An Introduction to Projection Predictive Variable Selection

Frank Weber, Research assistant at Rostock University Medical Center The projection predictive variable selection for Bayesian regression models, implemented in the R package projpred, has been shown to possess excellent properties in terms of the trade-off between predictive performance and sparsity. It also allows for valid post-selection inference by retaining all uncertainty inherent to the so-called reference model, also known as the actual belief model, that serves as a yardstick in terms of predictive performance and guides the projection, thereby filtering noise from the observed response values. Unfortunately, projection predictive variable selection still seems to be applied only rarely in practice. Thus, the goal of this presentation is to introduce projection predictive variable selection to a wider audience with the help of a real-world example. Main Sections 0:00 Introduction 3:24 Example 5:50 Search and Evaluation I 8:28 Search and Evaluation II 9:11 Performance evaluation plot 10:24 Model size decision 10:50 Identification of the selected submodel 11:28 Final projection 12:22 Marginals of the projected posterior I 14:08 Predictions based on the projected posterior: proj_predict() 15:34 Supported models 17:17 Closing remarks and Q&A More Resources R/Medicine Event site: https://events.linuxfoundation.org/r-medicine/ Main Site: https://www.r-consortium.org/ News: https://www.r-consortium.org/news Blog: https://www.r-consortium.org/news/blog Join: https://www.r-consortium.org/about/join Twitter: https://twitter.com/Rconsortium LinkedIn: https://www.linkedin.com/company/r-consortium/ Mastodon: https://fosstodon.org/@RConsortium