A0588
Title: Classification under outcome misclassification: Reliability quantification and partial identification
Authors: Muxuan Liang - University of Florida (United States) [presenting]
Abstract: Misclassification of outcomes or labels presents a prevalent challenge in classification problems. In many applications, the underlying outcome may not be directly accessible, while a surrogate outcome subject to misclassification can be observed. Directly using the surrogate outcome may lead to a biased estimation of the optimal classification rule. The sensitivity and specificity of the surrogate outcome at an individual level can be used to remove such bias. However, with limited accessibility of the underlying outcomes, point identification of individual-level sensitivity and specificity is difficult or even impossible. For classification problems, a range of individual-level sensitivity and specificity is assumed to be the reliable quantification of surrogate outcomes. With this partial information on sensitivity and specificity, partial identification is established for the distribution of the underlying outcome as well as the optimal classification rule using the surrogates. Based on this result, a robust classification framework and a novel estimation procedure are proposed to estimate a robust classification rule without requiring point identification of individual-level sensitivity and specificity.
