10 Fairness

 

Sospital is a leading (fictional) health insurance company in the United States. Imagine that you are the lead data scientist collaborating with a problem owner in charge of transforming the company’s care management programs. Care management is the set of services that help patients with chronic or complex conditions manage their health and have better clinical outcomes. Extra care management is administered by a dedicated team composed of physicians, other clinicians, and caregivers who come up with and execute a coordinated plan that emphasizes preventative health actions. The problem owner at Sospital has made a lot of progress in implementing software-based solutions for the care coordination piece and has changed the culture to support them, but is still struggling with the patient intake process. The main struggle is in identifying the members of health plans that need extra care management. This is a mostly manual process right now that the problem owner would like to automate.

10.1  The different definitions of fairness

10.2  Where does unfairness come from?

10.3  Defining group fairness

10.3.1    Statistical parity difference and disparate impact ratio

10.3.2    Average odds difference

10.3.3    Choosing between statistical parity and average odds difference

10.3.4    Average predictive value difference

10.3.5    Choosing between average odds difference and average predictive value difference

10.3.6    Conclusion

10.4  Defining individual and counterfactual fairness

10.4.1    Consistency

10.4.2    Counterfactual fairness

10.4.3    Theil index

10.4.4    Conclusion

10.5  Mitigating unwanted bias

10.5.1    Pre-processing

10.5.2    In-processing

10.5.3    Post-processing

10.5.4    Conclusion

10.6  Other considerations

10.7  Summary