAI produced no detectable effect. The effect sizes on distress reduction are worth examining directly against conventional benchmarks:

These are not numbers that failed to reach significance because the sample was small. They are genuinely tiny. The paper defends this by noting it was never designed to test efficacy — which is true. But the 29% behavior adoption rate gives reason to question the mechanism itself. Only one in three participants actually sought support after receiving the JITAI, even under the most favorable possible study conditions: motivated participants, financial compensation, researcher contact, and qualitative interviews framing the whole experience. If the core chain — prompt leads to support-seeking, support-seeking leads to feeling better — only activated partially even here, scaling this up to a real-world deployment raises serious questions.