Yang Ling, Li Bofeng, Shen Yunzhong, Zhang Zhiteng, Chris Rizos
Minimal detectable bias (MDB) describes a system's general ability to detect the existence of a single outlier. However, successful identification probability is closely related to the correlations between w-statistics. When the correlation is high, the validity of the MDB as a system internal reliability index is weakened. Based on the correlation coefficient, the probabilities of committing missed detection and wrong exclusion can also be estimated; however, the latter might be significant, inducing obvious discrepancies between the actual successful identification rate and the preset value dominated by the MDB. In this paper, given a threshold to the probability of wrong exclusion, a minimal separability bias (MSB), usually larger than the MDB, is obtained using an analytic method, which can control both the probability of missed detection and of wrong exclusion. Compared with previous work, the MSB described here is expressed by an algebraic formula, which simplifies the calculation and promotes its practical application.