Presentation (CTI)

OBJECTIVES: Treatment switching is common in oncology trials, yet network meta-analyses (NMAs) rarely account for the differential contamination it introduces across arms, distorting survival estimates and potentially reversing treatment rankings. We developed and validated a parametric multistate g-computation framework to recover hypothetical estimands - the treatment effects had no switching occurred - from networks mixing individual patient data (IPD) and aggregate data (AgD).
METHODS: We specified a four-state semi-Markov illness-death model with parametric transition intensities, representing control-arm crossover as an explicit post-progression state. A joint likelihood accommodated both IPD and AgD trials through marginalized likelihoods, and the hypothetical regime was recovered by setting the switching transition to zero. The simulation study was pre-specified using the ADEMP framework, with estimands and performance measures defined a priori. Framework performance was assessed across 500 simulation replicates of a three-arm network (Drugs B and C versus a common control, Treatment A) under heavy differential switching (50% versus 75%), with Drug C designated as the truly superior treatment. Robustness under parametric misspecification will be examined as a planned exploratory secondary analysis.
RESULTS: Differential switching markedly compressed apparent survival gains relative to the common control, and standard NMA propagated this distortion, incorrectly concluding that Drug B was the superior active therapy. The proposed framework eliminated this artifact, recovering direct and indirect hypothetical estimands with near-zero bias and nominal 95% confidence interval coverage across all 500 replicates, restoring the correct ranking in favour of Drug C.
CONCLUSIONS: Estimand heterogeneity across trials can invert clinical conclusions with direct consequences for reimbursement decisions. This framework offers a principled, validated approach to adjusting for treatment switching without IPD, grounded in the causal structure of the underlying disease process. Simulation findings highlight the importance of estimand alignment in oncology NMAs, particularly when switching rates differ meaningfully across trials.