|
Abstract:
|
Receiver operating characteristic curves, and subsequently the area under the curve (AUC), are often used to evaluate the performance of a diagnostic test as a binary classifier. It is often of interest to determine if, and to what extent, two diagnostic tests differ in predictive ability. This is usually accomplished by comparing the AUC of diagnostic tests via the method of DeLong et al. (1988), which is appropriate when the observed units are independent, and can be done with traditional software packages. An extension to the DeLong method (Obuchowski, 1997), which allows for the comparison of diagnostic tests when the observed units are clustered, is computationally expensive, and for which up to this point adequate software did not exist. This paper aims to further evaluate the scenarios in which the DeLong and Obuchowski methods differ, in terms of power and type 1 error, when analyzing data which is clustered, and additionally provide recommendations for scenarios in which the Obuchowski method provides benefit. Applications to imaging and survey data will be discussed followed by the introduction of an R package which can be used to quickly analyze such data sets.
|
