AkiTutorial
summary
Covers concrete methods and failures: uses PSIS-LOO via the loo package, reports pareto k > 0.7 as a badness threshold and offers loo+ exact re-computation for those points; shows rstanarm examples, a beta‑blocker binomial case (674 vs 680, 39 vs 22 deaths, posterior p(treatment helps) ≈ 99%) and roach counts where Poisson fails but negative binomial fixes fit; introduces projpred with horseshoe for model projection/selection that yields much sparser predictors and compares numeric ELPD differences and standard errors.
AI-written summary · about
chapters (9)
- 0:00Start
- 2:53Information Criteria
- 21:50Alternative Model
- 22:30Null Model
- 38:55Cross Validation Comparison for Coverage
- 42:06Negative Binomial Model
- 48:10Marginal Distribution
- 50:39Accuracy of Cross-Validation
- 54:38Hierarchical Model
related talks
01h 19mModel criticism with ArviZ
42mIs Your LLM Evaluation Missing the Point
46mEnhancing Agent Reliability Through Evaluation-Driven Development
18mEfficient Exposure of Partial Failure Bugs in Distributed Systems with Inferred...
32mStatistics for Engineers
42mDo Your Homework! Writing tests for Data Science and Stochastic Code
01h 00mThe Elephant in your Dataset: Addressing Bias in Machine Learning
44mPractical LLM Evals: Down-to-Earth Tooling for Everyone