Title: Assessing the sensitivity of a Youden posterior to extremes in the classification variable
Authors: Catherine Forbes - Monash University (Australia) [presenting]
Abstract: The two-class Youden index corresponds to an optimal threshold value for classifying subjects into one of two distinct groups. The value of the index maximises the sum of the two correct classification probabilities for each group. However, the population distributions of each group must be estimated since they are usually unknown. The empirical distribution functions for each have been used to estimate the threshold. A Bayesian bootstrap has been used to obtain a posterior distribution for the receiver operating characteristic (ROC) curve, which in turn may be used to obtain a posterior distribution for the threshold value. This Youden posterior effectively accommodates distributional uncertainty; however, it may be sensitive to extremes and asymmetry in the classification variable. To explore the impact of any potential sensitivity, we systematically re-weight individual sample observations to determine a range of posterior point and interval estimates of the Youden threshold. We illustrate our approach in the context of a derived (subjective) housing affordability indicator for an urbanised state in Malaysia.