Predicting inpatient hospital payments in the United States: a retrospective analysis

The Affordable Care Act (ACA) has increased rates of public and private health insurance in the United States. Increasing coverage could raise hospital revenue and reduce the need to shift costs to insured patients. The consequences of ACA on hospital revenues could be examined if payments were known for most hospitals in the United States. Actual payment data are considered confidential, however, and only charges are widely available. Payment-to-charge ratios (PCRs), which convert hospital charges to an estimated payment, have been estimated for hospitals in 10 states. Here we evaluated whether PCRs can be predicted for hospitals in states that do not provide detailed financial data.

We predicted PCRs for 5 payer categories for over 1,000 community hospitals in 10 states as a function of state, market, hospital, and patient characteristics. Data sources included the Healthcare Cost and Utilization Project (HCUP) State Inpatient Databases, HCUP Hospital Market Structure file, Medicare Provider of Service file, and state information from several sources. We performed out-of-sample prediction to determine the magnitude of prediction errors by payer category.

Many individual, hospital, and state factors were significant predictors of PCRs. Root mean squared error of prediction ranged from 32 to over 100 % of the mean and varied considerably by which states were included or predicted. The cost-to-charge ratio (CCR) was highly correlated with PCRs for Medicare, Medicaid, and private insurance but not for self-pay or other insurance categories.

Inpatient payments can be estimated with modest accuracy for community hospital stays funded by Medicare, Medicaid, and private insurance. They improve upon CCRs by allowing separate estimation by payer type. PCRs are currently the only approach to estimating fee-for-service payments for privately insured stays, which represent a sizable proportion of stays for individuals under age 65. Additional research is needed to improve the predictive accuracy of the models for all payers.