Joint Modeling of Incomplete Data with Diverse Variable Types using Latent-Variable Models

CHIPTS Methods Seminar – UCLA-Semel Institute Center for Community Health

Joint Modeling of Incomplete Data with Diverse Variable Types using Latent-Variable Models

Presented by:

Thomas R. Belin, Ph.D.

Professor, UCLA Department of Biostatistics

Tuesday, October 8, 2013
2pm – 3pm

Abstract:  In incomplete data sets with many variables and diverse variable types (e.g., continuous, ordinal categorical, nominal categorical), it is challenging to develop general-purpose strategies for handling missing data.  After reviewing sequential regression imputation methods (e.g., IVEWare, ICE, MICE, MIDAS) that might be viewed as competitors, this presentation will discuss joint modeling strategies based on latent-variable models that allow for the inclusion of diverse data types.  In particular, we will focus on the use of models that can be fit with the help of a parameter-extended Metropolis-Hastings strategy for drawing correlation matrices in an MCMC inference framework.  Illustrative examples will be presented and future directions for research in this area will be considered.