Methodology

pumsdata turns the U.S. Census Bureau's American Community Survey (ACS) Public Use Microdata Sample (PUMS) into interactive maps, rankings, and custom population estimates. This page explains exactly where the numbers come from, how they are computed, and the limitations you should keep in mind.

1. Data source & vintage

All current figures come from the ACS 2024 1-Year PUMS, published by the U.S. Census Bureau — the most recent complete PUMS release. It covers 51 geographies (the 50 states plus the District of Columbia) and 2,462 Public Use Microdata Areas (PUMAs) — Census-defined sub-state areas of roughly 100,000+ residents each, the most detailed geography PUMS supports.

The release is a sample of 3,422,888 de-identified person records (plus the associated housing records), weighted to represent a total population of 340,110,990 people.

2. Microdata & weighting

Unlike the Census Bureau's published summary tables (which are pre-computed totals), PUMS is record-level microdata: one de-identified row per sampled person or housing unit, each with hundreds of variables. That is what lets pumsdata answer questions the standard tables don't — arbitrary cross-tabulations like "renters aged 25–34 without a vehicle."

Because PUMS is a sample, every record carries a statistical weight(person weight PWGTP, housing weight WGTP) indicating how many people or households it represents. pumsdata applies these weights on every query — a population estimate is the sum of the weights of the matching records, not a raw record count. Statistics are computed over the correct universe: for example, the bachelor's-degree rate is taken over adults 25 and older, not the whole population, and medians (income, rent, home value) are weighted medians.

Our medians are true weighted medians of the microdata— the value at the 50th percentile of the weighted records. The Census Bureau's publishedmedian tables instead interpolate within coarse value bands (for example a single "$500,000–$749,999" home-value band), which smooths over the fact that respondents round their answers. Because home values especially pile up at round figures (many owners report exactly "$500,000"), our microdata median home value can read a few percent belowthe Bureau's published figure for the same place; income and rent medians track the published numbers closely. For an exact match to a specific published table, consult the source table on data.census.gov.

3. Sampling error & reliability

PUMS is survey-based, so all estimates are subject to sampling error. National and state figures are precise; individual PUMAs and small subgroups are noisier, and a very small matching population should be read as approximate. Estimates are also subject to non-sampling error (measurement, non-response, and processing), which the Census Bureau works to minimize through extensive quality control.

pumsdata does not yet display margins of error.Formal margins require the Census Bureau's replicate weights; adding per-estimate confidence intervals is a planned upgrade. Until then, treat small-area and small-subgroup figures as indicative, and consult the Census Bureau ACS methodology for the official reliability guidance.

4. De-identification & privacy

PUMS is de-identified by the Census Bureau and is designed so that individuals cannot be identified; attempting to re-identify any person in the microdata is prohibited. pumsdata does not collect or hold personal information about the people represented in the Census data. For what we collect about you as a user, see our Privacy Policy.

5. Not affiliated with the Census Bureau

pumsdata is an independent product and is not affiliated with, endorsed by, or sponsored by the U.S. Census Bureau or any government agency. We repackage public-domain Census data for easier exploration. The underlying data is a work of the U.S. government and is in the public domain.

6. Limitations & appropriate use

Estimates are provided for informational and analytical purposes onlyand should not be the sole basis for financial, legal, lending, housing, employment, or other consequential decisions. For critical applications, confirm against the Census Bureau's official published tables.

7. Historical data & year-over-year

pumsdata includes prior ACS years (currently 2023and earlier) for year-over-year comparison. Each year uses the same weighting and methodology described here. Dollar figures within a year reflect the Census Bureau's income/housing adjustment factors; cross-year comparisons are nominal unless otherwise noted.

8. Other data sources & licenses

Beyond ACS PUMS, pumsdata serves several other federal datasets, each verified against itsagency's published totals: the Census Bureau's Building Permits Survey, the BLS American Time Use Survey and Consumer Expenditure Survey microdata, HUD's Fair Market Rents (the Section 8 payment-standard basis, by county and metro), and public K-12 school statistics from the National Center for Education Statistics (Common Core of Data, EDFacts, and the Civil Rights Data Collection, supplemented from ED Data Express).

