Why three Census figures for the same county disagree
The decennial count, the annual estimates and the American Community Survey are three different products measuring three different things.
Ask how many people live in a county and the Census Bureau has three answers for you. All three are official, all three are current, and for nearly every county in the country they disagree. None of them is wrong. They answer different questions, are built by different methods on different schedules, and the gaps between them are a property of how the numbers are made rather than a sign that something has gone bad.
The three are the Decennial Census, the Population Estimates Program, and the American Community Survey. Knowing which one a figure came from matters more than the figure itself, because a number quoted without its source cannot be checked or updated.
The Decennial Census: the constitutional count
Every ten years the Bureau attempts to count every person living in the United States. This is the only one of the three products required by the Constitution, and it exists to apportion seats in the House of Representatives among the states. It is also the basis for redistricting and for a long list of formula-based funding programs.
Two things follow from that. First, the decennial count is a complete enumeration rather than a sample, so it carries no margin of error in the statistical sense. Second, once published it is fixed. The 2020 count for a county is the 2020 count forever, and it does not get quietly updated as better information arrives.
Its weakness is age. By the eighth or ninth year of a decade the decennial figure describes a country that has moved on, sometimes sharply, in fast-growing suburban counties and shrinking rural ones alike. There is also a geography trap: decennial products are published on the boundaries in force at the time of the count. Connecticut’s 2020 decennial data sits on the eight legacy counties (Fairfield, Hartford, Litchfield, Middlesex, New Haven, New London, Tolland and Windham), not the nine planning regions that replaced them with the 2022 vintage. Because the regions are groups of towns and are not coterminous with the old counties, no honest county-to-region crosswalk exists; only a town-level one works. A full account of that change, and of the other oddities in the national county list, is in how many counties there are in the US.
The Population Estimates Program: the annual roll-forward
Each year the Bureau publishes a new vintage of population estimates. Vintage 2025 was released on 26 March 2026. The method is a demographic accounting exercise rather than a count: start from the decennial base and carry it forward one year at a time, adding births, subtracting deaths and applying net domestic and international migration.
That gives a current figure for every county, every year, with no sampling error attached, which is why the Estimates Program is the right tool for almost any straightforward “how big is this county now” question.
The catch is that each vintage is a fresh reconstruction of the entire series, not an extra row bolted onto last year’s file. The 2023 estimate in Vintage 2025 is not the same number as the 2023 estimate published in Vintage 2023, because two more years of data have revised the whole chain. Comparing a county’s population across vintages produces change that is partly real and partly methodological. Always take every year in a time series from the same vintage file.
The American Community Survey: everything except the headcount
The ACS is a continuous sample survey. It is where income, age, housing, commuting, language and educational attainment come from, none of which the decennial census asks about in any comparable depth.
Because it is a sample, its reliability depends on how much sample a place accumulates, and that produces two products. The 1-year estimates cover areas with populations of 65,000 or more. The 5-year estimates pool five years of interviews and are published for every geography, down to block groups.
The threshold bites hard at county level. Only 845 of the 3,144 county-equivalents in the 50 states and the District of Columbia reach 65,000 people. The other 2,299 have no 1-year estimate at all, ever. For roughly three counties in four, the 5-year file is the only ACS data that exists, and that includes the places people are often most curious about precisely because they are small: Kalawao County, Hawaii, with 82 residents, or Loving County, Texas, with 52.
Margins of error, and why small counties suffer
Every ACS figure ships with a margin of error, published at the 90 percent confidence level. This is not a disclaimer to be stripped out before publication; it is half the number.
The margin widens as the sample shrinks, and sample size tracks population. A median household income estimate for Los Angeles County, California, with 9,694,934 residents, rests on a very large pool of responses, and the interval around it is narrow. The same estimate for McPherson County, Nebraska, with 369 residents, may rest on a few dozen households, and the interval can be wide enough that the county is statistically indistinguishable from much of its state. Subgroups make it worse: median income for a single age band in a small county can carry a margin of error large enough to make the estimate close to useless on its own.
Two practical rules follow. Do not rank counties on differences smaller than their combined margins of error, a habit that quietly destroys most “poorest county in the state” lists. And remember that a 5-year estimate is not a current-year figure: it describes a five-year window with its centre in the past, so in a rapidly changing county it lags reality by design.
The ESTIMATESBASE2020 trap
This one catches careful people. The Vintage 2025 estimates file, co-est2025-alldata, contains a column called ESTIMATESBASE2020. It looks exactly like the 2020 Census count, and it is not the same number. The estimates base incorporates Count Question Resolution corrections and boundary updates made since census day.
Subtract it from the published 2020 Census figure and the difference you get is not migration and not error. It is method. Use the decennial file when you want the decennial count and the estimates file when you want the estimates series, and never treat the base column as a substitute for the census.
Which product answers which question
| Question | Product | Geography available | Margin of error |
|---|---|---|---|
| Apportionment, redistricting, legal population | Decennial Census | All counties, on the boundaries in force at the count | None (complete enumeration) |
| Current population, year-to-year change, growth ranking | Population Estimates Program | All counties, annually | None published |
| Median household income, median age, housing, commuting | ACS 5-year | All 3,144 county-equivalents | Published, and widens sharply in small counties |
| Same topics, most recent single year | ACS 1-year | 845 counties of 65,000 or more only | Published, and wider than the 5-year for the same area |
What this site uses
Population figures on countymapsus.com come from the Population Estimates Program, Vintage 2025. Land area comes from the 2025 Gazetteer, which this site treats as the authoritative area figure; it will not always agree with an area computed from the mapped polygons, which are drawn from the cartographic boundary file cb_2025_us_county_500k and simplified for display. Income and median age come from the ACS 2020-2024 5-year estimates, chosen over the 1-year precisely because they cover all 3,144 entities rather than 845 of them.
Every figure on a county page is labelled with the product it came from, because an unlabelled number invites exactly the mismatch this article is about. The full build process, including how the mapped geometry is derived, is set out in the methodology, and each file with its release date is listed under sources.
One last joining hazard. Products are matched on the five-digit FIPS code, and those codes are not permanent. They are assigned alphabetically within a state, so a rename usually forces a new code: Shannon County, South Dakota, 46113, became Oglala Lakota County, 46102, in May 2015. Joining an older ACS extract to a current estimates file on FIPS alone will silently drop counties like that. The stable alternative is COUNTYNS, the eight-digit GNIS or ANSI identifier, which survives a rename. Both identifiers are explained in understanding FIPS codes.