OBJECTIVES

Model-based meta-analysis is a technique increasingly used in drug development for synthesising results from multiple studies, allowing pooling of information on treatment, dose-response and time-course characteristics, which are often non-linear. Such analyses are used in drug development to inform future trial designs. Network meta-analysis is used frequently in Health Technology Appraisals and by reimbursement agencies for simultaneously comparing effects of multiple treatments. Recently, a framework for dose-response model-based network meta-analysis (MBNMA) has been proposed. We aim to develop a MBNMA framework for time-course models.

METHODS

We propose a Bayesian time-course MBNMA modelling framework that allows for non-linear modelling of multi-parameter time-course functions for comparative effectiveness, and that can account for residual correlation between observations using a multivariate likelihood. These methods preserve randomisation by aggregating within-study relative effects and, by modelling consistency equations on the time-course parameters, they allow for testing of inconsistency between direct and indirect evidence. We demonstrate our modelling framework using an illustrative dataset of 24 trials investigating treatments for pain in osteoarthritis.

RESULTS

For our dataset, we report results from 10 different models. An E_max function allowed for the greatest degree of flexibility, both in the time-course shape and in the specification of time-course parameters (E_max and ET). Our final model had a posterior mean residual deviance of 291.4 (compared to 345 data points), indicating a good fit to the data. Some simplifying assumptions were needed to identify ET, as studies contained few observations at earlier follow-up times. Treatment estimates were robust to the choice of likelihood (univariate/multivariate).

CONCLUSIONS

By preserving randomisation and allowing for testing of inconsistency, time-course MBNMA can be used to incorporate multiple study time points into analyses and has the potential to be used both in helping design and predict studies in drug development, as well as for decision-making by reimbursement agencies.