FRACTIONAL POLYNOMIAL NMA MODELS FOR SURVIVAL ANALYSES- RESULTS FROM A SIMULATION STUDY

OBJECTIVES:

When conducting a Network Meta-Analysis (NMA) of survival data, the proportional hazard assumption has been increasingly challenged, especially in oncology. The fractional polynomial NMA model has been proposed to tackle this issue and has been used in HTA submissions. However, using this type of model in practice raises several methodological and computational issues. Our aim was to assess the behaviour of the fractional polynomial NMA based on number of included data points and time windows for corresponding hazards.

METHODS:

Simulation studies were conducted to investigate the behaviour of the fractional polynomial model. First, the influence of the number of data-points from the Kaplan-Meier curves was investigated, aiming to provide guidance on the number of data points to optimise model convergence while minimising computation time. In a second step, model robustness to different time windows selected for the corresponding hazards was assessed.

RESULTS:

Simulation results showed a decreasing marginal benefit of adding further data points, whereas computation time increased exponentially and the local hazard rate estimation noise increased. Optimal number of points should be selected based on available running time and sample size of included trials. A sample of ten to fifteen data points provided a good trade-off between computation time, uncertainty reduction and local hazard rate estimations across most simulations. The model was robust to time window specification, which should be selected based on each study and not across studies.

CONCLUSIONS:

In conclusion, the fractional polynomial NMA model has proven to be a valuable method allowing incorporation of a fully flexible time-dependent hazard function, addressing an important source of structural uncertainty in NMAs for survival outcomes. This study provides an overview of the model behaviour in various scenarios. However, additional research is needed to provide guidance on how to integrate such models into cost-effectiveness model.