Spending by county and HCPCS codes
This map visualizes the Medicaid Provider Spending dataset published by the U.S. Department of Health & Human Services (HHS), Centers for Medicare & Medicaid Services (CMS). The dataset contains 238 million rows covering 21.8 billion claim lines from January 2018 through December 2024, representing approximately $21.8 trillion in total Medicaid payments. Each row aggregates claims by billing provider, servicing provider, HCPCS code, and month.
The dataset contains 12,263 unique HCPCS (Healthcare Common Procedure Coding System) codes. To keep the interactive map performant, only the top 35% most popular codes by claim volume are included (3,893 codes). These codes account for 99.8% of all claim lines and over 99.9% of all dollars paid. Code descriptions are sourced from the CMS 2026 Physician Fee Schedule Relative Value Unit (RVU) file.
The spending dataset identifies providers by National Provider Identifier (NPI) numbers but does not include geographic information. To map spending to counties, a three-step lookup is performed:
Each row in the spending dataset may contain two NPI numbers:
For geographic assignment, the servicing provider’s location is used when available, with the billing provider’s location as a fallback. This reflects where the care was delivered rather than where the claim was filed.
Among the 218 million rows where both NPIs could be resolved to a ZIP code via NPPES:
Using the servicing-then-billing NPI fallback strategy:
The unmapped claims fall into four failure categories in the NPI → ZIP → County lookup chain:
| Failure reason | % of unmapped claims | Description |
|---|---|---|
| Both NPI fields empty | 85.1% | Neither billing nor servicing NPI is populated. These are high-volume aggregated rows (averaging ~7,500 claims each) that represent managed care organization (MCO) bulk-reported claims where individual providers are not identified. |
| NPIs not found in NPPES | 7.4% | NPI values are present but do not appear in the NPPES registry. These include placeholder values (e.g., “0000000000”), atypical identifiers beginning with letters, and state Medicaid agency codes that are not standard NPI numbers. |
| ZIP codes not in Census ZCTA crosswalk | 5.6% | NPIs resolve to valid ZIP codes in NPPES, but 812 of those ZIP codes do not appear in the Census ZCTA-to-County relationship file. These are predominantly institutional or campus ZIP codes assigned by USPS to large organizations (e.g., 44195 — Cleveland Clinic, 94143 — UCSF, 01199 — Brigham & Women’s, 84132 — University of Utah Health) that have no corresponding residential ZCTA. |
| Mixed failures | 2.0% | One NPI is absent from NPPES while the other’s ZIP code is not in the crosswalk, so neither path yields a county. |
Although the 2.4 billion unmapped claim lines cannot be placed on the map, they reveal distinctive patterns about what types of Medicaid services fall outside the NPI-based geolocation chain.
The unmapped claims are dominated by services that operate outside the traditional healthcare facility model:
Unmapped rows are structurally different from mapped rows:
TOTAL_PAID field in unmapped rows exhibits extreme outliers — a small number of rows (concentrated in 2018) carry implausibly large dollar values that account for the vast majority of the dataset’s nominal dollar total. Because of this, dollar-based statistics for unmapped claims are unreliable and are not reported here. The per-enrollee spending values shown on the map (derived from mapped claims only) are not affected.The share of unmapped claims has gradually declined from 12.9% in 2018 to 10.4–11.4% in 2022–2024, suggesting improving coverage of the NPPES provider registry over time.
Dollar values on the map are per-enrollee amounts, computed as:
The TOTAL_PAID field from the spending dataset is summed by county, HCPCS code, and calendar quarter. County-level Medicaid enrollment counts are derived from the American Community Survey (ACS) 2023 5-year estimates, Table C27007 (Medicaid/Means-Tested Public Coverage Status by Sex by Age). The six “With Medicaid/means-tested public coverage” age-sex subgroups are summed to produce a single enrollment figure per county. A single enrollment figure is used across all quarters (2018–2024).
Note: The ACS measure includes CHIP and other state means-tested programs alongside Medicaid. Survey-based estimates are known to undercount Medicaid enrollment by roughly 10–20% compared to administrative records. Total county population from the Census Bureau’s 2024 population estimates is also shown for reference.
The source data is monthly. Claims are grouped into calendar quarters (Q1 = Jan–Mar, Q2 = Apr–Jun, Q3 = Jul–Sep, Q4 = Oct–Dec) based on the CLAIM_FROM_MONTH field. The dataset spans Q1 2018 through Q4 2024 (28 quarters).
The choropleth uses a sequential blue color scale with linear interpolation. To prevent extreme outliers from compressing the visible color range, the scale maximum is set at the 98th percentile of per-enrollee values across all counties for the selected code and quarter. Counties above this threshold are shown in the darkest blue. Counties with no data for the selected code/quarter are shown in gray.