Use of Generalized F Parameterization in Long-Term Extrapolation for Adjuvant Cancer Therapies

OBJECTIVES: In the adjuvant oncology setting, there is often an initial high risk of relapse or death after surgery, but patients who remain relapse free for a longer period may experience outcomes similar to the general population. Flexible survival modelling approaches, such as spline or semi-parametric analysis, are required to provide good fits to observed data while predicting plausible long-term outcomes. In contrast to spline models, which rely on continuity assumptions and may not be limited to positive hazards, the generalized F (genF) function is a flexible, fully parametric extension to the most common survival distributions, and so shares the same extrapolative assumptions as these well-accepted models. As such, this approach was explored and applied in two recent NICE submissions for adjuvant therapies. This paper highlights key challenges associated with the genF distribution and describes potential solutions.

METHODS: The standard evaluation of the genF probability function can make invalid predictions during long-term extrapolation due to computing errors, typically visualized as a sharp truncation of the survival extrapolation. These errors can also prevent generation of an uncertainty matrix required for probabilistic sensitivity analysis, as required for health technology assessments (HTA).

RESULTS: An alternative evaluation of the genF probability distribution was developed to avoid these errors, improving accuracy of predictions and facilitating generation of uncertainty matrices. This alternative genF evaluation can be used to support economic modelling for HTA.

CONCLUSIONS: Although not a required approach for some HTA bodies, including in the UK, the genF distribution should be considered, particularly for adjuvant cancer therapies where survival data may be highly immature and require flexible approaches to extrapolation. This study provides solutions to the most common numerical challenges when using genF distributions.