Methodology

How we identify unusual billing patterns across 522,900 Medicaid providers.

Important Disclaimer

This dashboard identifies statistically unusual billing patterns — it does not accuse anyone of fraud. There are many legitimate reasons a provider may be flagged: specialty drugs, per-diem billing, large multi-provider organizations, academic medical centers, and county health systems all routinely trigger statistical outlier detection. Every flag should be interpreted as a question (“Why is this unusual?”), not an answer. Proper investigation requires access to claim-level detail, clinical records, and subject-matter expertise that is beyond the scope of this aggregate analysis.

Data Source

Dataset: CMS Transformed Medicaid Statistical Information System (T-MSIS) Analytic Files, specifically the “Top 500 Most Frequently Billed HCPCS Codes by Provider” file.

Source: data.medicaid.gov (CMS Open Data)

Records: 227 million rows covering the top 500 HCPCS codes billed by each provider, 2018-2024.

Coverage: All 50 states + District of Columbia — 522,900 unique providers, $967.5 billion in total spending, 16.5 billion claims.

Provider enrichment: NPIs cross-referenced against the NPPES National Plan and Provider Enumeration System (8.96M US healthcare providers) for provider names, entity types, and taxonomy codes. Taxonomy codes translated to human-readable specialties using the NUCC Healthcare Provider Taxonomy Code Set (Version 25.1, 883 codes).

Limitations: Data is aggregated at the provider-state-HCPCS level (not individual claim lines). We cannot see individual patient encounters, diagnoses, or modifier codes. This limits the precision of volume-based detection methods.

Statistical Approach

Robust statistics: We use median and Median Absolute Deviation (MAD) instead of mean and standard deviation. MAD is resistant to the very outliers we're trying to detect — using mean/std would let a single extreme provider inflate the threshold and escape detection.

Code-specific benchmarks: Rather than comparing all providers against a single threshold, we compute separate benchmarks for each of 6,752 HCPCS codes. A dermatologist billing $200/claim is compared to other dermatologists billing the same code, not to pharmacies dispensing $5 prescriptions. Full percentile distributions (p10 through p99) are stored in the code_benchmarks table.

Multi-flag requirement: A provider must be flagged by at least 2 independent tests to appear on the watchlist. Single-flag providers are excluded except for high-confidence signals (LEIE exclusion). This dramatically reduces false positives — if each test has a 10% FP rate, the probability of 2+ false flags is 1%.

Detection Tests (13 total)

Spending

High Total Spending

Flags providers whose total Medicaid payments exceed the state p99 (99th percentile).

Threshold: Total paid > state 99th percentile

Catches: Unusually large billing operations within a state

Code-Specific Cost Outlier

Compares each provider's average cost per claim against the national median for the same HCPCS code, using Median Absolute Deviation (MAD) instead of standard deviation for robustness.

Threshold: MAD score > 5.0 (5+ MADs above the median for their top HCPCS code)

Catches: Upcoding, billing at rates far above peers for the same procedure

Multi-Code Rate Outlier

Extends cost outlier analysis across all HCPCS codes a provider bills, weighted by volume. Providers whose volume-weighted average MAD score is extreme are flagged.

Threshold: Weighted MAD score > 3.0 across all billed codes

Catches: Providers systematically overbilling across their entire practice, not just one code

Volume

High Claims per Beneficiary

Identifies providers where the ratio of claims to unique beneficiaries is abnormally high.

Threshold: Claims/beneficiary > state median + 5 * MAD, AND > 100 claims/beneficiary

Catches: Overutilization — performing unnecessary services on the same patients

Billing Velocity

Calculates claims per working day (22 days/month) to identify physically impossible billing rates.

Threshold: Individual providers: >50 claims/day; Organizations: >5,000 claims/day

Catches: Phantom billing — submitting claims for services never rendered

Pattern

Single Code Mill

Flags providers who bill 90%+ of their claims under a single HCPCS code, suggesting a "mill" operation.

Threshold: >90% of claims under one code AND >1,000 total claims

Catches: Procedure mills (e.g., urine drug testing mills, therapy mills)

E&M Level Upcoding

Compares each provider's E&M code level distribution (99211-99215) against their specialty peers in the same state. Upcoding patterns are among the most common findings in OIG program integrity reviews.

Threshold: Average E&M level > peer average + 0.5, chi-squared > 100, high-level code % > peer baseline + 20%

Catches: Systematic billing of higher-complexity office visit codes than justified — upcoding

Growth

Billing Spike

Detects months where a provider's billing jumps >3x their median monthly billing.

Threshold: Any month > 3x the provider's median monthly paid amount

Catches: Sudden ramp-up in billing -- a pattern that warrants further review

Explosive Growth

Compares the last 12 months of billing to the previous 12 months.

Threshold: Second-half billing > 3x first-half billing

Catches: Rapidly growing billing operations

New High-Volume Entrant

Flags providers who appear in only the last 2 years of data but already bill above the state median.

Threshold: Active ≤ 24 months AND total paid > state median

Catches: New providers who immediately bill at high volumes — potential shell companies

Temporal

CUSUM Change Point

Uses Cumulative Sum (CUSUM) control charts to detect the exact month when a provider's billing pattern underwent a permanent, sustained shift.

Threshold: Post-change mean > 3x pre-change mean, minimum 6 months of data

Catches: Abrupt regime changes in billing behavior, pinpointing the exact month

External

LEIE Exclusion

Cross-references all providers against the OIG List of Excluded Individuals/Entities (LEIE). Excluded providers are legally barred from billing federal healthcare programs.

Threshold: NPI match in LEIE database (82,714 excluded entities)

Catches: Providers billing Medicaid while officially excluded — potentially illegal

Risk Tiers

TierFlags RequiredProvidersFalse Positive Estimate
Highest Confidence5+~1,7000.001% (1 in 100,000)
Critical4~3,9000.01% (1 in 10,000)
High3~11,3000.1% (1 in 1,000)
Elevated2~37,3001% (1 in 100)
Moderate1 (high-confidence)~370LEIE match only

False positive estimates assume test independence. In practice, some tests are correlated (e.g., high spending and cost outlier), so actual FP rates may be higher than shown.

Known Limitations

  • Government entities: County health departments, state hospital systems, and public universities (e.g., “CITY OF NEW YORK”, “NYC HEALTH AND HOSPITALS CORPORATION”) are large by design and frequently trigger volume-based flags. They appear on the watchlist but are expected given their scale and mission.
  • Per-diem and bundled billing: Facilities that bill per-diem rates (e.g., nursing facilities) may show high claims-per-beneficiary ratios that are normal for their billing model.
  • Specialty drugs: Providers dispensing high-cost specialty medications will have legitimately high cost-per-claim figures.
  • Fiscal intermediaries: Some NPIs represent billing intermediaries, not individual practitioners. They aggregate billing for many downstream providers, inflating their apparent volume.
  • Aggregate data only: This dataset provides provider-level aggregates, not claim-line detail. We cannot assess medical necessity, verify services were rendered, or examine modifier codes.
  • LEIE name matching: We match LEIE entries by NPI only. Providers who were excluded under a different NPI or before receiving an NPI may not be detected.

Explore the Analysis