OPAB Method Detects Outcome Performativity in Predictive Systems
A new paper on arXiv introduces Outcome Performativity A/B Detection (OPAB), a formal method for detecting outcome performativity—where predictions causally influence the outcomes they predict. The approach assesses dissimilarity in outcome distributions across different prediction groups (interventions). Significant dissimilarity indicates performativity. The authors derive sample complexity bounds under various performative assumption classes and validate them empirically. Results show OPAB can detect performativity in many cases, but also reveal regions of indistinguishability where the number of interventions is insufficient. The method applies to domains like palliative care, credit assignment, and recommender systems.
Key facts
- arXiv paper 2607.26908 introduces OPAB for detecting outcome performativity
- Outcome performativity occurs when predictions causally influence predicted outcomes
- OPAB compares outcome distributions across different prediction groups (interventions)
- Significant dissimilarity in distributions indicates outcome performativity
- Sample complexity bounds derived for various performative assumption classes
- Empirical validation shows OPAB detects performativity in numerous cases
- Regions of indistinguishability exist where intervention count is insufficient
- Applicable to palliative care, credit assignment, and recommender systems
Entities
Institutions
- arXiv