Events Calendar

18 Jun
Pitt Public Health dissertation defenses
Event Type

Defenses

Topic

Research

Target Audience

Faculty, Graduate Students, Postdocs

Website

https://publichealth.pitt.edu/defenses

University Unit
Department of Biostatistics, Graduate School of Public Health: Dean's Office
Subscribe
Google Calendar iCal Outlook

Xiaotian Gao: Joint Modeling of Longitudinal and Survival Data and Robust Nonparametric Regression

Doctoral candidate  Xiaotian (Steven) Gao of the Department of Biostatistics will defend his dissertation entitled, "Joint Modeling of Longitudinal and Survival Data and Robust Nonparametric Regression."

Advisors:  Abdus S. Wahed, Chaeryon Kang

Committee members:  Abdus S. Wahed, Chaeryon Kang, Zhao Ren, and Gong Tang 

Abstract

This dissertation covers two areas: joint modeling of longitudinal and survival data and robust nonparametric regression. For joint modeling, the first chapter focuses on developing a statistical inference tool to estimate the mean quality-adjusted lifetime (QAL). QAL is a “quality” weighted survival lifetime that summarizes both quantitative and qualitative health aspects. It is especially of interest when treatments with side effects are compared. Traditionally, QAL is calculated based on discrete health status. Each status is assigned to a prefixed weight, and QAL is defined as the weighted average of time spent in each of the statuses. However, there are several issues in defining QAL based on discrete health status: 1. possible heterogeneity of quality of life within one status and 2. exact transitioning time between status not known. To overcome both barriers, we proposed to incorporate continuous quality of life scores in calculating QAL and use joint models with inverse probability weighting to estimate mean QAL. Asymptotic properties of the estimator were studied, and finite sample performance was evaluated via simulations. The method was also applied to the Virahep-C data set. We continue to develop a joint model that can handle multiple longitudinal outcomes and semi-competing risk survival data.

 

In the second chapter of the dissertation, we develop a robust method for nonparametric additive regression. Assuming only an additive structure, the nonparametric additive model relaxes the strict linearity assumption of the regular linear regression but preserves the interpretability and suffers less from the curse of dimensionality compared with fully nonparametric models. The goal of this chapter is to develop a robust nonparametric additive model that produces stable estimates even when the errors have asymmetric and heavy-tail distributions. This is accomplished using an adaptive Huber loss function, with the value of its parameter increasing with the sample size at a certain rate. Compared with traditional robust regressions, such as the quantile and median regression, the proposed method balances the robustness-to-unbiasedness trade-off and produces stable and consistent estimates. Non-asymptotic deviation results are studied, and stable algorithms are developed under both low and high dimension regimes. The method is applied to the NCI-60 data set to explore the relationship of expression levels between proteins (heavy-tail) and genes.

 

PUBLIC HEALTH SIGNIFICANCE: In this dissertation, the chapter on quality-adjusted lifetime provides useful tools to compare treatments with side effects. Incorporating continuous quality of life scores, the proposed method together with its sensitivity analysis could potentially help clinicians make individualized recommendations to different patients with different preferences over treatment effects and tolerance of side effects. The chapter on robust nonparametric regression is focused on analyzing the outcomes with heavy-tail distributions, such as protein expression levels. The proposed method is able to identify weaker signals while maintaining a similar false discovery rate compared with traditional methods. This will help clinical researchers identify key but weak genes or biomarkers with more confidence.

Thursday, June 18 at 2:30 p.m. to 4:30 p.m.

Virtual Event

Xiaotian Gao: Joint Modeling of Longitudinal and Survival Data and Robust Nonparametric Regression

Doctoral candidate  Xiaotian (Steven) Gao of the Department of Biostatistics will defend his dissertation entitled, "Joint Modeling of Longitudinal and Survival Data and Robust Nonparametric Regression."

Advisors:  Abdus S. Wahed, Chaeryon Kang

Committee members:  Abdus S. Wahed, Chaeryon Kang, Zhao Ren, and Gong Tang 

Abstract

This dissertation covers two areas: joint modeling of longitudinal and survival data and robust nonparametric regression. For joint modeling, the first chapter focuses on developing a statistical inference tool to estimate the mean quality-adjusted lifetime (QAL). QAL is a “quality” weighted survival lifetime that summarizes both quantitative and qualitative health aspects. It is especially of interest when treatments with side effects are compared. Traditionally, QAL is calculated based on discrete health status. Each status is assigned to a prefixed weight, and QAL is defined as the weighted average of time spent in each of the statuses. However, there are several issues in defining QAL based on discrete health status: 1. possible heterogeneity of quality of life within one status and 2. exact transitioning time between status not known. To overcome both barriers, we proposed to incorporate continuous quality of life scores in calculating QAL and use joint models with inverse probability weighting to estimate mean QAL. Asymptotic properties of the estimator were studied, and finite sample performance was evaluated via simulations. The method was also applied to the Virahep-C data set. We continue to develop a joint model that can handle multiple longitudinal outcomes and semi-competing risk survival data.

 

In the second chapter of the dissertation, we develop a robust method for nonparametric additive regression. Assuming only an additive structure, the nonparametric additive model relaxes the strict linearity assumption of the regular linear regression but preserves the interpretability and suffers less from the curse of dimensionality compared with fully nonparametric models. The goal of this chapter is to develop a robust nonparametric additive model that produces stable estimates even when the errors have asymmetric and heavy-tail distributions. This is accomplished using an adaptive Huber loss function, with the value of its parameter increasing with the sample size at a certain rate. Compared with traditional robust regressions, such as the quantile and median regression, the proposed method balances the robustness-to-unbiasedness trade-off and produces stable and consistent estimates. Non-asymptotic deviation results are studied, and stable algorithms are developed under both low and high dimension regimes. The method is applied to the NCI-60 data set to explore the relationship of expression levels between proteins (heavy-tail) and genes.

 

PUBLIC HEALTH SIGNIFICANCE: In this dissertation, the chapter on quality-adjusted lifetime provides useful tools to compare treatments with side effects. Incorporating continuous quality of life scores, the proposed method together with its sensitivity analysis could potentially help clinicians make individualized recommendations to different patients with different preferences over treatment effects and tolerance of side effects. The chapter on robust nonparametric regression is focused on analyzing the outcomes with heavy-tail distributions, such as protein expression levels. The proposed method is able to identify weaker signals while maintaining a similar false discovery rate compared with traditional methods. This will help clinical researchers identify key but weak genes or biomarkers with more confidence.

Thursday, June 18 at 2:30 p.m. to 4:30 p.m.

Virtual Event

Event Type

Defenses

Topic

Research

Target Audience

Faculty, Graduate Students, Postdocs