Remove the impressive leakage
Self-harm is removed as a predictor because it overlaps the target phenomenon.
Trade-off: Cross-validated performance drops sharply, but the remaining signal is more honest and useful.
All projects
Data product · Honest modeling
Mental Health Viz — Reproducible Global Analytics
A reproducible mental-health analytics project that makes model limitations as visible as its findings.
01
The problem
Global mental-health datasets can create persuasive but misleading models when target leakage, source differences and uncertainty are hidden behind a polished visualization.
Constraints
- WHO, IHME and World Bank data use different definitions and coverage.
- The experience must be educational and explicitly not clinical advice.
- Predictions need uncertainty and leakage checks, not a single confident number.
02
My contribution
- Built the data pipeline, exploratory analysis, baseline and advanced models, calibrated predictor and bilingual presentation layer.
- Documented the leakage discovery and rebuilt the model on independent covariates rather than hiding the performance drop.
03
Technical decisions
Show calibrated uncertainty
Split-conformal intervals accompany the predictor instead of presenting a point estimate alone.
Trade-off: Intervals are less visually simple, but communicate the limits of the estimate.
04
Verifiable outcomes
- The live bilingual site exposes the predictor, dashboards, methods and limitations together.
- The documented analysis shows how removing the circular feature changes validation performance.
- WHO/IHME agreement and model uncertainty remain inspectable rather than reduced to marketing claims.
05
Product evidence
a4aece9e
a4aece9e
a4aece9eCountry predictor
Educational estimator with an explicit non-clinical warning and uncertainty context.
View evidence06
What I learned
- A weaker honest model is more credible than a strong circular one.
- Methodology and warnings belong in the primary interface, not in fine print.
07
What I would change
- Design the cross-source data contract before building successive analysis versions.
- Add automatic data-card generation for every published dataset snapshot.