Research in use of evidence synthesis in the planning and interpretation
WP4 will be led by Friede at Goettingen. In WP4 we will develop evidence synthesis methods for designing, analysing and interpreting randomised clinical trials in the context of additional non-randomised data. Whereas in large populations usually two independent confirmatory trials are required to demonstrate efficacy and safety for regulatory purposes, in small populations the conduct of even a single large-scale confirmatory trial might be extremely difficult or not feasible. In this situation the synthesis of all available data from different sources including observational data from disease registries, uncontrolled trials and randomised controlled trials, is extremely important since it facilitates extrapolation e.g. from one subgroup to another.
The focus of work package WP4 will be the formal integration of data from registries and uncontrolled studies for the planning and interpretation of a confirmatory randomised controlled study in small populations and rare diseases, linking in with work packages WP2 and WP3. Hierarchical models provide a natural framework for the synthesis of data from various sources and extend traditional methods for meta-analyses of randomised controlled trials to network meta-analyses (also called indirect or multiple treatment comparison methods) and beyond. In what is sometimes referred to as generalised evidence synthesis, or cross-design synthesis, data from different types of data sources (e.g. randomized and non-randomized studies) are combined by explicitly modelling potential biases. For instance, in a fairly simple hierarchical model estimates of a parameter of interest from studies of the same study type might be considered exchangeable. In the statistical model this can be achieved by including a random study-type effect in addition to the usual random study effect in meta-analysis. These models can be extended to account for potential biases and also allow for different degrees of discounting of information from certain studies or study types.
When combining data from different sources the hierarchy in the model might be fairly clear. However, when combining data from subpopulations it might be less clear which subpopulations are exchangeable and which are not, meaning that the hierarchy in the model is less well defined. This additional uncertainty has to be incorporated in the model for proper inference. Flexible distributions of the random effects have been proposed to extend the models discussed above and we will build on these methods and investigate alternatives when developing new methodology particularly suited to small populations and rare events considered in this project.
The hierarchical models, which can be fitted either using Bayesian or likelihood approaches, allow modelling of heterogeneity between studies and study types by including appropriate variance components as explained above. This is of particular importance in small populations because of the relatively small number of studies, small study sizes, larger heterogeneity in studied populations, and variations in study designs that are often more pronounced than in larger populations. The computational issues arising from these circumstances will also be dealt with in WP4.
The work will extend previous methods for evidence synthesis focussing particularly on problems in small populations and rare diseases where there is very little existing work.
Links: UMG Department of Medical Statistics