OBJECTIVES: Network meta-analysis (NMA) methods for time-to-event (TTE) outcomes that allow for time-varying treatment effects are important for extrapolations when the proportional hazard (PH) assumption is uncertain. We propose a one-step fully Bayesian parametric NMA model that fits TTE outcome data based on (reconstructed) individual patient data (IPD) with the exact likelihood and allows for time-varying treatment effects.

METHODS: We define fixed or random effects model assuming the following distributions: Weibull, Gompertz, log-normal, log-logistic, gamma, or generalized gamma distributions. With an artificial dataset, we demonstrate how the one-step model leverages exact event and censor times and does not require the PH assumption. We apply the one-step NMA models to a network of RCTs evaluating multiple interventions for advanced melanoma and compare results with those obtained with a previously proposed two-step approach (Cope et al., 2020).

RESULTS: In the artificial dataset, the NMA estimates, and 95% credible intervals aligned well with the “true” values used to simulate the data, as well as with estimates from frequentist study-level fits using . In the illustrative case study, the approximate leave-one-out information criterion suggested that the log-logistic distribution was most appropriate, and parameter estimates were consistent with the two-step approach.

CONCLUSIONS: We provide a one-step NMA model that uses (reconstructed) IPD that may be more accurate than methods using discrete hazards, which require reparameterizations. Model selection among the ‘standard’ distributions is straightforward, with treatment effects on either the scale alone, or with multivariate treatment effects. Inclusion of the g The one-step NMA model reduces required assumptions versus the two-step model and provides a more generalizable framework for advanced analyses.