A Comparison of Methods to Address Unmeasured Confounding When Internal or External Validation Data Are Available

• Activity Number: 554
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Dane R. Van Domelen* and Robert H. Lyles and Sunni L. Mumford and Emily Mitchell and Enrique F. Schisterman
• Companies: Emory University and Emory University and Eunice Kennedy Shriver National Institute of Child Health and Human Development and and NIH
• Keywords: propensity score calibration ; regression calibration ; measurement error ; maximum likelihood ; unmeasured confounding

Abstract

Epidemiologists tasked with estimating an exposure-disease association often find that a potentially important confounder was not measured. Failure to account for such a confounder can bias parameter estimates and their standard errors in an unpredictable direction. The purpose of this simulation study was to compare methods for restoring validity when validation data are available for the unmeasured confounder. The methods include maximum likelihood and regression calibration, popular approaches from the measurement error literature, and propensity score calibration, which was developed to handle unmeasured confounding. These approaches are based on two assumed models: a disease model relating the outcome to a binary exposure and a set of confounders, one of which is missing in the main study; and a measurement error model relating the confounder to a set of variables that explain some of its variability. These methods are applicable to many disease models, but we focus on linear and logistic regression. Assumptions for each method, some of which depend on the nature of the validation data, are outlined. Simulation results are provided to compare validity and efficiency.