OBJECTIVES : FP NMA models estimate comparative effectiveness of time to event outcomes accounting for time-varying hazards. However, as many power and order combinations can be fitted to the data, the optimal model choice is not straightforward. We developed an algorithm that improves model selection based on predictive accuracy and clinical plausibility, in a case study in first-line advanced renal cell carcinoma (1L aRCC).

METHODS : Forty-four candidate FP models were considered. Convergent models were ordered with respect to deviation information criteria (DIC). Next, face validity of time varying hazard ratios was assessed to determine any exclusion of first-or-second order models, based on a set of prespecified criteria. Model fit was assessed by comparing the predictive accuracy of median survival, landmark survival at 24 months, and restricted mean survival time with observed trial data. Clinical plausibility of long-term survival extrapolations and hazard functions were examined to complete the selection process of viable models, among which the model with the lowest DIC was chosen.

RESULTS : The algorithm was applied in the FP NMA of randomized controlled trials evaluating progression-free survival (PFS) and overall survival (OS) in 1L aRCC. Indicated by the lowest DIC only, the second-order model (P1=-2, P2=-2) for PFS and the first-order model (P1=-2) for OS led to clinically implausible survival extrapolations. For PFS, 6 models were considered viable, although all these models performed imperfectly against available trial data. For OS, second-order models overfitted the relatively immature trial data, generating implausible hazard patterns and were therefore excluded altogether; 3 models were considered viable. The first-order model (P1=-1) was selected as most plausible for PFS and OS.

CONCLUSIONS : While DIC remains an effective measure for assessing fit to observed data, FP models with low DICs may be clinically implausible. Applying this decision algorithm improved the predictive accuracy of model estimates, aligning with clinical expectations.