Shannon Woolley of the Department of Biostatistics defends her dissertation on "Tests for Random Signs Censoring in Competing Risks"
Graduate faculty of the University and all other interested parties are invited to attend.
In the setting of competing risks, the marginal survival functions of the latent failure times are nonidentifiable without making further assumptions about the joint distribution, the majority of which are untestable. One exception is the random signs censoring assumption which assumes the main event time is independent of the indicator that the main event preceded the competing event. Few methods exist to formally test this assumption, and none consider a stratified test, which detects whether random signs censoring is met within subgroups of a categorical covariate. We develop a nonparametric stratified test for random signs censoring that is easy to implement. In addition, it is often of interest to model the effects of several covariates in relation to the cause of interest. Thus, as an extension of the stratified test, we also propose a test for conditional random signs censoring, which allows for the random signs censoring assumption to be met after adjusting for categorical and/or continuous covariates. Through Monte Carlo simulations, we show our proposed test statistics have empirical levels close to the nominal level and maintain adequate power even with relatively small sample sizes and random right censoring. Compared to the standard test, both of our proposed tests have nearly equivalent power under random signs censoring and are superior in situations of stratified or conditional random signs censoring, where the standard test fails to detect random signs censoring within subgroups or after adjusting for covariates, respectively. Their ease of implementation and utility are illustrated through an application to liver transplant data from the United Network for Organ Sharing. Public Health Significance: Clinicians must make decisions affecting patients’ lives using the information available to them. Relying on research results based on models that use unverifiable assumptions can lead to inaccurate conclusions. The methods proposed here offer a solution to allow for more accurate modeling of marginal survival functions with competing risk data. Through use of these new methods, patient outcomes can be improved over time.