Master's Thesis · MS in Education Data Science, Stanford
Who Responds to Digital Friction? Heterogeneous Effects of Wellbeing Prompts on Adolescent Smartphone Use
Methods: mixed-effects models, behavioral trajectory clustering, longitudinal causal inference on 1M+ smartphone events.
Behavioral clusters are groups of users with similar phone-usage patterns, found via unsupervised clustering on baseline data. Intervention prompts are the three friction nudges described below.
Overview
Adolescents spend a growing share of waking hours on smartphones, and a growing class of digital wellbeing interventions tries to curb problematic use. Evidence on whether these interventions actually shift behavior, and for whom, is mixed.
The interactive figures below are drawn from a Danish randomized controlled trial (RCT, N≈264, adolescents aged 13–18) of three friction-based prompts embedded in One Sec, a digital wellbeing app that inserts a brief pause before the user can open a targeted app. A parallel German RCT (N≈165) is being analyzed separately and is not pooled with these results.
The app asks the user to pause and reflect on why they're about to open the app.
The user is prompted to set an intention for what they'll do in the app before entering.
A short timed delay (the app's default friction) before access is granted.
Who falls into which cluster
Each behavioral cluster has a distinct psychological fingerprint. The radar plots mean baseline trait scores per cluster, standardized so 0 is the sample average. Click a cluster in the legend to highlight it; the others dim but stay visible for comparison.
A day in the life
Each participant below is a representative user from one of the four behavioral clusters introduced above. Pick a participant and toggle between a typical weekday and weekend to see when they use which apps. Identifiers are stripped, and each user illustrates the cluster's pattern rather than a specific person.
Cluster comparison
Mean minutes-per-hour by behavioral cluster, averaged across all users and days post-baseline. Clusters differ less in how much they use phones than in when they do. Click any cluster to jump to its representative participant above.
App spotlight
Per-app daily-use averages across the post-baseline period. Pick an app to see how its use varies across behavioral clusters and intervention conditions.
By behavioral cluster
By condition
Intervention effects over time
Mean daily opens (the number of times a participant launched any tracked app on their phone in a 24-hour window) by condition, week by week. Shaded bands show 95% confidence intervals. Switch cohorts to compare the Danish and German RCTs; click a condition to highlight it while the others stay visible.
Methods
- Data: longitudinal smartphone event logs (~350K events, Danish sample) paired with survey-based psychological measures.
- Modeling: mixed-effects / GEE models for average intervention effects; behavioral trajectory clustering for heterogeneity.
- Validation: bootstrap stability for clusters, FDR correction across subgroup tests, attrition-bound sensitivity analyses.
Materials
- Code — GitHub