Article type
Year
Abstract
Objectives: To identify characteristics that drive the choice for subgroup analyses in individual patient data (IPD) meta-analyses.
Methods: IPD meta-analyses were identified by a comprehensive literature search. Data were extracted regarding publication year, number of included trials, number of included patients, domain, outcome, effect measures, fixed and/or random effect models, testing heterogeneity, testing interaction, and one- versus two-stage approach of IPD meta-analyses. Multivariate regression analyses were used to identify characteristics that could predict whether or not subgroup analyses were reported. Discrepancies between predictions and observations were examined.
Results: In total, 171 IPD meta-analyses were identified. A number of study characteristics were related to the choice for reporting subgroup analyses, namely: number of included patients (odds ratio (OR) 1.68; 95% confidence interval (CI) 1.13; 2.52), fixed effect analyses (OR 7.22; 95% CI 1.98; 26.31), testing heterogeneity (OR 3.48; 95% CI 1.34; 9.05), and testing interaction (OR 10.35; 95% CI 2.24; 47.74). For 31 IPD meta-analyses (18%) these characteristics could not accurately predict whether subgroup analyses were reported or not: for 13 IPD meta-analyses subgroup analyses were predicted but not performed, for 18 the reverse was true.
Conclusion: In this analysis of 171 IPD meta-analyses the number of included patients, usage of fixed effect analyses, testing heterogeneity and interaction drive the choice for reporting subgroup analyses. However, predictions were not perfect for 18% of the studies. The discrepancies between observed and predicted are most likely due to inconsistent use of subgroup analytic methods.
Methods: IPD meta-analyses were identified by a comprehensive literature search. Data were extracted regarding publication year, number of included trials, number of included patients, domain, outcome, effect measures, fixed and/or random effect models, testing heterogeneity, testing interaction, and one- versus two-stage approach of IPD meta-analyses. Multivariate regression analyses were used to identify characteristics that could predict whether or not subgroup analyses were reported. Discrepancies between predictions and observations were examined.
Results: In total, 171 IPD meta-analyses were identified. A number of study characteristics were related to the choice for reporting subgroup analyses, namely: number of included patients (odds ratio (OR) 1.68; 95% confidence interval (CI) 1.13; 2.52), fixed effect analyses (OR 7.22; 95% CI 1.98; 26.31), testing heterogeneity (OR 3.48; 95% CI 1.34; 9.05), and testing interaction (OR 10.35; 95% CI 2.24; 47.74). For 31 IPD meta-analyses (18%) these characteristics could not accurately predict whether subgroup analyses were reported or not: for 13 IPD meta-analyses subgroup analyses were predicted but not performed, for 18 the reverse was true.
Conclusion: In this analysis of 171 IPD meta-analyses the number of included patients, usage of fixed effect analyses, testing heterogeneity and interaction drive the choice for reporting subgroup analyses. However, predictions were not perfect for 18% of the studies. The discrepancies between observed and predicted are most likely due to inconsistent use of subgroup analytic methods.