Cancer Basemap

Browse the releases Try it in your browser Cite a release View on GitHub

Cancer Basemap publishes population cancer data — screening, risk factors, the social and environmental context around them, and the places people get care — as Apache Iceberg tables keyed on geography and time. Every release every source has ever published stays queryable, from DuckDB, R, Python, or a browser, with no account and no token.

The lake exists so a cancer center can write, and defend, one sentence: in our catchment, as reported by source S in release R, the measure was X. Today that sentence usually rests on a screenshot, a downloaded CSV of unknown vintage, or a dashboard that has since silently changed underneath it. Here, the release that produced a number stays attached to it.

Live county map

The map below reads a small, versioned extract of measure.observation written at build time — the same numbers you’d get querying the catalog directly, just not re-queried on every page load. “Run this live in your browser” runs the same measure/release query against the real, 25-million-row table through DuckDB-WASM — scoped to one state at a time rather than the whole country, for the reason given on that page.

Note

Two known gaps in this map, verified against the data:

• Connecticut doesn’t join. PLACES 2024/2025 report Connecticut’s nine 2022-vintage planning regions; the county boundary shapes used to draw this map (us-atlas, still on legacy county lines) only have the eight counties those regions replaced. Connecticut draws as no-data here — a boundary-vintage mismatch worth surfacing, not a bug to hide (see Concepts › Boundary vintages).
• Kentucky and Pennsylvania show only 2 of these 6 measures for release 2025 — mammography use and colorectal cancer screening — with cancer (non-skin)/melanoma, current smoking, obesity, and annual checkup all missing. Verified directly against raw.places__county: both states report 40 distinct measures for 2024 but only 5 for 2025 (PLACES-wide, not specific to this map’s 6), and CDC’s own PLACES release notes confirm why — “Estimates for Kentucky and Pennsylvania were not available for measures based on the 2023 BRFSS. However, the five measures based on the 2022 BRFSS were carried over for Kentucky and Pennsylvania” (mammography use, colorectal cancer screening, dental visits, short sleep duration, and all teeth lost — only the first two are among this map’s 6) — the two states did not collect enough months of 2023 BRFSS data to meet PLACES’ own inclusion threshold, so those five measures are simply carried forward from the prior data year rather than missing.

mapData = FileAttachment("extracts/home_map.json").json()

viewof measureId = Inputs.select(mapData.measures, {
  label: "Measure",
  format: (d) => d.label,
  value: mapData.measures[0]
})
viewof release = Inputs.select(mapData.releases, { label: "Release" })

STATUS_LABEL = ({
  R: "reported", S: "suppressed (small count)", C: "suppressed (complementary)",
  U: "suppressed (unreliable)", N: "not available", X: "not applicable"
})

rowsForSelection = mapData.obs[release]?.[measureId.id] ?? []
byFips = new Map(rowsForSelection.map((r) => [r[0], r]))

// County boundaries live nowhere in the lake (it holds tabular data, not
// geometry) — loaded once, at view time, from a public TopoJSON CDN.
countyGeo = {
  const topojsonMod = await import("https://cdn.jsdelivr.net/npm/topojson-client@3/+esm");
  const us = await fetch("https://cdn.jsdelivr.net/npm/us-atlas@3/counties-10m.json").then((r) => r.json());
  return topojsonMod.feature(us, us.objects.counties);
}

joined = countyGeo.features.map((f) => {
  const row = byFips.get(f.id);
  const name = mapData.counties[f.id] ?? f.properties.name;
  return { ...f, properties: { ...f.properties, fips: f.id, name, row } };
})

Plot.plot({
  projection: "albers-usa",
  color: {
    type: "sequential",
    scheme: "blues",
    label: measureId.label,
    legend: true,
    nice: true
  },
  marks: [
    Plot.geo(joined, {
      fill: (d) => (d.properties.row && d.properties.row[4] === "R") ? d.properties.row[1] : null,
      stroke: "white",
      strokeWidth: 0.3,
      title: (d) => {
        const row = d.properties.row;
        if (!row) return `${d.properties.name} (${d.properties.fips}): no data`;
        if (row[4] !== "R")
          return `${d.properties.name} (${d.properties.fips}): ${STATUS_LABEL[row[4]]} — no value`;
        return `${d.properties.name} (${d.properties.fips})\n${measureId.label}: ${row[1]}% (CI ${row[2]}–${row[3]})\nrelease ${release}`;
      }
    }),
    // Suppressed, unavailable, or unmatched counties: drawn grey, never left
    // to read as zero (AGENTS.md's suppression gate applies to how the map
    // draws, not only to the value column).
    Plot.geo(
      joined.filter((d) => !d.properties.row || d.properties.row[4] !== "R"),
      { fill: "#ddd", stroke: "white", strokeWidth: 0.3, fillOpacity: 0.6 }
    )
  ],
  width: 900,
  height: 550
})

no data, suppressed, or not applicable  ·  Run this live in your browser

Where to go next

• Want the data files? Releases lists every release on R2 and Hugging Face; the playground queries the latest one in your browser, and Versions and citation says which version to pin and cite.
• New here? Start here has the exact ATTACH statement for R, SAS, SEER*Stat, Python, and the DuckDB CLI.
• Building a catchment report? Concepts covers the three time axes, why suppression is a status rather than a blank cell, and boundary vintages — the things that make a query wrong in a way that still runs.
• Looking for a specific measure or table? The Dictionary is generated straight from the catalog and the declared schemas, regenerated at each site build — not hand-maintained, though (like this page’s map) it’s a snapshot as of that build, not a live query.