Source-linked AI summary
To Stop or Not to Stop: Exploring the Intention-Behavior Gaps in Smartphone Usage
Jian Zheng, Eun Kyoung Choe
TL;DR
Problematic smartphone usage has been measured through intention or behavior, but either alone may misalign interventions with users’ needs. This study operationalizes PSU as the intention-behavior gap and models it alongside intention and behavior from phone-use data. IBG was associated with demographic and contextual factors, and prediction accuracy varied by target and data scope.
Problem
Prior smartphone-usage research measured problematic use through intention or behavior, while smartphone IBG had not been examined and prior IBG work typically relied on self-reported behavior at week- or month-scale.
Source: §2.1.2 · p. 3, §2.2 · p. 4Method
The researchers collected ESM reports and phone behavior from 37 Android users over at least two weeks, calculated IBG, and modeled intention, behavior, and IBG.
Source: §1 · p. 2, §3.4 · p. 7Results
IBG was associated with gender, time, app, input interactions, physical activity, and session length; intention was best predicted by a personal model, while behavior and IBG were best predicted by combined personal-global models.
Source: §1 · p. 2, §6 · pp. 20–21Takeaways & Limitations
The authors suggest that predicted IBG could inform future intervention tools that adapt their timing and intensity.
Source: §1 · p. 2, §6 · p. 21Participants were Android users interested in reducing excessive phone use, so the findings may not generalize to people without that interest or perception.
Source: §5.5 · pp. 19–20
Abstract
from arXiv · showhide
As smartphones become integral to daily life, researchers have sought to identify when the use becomes problematic. Previous studies have operationalized problematic smartphone usage (PSU) from either an intention or a behavior perspective. Both risk delivering interventions not welcomed by users. We propose a novel approach to operationalizing PSU as the intention-behavior gap (IBG). We collected self-reported data on intentions to stop phone usage, alongside usage behavior data, from 37 participants over two weeks. We calculated IBG, examined effects of demographic and contextual variables, and developed machine learning models to predict IBG in real time. We found that IBG was explained by gender, time, app, and input interactions, among other factors. Intention was predicted most accurately with only personal data, whereas behavior and IBG were predicted most accurately with both personal and global data. Our findings can inform the design of future intervention tools optimized for timing and adaptive intensity.
1 Introduction
The paper frames problematic smartphone usage as the gap between wanting to stop and continuing, seeking triggers better aligned with users’ intentions and behavior. It studies that gap and its predictors, then evaluates real-time prediction using data from 37 Android users.
1 Introduction
- Behavior-triggered interventions can disrupt needed phone use, while intention-triggered interventions may be unnecessary when users can stop unaided.The latter may also threaten users’ perceived autonomy and cause psychological reactance.Source: §1 · p. 1
- The study operationalizes problematic smartphone usage as an intention-behavior gap: users want to stop but continue using their phones.The authors argue interventions may be more appropriate in these “should but cannot stop” cases than when triggered by behavior or intention alone.Source: §1 · p. 2
- The study collected ESM intention reports and phone behavior from 37 Android users over at least two weeks, then interviewed 19 participants about observed gaps.It examined demographic and contextual factors and the feasibility of real-time machine-learning prediction.Source: §1 · p. 2
- IBG was associated with factors including gender, time, app in use, input interaction, physical activity, and session length.The paper reports these as factors that influence or explain IBG, not as causal effects.Source: §1 · p. 2
- The work contributes an IBG-based operationalization of PSU, machine-learning benchmarks, and a dataset of 5,772 usage sessions from 37 users.The dataset includes intention, behavior, contextual factors, and demographics and is available on OSF.Source: §1 · p. 2
2 Related Work
Prior work measures problematic smartphone usage at different levels and typically intervenes using behavioral rules, creating a mismatch with users’ subjective intentions. The paper motivates measuring session-level intention-behavior gaps and predicting intention, behavior, and PSU continuously.
