An evaluation of computational methods for aggregate data meta-analyses of diagnostic test accuracy studies.

BACKGROUND: A Generalized Linear Mixed Model (GLMM) is recommended to meta-analyze diagnostic test accuracy studies (DTAs) based on aggregate or individual participant data. Since a GLMM does not have a closed-form likelihood function or parameter solutions, computational methods are conventionally used to approximate the likelihoods and obtain parameter estimates. The most commonly used computational methods are the Iteratively Reweighted Least Squares (IRLS), the Laplace approximation (LA), and the Adaptive Gauss-Hermite quadrature (AGHQ). Despite being widely used, it has not been clear how these computational methods compare and perform in the context of an aggregate data meta-analysis (ADMA) of DTAs. METHODS: We compared and evaluated the performance of three commonly used computational methods for GLMM - the IRLS, the LA, and the AGHQ, via a comprehensive simulation study and real-life data examples, in the context of an ADMA of DTAs. By varying several parameters in our simulations, we assessed the performance of the three methods in terms of bias, root mean squared error, confidence interval (CI) width, coverage of the 95% CI, convergence rate, and computational speed. RESULTS: For most of the scenarios, especially when the meta-analytic data were not sparse (i.e., there were no or negligible studies with perfect diagnosis), the three computational methods were comparable for the estimation of sensitivity and specificity. However, the LA had the largest bias and root mean squared error for pooled sensitivity and specificity when the meta-analytic data were sparse. Moreover, the AGHQ took a longer computational time to converge relative to the other two methods, although it had the best convergence rate. CONCLUSIONS: We recommend practitioners and researchers carefully choose an appropriate computational algorithm when fitting a GLMM to an ADMA of DTAs. We do not recommend the LA for sparse meta-analytic data sets. However, either the AGHQ or the IRLS can be used regardless of the characteristics of the meta-analytic data.

An empirical comparison of statistical methods for multiple cut-off diagnostic test accuracy meta-analysis of the Edinburgh postnatal depression scale (EPDS) depression screening tool using published results vs individual participant data.

BACKGROUND: Selective reporting of results from only well-performing cut-offs leads to biased estimates of accuracy in primary studies of questionnaire-based screening tools and in meta-analyses that synthesize results. Individual participant data meta-analysis (IPDMA) of sensitivity and specificity at each cut-off via bivariate random-effects models (BREMs) can overcome this problem. However, IPDMA is laborious and depends on the ability to successfully obtain primary datasets, and BREMs ignore the correlation between cut-offs within primary studies. METHODS: We compared the performance of three recent multiple cut-off models developed by Steinhauser et al., Jones et al., and Hoyer and Kuss, that account for missing cut-offs when meta-analyzing diagnostic accuracy studies with multiple cut-offs, to BREMs fitted at each cut-off. We used data from 22 studies of the accuracy of the Edinburgh Postnatal Depression Scale (EPDS; 4475 participants, 758 major depression cases). We fitted each of the three multiple cut-off models and BREMs to a dataset with results from only published cut-offs from each study (published data) and an IPD dataset with results for all cut-offs (full IPD data). We estimated pooled sensitivity and specificity with 95% confidence intervals (CIs) for each cut-off and the area under the curve. RESULTS: Compared to the BREMs fitted to the full IPD data, the Steinhauser et al., Jones et al., and Hoyer and Kuss models fitted to the published data produced similar receiver operating characteristic curves; though, the Hoyer and Kuss model had lower area under the curve, mainly due to estimating slightly lower sensitivity at lower cut-offs. When fitting the three multiple cut-off models to the full IPD data, a similar pattern of results was observed. Importantly, all models had similar 95% CIs for sensitivity and specificity, and the CI width increased with cut-off levels for sensitivity and decreased with an increasing cut-off for specificity, even the BREMs which treat each cut-off separately. CONCLUSIONS: Multiple cut-off models appear to be the favorable methods when only published data are available. While collecting IPD is expensive and time consuming, IPD can facilitate subgroup analyses that cannot be conducted with published data only.