Statistics Seminar: Francis Hui (ANU) -- Joint Effect Selection in Mixed Models using CREPE

 Abstract: 

Generalized linear mixed models (GLMMs) are a widely used class of models for analyzing
correlated data in medicine and ecology.  As with all regression models though, a key
part of the inference process is variable selection.  I focus on penalized likelihood
methods for performing simultaneous selection of fixed and random effects in GLMMs.  In
particular, there is often an implicit belief that when fitting GLMMs with random
intercepts and slope, covariates should be included in the model as either a fixed
effect only, or a composite (i.e.  both fixed and random effect) effect.  

In this talk, I propose a penalty called CREPE that builds the aforementioned belief
into the variable selection process.  Specifically, CREPE accounts for the hierarchical
nature of the covariates, so that the final submodel involves only fixed and/or
composite effects.  Asymptotic properties of the CREPE estimator are discussed including
selection consistency and the oracle property.  Simulations demonstrate the strong
performance of CREPE compared to some other currently available penalized methods for
GLMMs.  

Biography: Francis Hui is a post-doctoral fellow at the Mathematical Sciences Institute
at ANU, working under the guise of Alan Welsh and USYD’s very own superstar Samuel
Muller.  His interests include mixed models, mixture models, variable selection, and
ecological statistics, all whilst drinking copious amounts of tea.