The NICE Decision Support Unit has recently published technical support document 21, focusing on flexible survival methods and considering excess mortality. In this study, we research these flexible survival methods by assessing the effect of treatment on the different parameters of the distributions and the implementation of these flexible excess survival models in network meta-analyses (NMA).
Standard parametric, mixture- (MCM) and non-mixture-cure (nMCM), piecewise, and splines are used for comparison. For the base case model, selection treatment coefficients were assigned to all parameters. The lowest leave-one-out information criterion (LOOIC) defined the base case parametric distribution per flexible model. The best-fitting distribution per model was rerun with treatment coefficients only for those larger than their standard deviation, and the one with the lowest LOOIC was considered the base case. Models were compared based on incremental mean survival and uncertainty. We used a network of four trials in previously treated advanced non-small cell lung cancer, comparing nivolumab vs. docetaxel, pembrolizumab vs. docetaxel, and two trials comparing atezolizumab vs. docetaxel.
The base case parametric distributions were log-logistic for standard parametric distribution, log-logistic for MCM, lognormal for nMCM, Weibull for mixture, log-logistic for piecewise, and log-logistic for spline models, with corresponding LOOICs 13418, 13417, 13420, 13417, 13418, and 13419. Regarding treatment effect parametrization, we applied no treatment effect on cure, for instance, but treatment effects on the scale for MCM and nMCMs. For piecewise, treatment effects were only applied on the second piece. After reducing the number of treatment effects, the mixture Weibull had the lowest LOOIC with 13407. The mean incremental survival with docetaxel as reference for mixture Weibull was 3.58[95%CI 1.2,5.26], 0.45[0.19,0.93] and 2.67[0.2,9.48] for nivolumab, pembrolizumab and atezolizumab, respectively. For standard log-logistic these were 1.67[0.94,2.55],1.01[0.55,1.61], and 0.43[0.17,0.77].
Treatment effect specification is important as outcomes and uncertainty differ over the tested models.
© 2021 Published by Elsevier Inc.