Background: The landscape of 1L aRCC treatment is rapidly evolving from tyrosine kinase inhibitor (TKI) monotherapy to their combination with immuno-oncology (IO) agents. Differences in mechanisms of action (MoA) of different IO+TKIs and of TKI monotherapy lead to differences in survival trends and generate non-constant hazards over time. To assess the cost-effectiveness of new treatments, trials comparing all relevant treatments are needed. The lack of these necessitate indirect treatment comparisons accounting for non-proportional hazards, such as a FP NMA. Multiple FP models are available to model the time behavior of hazards. Often model selection is solely based on statistical fit, which can result in clinically implausible long-term survival estimates. Methods: We sought to develop an algorithm that improves FP model selection for predictive accuracy, clinical plausibility and goodness of fit. The network included FDA-approved treatments avelumab + axitinib, pembrolizumab + lenvatinib (P+L), nivolumab + cabozantinib (N+C), and pembrolizumab + axitinib, with sunitinib (S) as a common comparator. Synthetic progression-free survival (PFS) and overall survival (OS) data were reconstructed from 4 phase 3 clinical trials and 44 FP models were explored. For model selection, a priori criteria (face validity and predictive accuracy against trial data [median survival, landmark survival at 24 months, and restricted mean survival time]) were used along with statistical fit criteria. Criteria for clinical plausibility, and long-term survival extrapolations and hazard functions of viable models were discussed with clinical experts to select the optimal models. Results: The selected FP models based solely on statistical fit criteria led to clinically implausible survival extrapolations. Using our proposed selection algorithm, three models were considered viable for PFS and two models for OS. The optimal FP model predicted the highest PFS for P+L and for N+C (medians of 23.2 and 19.0 months), and the lowest PFS for S (9.8 months). The PFS models overestimated outcomes for most treatments. The OS model was selected in line with expert opinion that subsequent treatment would cause IO + TKI vs. S survival curves to cross over time. OS was similar for all IO + TKI treatments. Conclusions: The use of additional model selection criteria based on face validity, predictive accuracy and expert opinion helped improve the clinical plausibility of survival outcomes compared to models based on statistical fit only. In our study, FP models generally performed imperfectly against some of the trial data. This likely reflects heterogeneity between trial populations, differing MoAs and subsequent treatments and highlights further research is needed on the integration of adjustment for heterogeneity in the fitting of FP models.