Objectives Systematic reviews and meta-analyses are well established as the highest level of evidence in health care, but can be costly in time, money and labour. Meta-analyses have historically advocated ‘more is better’ to approximate the true effect. However, restricting the number of studies in meta-analyses would reduce the time taken to do a review. It has been suggested that sample size is an adequate indicator of effect estimate and larger studies have estimates closer to the true effect. This research’s objective was to generate a minimum threshold number of studies based on study sample size, to use in meta-analyses without comprising their overall conclusions.

Method Using a derivation dataset from the Cochrane library, meta-analyses were ranked by sample size. Pooled estimates for the individual meta-analyses were recalculated to obtain a threshold number of larger studies (study sample size) where the change in the effect estimate 95% confidence interval width had stabilised and was minimal. This threshold number of studies was tested for concordance between the original meta-analysis and the restricted meta-analysis in a validation dataset of meta-analyses. Comparisons were made between the paired meta-analyses using two methods: levels of agreement in direction of effect and statistical significance, and correlation of effect estimates.

Results The research suggests where studies are ranked by study sample size, nine studies are sufficient to draw the same meta-analysis conclusion in terms of total agreement between restricted and the original meta-analysis 80% of the time. Correlation of the effect estimates was 0.96. Where restricted meta-analysis results are statistically insignificant caution should be taken as there may be a higher chance of disagreement between the paired meta-analyses.

Conclusions This research adds to the body of evidence that supports the view that it is possible to use restricted meta-analyses without jeopardising the integrity of the final findings for a given research question.