Semiparametric Regression Models for Disease Natural History and Multiple Events in Cancer Research.

[摘要] This dissertation is concerned with semiparametric joint models of disease natural history and its relationship with observed multiple events.A common disease progression process that generates both the disease natural history events and the observed survival time consists of the complete multivariate survival data structure that statistical models built upon.In the first project, the disease natural history process is observed through a current-status type data surrogate (mark variable).A semiparametric regression model is proposed to assess the covariate effects on the observed marked endpoint explained by a latent disease natural history.Constructed through a nested series of Cox proportional hazard (PH) models with time-dependent covariates, the proposed model can be represented as a transformation model in terms of mark-specific hazards, induced by a complex non-proportional process-based frailty. An estimating equation based approach and a nonparametric maximum likelihood estimation (NPMLE) approach are proposed for estimation.The second project deals with the case when the disease natural history is observable in principle, while the status of the dependent censoring by the terminal event is missing in a fraction of patients.The disease natural history thus may be right- or left-censored by the terminal event informatively, and the observed data become a mixture of semicompeting risks data and terminal event only data.A semiparametric illness-death model with PH assumptions is proposed to study the relationship between nonterminal and terminal events.The corresponding NPMLE is studied for statistical inference.The third project considers the scenario when disease progression always precedes death in advanced cancer settings, such that the cancer progression and death are sequentially observed on the complete data level.The relationship between covariate, progression and death is our main interest, and its evaluation is complicated by undetected or missing progression-related events.A semiparametric progressive multistate model with a shared-frailty and order constraint modeling the association between progression and death is thus proposed.An Expectation-Maximization (EM) approach is used to derive NPMLE of model parameters.All three methods are applied to completed randomized or observational studies in cancer research, and the large-sample and finite-sample properties of proposed estimators are studied and evaluated accordingly.

[发布机构] University of Michigan

[效力级别] Disease Natural History

[关键词] Survival Analysis;Disease Natural History;Cancer Research;Semiparametric Regression;Multiple Events;Missing Data;Public Health;Statistics and Numeric Data;Health Sciences;Science;Biostatistics