Robust Respondents and Lost Limitations: The Implications of Nonrandom Missingness for the Estimation of Health Trajectories.

Abstract

We offer a strategy for quantifying the impact of mortality and attrition on inferences from later-life health trajectory models. Using latent class growth analysis (LCGA), we identify functional limitation trajectory classes in the Health and Retirement Study. We compare results from complete case and full information maximum likelihood (FIML) analyses, and demonstrate a method for producing upper- and lower-bound estimates of the impact of attrition on results. LCGA inferences vary substantially depending on the handling of missing data. For older adults who die during the follow-up period, the widely used FIML approach may underestimate functional limitations by up to 20%. The most commonly used approaches to handling missing data likely underestimate the extent of poor health in aging populations. Although there is no single solution for nonrandom missingness, we show that bounding estimates can help analysts to better characterize patterns of health in later life.