Outlier Detection

The triage shortlist — providers ranked by how far they sit from their specialty-and-state peers on the metric you pick. A high score is a candidate to investigate, not a verdict.

Pick a metric — the table re-queries instantly in your browser (DuckDB-WASM, no server round-trip). Sort any column, or search by provider, specialty, or state.

What the columns mean (full math + what a flag does not mean: Methodology):

• Modified z (MAD): robust distance from the peer-group median, 0.6745 × (value − median) / MAD. Flags at |z| ≥ 3.5 — unitless, and genuinely huge on skewed spend (the top row's ~134,510 is not a typo).
• Peer n: providers in the same (NPPES taxonomy × state) peer group, floored at n ≥ 30.
• Metric peer n: how many of those peers have a value for the selected metric — a flag requires this ≥ 30, so no provider is scored against a thin median.
• Value: the selected metric — USD totals for total_drug_cost / total_medicare_payment, counts for claims / services, a 0–1 ratio for brand_cost_share.

The top total_drug_cost row is an Emergency Medicine provider in IL — $84M in Part D drug cost against a peer-group median that puts the modified z near 134,510. High Emergency-Medicine scores are typically facility-level prescribing aggregated to one attending NPI — see Methodology.

Modified-z distribution — total_drug_cost

The shape of the top-1,000 tail for the selected metric: