ABSTRACT
A commonly used procedure for comparing two survival distributions is the log rank test. One reason that the log rank test is often used is that it can handle censored observations. That is, it can account for subjects who have been randomized and treated but for whom the event of interest has not occurred at the time of the analysis. Doksum (1969), Janssen (1999), and Schoenfeld (1981) have investigated the power of the log rank test. The ability to handle censored observations also introduces a difficulty when interpreting the results of the log rank test. Oftentimes, the events of interest do not come from the whole range of the survival distribution but instead only arise from those subjects who are the first to have an event. We are therefore forced to assume that the experience for the subjects who had an event early can be extrapolated to the experience for subjects who will have an event late to draw conclusions related to how the two survival distributions as a whole compare.