Historical education data (1986–2024 school-level panels) and the MEPS school-poverty model are accessed through the Education Data Portal, Urban Institute, and used under the Open Data Commons Attribution License (ODC-By) v1.0. The most recent graduation rates (2023–24 onward) and the state report-card outcomes shown on each state page — proficiency, college-going, dropout, and current chronic absenteeism — are collected directly from the state education agencies (public records), which publish ahead of the federal series. Because states use their own assessments, definitions, and reporting windows, those state-reported outcomes are not comparable across states and are presented per state only, never as a cross-state ranking. pumsdata is independent and not affiliated with the Census Bureau, BLS, HUD, NCES, the Urban Institute, or any state agency.

9. Presidential election results (MEDSL)

Presidential election results on pumsdata come from the MIT Election Data & Science Lab (MEDSL): certified, official returns for every state (1976–2024, DOI 10.7910/DVN/42MVDX) and county (2000–2024, DOI 10.7910/DVN/VOQCHQ). We bucket votes into Democratic, Republican, and all other candidates, and report the marginas the Democratic minus the Republican share of all votes cast, in percentage points (D+/R+). Caveats: Alaska reports county-level results by state house district rather than by borough; Connecticut's county-equivalents changed to planning regions in 2022; and Kansas City, MO is reported separately from its surrounding counties. "Popular-vote winner" means the national popular-vote leader — not necessarily the winner of the presidency. pumsdata is an independent product and is not affiliated with, endorsed by, or sponsored by MIT or MEDSL.

10. Consumer Price Index (BLS CPI-U)

Inflation figures come from the BLS Consumer Price Index for All Urban Consumers (CPI-U), loaded from the Bureau's published cu.data flat files with the index values kept verbatim — we never recompute or rebase an index. Percent changes follow the BLS conventions: year-over-year change uses the not-seasonally-adjusted series and month-over-month change uses the seasonally adjusted series; annual averages are the file's own M13 values. The October 2025 data gap from the federal government shutdown is a real hole in the published record — we skip any change calculation that touches it and never interpolate across it. Series published only semiannually (some small-metro series) are excluded, and metro series that publish bimonthly simply show no value in the off months. pumsdata is an independent product and is not affiliated with, endorsed by, or sponsored by the Bureau of Labor Statistics.

11. County housing market (Redfin)

County housing-market figures come from the Redfin Data Center county exports: median sale price and price per square foot, median new-listing price, homes sold, new listings, pending sales, inventory, median days on market, months of supply, and the share of homes sold above list — for roughly 3,100 counties, monthly since January 2017. Values are loaded verbatim, as not-seasonally-adjusted levels; we never recompute, rebase, or seasonally adjust them. All change figures are same-month-prior-year comparisonscomputed from those levels — month-over-month comparisons of NSA housing data mostly measure seasonality, so we never show them. The export's own year-over-year columns are unused: at least one is mislabeled at the source, so every change is computed by us from the level series.

One editorial rule sits on top of verbatim: in a county-month with fewer than 3 homes sold, or where the median sale price falls outside a wide plausibility band, the median/ratio figures are suppressed (shown as no data, never zero) — a median of one or two sales is noise; sale and listing counts are always kept. Medians are also never pooled across geographies: multi-county views average the county medians and label the result an average. Geography caveats: Connecticut is reported by its legacy counties rather than the 2022 planning regions, so CT appears unshaded on 2024-boundary maps while remaining fully present in tables and on its own pages; and the dissolved Valdez-Cordova Census Area, AK (through 2019) appears in tables only. Citation, as the source requires: Data provided by Redfin, a national real estate brokerage. pumsdata is an independent product and is not affiliated with, endorsed by, or sponsored by Redfin.

12. References

For per-variable definitions and code values, see the PUMS data dictionary. For the authoritative source material, see the Census Bureau PUMS documentation and the Census Bureau ACS methodology.