jmobrien/ghost

Reproducible code for model projections paper

Code for model summary and projections paper

"Model interpretation through lower-dimensional posterior summarization"
by Spencer Woody, Carlos M. Carvalho, and Jared S. Murray

Contents

• toy-sigmoid-example.R contains code for the toy example in Section XX which illustrates how our summarization approach can approximate the partial effects in nonparametric regression models.
• sim-itx-search.R contains code for the simulation example in Section XX which demonstrates our method for discovering the most significant interactions present in nonparametric regression models. This script creates 20 simulations, while sim-itx-search-big.R contains 1000 simulations.
• crime-example.R contains code for the linear summary of linear regression model used the crime example regression analysis in Section XX.
• CA-example-[xxx].R these scripts replicate the extensive case study of Section 4, summarizing a nonparametric regression model for housing prices in California at the census tract level.
  • CA-example-01dataprep.R reads in and transforms the data.
  • CA-example-02GPfit.R fits a Gaussian process regression model to the data.
  • CA-example-03GPposterior.R computes MCMC draws from the posterior of the GP regression model and the error variance parameter.
  • CA-example-04globalsummaries.R computes global summaries (linear, additive, additive with bivariate interaction), and the interaction search mentioned in Section 4.1.
  • CA-example-05localsummaries.R computes local linear summaries for the GP regression model, at various resolutions of locality (metropolitan area, county, city, and neighborhood).
• R/ contains R functions for the rest of the analysis.
• src/ contains source C++ functions used in R scripts.
• data/ contains the California housing price data.

Details for these analyses can be found in the original paper.

Contact

Spencer Woody
Statistics PhD Candidate, UT-Austin
Email: [email protected]