Article type
Year
Abstract
Background: In systematic reviews, handling studies with rare events is a challenging issue. The generalized linear mixed models (GLMMs) based on a one-stage framework may have good performance in dealing with studies with rare events, especially when no events occur in a single trial. In this study, we compared the performance of two GLMMs with the generic two-stage methods.
Methods: We used simulations to generate trials of grouped data with the risk of control group as 0.01. We set the studies for each meta-analysis as 10, and the sample size in both group as a uniform distribution from 15-58. The statistical properties of random-intercept GLMM (Method 1) and the random-intercept and random-coefficient GLMM (Method 2) were compared with the fixed-effect inverse variance (Method 3), the random-effect inverse variance (Method 4), the fixed-effect MH (Method 5), the random-effects MH (Method 6), and the Peto method (Method 7). The inverse variance method was used through continuity correction. The percentage bias, mean square error (MSE), and coverage probability were used as performance indicators. We set 25 scenarios and each scenario generated 3000 loops of meta-analyses by simulation, the statistical properties of these methods were compared under different effect sizes (OR=1, 2, 3, 4, 5) and heterogeneity (tau as 0.2, 0.4, 0.6, 0.8. and 1.0 for mild to large heterogeneity).
Results: Our simulation suggested that when the heterogeneity was mild (Tau 95%) across these methods, except that the Peto method showed a poor coverage (< 90%) at a odds ratio of 1. In the presence of substantial heterogeneity (Tau >=0.8), the performance of these methods became declined, especially for the Peto method. The two GLMMs continued to show lowest bias and good coverage, although inverse variance (Method 3 and 4) and MH (Method 5 and 6) methods had lower MSE.
Conclusions: The generalized linear mixed model may be preferred over the generic two-stage methods when handling studies with no events. Empirical studies are warranted to confirm this finding.
Methods: We used simulations to generate trials of grouped data with the risk of control group as 0.01. We set the studies for each meta-analysis as 10, and the sample size in both group as a uniform distribution from 15-58. The statistical properties of random-intercept GLMM (Method 1) and the random-intercept and random-coefficient GLMM (Method 2) were compared with the fixed-effect inverse variance (Method 3), the random-effect inverse variance (Method 4), the fixed-effect MH (Method 5), the random-effects MH (Method 6), and the Peto method (Method 7). The inverse variance method was used through continuity correction. The percentage bias, mean square error (MSE), and coverage probability were used as performance indicators. We set 25 scenarios and each scenario generated 3000 loops of meta-analyses by simulation, the statistical properties of these methods were compared under different effect sizes (OR=1, 2, 3, 4, 5) and heterogeneity (tau as 0.2, 0.4, 0.6, 0.8. and 1.0 for mild to large heterogeneity).
Results: Our simulation suggested that when the heterogeneity was mild (Tau 95%) across these methods, except that the Peto method showed a poor coverage (< 90%) at a odds ratio of 1. In the presence of substantial heterogeneity (Tau >=0.8), the performance of these methods became declined, especially for the Peto method. The two GLMMs continued to show lowest bias and good coverage, although inverse variance (Method 3 and 4) and MH (Method 5 and 6) methods had lower MSE.
Conclusions: The generalized linear mixed model may be preferred over the generic two-stage methods when handling studies with no events. Empirical studies are warranted to confirm this finding.