Transporting Treatment Effects for Survival Outcomes Under Informative Censoring: A Simulation Study Using Summary-Level Data in the Target Population
Author(s)
Richard Yan, MSc1, Rebecca Metcalfe, BA, MA, PhD2, Tianyu Guan, PhD3, Haolun Shi, PhD4, Jay JH Park, PhD2.
1Department of Statistical and Actuarial Sciences, Simon Fraser University, Vancouver, BC, Canada, 2Core Clinical Sciences, Vancouver, BC, Canada, 3Mathematics and Statistics, York University, Toronto, ON, Canada, 4Department of Statistical and Actuarial Science, Simon Fraser University, Vancouver, BC, Canada.
1Department of Statistical and Actuarial Sciences, Simon Fraser University, Vancouver, BC, Canada, 2Core Clinical Sciences, Vancouver, BC, Canada, 3Mathematics and Statistics, York University, Toronto, ON, Canada, 4Department of Statistical and Actuarial Science, Simon Fraser University, Vancouver, BC, Canada.
OBJECTIVES: Transportability analysis can be used to assess the external validity of a source study for a target population. Most transportability analysis methods require access to patient-level data for both the source and target population. Transportability analyses that can use aggregate-level data (AgD) remains under-explored, especially for survival analyses that require adjustment for informative censoring. Using a simulation study and an application, we explored a novel two-stage weighting method called “Target Aggregate Data Adjustment” (TADA) that combines participation weights estimated via the method of moments (MoM) with time-varying inverse probability censoring weights.
METHODS: We conducted simulations to assess the performance (bias and coverage) of TADA compared MoM-based methods without adjustment for informative censoring. We also applied our TADA methods to a real case-study of a non-small-cell lung cancer trial (NCT00981058).
RESULTS: In general, when informative censoring and effect modifications were present, TADA with adjustment for time-varying censoring demonstrated superior performance in terms of bias and coverage than the MoM-based approach alone. The benefits of TADA became most prominent between overall censoring rates of 20-30%. For instance, under a 20% censoring rate, the adjusted bias is 0.037 versus an unadjusted bias of 0.128. However, the advantages became less pronounced when the censoring rates reached 50%. In the applied case-study, an increased estimate of overall survival when applying TADA illustrates how the unadjusted censoring and covariate distribution imbalance may lead to different clinical conclusions when transporting clinical trial outcomes to real-world populations.
CONCLUSIONS: Accounting for informative censoring via TADA can greatly improve both control of bias and estimator coverage compared to the MoM-based approach without adjustment for censoring. However, when censoring rates exceeded 50%, the benefit of TADA became less prominent, and both TADA and censoring unadjusted MoM-based approach performed poorly.
METHODS: We conducted simulations to assess the performance (bias and coverage) of TADA compared MoM-based methods without adjustment for informative censoring. We also applied our TADA methods to a real case-study of a non-small-cell lung cancer trial (NCT00981058).
RESULTS: In general, when informative censoring and effect modifications were present, TADA with adjustment for time-varying censoring demonstrated superior performance in terms of bias and coverage than the MoM-based approach alone. The benefits of TADA became most prominent between overall censoring rates of 20-30%. For instance, under a 20% censoring rate, the adjusted bias is 0.037 versus an unadjusted bias of 0.128. However, the advantages became less pronounced when the censoring rates reached 50%. In the applied case-study, an increased estimate of overall survival when applying TADA illustrates how the unadjusted censoring and covariate distribution imbalance may lead to different clinical conclusions when transporting clinical trial outcomes to real-world populations.
CONCLUSIONS: Accounting for informative censoring via TADA can greatly improve both control of bias and estimator coverage compared to the MoM-based approach without adjustment for censoring. However, when censoring rates exceeded 50%, the benefit of TADA became less prominent, and both TADA and censoring unadjusted MoM-based approach performed poorly.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR206
Topic
Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
Oncology