[Comparison of simple pooling and bivariate model used in meta-analyses of diagnostic test accuracy published in Chinese journals].

Objective: To investigate the use of simple pooling and bivariate model in meta-analyses of diagnostic test accuracy (DTA) published in Chinese journals (January to November, 2014), compare the differences of results from these two models, and explore the impact of between-study variability of sensitivity and specificity on the differences.

Methods: DTA meta-analyses were searched through Chinese Biomedical Literature Database (January to November, 2014). Details in models and data for fourfold table were extracted. Descriptive analysis was conducted to investigate the prevalence of the use of simple pooling method and bivariate model in the included literature. Data were re-analyzed with the two models respectively. Differences in the results were examined by Wilcoxon signed rank test. How the results differences were affected by between-study variability of sensitivity and specificity, expressed by I2, was explored.

Results: The 55 systematic reviews, containing 58 DTA meta-analyses, were included and 25 DTA meta-analyses were eligible for re-analysis. Simple pooling was used in 50 (90.9%) systematic reviews and bivariate model in 1 (1.8%). The remaining 4 (7.3%) articles used other models pooling sensitivity and specificity or pooled neither of them. Of the reviews simply pooling sensitivity and specificity, 41(82.0%) were at the risk of wrongly using Meta-disc software. The differences in medians of sensitivity and specificity between two models were both 0.011 (P<0.001, P=0.031 respectively). Greater differences could be found as I2 of sensitivity or specificity became larger, especially when I2>75%.

Conclusion: Most DTA meta-analyses published in Chinese journals(January to November, 2014) combine the sensitivity and specificity by simple pooling. Meta-disc software can pool the sensitivity and specificity only through fixed-effect model, but a high proportion of authors think it can implement random-effect model. Simple pooling tends to underestimate the results compared with bivariate model. The greater the between-study variance is, the more likely the simple pooling has larger deviation. It is necessary to increase the knowledge level of statistical methods and software for meta-analyses of DTA data.