COMPARISON OF NAIVE VERSUS DELAYED-ENTRY ADJUSTMENT METHODS IN REAL-WORLD SURVIVAL ANALYSIS OF FIRST-LINE METASTATIC NSCLC WITH PD-L1 TPS =50%
Author(s)
Samina Dhuliawala, MS1, Ryan Thaliffdeen, PharmD, MS2.
1IQVIA Inc., Durham, NC, USA, 2Associate Director, GHEOR, Gilead Sciences, Foster City, CA, USA.
1IQVIA Inc., Durham, NC, USA, 2Associate Director, GHEOR, Gilead Sciences, Foster City, CA, USA.
OBJECTIVES: Immortal time bias may arise in observational analyses when patients must survive long enough to undergo sequencing. This study evaluated how naive versus delayed-entry adjusted approaches influence overall survival (OS) and progression-free survival (PFS) in first-line (1L) metastatic non-small cell lung cancer (NSCLC) patients with PD-L1 tumor proportion score (TPS) ≥50% and no actionable genomic aberrations.
METHODS: This retrospective study used the Tempus Multimodal dataset (2016-2025), with index date of 1L metastatic therapy initiation. A naive approach was compared with a delayed-entry adjusted approach using risk-set adjustment (RSA), which accounts for left truncation by allowing entry at sequencing and relies on the independent delayed-entry assumption that time of sequencing is not associated with death. If the assumption was violated, naive models were compared with a prospective approach, excluding patients sequenced after treatment initiation. Kaplan-Meier methods and Cox proportional hazards models were used.
RESULTS: For OS-based models, the independent delayed-entry assumption was violated, and prospective-adjusted methods were employed. For OS, naive analysis (N=480) yielded a median of 19.30 months (95% CI: 16.96-21.6), while prospective analysis (N=368) yielded 20.52 months (95% CI: 17.75-23.67). For PFS, naive analysis (N=435) yielded a median of 8.52 months (95% CI: 7.46-9.9), compared with 8.71 months (95% CI: 7.53-10.49) using RSA (N=394). In OS-based Cox models, most factors significant in naive analyses (1L chemotherapy, sex, TPS >90%) lost significance after prospective adjustment, except ECOG performance status ≥2. In PFS-based Cox models, all factors significant in naive analyses retained significance under RSA.
CONCLUSIONS: While PFS estimates and associated predictors were largely unchanged, OS analyses showed greater sensitivity to adjustment, particularly in multivariable modeling. The observed differences between naive and delayed-entry adjusted analyses suggest sequencing-related delayed entry influenced time-to-event estimation, supporting routine use of adjustment approaches to assess the impact of immortal time bias.
METHODS: This retrospective study used the Tempus Multimodal dataset (2016-2025), with index date of 1L metastatic therapy initiation. A naive approach was compared with a delayed-entry adjusted approach using risk-set adjustment (RSA), which accounts for left truncation by allowing entry at sequencing and relies on the independent delayed-entry assumption that time of sequencing is not associated with death. If the assumption was violated, naive models were compared with a prospective approach, excluding patients sequenced after treatment initiation. Kaplan-Meier methods and Cox proportional hazards models were used.
RESULTS: For OS-based models, the independent delayed-entry assumption was violated, and prospective-adjusted methods were employed. For OS, naive analysis (N=480) yielded a median of 19.30 months (95% CI: 16.96-21.6), while prospective analysis (N=368) yielded 20.52 months (95% CI: 17.75-23.67). For PFS, naive analysis (N=435) yielded a median of 8.52 months (95% CI: 7.46-9.9), compared with 8.71 months (95% CI: 7.53-10.49) using RSA (N=394). In OS-based Cox models, most factors significant in naive analyses (1L chemotherapy, sex, TPS >90%) lost significance after prospective adjustment, except ECOG performance status ≥2. In PFS-based Cox models, all factors significant in naive analyses retained significance under RSA.
CONCLUSIONS: While PFS estimates and associated predictors were largely unchanged, OS analyses showed greater sensitivity to adjustment, particularly in multivariable modeling. The observed differences between naive and delayed-entry adjusted analyses suggest sequencing-related delayed entry influenced time-to-event estimation, supporting routine use of adjustment approaches to assess the impact of immortal time bias.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR29
Topic
Methodological & Statistical Research, Real World Data & Information Systems, Study Approaches
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
Oncology