Pitfalls of Using Composite Primary End Points in the Presence of Competing Risks

In randomized controlled trials (RCTs), event-free survival (EFS) end points are commonly employed as primary end points. Examples of EFS end points include overall survival, disease-free survival, progression-free survival, and other combinations of events that include death as a result of any cause. These are composite end points that are used to estimate power and determine the benefit of the therapy being tested. Multiple problems can arise when using EFS end points in the presence of competing risks. Events under competing risks preclude or interfere with the ability to observe events of primary interest in relation to the therapy being tested. Examples of competing risks include second malignancies, death as a result of comorbid diseases, and treatment-related adverse effects or mortality. Such events interfere with analysis of effects on disease-specific events, which are typically the events directly affected by cancer therapies. In competing risk settings, precision in estimating the probability of occurrence of primary events of interest is reduced, statistical power is reduced, and inferences on the basis of composite end points are prone to error. The quintessential competing risk for disease-specific events is death as a result of noncancer causes (ie, competing mortality). Patients at high risk for competing mortality relative to disease-specific events are less likely to benefit, on the margin, from more intensive cancer therapy. In addition to not being well served by treatment intensification, such patients represent deadweight in RCTs of cancer therapies, because the predominant events these patients contribute to the primary end point are not reduced by the treatment in question. Inclusion of patients at high risk for competing mortality in RCTs may increase inefficiency and cost. Strategies to reduce the incidence of competing mortality in RCTs are therefore important both in designing efficient studies and facilitating interpretation of treatment effects. Use of primary EFS end points in competing risk settings is complicated by the decoupling of treatment effects versus net benefits when the incidence of competing mortality is high. EFS end points are well defined with respect to net benefits but not to effects, because EFS end points are composite end points by construction. That is, the effect of a treatment on A B demonstrates nothing about its individual effects (or lack thereof) on A and B. A common problem arises when effects on EFS end points are misconstrued as properties of the treatment rather than conditions of the study. For example, a treatment may be effective at reducing cancer mortality, leading to improved survival in young healthy patients but not in older sicker ones. Although the effectiveness of the treatment may be the same in both groups, the effect on the EFS end point may not be. A concern with using EFS end points is the potential to attribute nonspecific effects to a treatment that affects death as a result of any cause, irrespective of its cause-specific effects. That is, if a treatment affects noncancer mortality, it is likely to affect EFS. Such phenomena have been observed in RCTs, whereby a treatment effect on overall survival was attributable to noncancer rather than cancer-specific mortality. Explicitly decomposing primary end points and net effects into cause-specific components may be a safeguard against this problem, but this is done infrequently. Examining cause specificity is especially crucial for biomarker and gene signature studies to draw valid mechanistic inferences. A more subtle pitfall can arise in studies purporting to show interactions of treatment effects on EFS end points. Irrespective of its cause-specific effects, a treatment will seem to interact with variables correlated with competing mortality, because as the incidence of competing mortality increases, the net effect on the EFS end point dampens (tends toward unity; Appendix, online only). In such instances, the purported interaction may be falsely identified as an intrinsic property of the treatment. For example, hormone therapy for prostate cancer has been reported to interact with comorbidity or age, even without evidence of a difference in cause-specific effects. This creates a misperception that such treatments have fundamentally different activity with respect to these covariates, whereas this may be simply a result of the increase in noncancer mortality as a function of said covariates. These problems seem to arise because of the paradigm that has evolved for testing treatments in RCTs, which is based on a single primary end point and hypothesized effect size. Because multiple end points are typically of interest in RCTs, the relevance of designating a primary end point is essentially to drive sample size calculations. With primary EFS end points, the key end points are combined into a single metric to gauge the net benefit of the therapy. This paradigm implicitly regards the treatment effect as homogeneous across both diseasespecific and nonspecific components of the end point. However, this is unlikely to be characteristic of the effects of cancer therapies, which typically reduce disease-specific events but not competing events. The validity of power estimates then becomes critically dependent on the assumption that the distributions of covariates predictive of both cause-specific and competing events in the sample represent those from which the treatment effect and incidence estimates were derived. When this assumption does not hold, power derivations that ignore JOURNAL OF CLINICAL ONCOLOGY COMMENTS AND CONTROVERSIES VOLUME 28 NUMBER 28 OCTOBER 1 2010