Department of Biostatistics Seminar/Workshop Series

Extending regression calibration to adjust for correlated errors in outcome and exposure

Pamela Shaw, PhD

Assistant Professor, Department of Biostatistics and Epidemiology, Perelman School of Medicine, University of Pennsylvania

Measurement error arises through a variety of mechanisms. A rich literature exists on the bias introduced by covariate measurement error and on methods of analysis to address this bias. By comparison, less attention has been given to errors in outcome assessment or differential error. This talk considers an extension of the regression calibration method, a popular method that adjusts for covariate measurement error, to allow for errors in the outcome that may be correlated with prognostic covariates or with covariate measurement error. The method is derived using continuous data and its extension to survival data is also considered. This method can be applied when a validation subset is available, on which the true (error-free) data are also observed, or a reliability subset, where a second measurement of the error prone measurements are available. The former situation may arise say from a study audit, often performed in clinical studies; the latter is the only option when measurements without error are not possible. For the linear case, we discuss conditions for which the proposed method leads to unbiased estimates of the regression parameters. The performance of the proposed method is examined using simulation for a variety of measurement error scenarios and varying sizes of the reliability subset. The method is further illustrated with a real data example.