Methods of Analysis for Failure Time Data That Adjust for Errors in Survival Outcomes Revealed by Audits

• Activity Number: 71
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: ENAR
• Author(s): Pamela A Shaw*+ and Bryan E Shepherd
• Companies: University of Pennsylvania and Vanderbilt University
• Keywords: Audit ; Measurement Error ; Multiple Imputation ; Regression Calibration ; HIV

Abstract

Time-to-event outcomes can be subject to error. For example, misclassification can occur due to the difficulty of making a correct diagnosis or mistimed events can occur due to errors in the data abstraction or recording process. Furthermore, errors in the survival outcome could be correlated with errors in other important patient covariates.These errors can introduce bias into estimates of the hazard ratio coefficient in standard Cox regression. For many clinical studies, audits and endpoint reviews are performed to ensure data quality. Validated data, often available only on a subset, can be used to adjust estimation for potential errors in the unaudited data. We extend estimation techniques typically used to correct analysis for covariate measurement error, namely multiple imputation and regression calibration, to accommodate potentially correlated errors present in survival outcome and covariates. The performance of the error-correction and naïve estimators is studied in simulations. We show the extent of bias and relative performance of the methods depend on the extent and nature of the error as well as the size of the audit. We consider an example from an HIV cohort study.