Beyond Simple
Correlation.
Traditional health apps track correlations. If you sleep poorly on nights you log stress, they may imply stress is the cause. But personal wellness patterns are often obscured by hidden factors. Causal uses statistical models to estimate which logged factors may influence your biometrics.
1. The Flaw of Manual Diaries
Health journals rely on manual, retroactive logging. This introduces three severe biases that degrade data quality:
- Recall Bias: Trying to remember how you felt 8 hours ago is highly inaccurate. People tend to overlay their current mood onto the past.
- Selection Bias: You are far more likely to open a health app and log data when you are feeling exceptionally good or exceptionally bad, completely missing the baseline.
- Logging Fatigue: Detailed text diaries create enough friction that many people stop logging before useful patterns can emerge.
Causal solves this through Ecological Momentary Assessment (EMA). By sending randomized, 3-second prompts during waking hours, we capture clean, low-bias health inputs in real-time.
2. Testing Confounders with a Bounded Agent
In statistics, a confounder is an unmeasured variable that influences both the supposed cause and effect. For instance, cold weather can contribute to both high heating bills and dry skin, making bills and dry skin highly correlated, but heating bills do not cause dry skin.
In health tracking, stress is a major confounder. High stress might lead to drinking more coffee and sleeping poorly. A standard correlation analysis may blame caffeine while missing stress. Causal uses an agent to register a small set of hypotheses and call deterministic context, multivariable, robustness, and sensitivity tools:
The model chooses tests; server code owns the data, statistics, evidence gates, and publication decision.
Exclusions never delete observations. A context-filtered analysis must name the reason and agree in direction with an all-data sensitivity run. Multiple hypotheses are corrected together, and all attempted tests remain auditable. The resulting estimates are personal wellness signals, not medical proof or diagnosis.
3. Graduated, Versioned Evidence
New personal findings begin as emerging only after minimum coverage, effect-size, false-discovery, and stability gates. Supported findings require longer coverage, bootstrap, placebo, and full-data checks. As new observations arrive, a finding can strengthen, remain stable, weaken, be contradicted, or retire. Causal never publishes a result merely because it was the strongest one tested.