OBJECTIVES

: Recently, we developed a two-step network meta-analysis (NMA) for time-to-event data using alternative parametric distributions often used in health technology assessments (HTAs), including exponential, Weibull, Gompertz, log-normal, and log logistic. With these models the hazard ratio does not have to be assumed to be constant over time, thereby reducing the possibility of violating consistency in indirect comparisons. However, it is also possible to extend this approach to evaluate fractional polynomial distributions, which are increasingly being used in an HTA setting.

METHODS

First, for each arm of every randomized controlled trial (RCT) connected in the network of evidence simulated patient data were fit to alternative parametric distributions, including fixed and random effects first and second order fractional polynomials. Additionally, we compared these results to Weibull, Gompertz, exponential, log-normal, and log logistic. For each distribution, the resulting scale and shape parameters per arm were then included in a multivariate NMA, which preserved randomization and accounted for the correlation between the parameters.

RESULTS

: An illustrative analysis is presented for a network of RCTs evaluating interventions for advanced melanoma. The NMA was assessed for overall survival using alternative distributions, which were compared using Akaike information criterion (AIC), which can facilitate model averaging to propagate structural uncertainty in a cost-effectiveness analysis. Based on the AIC, fractional polynomial provides a good fitting alternative to the more traditional parametric distributions that can all be compared in a straightforward manner based on goodness of fit as well as clinical plausibility for each trial.

CONCLUSIONS

: A two-step NMA of survival data for fractional polynomials allows for a straightforward and efficient comparison of alternative models using the individual event times in the frequentist framework in the first step rather than an approximation based on discrete hazards in Bayesian framework. This approach provides a more generalizable evidence synthesis framework for HTA.