Keywords: Missing data, registry data, joint replacement, random-effects models, pattern-mixture model, latent-variable
Clinical registries can be used to study the effects of medical interventions on a variety of outcomes, including patients’ perspectives on their own health. However, the amount of missing observations in these patient-reported outcomes (PROs) may be large, particularly in longitudinal studies due to patient drop-out. These missing data may be ignorable, but are more likely to be non-ignorable. We estimate the change in PROs from longitudinal registry data using different approaches that are based on assumptions of ignorable or non-ignorable missing data. Study data were from a clinical registry that captures > 90% of joint replacement surgeries for the province of Manitoba, Canada. Oxford Hip (OHS) and Knee (OKS) and twelve-item Short Form Survey version 2 Physical and Mental Component Summary (PCS, MCS) scores for the period from 2009-2015 were analyzed. PRO data were captured one month prior to surgery and at one year post-surgery. Change scores were estimated after adjusting for age and sex using: (a)fixed-effects model for complete cases, (b) multiple imputation model, (c) random-effects model, and (d) pattern-mixture model for non-ignorable missing data.