6F-03 - Immortal Time Bias Impacts Time-To-Event Outcomes in Cohort with Genomic Testing

Background: Immortal time bias is a critical bias to consider in the context of time-to-event analyses in cohorts with genomic testing.

Objectives: To assess the presence and magnitude of potential immortal time bias when estimating survival in a group of cancer patients identified by genomic testing.

Methods: We leveraged a de-identified nationwide (US-based) non-small cell lung cancer (NSCLC) clinico-genomic database to estimate real-world survival (rwOS) in patients with advanced NSCLC, diagnosed between 2011 and 2019, who tested positive for the KRAS G12C mutation. The results of two methods for addressing immortal time bias (the restricted method and Left Truncated adjustment method) are discussed, implemented, and compared to the unadjusted results for rwOS. The restricted method excluded patients whose genomic testing occurred > 21 days following the start of front line therapy to remove most patients with immortal time, while the Left Truncated method adjusted for time at risk in the model.

Results: When assessed from the start of front-line therapy without accounting for immortal time, the median rwOS was 17.6 months (95% confidence interval [CI]: 15.3-19.8) with 77% (95% CI 0.73 - 0.81) surviving 6 months and 60% (95% CI: 0.56 - 0.65) surviving 12 months. In the restricted method controlling for ITB, median rwOS was 12.0 months (95% CI: 9.6 - 15.3) with 68% (95% CI: 0.62 - 0.73) surviving 6 months and 49% (95% CI: 0.43 - 0.56) surviving 12 months. Finally, the Left Truncated model adjusting for time at risk estimated a median rwOS of 10.2 months (95% CI: 8.6 - 12.8) with 65% (95% CI: 0.60 - 0.71) surviving 6 months and 45% (95% CI: 0.40 - 0.51) surviving 12 months.

Conclusions: In this example, both approaches adjusting for immortal time bias produced similar results. When immortal time bias was ignored, rwOS was overestimated in this population of patients with genomic-sequenced, advanced NSCLC, highlighting the importance of adjusting for time-to-event outcome in analyses of populations who undergo genomic testing.