Random Effect and Latent Variable Model Selection

Random Effect and Latent Variable Model Selection In recent years, there has been a dramatic increase in the collection of multivariate and correlated data in a wide variety of ?elds. For example, it is now standard pr- tice to routinely collect many response variables on each individual in a study. The different variables may correspond to repeated measurements over time, to a battery of surrogates for one or more latent traits, or to multiple types of outcomes having an unknown dependence structure. Hierarchical models that incorporate subje- speci?c parameters are one of the most widely-used tools for analyzing multivariate and correlated data. Such subject-speci?c parameters are commonly referred to as random effects, latent variables or frailties. There are two modeling frameworks that have been particularly widely used as hierarchical generalizations of linear regression models. The ?rst is the linear mixed effects model (Laird and Ware , 1982) and the second is the structural equation model (Bollen , 1989). Linear mixed effects (LME) models extend linear regr- sion to incorporate two components, with the ?rst corresponding to ?xed effects describing the impact of predictors on the mean and the second to random effects characterizing the impact on the covariance. LMEs have also been increasingly used for function estimation. In implementing LME analyses, model selection problems are unavoidable. For example, there may be interest in comparing models with and without a predictor in the ?xed and/or random effects component.

Random effects and latent variable models are broadly used in analyses of multivariate data. These models can accommodate high dimensional data having a variety of measurement scales. Methods for model selection and comparison are needed in conducting hypothesis tests and in building sparse predictive models. However, classical methods for model comparison are not well justified in such settings.

This book presents state of the art methods for accommodating model uncertainty in random effects and latent variable models. It will appeal to students, applied data analysts, and experienced researchers. The chapters are based on the contributors’ research, with mathematical details minimized using applications-motivated descriptions.

The first part of the book focuses on frequentist likelihood ratio and score tests for zero variance components. Contributors include Xihong Lin, Daowen Zhang and Ciprian Crainiceanu.

The second part focuses on Bayesian methods for random effects selection in linear mixed effects and generalized linear mixed models. Contributors include David Dunson and collaborators Bo Cai and Saki Kinney.

The final part focuses on structural equation models, with Peter Bentler and Jiajuan Liang presenting a frequentist approach, Sik-Yum Lee and Xin-Yuan Song presenting a Bayesian approach based on path sampling, and Joyee Ghosh and David Dunson proposing a method for default prior specification and efficient posterior computation.

David Dunson is Professor in the Department of Statistical Science at Duke University. He is an international authority on Bayesian methods for correlated data, a fellow of the American Statistical Association, and winner of the David Byar and Mortimer Spiegelman Awards.