Statistical methods to adjust for measurement error in risk prediction models and observational studies

The first part of this dissertation focuses on methods to adjust for measurement error in risk prediction models. In chapter one, we propose a nonparametric adjustment for measurement error in time to event data. Measurement error in time to event data used as a predictor will lead to inaccurate predictions. This arises in the context of self-reported family history, a time to event covariate often measured with error, used in Mendelian risk prediction models. Using validation data, we propose a method to adjust for measurement error in this setting. We estimate the measurement error process using a nonparametric smoothed Kaplan-Meier estimator, and use Monte Carlo integration to implement the adjustment. We apply our method to simulated data in the context of Mendelian risk prediction models and multivariate survival prediction models, and illustrate our method using a data application for Mendelian risk prediction models. Results show our adjusted method corrects for measurement error mainly in two aspects; by improving calibration and total accuracy. In some scenarios discrimination is also improved. In chapter two, we use the methods proposed in chapter one to extend Mendelian risk prediction models to handle misreported family history. The second part of this dissertation focuses on methods to adjust for measurement error in observational studies. In chapter three, we propose various methods to adjust for a mismeasured exposure using validation data and propensity scores. Propensity score methods assume that the treatment assignment is error-free, but in reality these variables can be subject to measurement error. This arises in the context of comparative effectiveness research, in which accurate procedural codes are not always available. When using propensity score based methods, this error affects the treatment assignment variable directly, as well as the propensity score. We propose a two step maximum likelihood approach using validation data to adjust for measurement error. In addition, we show the bias introduced when using error-prone treatment in the inverse probability weighting estimator and propose an approach to eliminate this bias. Simulations show our proposed approaches reduce bias and mean squared error of the treatment effect estimator compared to using the error-prone treatment assignment.