The research behind
Stilla
Every feature in Stilla is grounded in peer-reviewed research. Here are the datasets and studies that inform our approach.
Social media addiction
by the numbers
people worldwide are estimated to be addicted to social media.
University of Michigan, 2022.of Americans aged 18-22 self-report being addicted to social media.
Statista, 2019.of adolescent social media users show clinically significant anxiety levels.
Mendeley SM Addiction Dataset, N=258. STAI Y-1 ≥ 40.score above the addiction cutoff on the Bergen Social Media Addiction Scale.
Mendeley SM Addiction Dataset, N=258. BSMAS ≥ 19.Datasets we use
Stilla's prediction engine is trained on open, peer-reviewed datasets. No proprietary data. No user surveillance.
Qwantify ESM Dataset
636 participants, 26,710 real-time desire episodes logged via experience sampling. Includes timestamps, emotional states, physical states, social context, and resistance outcomes.
Hasenkamp et al. (2022). Frontiers in Psychology.
osf.io/sxfrx | DOI
Mendeley SM Addiction Dataset
258 adolescents (ages 14-22) assessed with validated clinical scales: Bergen Social Media Addiction Scale, State-Trait Anxiety Inventory, Rosenberg Self-Esteem Scale.
Adolescent Social Media EMA
102 adolescents, 15 days of ecological momentary assessment tracking social media experiences and momentary affect.
Sequeira et al. (2024). OSF.
osf.io/3k9yt
Datasets overview
| Dataset | N | Type | Used for |
|---|---|---|---|
| Qwantify ESM | 26,710 episodes | Experience sampling | Model training, predictions |
| Mendeley SM Addiction | 258 participants | Clinical scales (BSMAS, STAI) | Addiction prevalence, anxiety correlation |
| Adolescent SM EMA | 102 adolescents | EMA (15 days) | Youth affect patterns |
What actually predicts
cravings
Gradient Boosting classifier trained on 26,710 desire episodes. 28 features, 68.7% accuracy.
Source: Qwantify ESM dataset. Gradient Boosting feature importance. Time of day (<1%), weekday (0%), and social context (0%) are not shown.
Key findings from
our analysis
Rumination is the #1 predictor
Overthinking predicts cravings more than stress, hunger, time of day, or social context. Breaking the thought loop is the most effective intervention.
Stress drives desire
The second strongest predictor. When you're stressed, you reach for your phone. Stilla intervenes before you do.
Hunger matters
An unexpected finding: being hungry significantly increases craving probability. Eat first, decide later.
Desires peak when alone
Two-thirds of cravings happen in solitude. The urge isn't about the app — it's about what you're avoiding.
Time of day barely matters
Contrary to popular belief, cravings are nearly flat across the day. Your mental state matters far more than the clock.
Model accuracy
Our trained model predicts desire with 68.7% accuracy (vs 51.2% base rate) using 28 psychological, physical, and contextual features.
Key citations
Hofmann et al. (2012) — Everyday Temptations
7,827 desire episodes via experience sampling. Found media desires are MORE frequent and HARDER to resist than tobacco or alcohol cravings. Psychological Science, 23(6). DOI
Hunt et al. (2018) — Reducing Social Media Use
Controlled experiment at UPenn: limiting SM to 30 min/day led to significant reductions in loneliness and depression. Journal of Social and Clinical Psychology, 37(10). DOI
Lambert et al. (2022) — One Week SM Break
RCT: 1-week break reduced anxiety by 16% and depression by 25%. Cyberpsychology, Behavior, and Social Networking, 25(5). DOI
Allcott et al. (2020) — Facebook Deactivation
Stanford/NYU: 4-week FB deactivation gained 60 min/day, increased happiness, reduced polarization. American Economic Review, 110(3). DOI
Andreassen et al. (2012) — Bergen Scale
Development of the Bergen Social Media Addiction Scale — now the global standard. Psychological Reports, 110(2). DOI
Riehm et al. (2019) — SM Use and Mental Health
3+ hours/day doubles depression and anxiety risk. JAMA Psychiatry, 76(12). DOI
Twenge et al. (2018) — Teen Mental Health Decline
31% increase in teen suicide rate 2010-2015, correlating with smartphone adoption. Clinical Psychological Science, 6(1). DOI
Every in-app claim is traceable to its source. We believe in transparency, not vibes.
