How often the problem occurred depended on how it was defined

A problem statement usually carries a number: how many users have this problem, or how often it happens. A measurement study published on arXiv on 8 September 2026, and in the proceedings of MobileHCI 2026, shows how much that number can depend on a choice the statement rarely mentions: what exactly counts as having the problem.

The study

Jian Zheng and Eun Kyoung Choe at the University of Maryland studied problematic smartphone use, a field where the definitions are still evolving. Some research defines it by behaviour, such as long sessions or opening certain apps. Some defines it by what people say they want. The authors' argument is that both produce interventions people don't welcome, and that the problem is better defined as the gap between the two: wanting to stop and carrying on anyway.

To measure that, they built an Android app that did two things at once:

• Asked. During phone sessions it showed a one-question prompt: should I stop or continue using my phone now, on a 1 to 10 scale. Up to 16 prompts a day.
• Logged. It recorded unlocks, locks, app use and taps, scrolls and typing, but not content.

37 participants used it for at least two weeks, producing 5,772 sessions with both an answer and a measured outcome. The researchers then interviewed 19 of them, showing each about six sessions where what they said and what they did diverged.

What they found

The same data gave three very different prevalence figures. Using the authors' example thresholds:

Definition of "problematic"	Share of sampled sessions
Behaviour only: kept using the phone for more than 5 more minutes	51.4%
Intention only: answered below 5.5, leaning towards "stop"	28.3%
The gap: leaned towards "stop" and kept going for more than 5 minutes	14.9%

The behaviour-only definition flags people using the phone for reasons they consider justified, such as finishing an assignment. The intention-only definition flags people who would have stopped on their own a moment later. The authors argue both would trigger interventions that threaten users' sense of autonomy, the reaction psychologists call reactance.

What people said barely predicted what they did. Across all sessions the correlation between stated intention and remaining time on the phone was r = .07. It was stronger where people leaned towards stopping (r = .30 for answers 1 to 5) and close to nothing where they leaned towards continuing (r = .09).

"Continue" meant "let me finish", not "for a while". Interviewees said they chose continue when they were in the middle of a task, and often stopped soon after it was done. A stated intention was task-bound, not time-bound.

The gap was concentrated. It was larger at night than in the morning, larger in social, media, game and browser apps than in messaging, and larger when the previous minute involved heavy scrolling. Female participants had larger gaps than male participants, though the authors say demographic findings from 37 people should be read with caution.

People misjudged their own usage. Many interviewees said session length was what they used to decide whether to stop, yet the quantitative data did not show that pattern. The authors suggest people underestimated how long they had been on the phone. One participant, shown her own data:

I can't believe I used a half hour longer than what I intended.

The caveats, from the paper

• A small, self-selected sample. 37 people, all on Android, all in the US, all using their phones at least three hours a day and interested in using them less. Five more were excluded for answering "continue" to over 95% of prompts, which the authors say narrows further who the results apply to.
• The cut-offs are arbitrary. The authors use that word for the 5.5 and 5-minute thresholds behind the three percentages. Different cut-offs would give different shares.
• Sampled sessions were not typical sessions. Sessions that received a prompt averaged about 17 minutes; across all 31,349 logged sessions the average was about 8.
• One question per prompt. The authors acknowledge single-item measures, careless responses they could not reliably detect, and a question-behaviour effect, where being asked may itself change what people do.
• The gap is not the whole problem. Cases where someone should not have picked up the phone at all are outside this definition.
• No intervention was tested. The claim that gap-triggered interventions will be better received is a design proposal for future work, not a result.

What this changes for problem statements

The paper is about phones, but its first result is about every problem statement that cites a frequency.

"How often" is a function of the definition. 51.4% and 14.9% describe the same people and the same two weeks. A problem statement that says "half of sessions are problematic" without saying problematic by what rule has hidden the most important choice in the sentence. Write the definition next to the number, including the threshold, and if a reasonable alternative definition gives a very different figure, say so. That is the "how often" and "how you know" half of a problem statement that survives review, and it is the same discipline as writing down counting rules for a metric.

Whose problem includes when. The gap was concentrated at night and in a handful of app categories. "Users struggle to stop scrolling" is a weaker statement than one naming the situation, because the situation is what the design will have to detect.

The definition becomes the trigger. In this study the problem definition is literally the rule that would decide when an intervention fires. Define the problem by behaviour and you interrupt people finishing their homework. That cost belongs in the problem statement before anyone designs against it.

Neither the survey nor the log is enough on its own. Stated intention and measured behaviour correlated at r = .07. Teams that size a problem only from what users say, or only from what the dashboard shows, are measuring different things and calling both of them the problem. That is a version of the question every reviewer should ask: who asked for this, and how do you know?

FAQ

What did it measure? Stated intention to stop and actual continued use, across 5,772 sessions from 37 Android users.

How much did the definition matter? 51.4% of sessions by behaviour, 28.3% by intention, 14.9% by the gap.

Did words match actions? Barely: r = .07 overall.

Main limits? A small, self-selected sample, arbitrary thresholds, and no intervention tested.