SLIDE 1 Sherri Rose, Ph.D.
Associate Professor Department of Health Care Policy Harvard Medical School Co-Director Health Policy Data Science Lab drsherrirose.org @sherrirose
Fair Regression for Health Care Spending
Paper: arxiv.org/abs/1901.10566 Code: github.com/zinka88/Fair-Regression
SLIDE 2 Plan Payment Risk Adjustment
Over 50 million people in the United States currently enrolled in an insurance program that uses risk adjustment
◮ Redistribute funds based on health ◮ Encourage competition based on
efficiency and quality
◮ Massive financial implications
SLIDE 3 Plan Payment Risk Adjustment
Over 50 million people in the United States currently enrolled in an insurance program that uses risk adjustment
◮ Redistribute funds based on health ◮ Encourage competition based on
efficiency and quality
◮ Massive financial implications
Y = θ X
SLIDE 4 Plan Payment Risk Adjustment
Over 50 million people in the United States currently enrolled in an insurance program that uses risk adjustment
◮ Redistribute funds based on health ◮ Encourage competition based on
efficiency and quality
◮ Massive financial implications
Spending outcome Y = θ X
SLIDE 5 Plan Payment Risk Adjustment
Over 50 million people in the United States currently enrolled in an insurance program that uses risk adjustment
◮ Redistribute funds based on health ◮ Encourage competition based on
efficiency and quality
◮ Massive financial implications
Spending outcome Y = θ X Input vector
SLIDE 6 Plan Payment Risk Adjustment
Over 50 million people in the United States currently enrolled in an insurance program that uses risk adjustment
◮ Redistribute funds based on health ◮ Encourage competition based on
efficiency and quality
◮ Massive financial implications
Spending outcome Y = θ X Coefficient vector Input vector
SLIDE 7
Improving Mental Health Care, 1950-2000
Changes in financing and organization of mental health care, not new treatment technologies, made the difference “Improvements ... evolved through ... more money, greater consumer choice, and the increased competition among ... providers that these forces unleashed”
SLIDE 8
SLIDE 9
SLIDE 10
SLIDE 11
Mental Health and Substance Use Disorders (MHSUD)
Risk adjustment in the Marketplaces recognizes only 20% of enrollees with MHSUD Individuals with MHSUD can be systematically discriminated against
SLIDE 12 Algorithmic Fairness
Typical algorithmic fairness problem in computer science has
◮ outcome Y ◮ vector X that includes a protected class or sensitive attribute A ⊂ X
Goal:
Create estimator for f(X) = Y while ensuring the function is fair for A Common measures of fairness are based on the notion of group fairness, striving for similarity in predicted outcomes or errors for groups
SLIDE 13
Can we improve fairness for undercompensated groups in plan payment risk adjustment?
SLIDE 14 Can we improve fairness for undercompensated groups in plan payment risk adjustment? Challenges:
◮ Current formulas created with parametric regression without
built-in fairness criteria
◮ Much of the fairness literature considers binary decision-making
SLIDE 15
Fairness for Risk Adjustment
SLIDE 16
Fairness for Risk Adjustment
Data transformations
Health Economics Computer Science & Statistics
Newhouse (2002) Layton et al. (2016) Bergquist et al. (2018) Kamiran and Calders (2009) Zliobaite et al. (2011) Zemel et al. (2013) Calmon et al. (2017) Johndrow and Lum (2017)
SLIDE 17
Fairness for Risk Adjustment
Adding variables, separate formulas, statistical learning
Health Economics Computer Science & Statistics
van Kleef et al. (2013) Rose (2016) van Kleef et al. (2017) Shrestha et al. (2018) van Kleef et al. (2018) Kamishima et al. (2012) Berk et al. (2017) Zafar et al. (2017a,b) Bechavod and Ligett (2018) Dwork et al. (2018)
SLIDE 18
Fairness for Risk Adjustment
Reinsurance, differing thresholds
Health Economics Computer Science & Statistics
e.g., McGuire et al. (2018) Bansal et al. (2014) Hardt et al. (2016) Kleinberg et al. (2018) El Mhamdi et al. (2018)
SLIDE 19
Global vs. Group Fit
Group undercompensation is a result of large residuals in the risk adjustment formula; we define fairness as a function of these residuals
SLIDE 20
Global vs. Group Fit
Group undercompensation is a result of large residuals in the risk adjustment formula; we define fairness as a function of these residuals
SLIDE 21
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 22
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 23
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 24
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 25
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 26
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 27
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 28
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 29
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 30
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 31
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 32
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 33
Global vs. Group Fit Demonstration
(Rose & McGuire 2019)
SLIDE 34 Can we improve fairness for undercompensated groups in plan payment risk adjustment? Overview of Contributions:
◮ Develop new statistical machine learning methods for group
fairness with continuous outcomes
◮ Propose new evaluation metrics for group fairness with continuous
◮ Demonstrate techniques in plan payment risk adjustment and
simulation settings
SLIDE 35
Large Gains in Group Fairness vs. OLS
(Zink & Rose 2019)
SLIDE 36
Large Gains in Group Fairness vs. OLS
(Zink & Rose 2019)
SLIDE 37
Large Gains in Group Fairness vs. OLS
(Zink & Rose 2019)
SLIDE 38
Large Gains in Group Fairness vs. OLS
(Zink & Rose 2019)
SLIDE 39
Large Gains in Group Fairness vs. OLS
(Zink & Rose 2019)
SLIDE 40
Large Gains in Group Fairness vs. OLS
(Zink & Rose 2019)
SLIDE 41 Future Directions
◮ Multiple groups and detection of new groups
SLIDE 42 Future Directions
◮ Multiple groups and detection of new groups ◮ Integration with approaches from pre- and postprocessing phase ◮ Evaluating social impact of fair regression if deployed
SLIDE 43
Does Your Algorithm Have a Social Impact Statement?
“In both private enterprise and the public sector, research must be reflective of the society we’re serving.”
