Econometrics and Business Statistics Seminar: Paul Bürkner, University of Stuttgart

Title: Bayesian model and variable selection


Date Wed 28 Oct
Time 14:15 15:15
Location Online

Presenter: Paul Bürkner, University of Stuttgart

Title: Bayesian model and variable selection


An important aspect of almost all projects involving statistical modeling of real world data is the comparison and selection of models or variables. One widely applied approach for that purpose is cross-validation. Unfortunately, fitting Bayesian models via MCMC sampling or comparable algorithms takes a lot of time. As a result, exact cross-validation of these models takes a lot of time as well, which effectively renders it impossible for larger and more complex models. For this reason, we have developed several methods to perform model or variable selection using approximate cross-validation procedures that drastically reduce the required computation time. Two of these methods, which I am going to introduce in the talk, are approximate leave-one-out cross-validation via Pareto smoothed importance sampling (PSIS-LOO-CV) and variable selection via projective predictions (projpred). All of the statistical modeling is powered by the probabilistic programming language Stan and related R packages.

Area of research: Bayesian multilevel models


Organizer: Mikkel Bennedsen and Jesper Wulff

Econometrics and Business Statistics Seminar Series, CREATES