In oncology research, with the advent of new therapeutic agents, many long-term survivors are deemed permanently cured. The cure rate survival model is a popular model that has been used to represent data generated from cancer studies where a substantial fraction of the population does not experience the event of interest even after a sufficient long observation time. This model is primarily used to evaluate the cure fraction in view of observed data. Existing testing methods have relied on the so-called constancy assumption that neglects potential covariate effect on the cure fraction under the alternatives to no cure. In this thesis, I extend the literature by proposing a score-type test that incorporates information from categorical covariates in detecting cure. I show that the limiting null distribution of the proposed test statistic is a mixture chi-squared distribution which, in practice, can be well approximated by a resampling method that uses normal variates to perturb the associated influence function. The numerical simulations show that the proposed test can greatly improve efficiency over the classical score test that neglects covariate information under heterogeneity, in settings where the true cure fraction depends on covariates. Ovarian cancer data collected in Los Angeles from the SEER registry is used to illustrate the real life application of the methodology.
Read