2.1 Problematic Smartphone Usage
- Individual-level PSU questionnaires can identify users’ risk, but a high score does not establish that each phone-use instance is problematic or guide interventions.Problematic use can also depend on the activity or context, including use while driving, walking, or during face-to-face interaction.Source: §2.1.1 · p. 2, §2.1 · p. 2
2.1.1 Operationalization of Problematic Smartphone Usage.
- Session- and app-episode-level studies distinguish undesirable use from acceptable use, but have continued to rely on users’ self-reports.Sessions run between device unlock and lock; app episodes run between launching and closing an app.Source: §2.1.1 · pp. 2–3
2.1.2 Intervention against Problematic Smartphone Usage.
- Behavior-based triggers may interrupt justified use, while intention-based triggers may prompt users who can stop unaided and threaten autonomy.The proposed gap-based trigger targets cases where users intend to stop but continue.Source: §2.1.2 · p. 3
2.2 Intention-Behavior Gap
- Smartphone IBG had not been examined, and prior IBG studies typically used self-reported behavior and week- or month-scale measurements.The paper applies the intention-behavior matrix at the usage-session level, focusing on inclined abstainer sessions where users want to stop but do not.Source: §2.2 · p. 4
2.3 Predicting Smartphone Usage
- Prior work predicted app choice or subjective usage ratings, but had not targeted real-time session-length prediction or jointly predicted intention and behavior.Existing intention labels were usually binary; continuous prediction is proposed to support interventions with varying intensity.Source: §2.3 · p. 4
3 Methods
The study used AwareMind to collect phone-use intentions and behavior from Android users, then combined quantitative analysis with interviews about observed gaps. Its methods define the measured variables, describe recruitment and study procedures, and specify exclusions and coding approaches.
3 Methods
- AwareMind collected usage intentions through experience-sampling surveys alongside smartphone behavior in a study of 37 Android users lasting at least two weeks.Nineteen participants were subsequently interviewed about gaps observed in their data.Source: §3 · p. 5
3.1 Design of AwareMind
- The ESM survey asked whether participants should continue using their phone, rated from 1 (stop) to 10 (continue), with up to 16 responses per day.Unanswered surveys auto-dismissed after 30 seconds, and participants could dismiss them themselves.Source: §3.1.1 · p. 5
3.1.1 Intention Data Collection.
- AwareMind recorded screen locking and unlocking, app use, and taps, scrolls, and text edits without recording app content or typed text.For text edits, it recorded only the number of characters typed.Source: §3.1.2 · p. 5
3.2 Procedure
- Participants installed the app and completed a tutorial before data collection; after two weeks, those below 200 reports could stop or continue for up to one more week.Those who had reached 200 reports were told to stop ESM collection.Source: §3.2 · p. 6
3.1.2 Behavior Data Collection.
- Interviews used reconstructed sessions with a usage table and plots of app history, per-minute interactions, and detected physical activity to discuss intention-behavior mismatches.Nineteen selected participants discussed about six cases each; interviews lasted about 30 minutes.Source: §3.2 · p. 6, Figure 2 · p. 6
3.3 Participants
- The final sample comprised 37 participants after six technical dropouts and five exclusions for answering “continue” on over 95% of ESM questions.Recruitment targeted U.S.-resident adult Android users who spent at least three hours daily on their phones and were interested in reducing excessive use.Source: §3.3 · p. 7
3.4 Data Analysis
- The analysis defined intention as self-reported willingness to continue, behavior as remaining session length, and IBG as behavior minus intention.Intention was measured on a Likert scale and behavior as a continuous ratio variable; 202 inactive sessions were excluded using app-use and interaction data.Source: §3.4 · p. 7
- Interview transcripts were coded into topic summaries using deductive and inductive approaches, prioritizing participants’ explicit statements.The qualitative analysis was intended to help interpret quantitative findings, including why participants continued after reporting willingness to stop.Source: §3.4 · p. 7
4 Results
The results section reports the available ESM and session data, then lays out analyses of intention, behavior, and IBG. It uses mixed-effects models to examine associated factors and machine-learning models to predict the three variables in real time.
4 Results
- 5,772 valid ESM reports were collected, alongside 31,349 usage sessions overall; the analyses generally use the 5,772 sessions with ESM questions.Participants averaged 156 valid reports each; 591 surveys were manually dismissed and 1,017 auto-dismissed.Source: §4 · p. 8
- The analysis first describes intention, behavior, and IBG, then uses linear mixed-effects models for demographic and contextual factors and machine learning for real-time prediction.The prediction targets are intention, behavior, and IBG.Source: §4 · p. 8
4.1 Describing the Intention-Behavior Gap
Intentions skewed toward continuing, while remaining phone use varied widely and was only weakly related to intention. The section combines these dimensions into a continuous intention-behavior gap (IBG) to distinguish overuse from justified use.
