Medicare Provider Outliers

Cost & volume outlier detection across 7.06M CMS Part D / Part B providers, scored within (NPPES taxonomy × state) peer groups using a robust MAD modified z-score. Modeled in dbt on Snowflake; this site is served static from committed extracts — no live warehouse.

Who this is for — and why

$213B in Part D drug spend and $73B in Part B payments flow through these providers in a single data year. Finding the few hundred-thousand genuinely anomalous ones — without drowning in false positives from skew — is the difference between a usable triage list and noise. Three audiences read the mart differently:

• Payer & state Medicaid program-integrity teams — treat a flag as a triage signal to combine with their own beneficiary-overlap and audit tooling, then trim the noisy long tail using the per-metric peer-coverage columns.
• Healthcare journalists & policy researchers — start from the ranked table and the state rollup to find defensible, public-data leads worth investigating; the Methodology disclaimer guides the framing.
• Analytics & hiring reviewers — a worked example of robust outlier detection (MAD vs. classical z) on a real, messy, 43.7M-row public dataset.

Robust beats classical on skewed cost data

The MAD method flags 5.23% of providers versus the z-score's conservative 2.01% — the right call for right-skewed Medicare spend, where the mean and standard deviation get dragged toward the very tails we're trying to detect.

Explore:

• Outlier Detection — the parameterized, ranked outlier table
• Provider Drill-Down — a provider against their peer-group distribution
• Geography — outlier rate by state
• Methodology — peer groups, z vs MAD, and what a flag does not mean