Published in:
Open Access 01-12-2013 | Research article
Quantitative summaries of treatment effect estimates obtained with network meta-analysis of survival curves to inform decision-making
Published in: BMC Medical Research Methodology | Issue 1/2013
Login to get accessAbstract
Background
Increasingly, network meta-analysis (NMA) of published survival data are based on parametric survival curves as opposed to reported hazard ratios to avoid relying on the proportional hazards assumption. If a Bayesian framework is used for the NMA, rank probabilities associated with the alternative treatments can be obtained, which directly support decision-making. In the context of survival analysis multiple treatment effect measures are available to inform the rank probabilities.
Methods
A fractional polynomial NMA of overall survival in advanced melanoma was performed as an illustrative example. Rank probabilities were calculated and presented for the following effect measures: 1) median survival; 2) expected survival; 3) mean survival at the follow-up time point of the trial with the shortest follow-up; 4) hazard or hazard ratio over time; 5) cumulative hazard or survival proportions over time; and 6) mean survival at subsequent time points. The advantages and disadvantages of the alternative measures were discussed.
Results
Since hazard and survival estimates may vary over time for the compared interventions, calculations of rank probabilities for an NMA of survival curves may depend on the effect measure. With methods 1–3 rank probabilities do not vary over time, which are easier to understand and communicate than rank probabilities that vary over time as obtained with methods 4–6. However, rank probabilities based on methods 4–6 provide useful information regarding the relative treatment effects over time.
Conclusions
Different approaches to summarize results of a NMA of survival curves with rank probabilities have pros and cons. Rank probabilities of treatment effects over time provide a more transparent and informative approach to help guide decision-making than single rank probabilities based on collapsed measures, such as median survival or expected survival. Rank probabilities based on survival proportions are the most intuitive and straightforward to communicate, but alternatives based on the hazard function or mean survival over time may also be useful.