Background: The current methodology for diagnostic reviews relies on hierarchical models with normal random effects for the true logit sensitivity and specificity. The assumption of normally distributed random effects can lead to overestimation of between-study variances in the case of outliers, tails and mixtures, and this can results in extremely wide prediction intervals. Furthermore, a considerable number of studies is required to estimate the random effects reliably (Diaz 2015). This results in optimistic confidence intervals and large prediction intervals when substantial heterogeneity is present. Thus it is vital to expand the current methodology to address these issues.
Objectives: i) to develop new Bivariate Models (BM) for diagnostic reviews incorporating non-parametric random effects; ii) to provide a comprehensive review of all the methodological extensions of the BMs for diagnostic reviews within the framework of generalized latent variable modelling.
Methods: Several specifications of non-parametric random effects will be formulated for the BM. These will include: i) ordinal random effects; ii) nominal random effects; iii) latent classes. Data from two published meta-analyses were analysed for illustrating the properties of the models: 'Telomerase as a diagnostic marker for bladder cancer' (Glas 2003), and 'Lymphangiography for the diagnosis of lymph node metastasis in women with cervical cancer' (Scheidler 1997).
Results: In Telomerase review data, BM with ordinal random effects showed a better fit compared to standard BM, providing similar point estimates and confidence intervals of sensitivity and specificity, while prediction intervals were narrower. In Lymphangiography review data, both BM with nominal random effects and BM with latent classes showed a better fit compared to standard BM, providing us with a slightly higher estimate of specificity and, again, with narrower prediction intervals.
Conclusions: Although simulation studies are needed, we can conclude that BM with non-parametric random effects could represent a valid alternative to the standard BM to obtain more accurate prediction intervals of sensitivity and specificity.