טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentHershkovitz Yael
SubjectNew class of powerful non-parametric consistent tests
for survival data
DepartmentDepartment of Industrial Engineering and Management
Supervisor Professor Malka Gorfine-Orgad
Full Thesis textFull thesis text - English Version


Abstract

For testing treatment effects with time-to-event data, the logrank test is the most popular choice and has some optimality properties under proportional hazards alternatives. Various other novel testing procedures have been proposed, which generally are derived by assuming a class of specific alternative hypotheses with respect to the hazard functions. These tests are powerful when the survival curves do not cross each other, but when they do cross, the power of these tests decreases. There are variety of situations in which the hazard functions are non-proportional. The longer the follow-up period is, the more likely it is for various non-proportional scenarios to develop. We introduce a test, which is based on the test of Heller, Heller and Gorfine (2013), named HHG, which is powerful even when the survival curves are crossing. HHG is a powerful test for discovering associations between any two variables, especially for non-linear associations and small datasets. We developed a powerful modification of HHG for survival data with right censoring. We performed an extensive simulation study to compare our new test with the most powerful or popular existing tests. We studied various shapes of hazard functions, different sample sizes with equal sizes of treatment and control groups, and three levels of censoring rates. We found out that our new HHG-type test for right-censored data is often more powerful than the most popular existing tests under many scenarios with crossing hazards, and when the hazard curves are proportional or close to proportional, our test is comparable with the other tests. We applied our new test to several publicly-available real datasets, and in most cases, our test results were better than the other existing tests.