Improving Collection and Analysis of Overall Survival Data

Abstract

Advances in anticancer therapies have provided crucial benefits for millions of patients who are living long and fulfilling lives. Although these successes should be celebrated, there is certainly room to continue improving cancer care. Increased long-term survival presents additional challenges for determining whether new therapies further extend patients' lives through clinical trials, commonly known as the gold standard endpoint of overall survival (OS). As a result, an increasing reliance is observed on earlier efficacy endpoints, which may or may not correlate with OS, to continue the timely pace of translating innovation into novel therapies available for patients. Even when not powered as an efficacy endpoint, OS remains a critical indication of safety for regulatory decisions and is a key aspect of the FDA's Project Endpoint. Unfortunately, in the pursuit of earlier endpoints, many registrational clinical trials lack adequate planning, collection, and analysis of OS data, which complicates interpretation of a net clinical benefit or harm. This article shares best practices, proposes novel statistical methodologies, and provides detailed recommendations to improve the rigor of using OS data to inform benefit-risk assessments, including incorporating the following in clinical trials intending to demonstrate the safety and effectiveness of cancer therapy: prospective collection of OS data, establishment of fit-for-purpose definitions of OS detriment, and prespecification of analysis plans for using OS data to evaluate for potential harm. These improvements hold promise to help regulators, patients, and providers better understand the benefits and risks of novel therapies.

Figure 1.. Potential strategies to establish prospective thresholds to rule out harm and inform drug development decision making when OS data are underpowered for efficacy.

A) Definitions and hypothetical decision paths when using OS data to analyze efficacy or evaluate for harm. When OS is analyzed for efficacy, it is important to rule out a hazard ratio (HR) of 1 with a robust statistical test, such as excluding 1 with the upper bound of the 95 percent confidence interval (CI). In traditional analyses of OS, detriment is confirmed when the lower bound of the CI is greater than 1, which continues to be true. Additionally, when OS is not statistically powered for efficacy, one or more prespecified upper bound of the appropriate CI thresholds greater than 1 can be used to evaluate for a high probability of substantial detriment and inform drug development and benefit-risk decisions. When OS data are mature, failure to exclude hypothetical threshold Y with the upper bound of the CI indicates the possibility that the investigational therapy is harmful to patients relative to the control and halting drug development may be appropriate. Excluding Y, but failing to exclude X, would suggest additional OS data are needed. Excluding hypothetical threshold X would indicate a low probability of substantial detriment and that proceeding with the drug development process is appropriate if early endpoints are favorable. If a binary decision framework is preferred, a single threshold could be prespecified. B) Three hypothetical scenarios that highlight how excluding an OS HR threshold X with a 95 percent CI (red plot area) can be accomplished much earlier than ruling out 1 during a clinical trial. A separate strategy could entail allowing less stringent Type I error and prespecifying a confidence interval less than 95 percent (grey plot area) with which to exclude an OS HR of 1 at an earlier time point. Plot axes are for illustrative purposes and not to scale.