|
Abstract:
|
In many applications in quantitative biology, one seeks to compare different statistical or laboratory procedures on the basis of sensitivity and specificity. Frequently, large "gold standard" test sets are not available, but contaminated test sets---in which putative true positives contain some true negatives, and vice versa---are plentiful. Various authors have proposed methods for estimation of sensitivity and specificity, or alternatively, relative true and false positive rates, in this setting. Most assume conditional independence of procedures given the true status of each test set member, but this is often unrealistic. Here, we describe a novel "pseudo-ROC" approach which is intuitively appealing and is valid under weaker independence assumptions. Examples from ChIP-chip microarray experiments are provided, along with a simple statistical test for equivalence of pseudo-ROC curves.
|
