Presentation (CTI)

OBJECTIVES: In cost-effectiveness analyses for oncology products, survival extrapolation often contributes the majority of incremental life-years and QALYs. Current reliance on partitioned survival analysis (PartSA) risks a “tail wagging the dog” problem, whereby long-run extrapolation behaviour outweighs observed survival. We revisit contemporary recommendations for PartSA, contrasting it with multi-state and cure-/responder-based models, and examine how structural choices shift evidential weight from data fit to extrapolated assumptions.
METHODS: We compared three modelling families: (i) PartSA, (ii) multi-state models (MSMs), and (iii) cure and responder/mixture approaches, with explicit reference to the balance between the data-supported period and the extrapolation horizon. For each, we assessed: (a) the proportion of events accruing beyond observed follow-up; (b) the degree to which post-progression and long-term survival are structurally implied rather than empirically estimated; (c) coherence of state occupancy over time; and (d) sensitivity of incremental results to alternative long-run assumptions under comparable in-sample fit. Particular attention was paid to delayed treatment effects, subsequent therapy, and durable response patterns increasingly observed in oncology.
RESULTS: PartSA, with its focus on statistical fit to the data, commonly results in the extrapolation period driving the analysis, especially when survival data are immature, allowing modest differences in tail behaviour to dominate results. MSMs and cure-/responder-based models with their increased structure can reduce tail dominance, but may not fit the observed data so well. Across modelling types, structurally similar in-sample fit frequently masked substantial divergence in extrapolated outcomes.
CONCLUSIONS: The “tail wagging the dog” is a structural risk in oncology modelling. HTA guidance should move beyond default endorsement of PartSA based on fit to the observed data and require explicit demonstration that extrapolated survival does not dominate. Routine reporting should quantify the contribution of extrapolation to incremental outcomes and benchmark results across alternative structural models.