Optimizing the Use of Excess Hazard Models in Survival Analysis: Considering Disease Severity and Follow-up Maturity
Author(s)
Jessica E. Forsyth, PhD1, Shubhodeep Mitra, MSc2, Tushar Srivastava, MSc3, Kate Ren, PhD4.
1Sheffield Centre for Health and Related Research (SCHARR), University of Sheffield, Sheffield, United Kingdom, 2ConnectHEOR Limited, Delhi, India, 3ConnectHEOR Limited, London, United Kingdom, 4University of Sheffield | ConnectHEOR Limited, Sheffield | London, United Kingdom.
1Sheffield Centre for Health and Related Research (SCHARR), University of Sheffield, Sheffield, United Kingdom, 2ConnectHEOR Limited, Delhi, India, 3ConnectHEOR Limited, London, United Kingdom, 4University of Sheffield | ConnectHEOR Limited, Sheffield | London, United Kingdom.
OBJECTIVES: Survival extrapolation is a critical component of Health Technology Assessment (HTA) and often conducted using standard parametric models, which do not account for background population mortality. This can lead to implausible hazard projections falling below age- and sex-matched mortality rates. In economic models, this is sometimes corrected by using artificial “clamping”, which introduces artificial discontinuities in the hazard. Excess hazard (EH) models offer an alternative approach by explicitly incorporating background mortality into survival projections. This study explores when EH models are most appropriate and how much follow-up is needed for reliable results.
METHODS: Several case studies were selected to represent diseases of varying severity, defined by differences in expected overall survival (OS). Standard parametric and additive EH models were applied to each. Published Kaplan-Meier OS curves were digitised, and individual patient data (IPD) were reconstructed. The impact of limited follow-up was assessed by artificially truncating survival data to simulate shorter observation periods.
RESULTS: In high-severity diseases with short survival, EH models produced results comparable to standard parametric models, as disease-specific hazards dominated background mortality. In contrast, EH models often failed to converge or yielded unstable estimates in low-severity settings where disease-related hazards were low. In cases where background mortality contributed notably to the overall hazards, EH models generated narrower and more stable extrapolations than standard models, particularly when at least 24 months of follow-up data were available. However, when follow-up was limited to 6 and 12 months, model variability and uncertainty increased markedly.
CONCLUSIONS: EH models are most valuable in settings where background mortality meaningfully contributes to the overall hazard. However, its use requires sufficient follow-up data to ensure model stability and accuracy. This work highlights specific contexts where EH modelling may provide benefits within HTA, despite limitations associated with reconstructed IPD.
METHODS: Several case studies were selected to represent diseases of varying severity, defined by differences in expected overall survival (OS). Standard parametric and additive EH models were applied to each. Published Kaplan-Meier OS curves were digitised, and individual patient data (IPD) were reconstructed. The impact of limited follow-up was assessed by artificially truncating survival data to simulate shorter observation periods.
RESULTS: In high-severity diseases with short survival, EH models produced results comparable to standard parametric models, as disease-specific hazards dominated background mortality. In contrast, EH models often failed to converge or yielded unstable estimates in low-severity settings where disease-related hazards were low. In cases where background mortality contributed notably to the overall hazards, EH models generated narrower and more stable extrapolations than standard models, particularly when at least 24 months of follow-up data were available. However, when follow-up was limited to 6 and 12 months, model variability and uncertainty increased markedly.
CONCLUSIONS: EH models are most valuable in settings where background mortality meaningfully contributes to the overall hazard. However, its use requires sufficient follow-up data to ensure model stability and accuracy. This work highlights specific contexts where EH modelling may provide benefits within HTA, despite limitations associated with reconstructed IPD.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
P25
Topic
Methodological & Statistical Research
Disease
Oncology