The Effects of Pooling Treatment Effects Targeting Treatment Policy and Hypothetical Estimands With Rank-Preserving Structural Failure Time Model in Oncology Aggregate-Level Meta-Analyses
Author(s)
Rebecca Metcalfe, MA, PhD1, Shomoita Alam, PhD2, Antonio Remiro Azócar, PhD3, Richard Yan, MSc4, Jay JH Park, PhD1.
1Core Clinical Sciences, Vancouver, BC, Canada, 2Department of Epidemiology, Biostatistics, and Occupational Health, McGill University, Montreal, QC, Canada, 3Novo Nordisk, Madrid, Spain, 4Department of Statistical and Actuarial Science, Simon Fraser University, Vancouver, BC, Canada.
1Core Clinical Sciences, Vancouver, BC, Canada, 2Department of Epidemiology, Biostatistics, and Occupational Health, McGill University, Montreal, QC, Canada, 3Novo Nordisk, Madrid, Spain, 4Department of Statistical and Actuarial Science, Simon Fraser University, Vancouver, BC, Canada.
OBJECTIVES: The implications of the estimands framework, which emphasizes the importance of post-randomization (intercurrent) events and their analytical strategies in randomized clinical trials (RCTs), have been under-explored for meta-analysis. For RCTs in oncology, rank-preserving structural failure time modelling (RPSFTM) is often recommended to adjust for the intercurrent event of treatment switching, since failure to account for treatment switching can produce misleading results for overall survival (OS). Using simulations, we examined the bias and coverage that can be caused by combining trial evidence that estimates different target estimands in a meta-analysis of RCTs.
METHODS: We simulated OS data for eight RCTs that allowed patients in the control group to switch to the intervention treatment after disease progression. We estimated treatment effects that ignored treatment switching (treatment policy estimand) and another that accounted for switching with RPSFTM (hypothetical estimand). These results were pooled via aggregate-level meta-analyses with varying the proportions of treatment policy and hypothetical effect estimates.
RESULTS: On average, meta-analyses pooling only hypothetical estimates derived via RPSFTM produced larger treatment effects than the meta-analyses that pooled only the treatment policy estimates. This was consistently observed across all scenarios with different randomization ratios and treatment switching rates. Bias and coverage were directly influenced by the concordance between the estimands pooled and the meta-analytic target, with greater discordance resulting in more bias and poorer coverage.
CONCLUSIONS: We found that combining treatment-policy and RPSFTM hypothetical estimates yields pooled effects that correspond to neither estimand, potentially leading to misleading conclusions even with random-effects models. Applying the estimands framework and ensuring that trial-level estimands are aligned with meta-analysis should improve the relevance and validity of synthesized evidence.
METHODS: We simulated OS data for eight RCTs that allowed patients in the control group to switch to the intervention treatment after disease progression. We estimated treatment effects that ignored treatment switching (treatment policy estimand) and another that accounted for switching with RPSFTM (hypothetical estimand). These results were pooled via aggregate-level meta-analyses with varying the proportions of treatment policy and hypothetical effect estimates.
RESULTS: On average, meta-analyses pooling only hypothetical estimates derived via RPSFTM produced larger treatment effects than the meta-analyses that pooled only the treatment policy estimates. This was consistently observed across all scenarios with different randomization ratios and treatment switching rates. Bias and coverage were directly influenced by the concordance between the estimands pooled and the meta-analytic target, with greater discordance resulting in more bias and poorer coverage.
CONCLUSIONS: We found that combining treatment-policy and RPSFTM hypothetical estimates yields pooled effects that correspond to neither estimand, potentially leading to misleading conclusions even with random-effects models. Applying the estimands framework and ensuring that trial-level estimands are aligned with meta-analysis should improve the relevance and validity of synthesized evidence.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR197
Topic
Clinical Outcomes, Epidemiology & Public Health, Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
Oncology