Accounting for the Additional Uncertainty Following Rank-Preserving Structural Failure Time Model (RPSFTM) Crossover Adjustment in Survival Analysis: A Sage Approach
Author(s)
Rebecca K. Judge, MSc1, Rachel H. Tao, MPH1, Tristan H. Curteis, MSc2.
1Costello Medical, London, United Kingdom, 2Costello Medical, Manchester, United Kingdom.
1Costello Medical, London, United Kingdom, 2Costello Medical, Manchester, United Kingdom.
OBJECTIVES: Crossover in randomised control trials (RCTs) occurs when patients switch from their allocated treatment, which may bias results from an intention-to-treat (ITT) analysis. The Rank Preserving Structural Failure Time Model (RPSFTM) approach is a common method to account for the impact of crossover on survival outcomes. When estimating RPFSTM-adjusted hazard ratios (HRs), the additional uncertainty introduced by RPSFTM must be captured. Conventional approaches include retaining the ITT p-value (adjusting the HR standard error to preserve the ITT p-value) and bootstrapping (repeating RPSFTM for resampled datasets). The objective of the present work was to evaluate whether our new approach denoted ‘Sampling After G-Estimation’ (SAGE) can more efficiently capture the uncertainty in RPSFTM-adjusted HR estimation.
METHODS: We applied RPSFTM to account for crossover using a simulated dataset based on a historic RCT. Our proposed approach, SAGE, derived confidence intervals (CIs) for the RPSFTM-adjusted HR by sampling (assuming a Normal distribution) the test-statistics from the RPSFTM g-estimation process. Sampled test statistics were mapped to produce multiple adjusted survival times and corresponding HRs and CI limits, with the limits taken as the 2.5th and 97.5th percentiles over all calculated lower and upper CIs, respectively. This SAGE CI was compared with a: naïve CI without uncertainty inflation; CI retaining the ITT p-value; and CI using bootstrapping.
RESULTS: Compared with the naïve CI (0.613, 0.959), retaining the ITT p-value resulted in limited uncertainty inflation (0.614, 0.958), whereas the SAGE approach moderately increased the width of the CI (0.592, 1.006) and the bootstrapping approach led to the widest CI of the four approaches (0.505, 1.150).
CONCLUSIONS: Our SAGE method has the potential to more fully reflect the uncertainty in RPSFTM than retaining the ITT p-value; additionally, whilst bootstrapping results are more conservative, SAGE is less computationally intensive to conduct. Further work should assess performance across multiple datasets.
METHODS: We applied RPSFTM to account for crossover using a simulated dataset based on a historic RCT. Our proposed approach, SAGE, derived confidence intervals (CIs) for the RPSFTM-adjusted HR by sampling (assuming a Normal distribution) the test-statistics from the RPSFTM g-estimation process. Sampled test statistics were mapped to produce multiple adjusted survival times and corresponding HRs and CI limits, with the limits taken as the 2.5th and 97.5th percentiles over all calculated lower and upper CIs, respectively. This SAGE CI was compared with a: naïve CI without uncertainty inflation; CI retaining the ITT p-value; and CI using bootstrapping.
RESULTS: Compared with the naïve CI (0.613, 0.959), retaining the ITT p-value resulted in limited uncertainty inflation (0.614, 0.958), whereas the SAGE approach moderately increased the width of the CI (0.592, 1.006) and the bootstrapping approach led to the widest CI of the four approaches (0.505, 1.150).
CONCLUSIONS: Our SAGE method has the potential to more fully reflect the uncertainty in RPSFTM than retaining the ITT p-value; additionally, whilst bootstrapping results are more conservative, SAGE is less computationally intensive to conduct. Further work should assess performance across multiple datasets.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR12
Topic
Methodological & Statistical Research
Disease
No Additional Disease & Conditions/Specialized Treatment Areas