Loading Events

David Rossell (University of Warwick)

16 October 2015 @ 12:00

 

  • Past event

Details

Date:
16 October 2015
Time:
12:00
Event Category:

Consistency of posterior model probabilities in high-dimensional model selection

In recent years there has been an increasing interest in developing Bayesian formulations that remain effective in high-dimensional and non-standard problems. We focus on high-dimensional model selection problems where the number of parameters may grow with the sample size, and review results characterizing the situations under which one can hope for the posterior probability assigned to the data-generating truth to eventually converge to 1 (often termed “strong consistency”). As it turns out, under certain regularity conditions for the canonical linear model such consistency can only be achieved if the prior formulation is based on so-called non-local priors. We will review the non-local prior framework and present some recent extensions for high-dimensional estimation. Our theoretical and empirical results show that inducing model selection parsimony via non-local priors allows working with simpler models and improving prediction accuracy simultaneously. Although for simplicity we shall focus on variable selection settings, time allowing we will touch upon extensions to more complex settings such as graphical models or mixture models.