Missing Data and Variable Selection Methods for Cure Models in Cancer Research

cure models;multiple imputation;cancer modeling;Public Health;Health Sciences;Biostatistics

Beesley, LaurenSchipper, Matthew Jason ;

University of Michigan

关键词: cure models;    multiple imputation;    cancer modeling;    Public Health;    Health Sciences;    Biostatistics;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/144010/lbeesley_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

In survival analysis, a common assumption is that all subjects will eventually experience the event of interest given long enough follow-up time. However, there are many settings in which this assumption does not hold. For example, suppose we are interested in studying cancer recurrence. If the treatment eradicated the cancer for some patients, then there will be a subset of the population that will never experience a recurrence. We call these subjects ;;cured.”The Cox proportional hazards (CPH) mixture cure model and a generalization, the multistate cure model, can be used to model time-to-event outcomes in the cure setting. In this dissertation, we will address issues of missing data, variable selection, and parameter estimation for these models. We will also explore issues of missing covariate and outcome data for a more general class of models, of which cure models are a particular case.In Chapter II, we propose several chained equations methods for imputing missing covariates under the CPH mixture cure model, and we compare the novel approaches with existing chained equations methods for imputing survival data without a cured fraction.In Chapter III, we develop sequential imputation methods for a general class of models with latent and partially latent variables (of which cure models are an example). In particular, we consider the setting where covariate/outcome missingness depends on the latent variable, which is a missing not at random mechanism.In Chapter IV, we develop an EM algorithm for fitting the multistate cure model. The existing method for fitting this model requires custom software and can be slow to converge. In contrast, the proposed method can be easily implemented using standard software and typically converges quickly. We further propose a Monte Carlo EM algorithm for fitting the multistate cure model in the presence of covariate missingness and/or unequal censoring of the outcomes.In Chapter V, we propose a generalization of the multistate cure model to incorporate subjects with persistent disease. This model has many parameters, and variable selection/shrinkage methods are needed to aid in estimation. We compare the performance of existing variable selection/shrinkage methods in estimating model parameters for a study of head and neck cancer.In Chapter VI, we develop Bayesian methods for performing variable selection when we have order restrictions for model parameters. In particular, we consider the setting in which we have interactions with one or more order-restricted variables. A simulation study demonstrates promising properties of the proposed selection method.