Background: Systematic reviews of diagnostic accuracy studies must contend with a great deal of variability. The more variability in accuracy between studies beyond chance not explained, the more difficult it is to come to robust conclusions about the clinical implications of the results of the meta-analysis. A recent meta-epidemiologic study evaluating diagnostic reviews shows that most reviews still use univariate methods not routinely accepted in Cochrane diagnostic test accuracy (DTA) reviews such as Cochran’s Q test and I2 to quantify heterogeneity. Numerical estimates of the random-effect terms in the hierarchical models do quantify the amount of heterogeneity in terms of between-study variances (τ2s) and prediction regions on log odds scale which are rarely reported.
Objectives: To implement a modified approach to quantify heterogeneity in diagnostic reviews: the calculation of the area of the ellipse prediction on proportion scale and compare its behaviour in different situations of heterogeneity.
Methods: We will show the results of the meta-re-analysis using several datasets of published diagnostic reviews that show different heterogeneity scenarios: i. No correlation, low heterogeneity; ii. No correlation; moderate heterogeneity; iii. No correlation; highly heterogeneous (sensitivity or specificity); iv. Moderate correlation; moderate heterogeneity and v. High correlation; high heterogeneity. We will show the Area of the prediction region within the ROC plane to visually illustrate its relationship with between study variances (τ2s) and different proposed I2 statistics.
Results: We will provide routine application to different heterogeneity scenarios. Routines in R or STATA will be discussed.