Hierarchical Summary ROC Analysis: A frequentist-bayesian colloquy in Stata

Meta-analysis of diagnostic accuracy studies requires the use of more advanced methods than meta-analysis of intervention studies. Hierarchical or multilevel modelling accounts for the bivariate nature of the data, both within and between study heterogeneity and threshold variability. The hierarchical summary receiver operating characteristic (HSROC) and the bivariate random-effects models are currently recommended by the Cochrane Collaboration. The bivariate model is focused on estimating summary sensitivity and specificity and as a generalized linear mixed model is estimable in most statistical software including Stata. The HSROC approach models the implicit threshold and diagnostic accuracy for each study as random effects and includes a shape or scale parameter which enables asymmetry in the SROC by allowing accuracy to vary with implicit threshold. As a generalized non-linear mixed model, it has not been previously/directly estimable in Stata though possible with WinBUGS and SAS Proc NLMIXED or indirectly extrapolating its parameters from the bivariate model in Stata. This talk will demonstrate for the first time how the HSROC model can be fitted in Stata using ML programming and the recently introduced bayesmh command. Using a publicly available dataset, I will show the comparability of Stata results with those obtained with WinBUGS and SAS Proc NLMIXED.