Duke 2008 | ASHE | Equity and Efficiency in Health and Healthcare

Prospective payment arrangements, coupled with value-based-purchasing or pay-for-performance provisions and competition promoting policies, have become dominant vehicles in the U.S. to promote incentives for cost-effective, quality health care, furthering technological innovation, and maximizing patient care access. One general approach to balancing payment along the continuum from fee-for-service and fully prospective payment entails risk adjustment methods which consist of risk instruments that organize diagnoses (e.g.) into risk categories, risk measures that proxy for care needs such as patient expenditure, and risk models that generate prediction results from estimating risk measures with risk instruments. Risk adjustment is the application of a risk model to payment or other goals. We develop a patient expenditure model intended to make optimum risk adjustment of patient based prospective payment, taking account of specific payment goals in our methods. This literature has focused on the accuracy of patient expenditure modeling, recently adding consideration of various types of General Linear Models (GLMs) as the risk model. However, there has been almost no attention in the literature to specific effects of risk adjustment to reduce adverse selection incentives and directly improve payment fairness to providers. General measures of predictive power for these models do not address the main concern that unobserved latent factors could cause predicted costs to deviate in ways that offer cream skimming opportunities through emphasis on particular services or barriers to particular populations. The concurrent modeling example in this paper will use U.S. Department of Veterans Affairs (VA) FY2000 expenditure data and a modification of the Diagnostic Cost Group/Hierarchical Condition Category (DCG/HCC) risk instrument to illustrate how payment can be adapted when one can identify specific types of patients who are systematically mis-predicted. The VA data is randomly sampled into two 100,000 patient groups for estimation and prediction intended to capture a common order of magnitude for local health plan estimation. In particular, the main point of this paper is to assert that symmetric loss functions where under-prediction and over-prediction are treated equally, adjusted for numerous functional forms that are tested here, creates opportunities for greater incentives for cream skimming if information can be used not readily added to the risk adjustment process. We accomplish this partly by using Pregibon Link Tests and adapted Hosmer-Lemeshow tests with 20 expenditure categories to accentuate parts of the distribution that over-predict and under-predict expenditures. Then the innovation of this paper is to use quantile regression with asymmetric loss functions that use the median as the optimum estimate. Our results illustrate a number of key points. First, that the main difference between OLS and GLM models is in what parts of the distribution of patients are over-predicted and under-predicted. Second, following Glazer and McGuire (AER, 2000), we can test quantile systems for over-paying on expectations of high cost and under-paying on expectations of low cost. Third, some people are high cost because they have many co-morbid HCCs while others are high cost because of a single expensive condition and these types of patients have very different implications for risk adjustments.