Comparing Missing Data Approaches in Structural Equation Modeling with Data Missing Not at Random

• Activity Number: 418
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Scientific and Public Affairs Advisory Committee
• Author(s): Jin-Wen Hsu* and Wansu Chen and Kristi Reynolds and Mary Helen Black
• Companies: Kaiser Permanente and Kaiser Permanente Southern California and Kaiser Permanente and Kaiser Permanente
• Keywords: missing data ; structural equation modeling ; FIML ; multiple imputation ; Monte Carlo simulation

Abstract

Most statistical methods for handling missing data assume ignorability (i.e. missing at random), which may not be reasonable in many situations. We compared 3 missing data methods: multiple imputation (MI), full information maximum likelihood (FIML) and complete case analysis (CC), in the context of structural equation modeling (SEM) when data are missing not at random (MNAR) using a Monte Carlo approach. Data were simulated based on variable distributions and relationships from a validation study of the Osteoporosis-Specific Morisky Medication Adherence Scale (OS-MMAS). We modeled a latent variable, measured by the 8 items of OS-MMAS, a self-reported medication adherence scale, as a predictor of medication possession ratio (MPR), a measure of medication adherence computed using electronic prescription refill data. Nonresponse in the OS-MMAS survey, i.e. missing data, is assumed to be based on age and education. The results show that FIML and MI produce similar estimates and both are less biased than CC when data are MNAR. However, FIML is preferred because it is simpler than MI and produces parameter estimates and model fit indices in a straightforward manner.