Effect of Model Misspecification on Regression Calibration Adjustment for Measurement Error

• Activity Number: 76 - Contributed Poster Presentations: Section on Statistics in Epidemiology
• Type: Contributed
• Date/Time: Monday, August 3, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Eunyoung Park* and Daniela Sotres-Alvarez and Paul Gustafson and Laurence S Freedman and Victor Kipnis and Pamela A Shaw
• Companies: University of Pennsylvania and University of North Carolina at Chapel Hill and University of British Columbia and Gertner Institute for Epidemiology and Health Policy Research, Sheba Medical Center and National Cancer Institute and University of Pennsylvania Perelman School of Medicine
• Keywords: Classical measurement error; Systematic measurement error; Regression calibration; Misspecification; Prediction Modeling

Abstract

Many epidemiologic studies have shown that measurement error (ME) in exposures can distort the observed exposure-outcome associations. Predominantly in nutrition epidemiology, dietary intakes from self-report tend to be subject to both systematic and random (unbiased) ME. For some nutrients, recovery biomarkers that have unbiased (classical) ME exist. Regression calibration (RC) can largely correct bias in risk estimates from outcome regression models induced by exposure ME, through application of a calibration model. Although it is well-known that misspecifying regression models results in biased parameter estimates, the robustness of RC to misspecification of the outcome model or calibration model has been less studied. In this study we have investigated how misspecification of either the calibration model or outcome model affects the performance of RC. We conducted extensive simulations under systematic ME and classical ME, and a logistic outcome model, examining the effects of omitting a covariate from 1) the calibration model, 2) the outcome model or 3) both, under varying scenarios of correlation between the omitted covariate and other predictors.