|Institution:||University of Michigan|
|Keywords:||Case-cohort Study; Clustered Data; Cox Model; Dependent Censoring; Outcome-dependent Sampling; Survival Analysis; Statistics and Numeric Data; Science|
|Full text PDF:||http://hdl.handle.net/2027.42/91398|
In this dissertation, we focus on the development of semiparametric methods for estimating proportional hazards models in the presence of non-standard data structures, namely clustering, outcome-dependent sampling, dependent censoring and external time-dependent covariate. In the first chapter, we propose methods based on estimating equations for case-cohort designs with clustered failure time data. We assume a marginal hazards model with a common baseline hazard and common regression coefficients across all clusters. Compared to their closest competitors in the literature, the proposed methods feature more tractable asymptotic derivations, variance estimation with reduced computational burden, and potentially increased efficiency. We apply these methods to the study of mortality among Canadian dialysis patients. In the second chapter, we propose methods for dealing with failure time data in the setting where the probability of sampling subjects depends on the outcome (e.g., death, survival) and where subjects are censored in a manner which is dependent on the failure rate. We employ a novel double-inverse-weighting scheme which combines weights arising from the probability of remaining uncensored and from the probability of being sampled. The proposed methods are applied to study the wait-list mortality among patients with end-stage liver disease. The third chapter is motivated by the challenges of fitting complex models to data from the smaller countries participating in the Dialysis Outcomes and Practice Patterns Study (DOPPS). We perform a comprehensive investigation of the association between the day-of-week-specific death rates and the dialysis schedule in the U.S., several European countries and Japan. Three Cox models are considered in which 'day of the week', 'day of dialysis schedule', or 'days since last dialysis' serves as a time-dependent covariate. The models are compared and contrasted, with special attention given to the setting where the sample size is small.