Multinma: A comprehensive R package for network meta-analysis of survival outcomes with aggregate data, individual patient data, or a mixture of both

Introduction
Survival or time-to-event outcomes are commonplace in disease areas such as oncology. Healthcare decision makers require estimates of relative efficacy between different treatment options, however treatments of interest are frequently not all compared in head-to-head randomised controlled trials, and so indirect comparison and network meta-analysis (NMA) methods are required to synthesise evidence from a connected network of trials and treatments. An extension of NMA, multilevel network meta-regression (ML-NMR), is increasingly used to account for differences in effect modifiers between populations where individual patient data are available from one or more trials. However, to date there has been no user-friendly software package that can perform NMA or ML-NMR with survival outcomes; instead analysts have needed to rely on complex bespoke modelling code.
Methods
A recent update to the multinma R package provides a user-friendly suite of models and tools for synthesising survival outcomes from multiple trials, with aggregate data, individual patient data, or mixtures of both. Models are fitted in a Bayesian framework using Stan. A full range of parametric proportional hazards and accelerated failure time survival distributions are implemented, along with flexible baseline hazard models via M-splines or piecewise exponential hazards with a novel random walk shrinkage prior that avoids overfitting. Shape parameters may be stratified or regressed on treatment arm and/or covariates to relax proportionality. Right, left, and interval censoring, and delayed entry are all supported.
Results
We present analyses of two case studies using the multinma package.
First, we performed a NMA of published aggregate data from a network of treatments for advanced non-small cell lung cancer using flexible M-spline baseline hazards. We introduced treatment effects onto the spline coefficients to account for non-proportional hazards, and produced estimated survival curves in a target population required for further economic modelling.
Second, we performed a ML-NMR using a mixture of individual patient data and aggregate data from a network of treatments for newly-diagnosed multiple myeloma. We adjusted for effect-modifying covariates, and produced population-adjusted estimates for target populations of interest to decision-making. Covariate adjustment removed evidence for non-proportional hazards that was present in unadjusted models.
Conclusions
The multinma package makes NMA and ML-NMR methods accessible to a broad audience. The latest update to include a suite of functionality for survival analysis facilitates application of these methods to widespread settings such as oncology, where until now there was no user-friendly software available.