Abstract

One hundred sixty-six samples from Medicare integrity investigations are displayed and described along with 156 associated payment populations. The samples support the All-or-Nothing mixture model for the data previously described in the literature. This model motivates certain Monte Carlo testing procedures for sampling plans as well as stratification methods based on anticipated model moments. A new stratification method is tested, along with existing methods from the literature, on 104 of the payment populations. The new method is found to be viable (i.e. delivering a confidence level consistent with federal requirements) more often than its best competitors while achieving comparable expected recovery. Unfortunately, no competitive algorithm was viable for even half of the 104 test populations in any setting.