Evidence based sample size calculations for future trials based on meta-analyses

Article type
Authors
Sutton A, Cooper N, Jones D, Lambert P, Thompson J, Abrams K
Abstract
Background: It is accepted that meta-analyses of RCTs provide the highest level of evidence regarding the effectiveness of interventions. Despite this, with few exceptions, meta-analyses are produced as observational bi-products of the existing literature. There is no consideration of future meta-analyses when individual trials are being designed including issues of sample size.

Objectives: A framework for sample size calculation for a future RCT based on the results of an existing meta-analysis is presented. It is argued that, in some circumstances, this approach is more logical than traditional sample size calculations since the resulting, updated, meta-analysis will be of more interest than the results of the new individual trial.

Methods: A simulation approach is described to calculate the power a new study will have regarding inferences of the updated meta-analysis in which it is to be included. Methods for both fixed and random effect meta-analyses models are discussed. The approach will be illustrated using examples from the Cochrane systematic review database. Different approaches to inference will be discussed, including the use of equivalence levels as an alternative to statistical significance. Since the method is simulation based, it is easy to modify for different situations. Extensions allowing the optimum sample size calculations for a series of trials are presented.

Results: Under a fixed effect model the power of the updated meta-analysis will often be greater than that of the new study singularly. Results are less intuitive under a random effect model, and a real example will be given when it is almost impossible to design a new study that will produce decisive inferences from a meta-analysis, irrespective of its size due to the heterogeneity present between existing studies. It is also shown that under certain conditions multiple small studies have more power than one large one.

Conclusion: An approach for calculating the power of including a new study in a meta-analysis has been developed. We believe this is novel, however, it does have similarities with cumulative meta-analysis and sample size calculations for multi-centre trial methodologies. Further work to allow the method to be placed in a fully decision theoretic framework is considered.