4.1.1 The Intention to Continue.
- 75% of intention ratings were 10, with a median of 8 (M=7.1, SD=2.8), showing a strong skew toward continuing.Participants described choosing stop for physical tasks or bedtime, while neutral ratings could reflect downtime that felt wasteful but not clearly worth stopping.Source: §4.1.1 · p. 8, Figure 3 · p. 9
4.1.2 The Behavior of Continuing.
- After ESM prompts, remaining session length averaged 898 seconds (SD=1,811) and had a median of 307 seconds, with a right-skewed distribution.The full-session data were also right-skewed; their median was 109 seconds, and the authors caution that ESM sessions are not representative of all sessions.Source: §4.1.2 · pp. 8–9
- Intention and remaining time were weakly related overall (r= .07), with a stronger positive correlation at intention values 1–5 (r= .30) than at 6–10 (r= .09).Source: §4.1.3 · p. 9
4.1.3 The Intention-Behavior Gap.
- The continuous IBG uses Winsorized, log10-transformed remaining time and standardized intention and behavior before calculating their difference.The authors describe this as a straightforward initial calculation, not the only or best way to calculate IBG.Source: §4.1.3 · pp. 9–10
- 14.9% of observations met both intention <5.5 and remaining session length >5 minutes, versus 51.4% behavior-only and 28.3% intention-only triggers.The combined thresholds aim to avoid false alarms from either dimension alone; the continuous IBG represents inconsistency along a continuum.Source: §4.1.3 · p. 9, Figure 4 · p. 10
- Falling-short gaps—intending to continue but stopping shortly—often reflected task completion or external interruption.The analysis focuses on overuse cases, where participants intend to stop but continue.Source: §4.1.3 · pp. 10–11
4.2 Explaining the Intention-Behavior Gap
Three mixed-effects models examined intention, behavior, and IBG using demographic and contextual predictors. The findings link larger gaps to gender, nighttime, app category, and input interactions, with additional patterns involving session length and movement.
4.2.1 The Linear Mixed-Effects Models.
- The IBG model had conditional R^2 = .535 and marginal R^2 = .325, similar to the intention model and above the behavior model.The behavior model had conditional R^2 = .401 and marginal R^2 = .263; the intention model had conditional R^2 = .583 and marginal R^2 = .354.Source: §4.2.1 · p. 11, Table 2 · p. 13
4.2 Explaining the Intention-Behavior Gap
- Gender, time, app category, and input interactions were associated with IBG, alongside session-length and movement patterns.The models included repeated observations nested within participants and standardized intention, behavior, and IBG as outcomes.Source: §4.2.2 · p. 11, §4.2.3 · p. 12, §4.2.4 · p. 14
4.2.2 Effects of Demographic Factors.
- Female participants had larger IBGs than male participants, associated with both lower intention and longer remaining session length.Gender also had a marginally significant effect on intention; age affected behavior, while full-time employment affected intention.Source: §4.2.2 · p. 11
4.2.3 Effects of Temporal Factors.
- Nighttime was associated with larger IBGs than morning, through lower intentions and longer remaining session length.Interviews described phone use around sleep and after waking, including difficulty stopping after checking news or notifications.Source: §4.2.3 · pp. 12–13, Figure 7 · p. 14
4.2.4 Effects of the App Category.
- Browser, game, social, and media use was associated with larger IBGs than message-app use, but those categories also included justified breaks and relaxation.One participant contrasted needing a phone for schoolwork with feeling she should stop when watching entertainment videos.Source: §4.2.4 · p. 14
- Longer phone use before the prompt was associated with longer continued use afterward, despite participants saying session duration informed their stop-or-continue decisions.The authors suggest uneven session lengths or inaccurate duration estimates could help explain the pattern.Source: §4.2.5 · p. 14
4.2.6 Effects of Behavioral Factors.
- High scrolling intensity was associated with larger IBGs, whereas high text-editing intensity was associated with smaller IBGs.Participants did not mention input interactions in interviews.Source: §4.2.6 · pp. 14–15
- Moving was associated with shorter remaining session length and smaller gaps than staying still.Possible explanations included safety awareness and interruptions from other people.Source: §4.2.6 · p. 15
4.3 Predicting the Intention-Behavior Gap
The study compared global, personal, and combined random-forest models for predicting intention, behavior, and IBG. Personal data gave the best intention predictions, while combined personal and global data performed best for behavior and IBG.
