Nonparametric paired tests for censored survival data incorporating prognostic covariate information.

In clinical trials, complications in the data structure can arise by design, as when treatment groups are dependent, or by happenstance, as when selection bias is present. In this dissertation, we develop nonparametric paired tests based on weighted integrated survival that incorporate prognostic covariate information, and adjust for selection bias as well as informative censoring. In the absence of these sources of bias, these new methods can improve power over current nonparametric tests in two ways. First, by extending existing methods that take advantage of covariate information to the paired setting, we gain power over methods that either ignore the covariates or the dependence between the treatment groups. Second, the methods we use to adjust for selection bias can produce additional efficiency gains when treatment groups are comparable with respect to baseline covariates. When information collected post-baseline is incorporated, additional efficiency gains are made as long as censoring is uninformative. Additionally, we develop paired stratified tests using both weighted rank-based and weighted integrated survival based methods. In addition to adjusting for potential bias from baseline covariate imbalances and weakening assumptions concerning uninformative censoring, stratification may improve power when treatment alternatives are more clearly detectable within each covariate stratum. We investigate the strengths and limitations of the different methods presented using a number of simulation scenarios. Data from the Early Treatment Diabetic Retinopathy Study are used throughout to illustrate methodology in comparing time to severe vision loss between treatment groups receiving either early or deferred photocoagulation therapy.Ph.D.Biological SciencesBiostatisticsUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/131477/2/3057914.pd