Summary
In a time-to-event (TTE) analysis of real-world overall survival (rwOS), some patients may not have a recorded date of death during the follow-up period. In these instances, survival methods often assume non-informative censoring, meaning loss-to-follow-up is not for reasons related to the treatment received and provides no information on the patient’s subsequent likelihood of survival; however, in real-world data (RWD), this assumption may not always be true. This study explores the utility of abstracted data in defining the censor date and signaling the potential for informative censoring.
Three cancer types—small cell lung, advanced head & neck, and metastatic pancreatic—were analyzed with data available through 2017. For patients without recorded dates of death, two definitions of last confirmed activity date were evaluated: (1) last confirmed structured activity date (LCSAD) and (2) latest of LCSAD or last abstracted date (e.g., biomarker test date, abstracted treatment date). Using definition 2, the source of the abstracted data was then assessed for prognostic value.
Findings indicate a minority of censored patients experienced a change in censor date when abstracted data was considered. However, abstracted data sources such as last clinic note date, disease recurrence dates, and hospice referrals did inform the censor date. While the overall impact of abstracted data inclusion on the censor date was modest, abstracted data elements in certain cancer types provided prognostic information, underscoring the importance of further investigation to mitigate bias on estimates of rwOS.
Why this matters
Ensuring accurate results for rwOS for use in research is paramount. This study highlights the potential for EHR-derived real-world data’s ability to signal potential biases impacting the validity rwOS analyses. Deepening understanding of this bias across common real-world endpoints in oncology will contribute to higher-quality real-world evidence across use cases.