CAMT paper: analysis scripts, results and audit coding sheet

This record contains the analysis scripts, result files and audit coding sheet that accompany the manuscript "A clustered AUC maximum test (CAMT) for 'best of K' model comparisons on patient-clustered diagnostic data" (Gul, A.). The manuscript studies the claim that the best of K candidate models has a higher AUC than a baseline on a test set with several images or records per patient. It proposes CAMT, a patient-level test that corrects for within-patient clustering and for selection of the winner, and evaluates it by simulation, on five public datasets (BreakHis, Parkinson's voice recordings, PAD-UFES-20, PTB-XL and the Kermany retinal OCT archive) and in a pilot audit of published claims. Contents: scripts/: simulation studies, experiments on the five datasets, the audit scripts and the research version of the test. data/: result files and run logs (exact nulls, power, simulations, within-patient correlation, second-metric and replication studies, numerical checks of the theory). data/audit/: the audit of published claims, with the coded entries and the sentence supporting each value, the audit results and the screening records. data/deep/: test-set scores and training logs of the deep models on PAD-UFES-20. paper/honest_inference/: scripts that build the tables and figures, and the numbers behind each figure. The original datasets, extracted features and the full texts of the screened papers are not included; the README lists the sources and the scripts that download and prepare the data. The released implementation of the test is a separate record (doi:10.5281/zenodo.23771271; Python package camt-test on PyPI). Code is released under the MIT licence; result files and the coding sheet under CC BY 4.0.