Article type
Year
Abstract
Background: The DerSimonian-Laird method has been the standard for random-effects meta-analysis for several decades. However, unfavorable statistical properties, especially in the case of very few studies, have been highlighted and discussed critically for some time now. A Cochrane working group recommended the use of the Knapp-Hartung method as the new standard approach when there are very few studies - say two to five. However, the Knapp-Hartung method, which in contrast to the DerSimonian-Laird method accounts for the uncertainty in estimating the between-study heterogeneity, can result in very wide confidence intervals, even if all studies are statistically significant in the same direction.
Objectives: To describe and discuss available approaches to perform meta-analyses in the case of very few studies.
Methods: Besides classical approaches for fixed-effect and random-effects meta-analysis, a number of alternative approaches are available including generalized mixed-effects models and Bayesian methods incorporating weakly informative prior distributions for the between-study heterogeneity. The basic features of these approaches are summarized and the required conditions for practical applications are discussed. The methods are illustrated by a variety of examples.
Results: The methods differ considerably in terms of their statistical properties, including coverage probabilities and lengths of confidence intervals for the combined effect. Furthermore, some methods require a better statistical understanding on the side of the practitioner than others. Some methods lend themselves more easily to sensitivity analyses than others. Currently, none of the available approaches can be considered to be a uniformly best method. Besides the classical approaches, the use of alternative methods such as generalized mixed-effects models seems to be useful.
Conclusions: Although meta-analyses with very few studies are very common, performing meta-analyses in the case of very few studies remains challenging. Currently no clear guidance exists on how best to proceed in these challenging scenarios. Further research in this field is required.
Objectives: To describe and discuss available approaches to perform meta-analyses in the case of very few studies.
Methods: Besides classical approaches for fixed-effect and random-effects meta-analysis, a number of alternative approaches are available including generalized mixed-effects models and Bayesian methods incorporating weakly informative prior distributions for the between-study heterogeneity. The basic features of these approaches are summarized and the required conditions for practical applications are discussed. The methods are illustrated by a variety of examples.
Results: The methods differ considerably in terms of their statistical properties, including coverage probabilities and lengths of confidence intervals for the combined effect. Furthermore, some methods require a better statistical understanding on the side of the practitioner than others. Some methods lend themselves more easily to sensitivity analyses than others. Currently, none of the available approaches can be considered to be a uniformly best method. Besides the classical approaches, the use of alternative methods such as generalized mixed-effects models seems to be useful.
Conclusions: Although meta-analyses with very few studies are very common, performing meta-analyses in the case of very few studies remains challenging. Currently no clear guidance exists on how best to proceed in these challenging scenarios. Further research in this field is required.