Plan-Analytics/medi-cal-synthetic

Synthetic Medi-Cal Dataset (v1)

Synthetic member, eligibility, claims/encounter and pharmacy data that behaves like California Medi-Cal (Medicaid) managed care data, with no real people in it. It was built for the PlanAnalytics YouTube tutorials on population health analytics: viewers can run every assessment shown on screen against the same data, using the same code.

• 500,000 members, CY2023–CY2025, 8 California regions, 5 eligibility segments with real DHCS aid codes
• 12.1M claim lines, 7.2M pharmacy fills, 11.4M member-months, about 330 MB of Parquet
• Calibrated to public DHCS, MACPAC, CHCF and GAO figures. See How realistic is it? below.
• Free to download and query, with no account or credentials needed: it is hosted on Hugging Face

Location: https://huggingface.co/datasets/Plan-Analytics/medi-cal-synthetic (DuckDB/Polars path hf://datasets/Plan-Analytics/medi-cal-synthetic/...). The v1 tag pins this release: hf://datasets/Plan-Analytics/medi-cal-synthetic@v1/...

Quickstart

DuckDB (0.10.3+) reads the files in place, so you only download the columns and partitions a query touches:

INSTALL httpfs; LOAD httpfs;

-- Medical paid per member per month (PMPM) by eligibility category, CY2024
WITH mm AS (
  SELECT eligibility_category, count(*) AS member_months
  FROM 'hf://datasets/Plan-Analytics/medi-cal-synthetic/member_month/year=2024/*/*.parquet'
  GROUP BY 1),
final_claims AS (   -- final-action version of each claim
  SELECT * FROM 'hf://datasets/Plan-Analytics/medi-cal-synthetic/claim_line/year=2024/*/*.parquet'
  QUALIFY adj_seq = max(adj_seq) OVER (PARTITION BY claim_id)),
paid AS (
  SELECT m.eligibility_category, sum(c.paid_amt) AS paid
  FROM final_claims c
  JOIN 'hf://datasets/Plan-Analytics/medi-cal-synthetic/member_month/year=2024/*/*.parquet' m
    ON c.member_id = m.member_id AND date_trunc('month', c.service_from_date) = m.month
  WHERE c.claim_status = 'PAID'
  GROUP BY 1)
SELECT eligibility_category, round(paid / member_months, 2) AS pmpm
FROM mm JOIN paid USING (eligibility_category) ORDER BY pmpm DESC;

Python (pandas via DuckDB, one region and year at a time):

import duckdb
la_2024 = duckdb.sql("""
    SELECT * FROM 'hf://datasets/Plan-Analytics/medi-cal-synthetic/claim_line/year=2024/region=LA/*.parquet'
""").df()

Polars: pl.scan_parquet("hf://datasets/Plan-Analytics/medi-cal-synthetic/claim_line/**/*.parquet", hive_partitioning=True).

Download everything (about 330 MB) to work offline:

pip install huggingface_hub
hf download Plan-Analytics/medi-cal-synthetic --repo-type dataset --local-dir medi-cal-synthetic

Heavy anonymous use can hit Hugging Face rate limits (HTTP 429). A free account raises them: run hf auth login once, then in DuckDB CREATE SECRET hf_token (TYPE huggingface, PROVIDER credential_chain);. Downloading once with hf download (above) and querying the local copy avoids the limits entirely.

Layout

(repo root)
  members/members.parquet                     one row per member
  provider/provider.parquet                   providers and facilities
  eligibility_span/year=YYYY/region=XX/       coverage spans (aid code, plan, FFS vs managed care, dual)
  member_month/year=YYYY/region=XX/           one row per member per enrolled month
  claim_line/year=YYYY/region=XX/             medical claims and encounters, one row per line per version
  rx_claim/year=YYYY/region=XX/               pharmacy fills (Medi-Cal Rx)
  data_dictionary.md  manifest.json  README.md  LICENSE

year is the year of the claim's first service date, the fill date, or the span or month start. region is the member's region: LA, BAY, SD_IMP, IE, ORANGE, CV, SAC, COAST_NORTH. DuckDB and Polars read them back as columns from the folder names (hive_partitioning). All tables join on member_id. Column definitions are in data_dictionary.md.

Things the data makes you handle, as real data does

• Claim versions. claim_id is stable across versions and adj_seq increases. For final-action claims, keep the highest adj_seq per claim_id and drop REVERSED and DENIED. About 0.3% of claims are exact duplicates under a new id (prefix D).
• Claim vs line. Facility claims carry several revenue-code lines. Count visits and admissions on distinct claim_id; sum dollars over lines.
• Run-out (IBNR). The extract date is 2026-06-30. Only claims paid by then are present, so late-2025 months are slightly incomplete.
• Enrollment churn. Gaps, partial years, mid-window enrollees and deaths. Use member_month for denominators, not a member count.
• Delivery systems. About 95% of member-months are managed care, where services arrive as source_system = ENCOUNTER. Some services come through separate systems: fee-for-service months (CLAIMS), county specialty mental health (COUNTY_MHP), Drug Medi-Cal (DMC_ODS), In-Home Supportive Services (IHSS), Regional Center developmental-disability waiver services (DDS_WAIVER), and pharmacy through Medi-Cal Rx (rx_claim).
• Duals. For members also on Medicare (dual_flag), Medicare pays first. paid_amt is Medi-Cal's share and other_payer_paid_amt is Medicare's. Use paid_amt for Medi-Cal spend.
• CalAIM. Enhanced Care Management (G9008-U1) and Community Supports appear in calaim_service; member_month.ecm_enrolled_flag marks ECM months.

How realistic is it?

Targets come from public sources, cited in the generator's docs/calibration_sources.yaml. The CY2024 build lands on them as follows:

Measure	Target	This build
Member mix: children / expansion / parents & pregnant / aged / disabled	33.1 / 34.4 / 15.5 / 10.0 / 7.0% (DHCS, Jul 2025)	33.9 / 34.3 / 15.3 / 9.7 / 6.7%
Medi-Cal paid per member-year: children / expansion / parents / aged / disabled	$3,820 / $7,080 / $4,840 / $18,070 / $37,250 (MACPAC FY2023, trended)	$3,845 / $6,798 / $4,733 / $17,121 / $37,480
ABD share of spend	~0.50	0.49
Top 1% / top 5% of members' share of spend	0.25 / ~0.55	0.25 / 0.56
Members with no spend in the year	0.12	0.12
Managed care share of member-months	0.94–0.95	0.95
Dual-eligible share	0.124	0.110
Births per 1,000 women 15–44	~44	44

It is still synthetic. Utilization rates, unit costs, provider networks and plan names are modeling assumptions, and every person, provider, plan and identifier is invented. The patterns are built to teach analytics, not to estimate anything about real Medi-Cal members, plans or providers. Do not use it for policy, rate-setting or clinical conclusions.

License and citation

Data: CC0 1.0 (public domain dedication). manifest.json lists row counts and content hashes for each table. Suggested citation: PlanAnalytics Synthetic Medi-Cal Dataset v1 (2026).