Work in Progress:

Performance vs Simplicity in Predictions for Public Policy: An Example from Medicare Advantage

Abstract Forthcoming

Risk Adjustment is Inaccurate for Complex Patients because Health Conditions have Increasing Marginal Effects

Background: Risk adjustment is used to adjust insurance premium payments for individuals’ health characteristics. However, current risk adjustment models consistently underestimate costs for multi-morbid patients.  Previous work has focused on the role of omitted variables bias in causing these inaccuracies. This project proposes and evaluates another potential explanation: incorrect functional form assumptions.

Theory: Most risk adjustment models assume that a health condition increases costs the same amount for every person, i.e. has a constant marginal effect.  However, the marginal effect of a health condition may increase in the presence of other health conditions.  This pattern can explain why current risk adjustment models underestimate costs for multi-morbid individuals and overestimate costs for healthy individuals.

Data and Methods: I look for evidence of increasing marginal effects in the Truven Marketscan Data 2014-2015 (n=9,574,907).  I estimate a risk adjustment model similar to models used by Medicare. I also calculate an empirical marginal effect using exact matching. I then compare model and empirical marginal effects.

Results: The 50 most common health conditions all display increasing marginal effects in the number of health conditions. In addition, the distribution of health conditions per person is highly skewed right. As a result, the risk adjustment model substantially underestimates the marginal effect a few highly morbid individuals and modestly misestimates the effect for many relatively healthy people.

Conclusion: Functional form assumptions must be relaxed to generate unbiased cost predictions for multi-morbid patients. Future work should focus on developing risk adjustment models which allow for heterogeneous marginal effects.


Miscellaneous Projects:

Best Subset Selection: Some Recommendations for Practitioners