Presentation (CTI)

OBJECTIVES: Health Technology Assessment (HTA) submissions usually require extrapolation of trial data to incorporate into health economic models to predict the long-term effects of an intervention based on short-term trial data. Model selection for extrapolation can be challenging due to different factors. In this abstract, we review methods to handle these challenges with a focus on proportional hazard assumption violation.
METHODS: The selection of an appropriate method for extrapolation relies on different factors such as trial data characteristics, and underlying assumptions about the survival distribution. The performance of standard extrapolation methods, traditional parametric models, such as Weibull and Log-Normal, and alternative models such as flexible parametric models (splines) were assessed based on different factors.
RESULTS: The comparison indicated that parametric methods may lack the flexibility to model the underlying hazard function shape and thereby may provide a poor fit to the data. However, alternative models such as flexible parametric models can provide greater flexibility and superior performance to model the data when restrictive assumptions from parametric models do not hold (e.g., proportional hazards [PH] assumption) as these models do not inherently assume PH).
CONCLUSIONS: When conducting extrapolation, there are various methods for extrapolating the long-term outcomes depending on the underlying assumptions about hazard rate and the data. The relative performance of these methods has not been evaluated. The evidence mentioned herein suggests that although there is limited guidance on how to explore and interpret alternative models for extrapolation of trial data, spline models provide more flexibility than standard parametric models, particularly when the relationship between time and the outcome (e.g., survival) is non-linear or the hazard function is unknown while parametric models remain useful for their simplicity.