Comparison of modelling approaches for network meta-analysis of time-to-event outcomes to aid decision making


Oral session: Overviews of reviews and network meta-analysis (2)


Monday 17 September 2018 - 15:00 to 15:10


All authors in correct order:

Freeman S1, Cooper N1, Sutton A1, Hawkins N2
1 University of Leicester; NIHR Complex Reviews Support Unit, United Kingdom
2 University of Glasgow, NIHR Complex Reviews Support Unit, United Kingdom
Presenting author and contact person

Presenting author:

Suzanne Freeman

Contact person:

Abstract text
Background: Synthesis of data from multiple trials to obtain an overall estimate of clinical effectiveness is a well-established component of decision making. Treatment effects for time-to-event (TTE) data are often summarised using a single hazard ratio (HR) which may be derived from semi-parametric Cox proportional hazard or parametric survival models. The estimated HRs are then synthesised in pairwise or network meta-analyses (NMA). The HRs represent an 'average' of the HR over the observed study time period. If the HR varies markedly over time, a single HR may not be a useful measure of treatment effect for decision-making. Comparison of HRs may be misleading and confounded by differences in study duration.

Objectives: To compare and contrast five different approaches for individual participant data NMA of TTE data when comparison of HRs may not be appropriate.

Methods: We considered five approaches to modelling TTE data for synthesising treatment effects: piecewise exponential models; fractional polynomial models; the Royston-Parmar flexible parametric model; generalised gamma model; and explicit time-treatment effect interactions. Each approach was applied to publicly available trial data for non-small cell lung cancer reporting overall survival (1). Methods were compared in terms of clinical interpretability of assumptions, potential to apply prior clinical beliefs, potential for over-fitting, practicality of using the approach, fit to observed data and credibility of extrapolation.

Results: We fitted all models to the dataset successfully. All models fitted the data reasonably well with some variation, however there was important variation in the extrapolations of the survival curve and the interpretability of the modelling constraints.

Conclusions: Deciding on the right approach for evidence synthesis of TTE outcomes is not straightforward. It is important to consider a wide range of factors, not just model fit. A considered holistic approach to model selection, including consideration of prior belief, can improve decision making.

Patient or healthcare involvement: Evidence-based decision making can affect the interventions that are available on the NHS which can change the care that is delivered to patients resulting in cost savings for the NHS and better quality of treatment for patients.

1) Jansen BMC Med Res Meth 2011;11:61

Relevance to patients and consumers: 

Clinical effectiveness is an important part of the decision making process which can affect the types of interventions that are available to patients on the NHS. Therefore it is important that estimates of clinical effectiveness are as accurate and reliable as possible. We applied a range of models for synthesising time-to-event outcomes (e.g. time-to-death) and identified variation between the models in terms of predicting survival beyond study duration. We encourage researchers to consider a wide range of factors to ensure accuracy and reliability in their estimates of clinical effectiveness.