Haoxuan (Charlie) Zhou
Title: Age-Specific Regression of Event Times, with Application to Childhood Cancer Survivor Study
Date: Wednesday, August 19th, 2026
Time: 10am
Location: ASB 10920 & Zoom
Supervised by: X. Joan Hu & Yi Xiong
Abstract:
Motivated by the premature menopause (PM) program of the Childhood Cancer Survivor Study (CCSS; e.g., Im et al., 2023), this thesis develops statistical methods for age-specific regression analysis of event time data with complex structures. Following a brief description of the CCSS-PM data in Chapter 2, Chapters 3, 4, and 5 present the problem formulations and proposed statistical learning procedures, followed by concluding remarks in Chapter 6.
Chapter 3 focuses on primary ovarian insufficiency (POI), a major cause of PM. The available information on age at POI in the CCSS-PM study is formulated as doubly censored event time data. Two approaches based on estimating functions are proposed for age-specific logistic regression analysis. To account for censoring, inverse probability of censoring weighting (IPCW) is incorporated into the estimation procedures. We establish the asymptotic properties of the proposed estimators, develop corresponding variance estimation procedures, and conduct extensive simulation studies to evaluate their performance. The proposed methods are further illustrated through an analysis of the CCSS-PM data.
Chapter 4 extends the proposed approaches to settings with competing risks. In addition to POI, surgical premature menopause (SPM) represents another major cause of PM; consequently, age at PM may be associated with either POI or SPM. To address this setting, the two approaches are extended to age-specific multinomial logistic regression models. The methods are further combined with Firth-type penalization to address potential data separation issues. Variance estimation procedures are developed, and extensive simulation studies are conducted to assess the performance of the proposed methods. A second analysis of the CCSS-PM data is presented to demonstrate the application of the proposed extensions.
Implementation of the proposed methods requires estimation of the conditional distribution of the censoring time involved in the IPCW weights. Chapter 5 investigates several candidate approaches, including a stratified empirical distribution estimator, estimation under the Cox proportional hazards model, and estimation based on the survival random forest algorithm. Extensive simulation studies are conducted to compare the performance of these approaches. We examine two methods for estimating the conditional censoring distribution under the Cox model, along with various hyperparameter configurations for the survival random forest. Twelve simulation settings are considered, representing varying levels of dependence between censoring times and covariates, different covariate dimensions, and moderate to heavy censoring rates. A third analysis of the CCSS-PM data is presented to illustrate the practical implications of our findings from the simulation.