Establishing noise exposure criteria for marine mammals has proven to be a difficult and contentious issue. Over the last decade, several attempts have been made to provide scientifically-based exposure criteria. While representing the "best available science" on the issue, these criteria, and the assumptions underpinning them, have led to considerable discussion among both scientists and policy-makers. However, one area where there has been little or no debate is around the use of appropriate statistical and other numerical procedures in the various criteria-establishing methodologies. A common issue, arising from a desire to include as much data as possible, is pseudoreplication. Examples from acoustic criteria are the use of many data points from a single animal to establish a value for one species and the use of several points from one species to set values for a functional hearing group. Less fundamental, but equally problematic for the application of the criteria to policy, is the failure to adequately represent uncertainty around proposed criteria through the use of confidence intervals. Other issues include the uneven treatment of different data in terms of transformation protocols and extrapolation, and the determination of which "outliers" to discard. Each of these errors introduces bias into the resulting criteria. Thus, despite the paucity of relevant data, we need to meet such statistical standards to truly provide objective advice that rises to the level of the "best available science."