Measures of survival benefit in cancer drug development and their limitations

Abstract

Background: A variety of measures of survival benefit are available to an investigator comparing outcomes across the various arms of a drug development trial. In this article, we systematically review the most common measures of comparative survival used in published studies.

Materials: We distinguish between relative and absolute survival differences, and measures of instantaneous and cumulative risk. We consider settings in which the end point is overall survival as well as those in which disease-specific end points are of primary interest.

Results: We note that different measures capture different aspects of benefit, and some may be more reliable than others or more representative of clinically relevant benefit.

Conclusions: Rather than simply using procedures that have become standard, analyses should identify the most clinically relevant measures of effect and apply procedures that reliably estimate these.

Keywords: Competing risks; Disease-specific mortality; Survival analysis.

Figure 1

Overall survival in a hypothetical randomized trial. (A) Treatment lowers the hazard of death by 40% and 12-month survival in the control group is 40%. (B) Treatment lowers the hazard of death by 40% and 12-month survival in the control group is 80%. (C) Median survival in the two treatment groups is similar but survival times in the treatment group are longer at the 25^th percentile.

Figure 2

Cumulative incidence of cancer death in a hypothetical randomized trial. (A) Competing risk of death is the same in both groups. (B) Competing risk of death is higher in the treatment group making the number of deaths observed in that group lower.

Figure 3

Disease-specific survival in a hypothetical randomized trial when risks of cancer-specific and other-cause death are positively correlated. (A) True underlying disease-specific survival. (B) Kaplan-Meier estimates of disease-specific death when competing deaths are censored. Because of the positive correlation, the competing deaths are precisely those at higher risk of disease-specific death. The individuals remaining in the sample are at a lower risk of disease-specific death, thus the estimated survival curves are more favorable than the true ones.