Year: 2008
Type: dissertation
Abstract: Statistical models are often used for the prediction of \nfuture random variables. There are two types of prediction, point \nprediction and probabilistic prediction. The prediction accuracy is \nquantified by performance measures, which are typically based on \nloss functions. We study the estimators of these performance \nmeasures, the prediction error and performance scores, for point and \nprobabilistic predictors, respectively. The focus of this thesis is \nto assess the prediction performance of survival models that analyze \ncensored survival times. To accommodate censoring, we extend the \ninverse probability censoring weighting (IPCW) method, thus \narbitrary loss functions can be handled. We also develop confidence \ninterval procedures for these performance measures. \n \n \nWe compare model-based, apparent loss based and cross-validation \nestimators of prediction error under model misspecification and \nvariable selection, for absolute relative error loss (in chapter 3) \nand misclassification error loss (in chapter 4). Simulation results \nindicate that cross-validation procedures typically produce reliable \npoint estimates and confidence intervals, whereas model-based \nestimates are often sensitive to model misspecification. The methods \nare illustrated for two medical contexts in chapter 5. The apparent \nloss based and cross-validation estimators of performance scores for \nprobabilistic predictor are discussed and illustrated with an \nexample in chapter 6. We also make connections for performance.
Source: UWSpace (University of Waterloo)
Author: Yan Yuan
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2244498332