Abstract:
|
In electronic health records (EHR)-related research, health status ascertained using phenotyping algorithms is sometimes error-prone. Ignoring misclassifications in EHR-derived phenotypes can lead to biased estimates of effect sizes of risk factors. To correct for such bias, manual chart reviews are usually conducted to obtain the true health status for a small validation set. Current methods to utilize the validation data for bias correction include direct estimation of misclassification rates or joint modeling of the validation data with the data without validation. There is lack of guideline on which method performs better under what scenarios. In this talk, we compare the relative performances of these two commonly used procedures under various real application motivated scenarios, and also propose two additional methods that can effectively incorporate the knowledge on misclassification rates in the bias correction. Simulation studies and case studies will be presented, in order to shed light on deciding the size of validation samples in practical EHR-based investigations.
|