OBJECTIVES: Unlike in partitioned survival models, competing risks present in multistate models (MSMs) mean that using Akaike information criterion (AIC)/ Bayesian information criterion (BIC) to determine best statistical fits (NICE TSD 14) is sometimes inappropriate. This paper aims to explore several alternative approaches to determine the best combination of parametric distributions to use.

METHODS: Individual patient-level data of the comparator arm (gemcitabine and cisplatin) of the SQUIRE trial conducted in non-small-cell lung cancer was obtained from Project Data Sphere (an open-access data-sharing platform). Time to death was split into three transitions (time to progression [TTP], pre-progression survival [PrePS], and post-progression survival incorporating two tunnel states [PPS1 and PPS2]). Standard parametric distributions were fitted based on a data-cut when 75% of patients had progressed. The model was optimised to find the best fitting combination of distributions using the smallest difference in 1) maximum area-under the curve (ΔAUCmax) and 2) mean survival time (ΔMST) of overall survival (OS), and compared to the best statistically fitting models based on AIC/BIC.

RESULTS: The best statistically fitting combination (TTP Weibull, PrePS Gompertz, PPS1 exponential, PPS2 Weibull) was outperformed in all metrics explored; including ΔAUCmax and ΔMST for OS. The best statistical fits based on AIC/BIC were also compared to the final data cut of the SQUIRE trial and found to underpredict survival in terms of ΔMST (4.5 months difference using best statistical fits versus 2.0 months difference using fits with the smallest ΔMST).

CONCLUSIONS: Optimising using either ΔAUCmax or ΔMST provides better fits to the data as well as better long-term survival predictions than using AIC/BIC values alone. More research is needed to validate on how best to optimise the choice of distributions used in MSMs.