(G) Selection of Censoring Schemes for Analyzing Overall Survival in Real-World Data with Linked Mortality Information

Publication Number: 665

Background: Real-world data (RWD) capture mortality in different ways. Some RWD only capture mortality within its data generation processes (e.g. disenrollment due to death/in-hospital death), while other RWD use external sources to supplement death information that occurs outside of the database (e.g. linkage of electronic health records (EHR) to Social Security Death Index). Accuracy of such mortality information varies by RWD source and is never 100% complete. There are conflicting recommendations for censoring overall survival (OS) in RWD.

Objectives: We present a simulation study to evaluate the impact of censoring schemes on estimates of median survival time and hazard ratios under conditions of incomplete mortality information.

Methods: Survival data was simulated to represent a real-world EHR database. OS was assumed to follow an exponential distribution with median OS time of 1.5 and 3 years for the control and treatment groups, respectively. The maximum follow up was truncated at 5 years, corresponding to ~25% of patients censored at 5 years. Last activity was correlated with death when death occurred within the EHR system, with median time from last activity to death being 14 days. The capture of linked mortality information was varied from 0% to 95% incomplete. 1,000 simulated datasets were generated with 200 patients assigned in 1:1 ratio to the treatment and control groups.
Two censoring schemes typically recommended were evaluated: (1) censor at last activity date in absence of an observed death, and (2) censor at maximum follow up date. Median OS times and hazard ratios were estimated in each dataset. Bias and coverage were used to measure performance of each censoring scheme under varying amounts of incomplete mortality information.

Results: Our simulations demonstrate censoring at last activity is most biased when linked mortality information is well captured. Estimates of median survival were biased down indicating shorter estimated survival compared to the true parameter (mean bias in years (control, treatment) when 0% incomplete: -.04, -.25; when 95% incomplete: -.005,-.03). Estimates of hazard ratios were biased upwards indicating a smaller effect than the true effect, however less impact was seen (mean bias in log HR when 0% incomplete: .03; when 95% incomplete: .0004). Censoring at the maximum follow up date regardless of last activity within EHR had low bias ( < .12 years) for median survival estimates when sensitivity of mortality was > 85%.

Conclusions: Understanding the mechanism and completeness of mortality information is important when choosing censoring schemes for analysis of OS in RWD. Substantial bias in estimates of median survival can occur if an inappropriate censoring scheme is selected.