Presentation (CTI)

OBJECTIVES: Trials in oncology often adopt and report results with different analytical strategies for treatment switching. Such reporting complicates meta-analyses of overall survival (OS), since the evidence base may contain different treatment effects (estimands) even when trials are aligned in terms of population, intervention, comparator, and outcome (PICO). To mitigate mixed-estimand and mixed-data challenges, we propose a causal multistate framework to facilitate network meta-analyses (NMAs) of OS.
METHODS: We formulate a structural multistate NMA survival model with explicit representation of pre-progression disease, post-progression treatment switching, and death. Estimands are defined marginally with respect to first-line treatment assignment under two regimes: (i) treatment-policy estimand that reflects as observed switching behaviour, and (ii) hypothetical estimand that assumes no post-progression switching. Each trial arm is mapped to the estimand it informs. A single structural model is fitted using a unified time-to-event likelihood accommodating both individual patient data (IPD) and/or aggregate data via integration over arm-specific covariate distributions. NMA is performed on transition hazards, allowing synthesis across trials while preserving estimand consistency. Counterfactual survival functions under unobserved regimes are obtained through causal intervention on the switching intensity using parametric g-computation.
RESULTS: The framework delineates how OS evidence from trials reporting different treatment switching strategies can be coherently synthesized by aligning trial-level estimands. By structuring evidence according to treatment-policy and hypothetical regimes, the approach clarifies the contribution of heterogeneous trial arms. Prior simulation studies of treatment switching in oncology meta-analysis demonstrate that ignoring estimand heterogeneity can lead to biased and poorly calibrated estimates, motivating the proposed framework.
CONCLUSIONS: Our causal multistate framework offers a principled approach to synthesizing heterogeneous trial evidence while maintaining alignment between estimands, modeling assumptions, and policy-relevant questions. This approach supports interpretable NMAs in settings with treatment switching and mixed-data availability, with relevance for health technology assessment and reimbursement decision-making.