Comparing Methods of Multiple Imputation for a Score-Variable Measured Repeatedly Over Time

• Activity Number: 313
• Type: Contributed
• Date/Time: Tuesday, August 2, 2016 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Elizabeth L. McCabe* and Joseph M. Massaro and Kathryn L. Lunetta and Susan Cheng and Joanne M. Murabito and Martin G. Larson
• Companies: Boston University and Boston University and Boston University and Framingham Heart Study and Framingham Heart Study and Boston University
• Keywords: multiple imputation ; fully conditional specification ; longitudinal ; score-variable

Abstract

We compare 4 methods of multiple imputation (MI) by fully conditional specification (FCS) to estimate means and correlations of the health aging index (HAI) over time. HAI is a score-variable based on 5 clinical components. We simulate multivariate normal data for each component at 4 time points across 14 years using within and across-time correlation patterns and percent of missingness representing observed Framingham Heart Study data. Our methods of MI are cross-sectional FCS (XFCS, imputation model uses other components at same time), longitudinal FCS (LFCS, uses same component at all times ignoring cross-component correlation), all FCS (AFCS, uses all components at all times) and 2-fold FCS (2fFCS, uses all components at current and adjacent times). We compare results for multiple sample sizes (n=100/1000), number of imputations (m=5/20) and mechanisms of missingness (MCAR/MAR/MNAR). All but XFCS produce unbiased estimates of means and correlations and yield nearly identical results. Increase in precision is small when increasing from 5 to 20 imputations. Ongoing work explores the MI by FCS effect on using linear mixed effects models to estimate the slope of HAI over time.