IMPACT OF ESTIMAND MISALIGNMENT IN INDIRECT TREATMENT COMPARISONS OF ONCOLOGY TRIALS WITH TREATMENT SWITCHING - A SIMULATION STUDY
Author(s)
Thilo Welz, Dr.1, Anna-Lena Roth, B.Sc.2, Alessandro Ghiretti, PhD3, Natalie Dennis, M.Sc.4, Fabio Pellegrini, M.Sc.5.
1Daiichi Sankyo Europe GmbH, Munich, Germany, 2Institute of Mathematics, University of Augsburg, Augsburg, Germany, 3Data and Statistical Sciences Centre for RWE and EG, Daiichi Sankyo Italia, Roma, Italy, 4Data and Statistical Sciences Centre for RWE and EG, Daiichi Sankyo Oncology France, Rueil-Malmaison, France, 5Data and Statistical Sciences Centre for RWE and EG, Daiichi Sankyo España, Madrid, Spain.
1Daiichi Sankyo Europe GmbH, Munich, Germany, 2Institute of Mathematics, University of Augsburg, Augsburg, Germany, 3Data and Statistical Sciences Centre for RWE and EG, Daiichi Sankyo Italia, Roma, Italy, 4Data and Statistical Sciences Centre for RWE and EG, Daiichi Sankyo Oncology France, Rueil-Malmaison, France, 5Data and Statistical Sciences Centre for RWE and EG, Daiichi Sankyo España, Madrid, Spain.
OBJECTIVES: Indirect treatment comparisons (ITCs) are widely used in health technology assessments when head-to-head evidence is unavailable. However, misalignment of target estimands across randomized controlled trials (RCTs) may introduce bias and compromise interpretability. A key driver of such misalignment is the handling of intercurrent events. Following the ICH E9(R1) estimand framework, interest has grown in understanding its implications for evidence synthesis. In oncology, treatment switching is a common intercurrent event, and overall survival (OS) may target either treatment policy or hypothetical estimands. The bias arising when ITCs combine trials using different strategies for treatment switching remains unclear. This study aims to quantify this bias.
METHODS: We conducted a Monte Carlo simulation of an anchored ITC comparing two RCTs based on the PROfound trial, with differing strategies for handling treatment switching. Outcomes were analyzed using network meta-analysis (NMA) and multilevel network meta-regression (ML-NMR) without covariate adjustment. Our primary trial targeted a treatment policy estimand. Hypothetical estimands were estimated for the competitor trial using (1) censoring at treatment switch and (2) rank-preserving structural failure time models (RPSFTM) with re-censoring. This work extends previous research by Metcalfe et al. (2026) conducted in the context of meta-analysis.
RESULTS: ITCs combining misaligned estimands produced biased estimates. Bias increased with higher proportions of treatment switching and remained stable across RCT sample sizes. For 50% treatment switching the OS HR bias was between -0.14 and 0.13 and for 75% treatment switching bias was between -0.3 and 0.3, depending on setting. Censoring-based approaches showed lower bias than RPSFTM-based methods. No consistent performance advantage was observed between NMA and ML-NMR.
CONCLUSIONS: Misalignment of estimands in ITCs can substantially bias results, particularly in the presence of treatment switching. Methodological adjustments may mitigate but not eliminate this bias, underscoring the importance of estimand alignment in evidence synthesis for HTA.
METHODS: We conducted a Monte Carlo simulation of an anchored ITC comparing two RCTs based on the PROfound trial, with differing strategies for handling treatment switching. Outcomes were analyzed using network meta-analysis (NMA) and multilevel network meta-regression (ML-NMR) without covariate adjustment. Our primary trial targeted a treatment policy estimand. Hypothetical estimands were estimated for the competitor trial using (1) censoring at treatment switch and (2) rank-preserving structural failure time models (RPSFTM) with re-censoring. This work extends previous research by Metcalfe et al. (2026) conducted in the context of meta-analysis.
RESULTS: ITCs combining misaligned estimands produced biased estimates. Bias increased with higher proportions of treatment switching and remained stable across RCT sample sizes. For 50% treatment switching the OS HR bias was between -0.14 and 0.13 and for 75% treatment switching bias was between -0.3 and 0.3, depending on setting. Censoring-based approaches showed lower bias than RPSFTM-based methods. No consistent performance advantage was observed between NMA and ML-NMR.
CONCLUSIONS: Misalignment of estimands in ITCs can substantially bias results, particularly in the presence of treatment switching. Methodological adjustments may mitigate but not eliminate this bias, underscoring the importance of estimand alignment in evidence synthesis for HTA.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR150
Topic
Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
Oncology