Article type
Year
Abstract
Background: systematic reviews of prevalence estimates are of critical importance for policymakers, health professionals and patients, since they describe the distribution of diseases among different populations and subgroups. Meta-analysis methods can be useful to pool these estimates and generate a mean prevalence estimate for different scenarios. However, there is still uncertainty and lack of standardized methods to conduct these analyses.
Objectives: to describe the meta-analysis methodology of recently published systematic reviews of prevalence of clinical conditions.
Methods: we searched Medline using the terms ‘prevalence’ and ‘systematic review’ in the title and limited the search to studies published between February 2017 and February 2018. We included systematic reviews on the prevalence of any clinical conditions published in English and extracted relevant data regarding the conduction of meta-analysis.
Results: our search identified 335 systematic reviews, of which we included 235 in our study. Of these, 152 (64.7%) conducted meta-analysis. Out of these 152 reviews, 151 (99.3%) used frequentist methods and only one (0.7%) summarized the evidence using Bayesian approach. Most reviews (n = 135, 89.4%) used a random-effects model in their main analysis, while one (0.7%) used a fixed-effect model, and two (1.3%) used arbitrary models. The majority of reviews (n = 116, 76.8%) did not report any details about the variance estimator used in the analysis; among the ones that reported, 30 (19.9%) used DerSimonian and Laird, four (2.6%) used HKSJ (Hartung, Knapp, Sidik and Jonkman) and one (0.7%) used restricted maximum-likelihood. Most reviews (n = 107, 70.4%) did not report how they transformed prevalence estimates; the most used transformation methods were Freeman-Tukey double arcsine (n = 32, 21.0%) and logit (n = 5, 3.3%). Only three reviews (2.0%) estimated prediction intervals. Heterogeneity was assessed with I2 statistics in 144 reviews (94.7%), and in 103 reviews (76.9%) it was 90% or higher. Subgroup analysis was conducted in 89 reviews (58.6%) and meta-regression in 57 (41.4%). Publication bias was examined in 56 reviews (36.8%) by visual inspection of funnel plots, Egger’s test in 54 (35.5%) and Begg’s test in 26 (17.1%).
Conclusions: recent systematic reviews of prevalence estimates conducted meta-analysis using heterogeneous methods, indicating lack of consensus of which methods are more appropriate. The use of inappropriate methods may limit the validity and applicability of results.
Patient or healthcare consumer involvement: appropriate and standard methods to conduct systematic reviews and meta-analysis of prevalence estimates are critical for patients and healthcare consumers since these data impact on healthcare decision-making, priority-setting definition, and public health policies.
Objectives: to describe the meta-analysis methodology of recently published systematic reviews of prevalence of clinical conditions.
Methods: we searched Medline using the terms ‘prevalence’ and ‘systematic review’ in the title and limited the search to studies published between February 2017 and February 2018. We included systematic reviews on the prevalence of any clinical conditions published in English and extracted relevant data regarding the conduction of meta-analysis.
Results: our search identified 335 systematic reviews, of which we included 235 in our study. Of these, 152 (64.7%) conducted meta-analysis. Out of these 152 reviews, 151 (99.3%) used frequentist methods and only one (0.7%) summarized the evidence using Bayesian approach. Most reviews (n = 135, 89.4%) used a random-effects model in their main analysis, while one (0.7%) used a fixed-effect model, and two (1.3%) used arbitrary models. The majority of reviews (n = 116, 76.8%) did not report any details about the variance estimator used in the analysis; among the ones that reported, 30 (19.9%) used DerSimonian and Laird, four (2.6%) used HKSJ (Hartung, Knapp, Sidik and Jonkman) and one (0.7%) used restricted maximum-likelihood. Most reviews (n = 107, 70.4%) did not report how they transformed prevalence estimates; the most used transformation methods were Freeman-Tukey double arcsine (n = 32, 21.0%) and logit (n = 5, 3.3%). Only three reviews (2.0%) estimated prediction intervals. Heterogeneity was assessed with I2 statistics in 144 reviews (94.7%), and in 103 reviews (76.9%) it was 90% or higher. Subgroup analysis was conducted in 89 reviews (58.6%) and meta-regression in 57 (41.4%). Publication bias was examined in 56 reviews (36.8%) by visual inspection of funnel plots, Egger’s test in 54 (35.5%) and Begg’s test in 26 (17.1%).
Conclusions: recent systematic reviews of prevalence estimates conducted meta-analysis using heterogeneous methods, indicating lack of consensus of which methods are more appropriate. The use of inappropriate methods may limit the validity and applicability of results.
Patient or healthcare consumer involvement: appropriate and standard methods to conduct systematic reviews and meta-analysis of prevalence estimates are critical for patients and healthcare consumers since these data impact on healthcare decision-making, priority-setting definition, and public health policies.