All ETDs from UAB

Advisory Committee Chair

George Howard

Advisory Committee Members

Inmaculada B Aban

Leslie A McClure

David C Naftel

Monika M Safford

Document Type

Dissertation

Date of Award

2011

Degree Name by School

Doctor of Philosophy (PhD) School of Public Health

Abstract

In the case of two or more competing risks when only the time to the event due to one risk is observable with times due to the other outcomes treated as censored observations, standard methods of survival analysis fall short of assessing the association of a set of predictors with the outcomes of interest. This problem is particularly severe when there is a differential association of a predictor with two competing risks (harmful for one risk and protective for the other risk) such as stroke and myocardial infarction and in situations where the predictor is associated with only one risk while not being associated with the other risk. This dissertation uses a non-parametric rank transform method by assigning appropriate scores such as the logrank scores for exponentially distributed survival times in a bivariate regression framework to detect these differential associations. In certain scenarios, it is more powerful than the Cox proportional hazards model while in some other scenarios it has a power that is only marginally less than that of the Cox proportional hazards model. In order to obtain a clinically meaningful interpretation quantifying the magnitude of the association between predictors and outcomes, this dissertation also looks at how recent advances in full non-parametric regression obtained by reducing the dispersion function of the residuals, can be extended to censored univariate survival data. A new modified algorithm is proposed that accounts for the censoring in the survival data by adjusting the Wilcoxon scores associated with the residuals. For low to moderate levels of censoring in the data, this modified method provides robust non-parametric estimates of the regression coefficient and demonstrates adequate power in detecting the association of the predictors with the response. The real life applications of the above mentioned methods are demonstrated by applying them to pancreatic cancer and cardiovascular-stroke data sets. The results throw new light on associations between covariates and the response that earlier went unnoticed.

Included in

Public Health Commons

Share

COinS