Multiple Imputation To Control for Unmeasured Confounding Using an Internal Validation Study

• Activity Number: 151
• Type: Contributed
• Date/Time: Monday, July 30, 2007 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Robert Glynn*+ and Til Stürmer
• Companies: Brigham and Women's Hospital and Brigham and Women's Hospital
• Keywords: confounding ; bias ; regression models ; multiple imputation ; missing data ; propensity scores

Abstract

Multiple imputation (MI) can be used to adjust main study estimates for unmeasured confounding if additional information on confounders is available from a validation study. We compared MI of a missing covariate and of the gold-standard propensity score (PSG). We simulated cohort studies with a binary exposure A, a Poisson outcome Y, and 3 independent, normally distributed confounders, of which 1 (C) was only observed in a validation sample. MI of C used linear regression of C on A, Y, and measured covariates; whereas MI of PSG used linear regression of PSG on A, Y and PSE (the error-prone propensity score without C). Both approaches performed well to control for unmeasured confounding and led to substantial bias reductions which were slightly less pronounced with MI of PSG. However, MI of PSG is easily implemented with multiple missing confounders and might be useful in that setting.