Dataset card for release 2026-10

README.md

@@ -0,0 +1,108 @@
+---
+pretty_name: RateCommons Cigna in-network rates
+license: other
+tags: [healthcare, price-transparency, transparency-in-coverage, insurance, parquet]
+configs:
+- config_name: rate_fact
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/rate_fact/*.parquet"
+- config_name: network_set
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/network_set/*.parquet"
+- config_name: network
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/network.parquet"
+- config_name: source_file
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/source_file.parquet"
+- config_name: code
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/code.parquet"
+- config_name: provider_set
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/provider_set/*.parquet"
+- config_name: provider_group
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/provider_group.parquet"
+- config_name: price_list
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/price_list/*.parquet"
+- config_name: price_term
+  data_files:
+  - split: release_2026_10
+    path: "release=2026-10/price_term/*.parquet"
+---
+
+# RateCommons: Cigna in-network rates
+
+Every in-network rate Cigna published under the Transparency in Coverage rule, parsed, deduplicated across files and stored as Parquet. Part of RateCommons, a free public dataset of negotiated health care prices. Code: https://github.com/Harshu-Pande/ratecommons
+
+Release `2026-10`: 124 source files, 304,709,442 distinct rate facts from 1,928,514,784 published rate rows, 11.6 GB.
+
+## Tables
+
+| Table | Rows | What it holds |
+| --- | ---: | --- |
+| `rate_fact` | 304,709,442 | One row per distinct rate fact: a billing code, a provider set and a price list, with the set of networks it appears in (`ns_id`). Sorted by code. 16 parts by billing code. |
+| `network_set` | 13,688 | `ns_id` -> the sorted list of `network_id`s sharing those facts. One file per `rate_fact` part, so a set can repeat across files. |
+| `network` | 124 | One row per derived network: distinct file content (size and gzip trailer). |
+| `source_file` | 124 | One row per source file: name, URL, size, gzip fingerprint, header, counts, parser flags, and its `network_id`. |
+| `code` | 116,566 | Billing codes as published (type, version, code, arrangement), keyed by `code_id`. Descriptor text is withheld where it is licensed (see below). |
+| `provider_set` | 1,659,291 | `ps_id` -> the provider groups (`pg_ids`) a rate applies to. |
+| `provider_group` | 664,715 | `pg_id` -> TIN (type, normalized value, raw values) and sorted NPI list; `tin_wide` when the file lists NPI 0 (the whole TIN). Business names are hints only. |
+| `price_list` | 28,290,701 | `pl_id` -> the price terms (`pt_ids`) of one negotiated_prices list. |
+| `price_term` | 9,390,706 | `pt_id` -> one negotiated price: type, amount (`rate_e4` = dollars x 10,000, plus `rate_raw` exactly as published), billing class, setting, service codes, modifiers, expiration date, additional information, parsed revenue-code ranges and flags. |
+
+## How the pieces fit
+
+IDs are content hashes (xxh3-128 over a canonical encoding of the values), so the same provider group, provider set, price term or price list gets the same ID in every file, release and payer. `code_id` also covers the payer's own wording (name, description, arrangement), so join codes across payers on `billing_code_type` and `billing_code`. A rate is the join
+
+```
+rate_fact -> code (code_id)
+          -> provider_set (ps_id) -> provider_group (pg_ids)
+          -> price_list (pl_id)   -> price_term (pt_ids)
+          -> network_set (ns_id)  -> network / source_file (network_ids)
+```
+
+Example with DuckDB (reads only what the query needs):
+
+```sql
+INSTALL httpfs; LOAD httpfs;
+CREATE VIEW fact AS SELECT * FROM 'hf://datasets/ratecommons/cigna/release=2026-10/rate_fact/*.parquet';
+CREATE VIEW ps AS SELECT ps_id, unnest(pg_ids) AS pg_id FROM 'hf://datasets/ratecommons/cigna/release=2026-10/provider_set/*.parquet';
+CREATE VIEW pg AS SELECT * FROM 'hf://datasets/ratecommons/cigna/release=2026-10/provider_group.parquet';
+CREATE VIEW pl AS SELECT pl_id, unnest(pt_ids) AS pt_id FROM 'hf://datasets/ratecommons/cigna/release=2026-10/price_list/*.parquet';
+CREATE VIEW pt AS SELECT * FROM 'hf://datasets/ratecommons/cigna/release=2026-10/price_term/*.parquet';
+
+SELECT pg.tin_value, pg.npis, pt.negotiated_type, pt.rate_e4 / 1e4 AS rate, pt.rate_raw,
+       pt.billing_class, pt.setting, pt.service_code, fact.ns_id
+FROM fact JOIN ps USING (ps_id) JOIN pg USING (pg_id) JOIN pl USING (pl_id) JOIN pt USING (pt_id)
+WHERE fact.billing_code_type = 'CPT' AND fact.billing_code = '99213'
+  AND pg.tin_value = '123456789';  -- a 9-digit EIN
+```
+
+`rate_fact` is split into 16 parts by `xxh3_64(billing_code_type + '\x1f' + billing_code) % 16`, so a lookup by code can read one part. Files are ZSTD Parquet sorted by code, then provider set.
+
+## Reading it correctly
+
+- A missing rate means "no published rate", never "out of network".
+- A network here is a derived one: one distinct source file. `network_set` says which files carry a fact; `source_file.name` names the file (its product or region). Plan-to-file mappings are in the payer's table of contents, archived daily in `ratecommons/raw-manifests`.
+- Rates keep the published string (`rate_raw`) next to the integer amount (`rate_e4`, ten-thousandths of a dollar). `percentage` rates are percents, not dollars. Read `negotiated_type`, `billing_class` and `setting` before comparing amounts.
+- Provider identity is TIN plus NPI list. Names in payer files (`business_names`, network names) are not identities.
+- Nothing is dropped for being odd: deviations from the schema are kept and counted as flags (`price_term.flags`, `source_file.flags`), for example placeholder amounts of $0.01 or several amounts under identical terms.
+
+## Code descriptions
+
+Code numbers are published; descriptor text is not always ours to republish. Text is withheld (`code.text_withheld = true`, `name` and `description` empty) for CPT (AMA), CDT (ADA), revenue codes and proprietary DRG groupers, including descriptions inside `bundle`. HCPCS Level II, MS-DRG, ICD, APC, HIPPS, NDC and payer-defined codes keep their published text.
+
+## Provenance and license
+
+Source: Cigna's machine-readable files published under 45 CFR 147.212 (Transparency in Coverage). `source_file` lists every file with its URL and fingerprint. RateCommons makes no claim to the underlying rates. This is not medical, legal or financial advice; a published rate is not a quote of what a patient will pay.