Logistic Regression Behind the Scenes
summary
Compares logistic implementations in Python and SAS with concrete examples: scikit‑learn uses liblinear and always regularizes, statsmodels uses Newton/IRLS and gives p‑values, and SAS proc logistic defaults to IRLS with a relative‑gradient stop. Shows how different convergence criteria produce different coefficients (9 vs 11 iterations or outright breakage) while predictions can remain similar, exposes multicollinearity and quasi‑separation failures, links L1/L2 penalties to Bayesian priors, and shares runnable notebooks on GitHub.
AI-written summary · about
chapters (10)
- 0:00Logistic regression basics
- 4:37Risks of blind use
- 6:50Choosing a solver
- 10:00Optimization differences
- 13:08Solver variations demo
- 16:44Missing values & categorical data
- 17:45P-values & inference limits
- 19:53Regularization impact
- 24:27Model improvement tips
- 24:49Q&A
related talks
01h 10mLies damned lies and statistics in Python
44mA Modern Maximum-Likelihood Theory for High-Dimensional Logistic Regression
49mSo you think you know about linear regression ...
14mProbabilistic Programming in Python
43mIt's Not Magic: Explaining Classification Algorithms
31mUsing Simpson’s Paradox to Discover Interesting Patterns in...
02h 10mLasso and Ridge regression models in R: Introduction
25mAki720