Model criticism with ArviZ
summary
ArviZ demonstrates concrete model-criticism tools: posterior predictive sampling (about 2,000 posterior predictive realizations in examples), marginal overlays and plot_ppc_pit ECDF diagnostics, calibration curves via plot_ppc_pava for binary and discrete outcomes, and a rotogram for count frequencies. Examples show radon and rugby data, under/overdispersion and bias detectable, category-level failures, and leave‑one‑out style checks from pointwise log‑likelihood approximations. Notebooks and a Pixie environment branch are provided and the demonstrations generally work while exposing when models misfit.
AI-written summary · about
chapters (12)
- 0:00Model criticism with ArviZ
- 1:54What is ArviZ and PPC
- 6:11Posterior predictive fundamentals
- 11:36Posterior vs observed data
- 21:31Visualizing model fit with PPC
- 32:00Hands-on continuous models
- 40:09Binary data model criticism
- 48:58Calibration plots for discrete data
- 54:37Hands-on discrete models
- 1:02:33Regression visualization in ArviZ
- 1:11:24LOO cross-validation
- 1:18:16Q&A
related talks
36mHow to review a model
01h 33mUnderstanding and diagnosing your machine-learning models
01h 18mAkiTutorial
17mMaking sense of marginal effects
15mInterpretable Machine Learning with R
43mProbabilistic Programming in the Real World
34mInterpreting Machine Learning Models: Why and How!
01h 24mInpsect and try to interpret your scikit-learn machine-learning models