|
|
Kernel estimates of hazard function
Selingerová, Iveta ; Horová, Ivanka (advisor) ; Prášková, Zuzana (referee)
Kernel estimates of hazard function Abstract This doctoral dissertation is devoted to methods for analysis of censored data in survival analysis. The main attention is focused on the hazard function that reflects the instantaneous probability of the event occurrence within the next time instant. The thesis introduces two approaches for a kernel esti- mation of this function. In practice, the hazard function can be affected by other variables. The most frequently used model suggested by D. R. Cox is presented and moreover two types of kernel estimates to estimate a condi- tional hazard function are proposed. For kernel estimates, there is derived some statistical properties and proposed methods of bandwidths selection. The part of the thesis is extensive simulation study where theoretical results are verified and the proposed methods are compared. The last chapter of the thesis is devoted to an analysis of real data sets obtained from different fields.
|
|
|
Joint Models for Longitudinal and Time-to-Event Data
Vorlíčková, Jana ; Komárek, Arnošt (advisor) ; Omelka, Marek (referee)
Title: Joint Models for Longitudinal and Time-to-Event Data Author: Jana Vorlíčková Department: Department of Probability and Mathematical Statistics Supervisor: doc. RNDr. Arnošt Komárek, Ph.D., Department of Probability and Mathematical Statistics Abstract: The joint model of longitudinal data and time-to-event data creates a framework to analyze longitudinal and survival outcomes simultaneously. A commonly used approach is an interconnection of the linear mixed effects model and the Cox model through a latent variable. Two special examples of this model are presented, namely, a joint model with shared random effects and a joint latent class model. In the thesis we focus on the joint latent class model. This model assumes an existence of latent classes in the population that we are not able to observe. Consequently, it is assumed that the longitudinal part and the survival part of the model are independent within one class. The main intention of this work is to transfer the model to the Bayesian framework and to discuss an estimation procedure of parameters using a Bayesian statistic. It consists of a definition of the model in the Bayesian framework, a discussion of prior distributions and the derivation of the full conditional distributions for all parameters of the model. The model's ability to...
|
|
|
Kernel estimates of hazard function
Selingerová, Iveta ; Horová, Ivanka (advisor) ; Prášková, Zuzana (referee)
Kernel estimates of hazard function Abstract This doctoral dissertation is devoted to methods for analysis of censored data in survival analysis. The main attention is focused on the hazard function that reflects the instantaneous probability of the event occurrence within the next time instant. The thesis introduces two approaches for a kernel esti- mation of this function. In practice, the hazard function can be affected by other variables. The most frequently used model suggested by D. R. Cox is presented and moreover two types of kernel estimates to estimate a condi- tional hazard function are proposed. For kernel estimates, there is derived some statistical properties and proposed methods of bandwidths selection. The part of the thesis is extensive simulation study where theoretical results are verified and the proposed methods are compared. The last chapter of the thesis is devoted to an analysis of real data sets obtained from different fields.
|
|
|
Regression analysis of recurrent events
Rusá, Pavla ; Kulich, Michal (advisor) ; Komárek, Arnošt (referee)
V této práci se zabýváme metodami pro regresní analýzu výskytu opako- vaných událostí, při které je třeba se vypořádat se závislostí čas· do události v rámci jednoho subjektu. V první části práce se zabýváme možným rozšířením Coxova modelu proporcionálního rizika, který se využívá při analýze cenzoro- vaných dat, pro analýzu výskytu opakovaných událostí. Hlavní část práce je věnována odhadu parametr· v marginálních modelech a jejich asymptotickým vlastnostem. Následně se zabýváme i odhadem parametr· v marginálních mo- delech pro mnohorozměrná cenzorovaná data. Vhodnost použití marginálních model· je zkoumána pomocí simulací. 1
|
|
|
Nové metody ve schvalování úvěrů
Rychnovský, Michal ; Arlt, Josef (advisor) ; Pecáková, Iva (referee) ; Veselý, Petr (referee)
This thesis contributes to the field of applied statistics and financial modeling by analyzing mathematical models used in retail credit underwriting processes. Specifically, it has three goals. First, the thesis aims to challenge the performance criteria used by established statistical approaches and propose focusing on predictive power instead. Secondly, it compares the analytical leverage of the established and other suggested methods according to the newly proposed criteria. Third, the thesis seeks to develop and specify a new comprehensive profitability-based underwriting model and critically reflect on its strengths and weaknesses. In the first chapter I look into the area of probability of default modeling and argue for comparing the predictive power of the models in time rather than focusing on the random testing sample only, as typically suggested in the scholarly literature. For this purpose I use the concept of survival analysis and the Cox model in particular, and apply it to a real Czech banking data sample alongside the commonly used logistic regression model to compare the results using the Gini coefficient and lift characteristics. The Cox model performs comparably on the randomly chosen validation sample and clearly outperforms the logistic regression approach in the predictive power. In the second chapter, in the area of loss given default modeling I introduce two Cox-based models, and compare their predictive power with the standard approaches using the linear and logistic regression on a real data sample. Based on the modified coefficient of determination, the Cox model shows better predictions. Third chapter focuses on estimating the expected profit as an alternative to the risk estimation itself and building on the probability of default and loss given default models, I construct a comprehensive profitability model for fix-term retail loans underwriting. The model also incorporates various related risk-adjusted revenues and costs, allowing more precise results. Moreover, I propose four measures of profitability, including the risk-adjusted expected internal rate of return and return on equity and simulate the impact of the model on each of the measures. Finally, I discuss some weaknesses of these approaches and solve the problem of finding default or fraud concentrations in the portfolio. For this purpose, I introduce a new statistical measure based on a pre-defined expert critical default rate and compare the GUHA method with the classification tree method on a real data sample. While drawing on the comparison of different methods, this work contributes to the debates about survival analysis models used in financial modeling and profitability models used in credit underwriting.
|
guest ::
