Assessing Alternatives to the Hazard Ratio for Non-inferiority
Richard Chappell, University of Wisconsin  *Jyotirmoy Dey, Abbvie, Inc.  Xiaotian Zhu, Abbvie, Inc. 

Keywords: Non-inferiority, restricted mean survival time, risk difference, hazard ratio

Non-inferiority (NI) outcome trials involving time-to-event (TTE) endpoints are traditionally designed with NI margins based on the hazard ratio (HR). The appropriateness of using HR as the primary measure is, however, questionable in situations where either assumptions underlying the Cox proportional hazards (PH) model do not hold or the background event rates are very low. The lack of a direct mechanistic interpretation of hazards was also noted by D. R. Cox himself. Recent publications have therefore proposed risk difference (RD) or difference in restricted mean survival times (RMST) as being more interpretable and clinically relevant alternatives. For example, in a safety trial where few events are expected to occur in a placebo or control arm, a very large and complex trial may be needed to establish HR-based non-inferiority whereas a lack of increased risk can be shown in a far less resource-intensive way using differences in risks or RMST. The critical issue is that the precision of a HR estimate depends solely on the number of events observed irrespective of the absolute rate at which these events accumulate. A comparative assessment of these approaches is therefore warranted. It is, however, intrinsically difficult to compare these measures since the key advantages of RMST and RD emerge in situations where the HR is not very effective. Also, specifying an appropriate NI margin remains a key challenge in designing NI trials for which RD and RMST based margins can provide results that are more readily understood. While regulatory agency guideline for specifying HR-based NI margins are currently available, determining NI margins in terms of RD or RMST need further understanding of these scales and discussion. In this presentation, we take a closer look at the comparability issue and attempt to gain insights into the applicability of these approaches through characterizing their relative efficiency under various conditions in real-life non-inferiority settings.