This paper shows that the current practice of filtering discordant U–Pb data based on the relative difference between the ²⁰⁶Pb/²³⁸U and ²⁰⁷Pb/²⁰⁶Pb ages is just one of several possible approaches to the problem and demonstrably not the best one. An alternative approach is to define discordance in terms of isotopic composition, as a log ratio distance between the measurement and the concordia line. Application to real data indicates that this reduces the positive bias of filtered age spectra.