| Summary: | 博士 === 國立陽明大學 === 公共衛生研究所 === 92 === Abstract
Statistical models play an important role in outcome prediction for modern epidemiologists and clinicians. Despite the wide use of such predictive models, several problems, including unpopular use of survival models in dealing with event history data, the failure of taking correlated data into account, lack of model diagnosis and validation, and mishandling of data with multi-state and repeated property, remains unsolved.
The aim of the thesis was thus to develop a computer-aided system to combine two conventional models (logistic regression models and survival models) with a recently developed multi-state model into a menu-driven, user-friendly and SAS-based package.
The overall framework includes two parts, data and model module. The former consists of data editing, data management, sampling for splitting data in cross-validation, and the computation of new variables such as centering in polynomial regression model. The model module consists of three models, logistic regression models for binary data, survival models for event history data, and multi-state model for delineating the disease natural history and estimating parameters with likelihood function formed by empirical data. Each model module also included model validation, checking, and prediction.
The two conventional models and multi-state model were then applied to two datasets, breast cancer screening form the Swedish Two-county Trial and a separate one from Taiwan multi-centre cancer screening with this computer-aided system. The results indicated the system was an efficient tool for user to solve the problems. The multi-state model with Markov cohort and Monte Carlo approaches was used to evaluate a practical screening scheme, which is a part of the cost-effectiveness analysis of the screening program. The performance of the developed package was evaluated by a randomized trial design.
In conclusion, we demonstrated in this thesis that statistical prediction model commonly used today in the clinical world can be enhanced and made more user-friendly by combining conventional models with a new multi-state model using existing statistical computer language.
|
|---|
