Candidate for the Lee B. Lusted Student Prize Competition
Method: We describe methods that allow external information on various different survival metrics to be combined with trial data to put constraints on the resulting estimated survival model. We indicate which external sources may potentially provide relevant evidence. RCT and external data were fitted simultaneously within a Bayesian framework. Deviance Information Criterion (DIC) and graphical fit against Kaplan-Meier (KM) estimates were monitored to balance the fit between unconstrained and constrained models. The difference in expected survival times between arms, with its 95% credible interval, was also recorded to give an idea of the potential impact of the method on CEA results.
Result: We illustrate the methods using a RCT of cetuximab + radiotherapy vs. radiotherapy alone for patients with head and neck cancer. A US cancer database (SEER), a meta-analysis, mortality statistics from the general population and expert opinion were used to impose constraints on overall survival, conditional survival, and hazard ratio. The external data allowed us to rule out certain parametric models and fit flexible spline models beyond the length of the trial follow-up. In the final model, goodness-of-fit to all sources of evidence was achieved within one integrated model.
Conclusion: We have demonstrated that external data can be used for model choice and estimation of survival curves from RCTs. We believe that clinical experts are essential to guide how the external data sources constrain the survival models fitted. We also found that appropriate sets of initial values were critical to achieve convergence of the model; however, these sets could be very limited and precise. Further work should be undertaken to optimise the selection process of initial values.