Multilevel Conference

Pre-conference workshop 1: ‘Projection predictive variable selection for multilevel models’


**Unfortunately, due to circumstances this workshop had to be cancelled** 


The morning pre-conference workshop on “projection predictive variable selection for multilevel models” on March 11 is taught by Dr. Andrew Johnson. Andrew is a postdoc with Aki Vehtari’s group at Aalto University (Helsinki), who are the primary developers and maintainers of the projpred package. Andrew is also part of the development team for Stan, the increasingly popular program for Bayesian analysis.

With more data available than ever before, researchers and statistical modellers are often tasked with the problem of developing the best possible predictive model with the fewest predictors (feature selection). For researchers familiar with variable selection in a Frequentist paradigm (e.g., LASSO or elastic-net regression), it can be difficult to translate the same approaches to a Bayesian context. One of the key hurdles that many encounter is the additional computation required for Bayesian MCMC estimation can make common cross-validation approaches far too time-consuming to be practical.

A promising approach is projection-predictive selection (see, e.g. Piironen, Paasiniemi, and Vehtari, 2020 and McLatchie, Rögnvaldsson, Weber, and Vehtari, 2023). Under this paradigm researchers first specify and fit ‘reference model’, which represents the most complex and best-performing predictive model possible. The ‘projection’ then uses this reference model to construct an approximation of the results of a restricted submodels (i.e., models with different combinations of removed predictors) without needing to additionally estimate them. Researchers can then use these projected results to identify the smallest combination of features with predictive performance close to the full reference model.

In this pre-conference workshop, I’ll provide a hands-on introduction to performing projection-predictive selection with multilevel models using the R packages brms and projpred. We will use the brms package for specifying and estimating our reference models, and the projpred package for performing the projection-predictive selection. We’ll also be discussing different approaches to performing the selection and validating the results, as well as diagnostics for the quality of the model predictions.

More details on required preparation and availability of software syntax and data will follow later will follow later.

This morning workshop session runs from 9:00 till 12:30. Lunch is included.

Note that you can also register for this workshop without participating in the conference.