New thesis examines goodness-of-fit tests for sparse logistic regression models

A new thesis explores goodness-of-fit tests and calibration algorithms for logistic regression models, particularly when dealing with sparse data. The research compares approximately 30 statistical tests and machine-learning calibration methods, finding that tests like the GiViTI calibration test, McCullagh, Osius-Rojek, le Cessie, and Stute-Zhu offer a good balance of power and accuracy. The study emphasizes that relying solely on formal tests is insufficient, and visual diagnostics like calibration plots are crucial for detecting model deficiencies. An application to the Low Birth Weight dataset revealed that many tests struggle with real-world data complexity, underscoring the need for a combination of statistical tests and visual inspection for robust model assessment. AI

IMPACT This research provides advanced statistical methods for evaluating machine learning models, crucial for ensuring reliability in AI applications.

RANK_REASON The item is an academic paper published on arXiv detailing statistical methodology. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• Associazione GiViTI
• chi-square test
• deviance test
• Ebrahim Khaled Ebrahim
• le Cessie
• logistic regression model
• Low Birth Weight dataset
• McCullagh
• Osius-Rojek
• Stute-Zhu

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →