Resampling approach for estimating prediction error and for adjusting logistic regression models for covariate measurement error

Methods based on the resampling approach are proposed to address two issues related to prediction modeling: estimation of prediction error and adjustment for covariance measurement error.; Repartitioning k-fold cross-validation (CVK R) is proposed to enhance the estimation of the prediction error of classification models. Compared to cross-validation, the traditional method of choice, CVKR reduces the variability of estimated prediction error. In addition, it provides an empirical distribution of prediction error rather than a single estimate unaccompanied by an estimate of standard error for the point estimate. SAS macros are developed for the implementation of CVKR.; Bootstrap regression calibration (BRC) is proposed to adjust the coefficient estimates of logistic regression models when measurement error is present in model covariates. This method can be thought of as a bootstrap-smoothed version of the popular regression calibration method (Rosner et al. AM. J. Epidemiol. 1990, 1992). These two methods are evaluated and compared with respect to prediction accuracy, something not found in previous works. Receiver Operating Characteristic (ROC) methodology was employed to measure models' prediction accuracy and the area under ROC curve (AUC) was used as the index for prediction accuracy. The methods were also evaluated with respect to the attenuation (or bias) in the estimated coefficients. Results from simulation studies showed that BRC offers consistent enhancement over the regression calibration method in terms of improving the prediction accuracy and reducing bias in estimated coefficients.