Evaluation of Ml and MI Missing Data Methods in Health Care Setting

• Activity Number: 353
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jean Chantra*+ and Margo Sidell
• Companies: Kaiser Permanente and Kaiser Permanente
• Keywords: missing data ; ML ; MI ; healthcare ; lab data ; epidemiology

Abstract

Researchers often face missing data problems in epidemiological and clinical studies and such problems can potentially threaten the validity of research results, especially when the missing rate is high. Some common problems in clinical settings include but are not limited to the handling of missingness of biomarkers in observational studies. This study compared the performances of maximum likelihood (XMISS in the LogXact module) and multiple imputation (PROC MI in SAS) methods using real healthcare data and generated data with missingness. The multivariate data contain a binary outcome and 4 covariates to represent disease status, categorical biomarkers, and demographic information. Missingness was limited to one of the categorical covariates, representing missing lab data and the missingness was generated from the original data with 10%, 20% and 30% missing rate. We compared the results of the complete case analysis with the results of the two methods under each set of circumstances. This study investigated which of the methods was most appropriate in terms of the bias, variance, and the coverage of the estimated 95% confidence interval for each of the estimated coefficients.