Focus
Calibration under missing data
I distinguish ranking performance from probability error and calibration.
Research paper
Read the research report.
I use this website for an interactive view of selected results. In the report, I document the study design, simulated missingness, model comparisons, probability metrics, figures, limitations, references, and reproducibility records.
PDF report
Target-Set and All-Feature Missingness Stress Tests for Two Tabular Clinical Benchmarks
Use the report to examine the statistical methods, numeric results, uncertainty, and limits of the benchmark.
What it covers
- 01Scope-matched MCAR, MAR, and MNAR experiments plus a separate MCAR-all stress test
- 02Logistic regression, random forest, and gradient boosting comparisons
- 03ROC-AUC, accuracy, Brier score, ECE, and reliability diagrams
- 04Descriptive intervals, limitations, reproducibility, and claim-linked artifacts
Evidence
Saved benchmark outputs
I link numeric claims to saved metrics, predictions, figures, and summary tables.
Scope
Research, not diagnosis
I study model behavior on public benchmarks. I do not provide medical advice.