Detecting bad actors in value-based payment models

The U.S. federal government is spending billions of dollars to test a multitude of new approaches to pay for healthcare. Unintended consequences are a major consideration in the testing of these value-based payment (VBP) models. Since participation is generally voluntary, any unintended consequences may be magnified as VBP models move beyond the early testing phase. In this paper, we propose a straightforward unsupervised outlier detection approach based on ranked percentage changes to identify participants (e.g., healthcare providers) whose behavior may represent an unintended consequence of a VBP model. The only data requirements are repeated measurements of at least one relevant variable over time. The approach is generalizable to all types of VBP models and participants and can be used to address undesired behavior early in the model and ultimately help avoid undesired behavior in scaled-up programs. We describe our approach, demonstrate how it can be applied with hypothetical data, and simulate how efficiently it detects participants who are truly bad actors. In our hypothetical case study, the approach correctly identifies a bad actor in the first period in 86% of simulations and by the second period in 96% of simulations. The trade-off is that 9% of honest participants are mistakenly identified as bad actors by the second period. We suggest several ways for researchers to mitigate the rate or consequences of these false positives. Researchers and policymakers can customize and use our approach to appropriately guard VBP models against undesired behavior, even if only by one participant. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s10742-021-00253-9.