Title: Accommodating missing covariates via product partition models
Authors: Garritt Page - Brigham Young University (United States) [presenting]
Fernando Quintana - Pontificia Universidad Catolica de Chile (Chile)
Abstract: Missing observations are ubiquitous in data driven research studies. Statistical methods that accommodate different types of missingness have been developed and now have a long history. Many of these methods depend on assumptions about the missingness mechanism (e.g., missing completely at random ) that are difficult to verify in a data driven way. The default method of choice (particularly in a Bayesian setting) is to assume missing at random and carry out multiple imputation where values for missing observations are imputed and analysis are run on the filled in data set. To avoid making it difficult to verify assumptions about the missing mechinism, we develop a methodology that flexibly accommodates missingness based on covariate dependent random partition models. Without discarding subjects with partially observed covariate vectors, we simply employ those variables that have been measured when forming clusters and making predictions. We apply the methodology to data gathered in a study of osteonecrosis.