4.3. Predicting the Intention-Behavior Gap
- The models used 35 features and random-forest regression, selected because the dataset was too small for more complex models such as neural networks.Features covered timing, interactions, app use, physical activity, cumulative screen time, historical usage, and demographics.Source: §4.3.1 · p. 15
4.3.1 Model Design.
- Global models predict held-out participants without personal data, personal models use only an individual's accumulated data, and combined models add personal data to data from others.Personal and combined predictions began after the first week; the combined model weighted personal observations 10x.Source: §4.3.1 · p. 15
4.3.2 Model Performance.
- Personal data predicted intention best, while combined personal and global data performed best for behavior and IBG.The performance comparison is summarized in Table 3.Source: Table 3 · p. 16, §4.3.2 · p. 16
- Intention prediction had an MAE of 0.528 in z-score, corresponding to 1.514 on a 1–10 scale; behavior prediction had an MAE of 0.660 in z-score, corresponding to 10.3 minutes.These values describe intention and behavior prediction, respectively.Source: §4.3.2 · p. 16
5 Discussion
The discussion connects session-level intention–behavior gaps to self-regulation and argues for interventions triggered by predicted gaps with intensity scaled to their magnitude. It also notes limits in the study’s sample, measures, and coverage of problematic smartphone usage.
5.1 Connecting Intention-Behavior Gap to Theories of Self-Regulation
- Self-regulation theories frame IBG as a discrepancy between intention and behavior, explored here within usage sessions rather than across weeks or months.The discussion relates gaps to competing conscious and habitual systems, short- versus long-term goals, and the role of prompts.Source: §5.1 · pp. 16–17
5.2 Comparison with Previous Studies
- Larger IBGs were associated with female participants, several app categories, and sleep hours; scrolling increased with PSU while text edits decreased.The authors also suggest sleep problems may contribute to PSU, while noting that this interpretation draws on interviews.Source: §5.2 · p. 17
- The models’ predictive performance was broadly comparable with related work, although the study was the first to predict IBG.After binarizing intention at 5.5, the random forest reached F1 = .66 versus .67 in [63]; remaining session length prediction had MAE = 10.3 minutes versus 13.8 minutes for surgical case duration.Source: §5.2 · pp. 17–18
5.3 Measurement of Problematic Smartphone Usage with Experience Sampling Method
- Binary PSU labels sacrifice variance information, while continuous measures preserve gradations; the authors recommend an explicit “I don’t know” ESM option.Interviews revealed a neutral zone distinct from selecting the middle of a scale.Source: §5.3 · p. 18
5.4 Intervention according to the Intention-Behavior Gap in Smartphone Usage
- IBG-based triggering would deliver interventions when users want to stop but cannot, rather than relying on intention or behavior alone.The authors propose replacing an existing trigger, such as unlocking the screen, with predicted IBG reaching a threshold.Source: §5.4 · pp. 18–19
- The authors propose continuously scaling intervention strength with IBG magnitude instead of applying a binary intervention.A linear rule maps IBG to intensity, but its parameters require calibration to intervention goals and formats; user fine-tuning may support autonomy.Source: §5.4 · p. 19
5.5 Limitations and Future Directions
- Generalizability is constrained by the small, self-selected Android sample and recruitment of users interested in reducing use; activity recognition could not reliably distinguish drivers from passengers.The sample also excluded five participants who consistently reported intending to continue.Source: §5.5 · pp. 19–20
- Single-item ESM may miss intention variance and affect behavior, while session-based IBG excludes some PSU forms and prediction errors remain unavoidable.The framework omits “should-not-start” cases and use while the screen remains locked; the authors discuss confidence-based selective prediction as a possible mitigation.Source: §5.5 · p. 20
6 Conclusion
The conclusion presents intention–behavior gap as an operationalization of problematic smartphone usage and reports the factors associated with it. It also summarizes the models’ prediction patterns and identifies real-time adaptive interventions as a next step.
6 Conclusion
- IBG was associated with gender, time of day, app category, input interactions, physical activity, and elapsed session length.Scrolling and text editing showed opposite effects.Source: §6 · p. 20
- Intention was best predicted by a personal model, while behavior and IBG were best predicted by combined personal–global models.The personal model achieved intention MAE 1.514 on a 10-point scale; the combined model achieved behavior MAE 10 minutes.Source: §6 · p. 21
- The authors plan to build an intervention system that adapts intensity in real time using predicted IBG.Source: §6 · p. 21
A.1 Interview protocol
The interview protocol begins with consent and background questions, then uses participants’ study data to discuss phone-use intentions and behavior. It closes with questions about their study experience and the interview.
A.1 Interview protocol
- The interviewer introduces the study, assures participants of confidentiality, and requests permission to record.Source: §A.1 · p. 25
- Background questions cover participants’ phone habits, usage purposes, and aspects they want to control or reduce.Source: §A.1 · p. 25
- Data-engagement prompts ask participants to interpret app and response patterns, recall intention–behavior divergences, and describe their study experience.The protocol also asks whether answering surveys affected their intentions or behavior.Source: §A.1 · p. 26