TG-ROC plots with confidence

Edson Zangiacomi MARTINEZ[1]

Francisco LOUZADA-NETO[2]

§    ABSTRACT: TG-ROC (two-graph receiver operating characteristic) plots are defined by a graph of the sensitivity and specificity against a range of cut-off points related to the possible outcomes of a diagnostic test. This approach provides a didactical picture for helping the selection of an optimal cut-off point, and it has several advantages compared with the usual ROC curve. The optimal cut-off point may be defined by clinical criteria related to the practical utility of the test and not only by mathematical rules, and the visualization of the performance measures provided by TG-ROC plots can be a useful tool to the medical researcher. We proposed the presentation of TG-ROC plots jointly with its respective confidence bounds, that allow a description of the sample variability of Se and Sp, and thus, to visualize the confidence of the diagnostic performance for each cut-off point candidate to be optimal. An approach for constructing confidence bounds is described, and a Monte Carlo study evidenced that these confidence bounds  has coverage probability satisfactorily near to nominal value .95. This simulation used normal probability models for described the distribution of the continuous diagnostic test.

§    KEYWORD: TG-ROC plots; diagnostic tests; Monte Carlo method; interval estimation.



[1] Departamento de Medicina Social, Faculdade de Medicina de Ribeirão Preto, USP, Av. Bandeirantes 3900, 14049-900, Ribeirão Preto, SP, Brazil. E-mail: edson@fmrp.usp.br

[2] Departamento de Estatística, Universidade Federal de São Carlos – UFSCar, caixa postal 676, 13565-905, Brasil. E-mail: dfln@power.ufscar.br