Evaluating the estimation performance of Generalized Pairwise Comparisons methods for censored survival data

In analysis of time-to-event outcome, the comparison of treatment effects between two experimental arms mostly relies on the hazard ratio which is often estimated through a Cox proportional hazard model. However, it has been recently argued that this traditional measure is not always suitable and sufficient to report the results of clinical trials, especially when the condition of proportional hazards is not met. The proposals for new measures as alternatives or complements to hazard ratio are thus being very welcomed. On the other hand, results analyzed based on one single endpoint may sometimes fail to convey the proper clinical value of a treatment. To make use of all clinically important events, there is also a growing interest in novel methods which are able to incorporate multiple outcomes in the analysis. Among the recently proposed measures of treatment effect, net benefit estimated using Generalized Pairwise Comparisons (GPC) procedure seems to be highly attractive, since it not only analyzes time-to-event data but can also be applied to any type of outcomes. More importantly, GPC can easily extend to several prioritized outcomes and include the thresholds of clinical significance in the analysis. This method was first proposed in 2010 and still needs to be improved as it suffers some drawbacks when dealing with censored data. Currently, four GPC approaches have been suggested to overcome the challenges left by the standard GPC procedure: Peron original method (PO method); Peron method with correction (PC method), method based on the Extreme Value Theory (EVT method); and Restricted net benefit method (RNB method). Considering that each approach has its own limitation and may not applicable to all situations, this Master’s thesis has been conducted to compare these new estimators (mainly between PO, PC and EVT) and to identify which method would be the most appropriate in certain cases. To achieve this goal, we performed an extensive simulation study to examine the performance of net benefit estimators in various scenarios. We found that the estimation procedure using EVT tool performs well in most of the cases whether the hazard ratio is constant or not. On contrary, the usefulness of PC procedure is quite limited; it somewhat improves the PO estimation but only under its strong assumption. The result obtained for PO estimation is the most unexpected as this method resists almost perfectly in the case of early treatment effect regardless of high censoring rates.