Comparison of Missing Data Imputation Methods in Longitudinal Study of ADRD Patients

• Activity Number: 259 - SPEED: Missing Data and Causal Inference Methods, Part 2
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 3:05 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Yi Cao* and Roee Gutman and Heather Allore and Brent Vander Wyk
• Companies: Brown University and Brown University and Yale University and Yale University
• Keywords: Multivariate longitudinal; Multiple imputation; General location model; Mixed models; Bayesian semi-parametric and nonparametric models; Chained equations

Abstract

In longitudinal studies with persons with Alzheimer’s Disease and related dementias (ADRD), missing data may be a result of worsening cognition and health status. When estimating relationships between multiple patient-centered outcomes of mixed type, methods for handling multivariate longitudinal missingness are required. Multiple imputation is one approach to fill-in missing data. Joint modeling (JM) and fully conditional specification (FCS) are two possible strategies. JM fits a multivariate distribution for the entire set of variables, but it may be complex to define and implement. FCS imputes data variable by variable from a set of conditional distributions, which is easier to define and implement. However, FCS suffers from theoretical deficiencies. Using simulated data from the National Health and Aging Trends Study, we compared the performance of FCS and various JM models (e.g. the general location model, general linear mixed models and nonparametric Bayesian models) for imputing multivariate longitudinal outcomes. We examined different missing data patterns and outcome types to provide advice to researchers on suitable methods to handle different missing data scenarios.