Binocular sensitivity and specificity of screening tests in cross‐sectional diagnostic studies of paired organs

Abstract

We introduce new binocular accuracy measures as alternatives to conventional marginal measures that can be used to evaluate screening tests in diagnostic studies involving paired organs (e.g. eyes and ears). Specifically, we consider screening studies based on a cross-sectional design, where both diagnosis and disease status are determined after study enrolment or sampling, yielding paired binocular binary data described via two models, namely, the extended common correlation model and the Gaussian copula probit model. The first relies on the assumption of exchangeability of fellow organs, while the second is more flexible. Binocular versions of sensitivity and specificity are defined, respectively, as the probability of at least one correct positive diagnosis in patients with one or both organs truly diseased and the probability of two correct negative diagnoses for patients with both organs truly un-diseased. Comparisons between the conventional marginal and binocular sensitivities and specificities are illustrated for both models using data from a diabetic retinopathy study. We show that our methodology provides a viable alternative to conventional ways of assessing diagnostic accuracy of screening tests for paired organs. The binocular versions of sensitivity and specificity reflect the way screening tests are conducted in practice, and they overcome the shortcomings of conventional measures. Copyright © 2017 John Wiley & Sons, Ltd.

Inference about binocular sensitivity and specificity of screening tests for paired organs

“…Table 1 shows the discrepancy between the true bsen and bsp values we derived and those in Table 2 of Perera et al. 2…”

“…But, equations (6) and (8) of Perera et al. 2 have a plus sign in front of p1111 that might not be correct. The binocular specificity is simply p0000/p++00.…”

“…Perera et al. 2 studied the properties of the two new binocular measures under the cases of exchangeability, namely, pl1r1l2r2=pr1l1r2l2 and non-exchangeability. For the former, they adopted the ECCM that was espoused in de Leon et al.…”

“…For the case of non-exchangeability, Perera et al. 2 used GC as the model and derived the maximum likelihood (ML) estimates for pl1r1l2r2. The plug-in estimates for the binocular measures were correspondingly formed and their variations similarly assessed.…”

“…We first intended to reconstruct Table 2 of Perera et al. 2 by using exactly the same parameter setting including the same values of π1, π2, ρ, ρ1, ρ2 and their joint model (11). Here, π1=p1+++, π2=p++1+, ρ is the correlation between YiL+YiR and δiL+δiR, ρ1 is the correlation between YiL and …”