Immortal time bias reproduces the reported survival benefit of conversion surgery in stage IV gastric cancer: a simulation study
- 1Ruijin Hospital, Shanghai Jiaotong University School of Medicine
Abstract
BackgroundConversion surgery for stage IV gastric cancer is supported by a pooled overall survival hazard ratio of 0.36 (95% confidence interval 0.32-0.40) and, in the largest international cohort, median survival of 36.7 versus 12.5-13.8 months on chemotherapy. Survival is measured from diagnosis; the median diagnosis-to-gastrectomy interval is 124 days, which patients must survive to be counted surgical.
MethodsWe simulated cohorts of 3,177 stage IV gastric cancer patients from published parameters: background median survival 14.5 months; median diagnosis-to-surgery interval 124 days (category-specific 92-174 days). Surgery had no effect (true hazard ratio 1.00 by construction). Data were analysed as the literature analyses them (exposure fixed at baseline, follow-up from diagnosis), and by time-varying Cox and landmark analysis. Confounding by indication was added in a second scenario.
ResultsUnder immortal time bias alone the naive analysis returned a hazard ratio of 0.794 (95% simulation interval 0.743-0.851), median survival 16.8 versus 12.8 months. Time-varying Cox recovered 1.000 and landmark analysis 1.000-1.004. Bias scaled with the interval: 0.849 at 92 days, 0.715 at 174 days. Adding confounding, the naive estimate fell to 0.601 (0.560-0.644) at strength 0.5 and 0.356 (0.323-0.385) at strength 1.5, overlapping the published estimate; median survival 21.9 versus 8.7 months. Correcting immortal time alone left residual bias (hazard ratio 0.439).
ConclusionsThe reported survival advantage of conversion surgery is reproducible where the operation does nothing; published estimates cannot distinguish benefit from bias. Resolving this requires individual patient data analysed with methods that assign person-time correctly, or completion of JCOG2301.
Related
- Predicting MCI-to-Alzheimer’s Conversion: A Systematic Benchmark of Classical, Machine Learning, and Deep Survival Models
- Simulated Placement and Rounds (SPaR): Improving Self-Reported Preparedness for Foundation Training through Ward-Round-Based Simulation
- Addressing Measurement Error of Machine-Learned Physical Activity in Nonlinear Dose-Response Survival Analysis: Development and Evaluation of Accelerated Failure Time, Spline, and Simulation-Extrapolation Method
- Survival benefits and prognostic models of surgical resection versus continued medical treatment in CNLC stage IIb/IIIa hepatocellular carcinoma after meeting resectability criteria following conversion therapy
- Supporting clinical guidelines for opioid conversion to methadone and tapering to prevent withdrawal in critically ill children using physiology based pharmacokinetic modeling and simulation
BibTeX
@misc{sah2026immortal,
title = {{Immortal time bias reproduces the reported survival benefit of conversion surgery in stage IV gastric cancer: a simulation study}},
author = {B. K. Sah and C. Li and J. Li and Z. Zhu},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.01.26361986},
url = {https://www.medrxiv.org/content/10.64898/2026.09.01.26361986v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Immortal time bias reproduces the reported survival benefit of conversion surgery in stage IV gastric cancer: a simulation study
AU - B. K. Sah
AU - C. Li
AU - J. Li
AU - Z. Zhu
PY - 2026
DA - 2026/09/01
PB - medRxiv
DO - 10.64898/2026.09.01.26361986
UR - https://www.medrxiv.org/content/10.64898/2026.09.01.26361986v1
AB - BackgroundConversion surgery for stage IV gastric cancer is supported by a pooled overall survival hazard ratio of 0.36 (95% confidence interval 0.32-0.40) and, in the largest international cohort, median survival of 36.7 versus 12.5-13.8 months on chemotherapy. Survival is measured from diagnosis; the median diagnosis-to-gastrectomy interval is 124 days, which patients must survive to be counted surgical. MethodsWe simulated cohorts of 3,177 stage IV gastric cancer patients from published parameters: background median survival 14.5 months; median diagnosis-to-surgery interval 124 days (category-specific 92-174 days). Surgery had no effect (true hazard ratio 1.00 by construction). Data were analysed as the literature analyses them (exposure fixed at baseline, follow-up from diagnosis), and by time-varying Cox and landmark analysis. Confounding by indication was added in a second scenario. ResultsUnder immortal time bias alone the naive analysis returned a hazard ratio of 0.794 (95% simulation interval 0.743-0.851), median survival 16.8 versus 12.8 months. Time-varying Cox recovered 1.000 and landmark analysis 1.000-1.004. Bias scaled with the interval: 0.849 at 92 days, 0.715 at 174 days. Adding confounding, the naive estimate fell to 0.601 (0.560-0.644) at strength 0.5 and 0.356 (0.323-0.385) at strength 1.5, overlapping the published estimate; median survival 21.9 versus 8.7 months. Correcting immortal time alone left residual bias (hazard ratio 0.439). ConclusionsThe reported survival advantage of conversion surgery is reproducible where the operation does nothing; published estimates cannot distinguish benefit from bias. Resolving this requires individual patient data analysed with methods that assign person-time correctly, or completion of JCOG2301.
ER -