Why cravings pass — the science of the urge
Most cravings are not linear "the longer you wait, the worse it gets" — they rise, peak, and subside on their own, usually inside 5 to 20 minutes. Every intervention in Stilla is built around this curve.
Marlatt (1985) — The urge-peak-subside curve
Original observation that cravings behave like ocean waves rather than steady-state signals. Foundation of Relapse Prevention and all urge-surfing interventions. Relapse Prevention: Maintenance Strategies in the Treatment of Addictive Behaviors. Guilford Press.
Bowen et al. (2009) — Mindfulness-Based Relapse Prevention RCT
RCT (N=168) showed MBRP significantly reduced craving and substance use at 4-month follow-up vs. treatment-as-usual. Urge surfing is a core component. Substance Abuse, 30(4). DOI
Hasenkamp et al. (2022) — Qwantify ESM dataset
Experience-sampling study (N=636, 26,710 desire episodes) measured craving emotions in real time: anxious 28%, restless 23%, lonely 22%, content 25%. 65% of cravings occur when alone. Drives Stilla's trigger predictor. Frontiers in Psychology. DOI
Killingsworth & Gilbert (2010) — Mind wandering and unhappiness
Real-time sampling showed the mind wanders ~47% of waking hours, and mind-wandering predicts lower happiness — explaining why we reach for the phone when bored or idle. Science, 330(6006). DOI
Witkiewitz, Bowen et al. (2013) — Mindfulness reduces craving reactivity
Mindfulness training specifically weakens the link between depressive symptoms and craving — a trigger pathway central to behavioral addiction. Drug and Alcohol Dependence, 126(1-2). DOI
Tiffany & Wray (2012) — Craving as a clinical target
Review of why craving intensity matters independently of substance/behavior consumption. Justifies treating the urge itself, not just the behavior. Annals of the New York Academy of Sciences, 1248(1). DOI
Interventions — what's behind them
Stilla offers 8 in-app interventions for craving moments. Evidence strength varies — we label it honestly here so you can decide what to trust.
| Intervention | Origin | Evidence |
|---|---|---|
| 4-7-8 Breathing | Pranayama + Benson Relaxation Response | Strong |
| Box Breathing | Equal-ratio paced breathing | Strong |
| Urge Surfing | Mindfulness-Based Relapse Prevention (Bowen, Marlatt) | Strong |
| Body Scan | MBSR (Kabat-Zinn) | Strong |
| Challenge the Thought | CBT — cognitive restructuring | Strong |
| 5-4-3-2-1 Grounding | DBT / trauma grounding | Moderate |
| 10-Minute Delay | Urge-delay / stimulus control (Marlatt) | Moderate |
| 10-10-10 Rule | Pop-psychology prompt (Welch, 2009) | Face-valid only |
Citations
Zaccaro et al. (2018) — Paced breathing systematic review
How Breath-Control Can Change Your Life. Front. Hum. Neurosci. — Slow paced breathing (~6 bpm) reduces stress and increases vagal tone. DOI
Brown & Gerbarg (2005) — Sudarshan Kriya breathing
J. Altern. Complement. Med. — Yogic breathing reduces anxiety and depression. DOI
Bowen et al. — Mindfulness-Based Relapse Prevention
MBRP for substance use disorder. Underpins urge surfing. Site
Kabat-Zinn (1982) — MBSR body scan
An outpatient program in behavioral medicine for chronic pain patients based on mindfulness meditation. Gen. Hosp. Psychiatry. DOI
Beck (2011) — Cognitive Behavior Therapy
Core cognitive restructuring underlies "Challenge the Thought" and urge-delay. Basics and Beyond, Guilford Press.
Alvarsson, Wiens & Nilsson (2010) — Nature sound and stress
Int. J. Environ. Res. Public Health — Faster physiological stress recovery with nature sound. Relevant to ambient audio in Stilla's breathing interventions. DOI
Buxton et al. (2021) — Natural sounds and health
PNAS — Systematic synthesis of nature-sound benefits on stress and mood. DOI
"Face-valid only" means an intervention feels plausible and may help some users, but there is no published RCT supporting its specific effect on craving reduction. We keep the 10-10-10 Rule in the app because users report it's useful, not because it's clinically proven.