This repo contains all of the course materials for the 1-day Stan course on 08/09/2017.
I'm Rob Trangucci, and I've been a Stan developer for the past 3 years. I've been funded by YouGov to do general research and software development for the Stan library. I focus on hierarchical modeling, Gaussian processes, and writing the matrix functions for the Stan math library.
Email: robert.trangucci@gmail.com GitHub: https://github.com/rtrangucci
Come to class with R installed on your machines and the following packages installed:
rstanrstanarmbayesplotdplyrggplot2
You can also just run this script to install the packages that you don't have.
if (!require("rstanarm")) install.packages("rstanarm")
if (!require("bayesplot")) install.packages("bayesplot")
These packages will install ggplot2 and dplyr if they're not
already installed. As for installing rstan, Mac and Linux users
with recent OS versions and Xcode installed (for Mac users) should
have no problem using install.packages("rstan"). If you do run
into problems, follow these instructions.
Windows users should follow these instructions
devtools- RStudio
- This will make viewing the Rmd files much easier
shinystan
I'll walk through a presentation from 9:30 to 10:30 that'll introduce everyone to Stan, and some bedrock principles needed to understand how Stan works. Then we'll walk through four RMarkdown documents, which you'll have access to, and we'll work through writing and fitting models to both real and generated data. I'll highlight key takeaways for each example and how they apply to running models over real data. When the class is over you'll have Stan and RStanArm code that'll generalize across datasets, along with all the data prep code and commentary in the RMarkdown docs.
| Time | Topic |
|---|---|
| 09:00 AM to 09:30 AM | Intro and coffee |
| 09:30 AM to 11:30 AM | Hierarchical models in Stan |
| 11:30 AM to 12:00 PM | Coffee and questions |
| 12:00 PM to 01:30 PM | Logistic regression, postrat in Stan and rstanarm |
| 01:30 PM to 02:00 PM | Lunch |
| 02:30 PM to 04:30 PM | Multinomial logit in Stan |
| 04:30 PM to 05:30 PM | Review and questions |
- Hierarchical models in Stan
-
One-way normal model
- How to write a Stan model
- What Stan does under the hood
- Sampler output and diagnostics
- What to do about divergences? (Non-centered parameterization)
- How to modify the target density manually for speed (only if there is time)
-
Linear regression with group-varying intercepts and coefficients
- Priors for group-varying intercepts and coefficients
- Priors for covariance matrices
- LKJ prior for correlation matrices
-
- Hierarchical logistic regression and post-stratification
- Stan control parameters (adapt-delta, max-treedepth)
- Poststrat in the generated quantities block
- QR decomposition for nicer posteriors
- Priors for hierarchical variance parameters
- Hierarchical logistic regression and post-strat
67DB
ification in Rstanarm
- Easier way to code hierarchical generalized linear models
- Poststrat using rstanarm
- Posterior predictive checks (PPC) in
bayesplot - Structured priors for interactions
- Hierarchical multinomial logit
- Model description
- Fake data generation for model checking
- Equivalent models
- Tricks for speed
- Priors for hierarchical variance parameters
- Fit to survey data
- Vectorize when possible
- Statistical efficiency and good model fit are the best antidotes to a slow model (reducing treedepth for each iteration will reduce the time it takes to fit a model by an order of magnitude)
- Use native Stan functions rather than writing your own in the Stan language
- Most of Stan's functions have been optimized a lot (see linear regression example)
- Can you use sampling independence to reduce the size of your dataset?
- The only expressions that matter are those that are in the kernel of the posterior. Do not worry about constants!
- Use sufficient statistics to reduce the size of the expression tree whenever possible (examples are IID normal observations at any level linear models)
- Explanation of
state_pres_votedata - Stan case study about why we fit hierarchical models (uses baseball data)
- GP presentation
- Intuition behind different covariance priors
- Information about ways to parameterize a covariance matrix prior and how to evaluate the properties of different priors. Lighter on math, lots of nice graphs to give a better idea as to what is happening at high dimensions.
- Conceptual introduction to Hamiltonian Monte Carlo
- Wonderful intro to why HMC is successful where other samplers fail and how and when HMC fails. Also lighter on math, lots of helpful graphics.
- Challenges sampling from Hierarchical models
- In-depth treatment of non-centered parameterization for hierarchical models. Lighter on math.