# Risk Adjustment: Predicting Future Expenditures

Many states rely on managed care organizations (MCOs) to provide medical services for their Medicaid beneficiaries.  Contracting out medical services to private providers relies on the government’s capacity to accurately predict expected cost of care for each beneficiary.  This is typically done through risk-adjusted capitation rates.

Which risk adjustment strategy works best?  The answer of course depends on the context.  A paper by Yu and Dick (2010) examines 5 predictors specifications to predict future expenditures for Medicaid eligible children.  I list each of the five specifications and their performance (measured as the R2) below:

• Age/Gender only: 0.2%
• Age/Gender + subjective health status measure: 3.9%
• Age/Gender + CSHCN: 7.3%
• Age/Gender + HCC: 12.1%
• Age/Gender + prior year expenditure: 43.5%

One can clearly see that the best predictor of a child’s current year expenditures is the child’s prior year’s expenditures.

In the list above, the subjective health status is based on a one-to-five scale.  CSHCN stands for a series of children with special health care needs (CSHCN) identified by the answers to five questions which include “need or use of prescription drug, having limitation, need or use more health care than is usual for other children of the same age, need or use special therapy, needing or using counseling for at least 12 months.”  Hierarchical Condition categories uses the patient’s previous year’s ICD-9 codes (i.e., diagnosis codes) to create a scoring system indicating the beneficiary’s relative illness.

Using data from the Medical Expenditure Panel Survey (MEPS), future expenditures are predicted using a two-step process.  The first equation indicates whether or not there is any medical service utilization, U:

• Logit(U) = α1 + β1Xi

The second equation is a continuous function measuring total expenditures,Y, conditional on any medical care utilization.

• Y = α2 + β2Xi

The authors estimate the first equation with a logit specification and the second equation with a GLB model with a Poisson distribution and a log link (for more information on GLM models, see this post).

Source: