Census ACS for Trade-Area Demographics: Rings, Block Groups, and Where Apportionment Breaks

A trade area is a shape a lender draws. The American Community Survey publishes estimates for shapes the Census Bureau drew. The two sets of shapes almost never coincide, and nearly everything that goes wrong with ACS trade-area demographics follows from that single fact. A three-mile ring around a proposed self-storage site, a fifteen-minute drive-time polygon around a medical office, a fire district, a school attendance zone: none of them is a geography the ACS tabulates. The numbers that appear in the memo are therefore never read off a table. They are apportioned, and the apportionment is done by software that rarely tells the reader what rule it used.

This article sets out what the ACS is for a trade area and what it is not, where the geography ladder stops, how large the margins of error become once the ladder reaches block groups, and exactly where apportionment breaks. It is the demographics chapter of the public-data stack for commercial real estate analysis, and it sits next to the pieces on reading LODES commuting flows for site selection and on geocoding accuracy in CRE analysis, because a trade-area number is only as good as the point it was drawn around and the daytime population it ignores.

The finding worth carrying into the credit file is uncomfortable, and it is arithmetic rather than opinion. The margin of error printed beside a ring estimate moves in the opposite direction to the error that actually dominates that estimate. Combining block groups shrinks the reported margin of error, because sampling errors are combined as a root sum of squares and the square root of a growing sum of squares grows more slowly than the sum of the estimates. So the more block groups a ring sweeps in whole, the tighter its published margin of error looks. Meanwhile the error that matters at small radii, the decision to count a block group in or out, has no margin of error attached to it and never appears in any output. Across the 3,221 county population centers published in the 2020 Census, a three-mile ring captures no block group at all at 545 of them, and at the 2,676 where it captures something the median ring is built from five block groups, one of which accounts for a median 17.8 percent of the answer. In the rings that return fewer than 5,000 people, the median single block group is half the total. The number that reads as plus or minus 18 percent can be wrong by 50 percent for a reason the 18 percent does not measure.

What an ACS number actually is

The American Community Survey is the continuous replacement for the decennial census long form, run under the authority of 13 U.S.C. 141(a) and protected by the confidentiality provisions of 13 U.S.C. 9(a). It selects roughly 3.54 million housing unit addresses a year and about 150,000 group quarters residents, and it publishes estimates covering more than 40 topics for more geographies than any other Census Bureau survey (U.S. Census Bureau, Creating Custom American Community Survey Geographies webinar transcript, April 9, 2026).

Every ACS figure is a period estimate. The 2024 1-year estimates describe January 1 to December 31, 2024. The 2020 to 2024 5-year estimates describe January 1, 2020 to December 31, 2024 as a whole, and the Bureau states plainly that a multiyear estimate describes the average characteristics over the entire period and cannot be used to describe any particular year within it (U.S. Census Bureau, When to Use 1-Year or 5-Year Estimates, 2026). Dollar values in a multiyear file are adjusted to the final year of the period, so the 2020 to 2024 release is in 2024 dollars, inflated using the national consumer price index because no regional index covers the whole country.

Two consequences bind immediately. First, consecutive 5-year releases are not a time series. The Bureau's own comparison guidance treats 2005 to 2009 against 2006 to 2010 as an improper comparison because the periods overlap by four years, and 2005 to 2009 against 2010 to 2014 as a proper one. A trade-area model that reports population growth by differencing the 2019 to 2023 and 2020 to 2024 releases is reporting four-fifths of the same data against itself. Second, the choice between 1-year and 5-year data is not a recency preference. It is a geography constraint. The 1-year file exists only for areas of 65,000 or more and the 1-year supplemental only for areas of 20,000 or more; the 5-year file is published for areas of every size and is the only file that carries census tracts and block groups at all.

The precision of any of it is a function of how many interviews the survey actually completes, and that number has been falling while the sample drawn has not. The Census Bureau has selected close to 3.54 million addresses every year since 2012, with 2020 as the pandemic exception at 2,872,402. Completed housing unit interviews over the same period fell from 2,375,715 in 2012 to 1,944,002 in 2024, a decline of 18.2 percent, which moves the yield from 67.1 completed interviews per 100 addresses selected to 54.9 (U.S. Census Bureau, ACS Sample Size, United States, 2026). Nothing about the block group margins of error later in this article is mysterious once that series is in view.

The group quarters side deserves a separate warning, because it is where trade-area population counts quietly break in college towns, prison counties, military markets and any submarket with a large nursing home. The group quarters sample was designed to support state-level estimates only. Since the 2011 release the Bureau has supplemented it with large-scale whole-person imputation into facilities that were not in sample, and it states the scale of that operation without euphemism: roughly as many group quarters persons are imputed as interviewed. In 2024 the file carried 127,363 actual group quarters interviews and 146,840 synthetic ones. The Bureau's standing position is that it will continue to produce only state-level estimates of characteristics of the group quarters population and does not plan to release characteristics for substate areas (U.S. Census Bureau, user note, Changes to ACS Group Quarters Small Area Estimation, September 2012). Total population for a small area still includes group quarters residents. Their characteristics do not come from that area.

The geography ladder, and where it stops

The 2020 to 2024 ACS 5-year release publishes a total population estimate for 616,690 geographic areas across 116 summary levels, a count taken directly from table B01003 of the published Summary File. That inventory is the reason the survey feels inexhaustible and the reason it disappoints at the scale a site screen works in. The ladder runs from the nation and 52 state-level areas down through 3,222 counties and county equivalents, 36,421 county subdivisions, 32,330 places, 33,772 ZIP Code Tabulation Areas, 85,382 census tracts and, at the bottom, 242,297 block groups.

The block group is the floor. It is a group of blocks inside a census tract, generally defined to hold between 600 and 3,000 people, sitting under tracts that generally hold 1,200 to 8,000 with an optimum of 4,000 (U.S. Census Bureau, geography glossary, 2026). Nothing smaller is published from the ACS. Census blocks exist, they carry decennial counts, and they carry no ACS characteristics at all.

Three restrictions on that floor matter to anyone building a screen. Block groups appear only in the 5-year file, so any block group figure is a five-year average by construction. Only a subset of detailed tables is published for them: the Bureau's own release note states that the 2020 to 2024 estimates are available down to census tract for all data tables and to block group for select detailed tables (U.S. Census Bureau, 2024 Data Release New and Notable, 2026), and the suppression rules exclude tables with more than 100 independent lines along with sensitive tables such as place of birth of the foreign born. And the block group boundaries themselves are 2020 vintage statistical lines drawn for tabulation, which the Bureau's TIGER/Line documentation warns are not a legal land description and carry no warranted positional accuracy.

Median area sizes explain why the ladder is a trap. The median block group in the 2020 to 2024 file holds 1,258 people. The median census tract holds 3,790. The median ZIP Code Tabulation Area holds 2,789, a figure that undercuts the common assumption that a ZIP-based trade area is the coarse option; ZCTAs are approximations of ZIP delivery areas built from blocks, they do not nest inside counties or places, and they change between vintages. A one-mile ring covers 3.14 square miles, and outside dense cities that is routinely smaller than one block group. At the 2020 population center of Douglas County, Colorado, a one-mile ring built from block group population centers captures exactly one of them, so the trade-area estimate is a subset of a single polygon and no amount of care in drawing the circle will produce a number that is about the ring rather than about the polygon.

The practical rule that follows is worth stating before the arithmetic arrives. Geographic detail and statistical precision are traded against each other at a punishing exchange rate, and the direction of the trade is the opposite of what a map suggests. Zooming in does not sharpen the estimate. It replaces a number the Bureau controlled with a number the Bureau sampled.

How much of a block group estimate is noise

Every published ACS margin of error is stated at 90 percent confidence, and the standard error is the margin of error divided by 1.645 (U.S. Census Bureau, ACS Accuracy of the Data, 2024). Applied to the block group file, that convention produces numbers most trade-area work never surfaces.

Across the 240,160 block groups in the 2020 to 2024 release that carry a positive total population estimate and a numeric margin of error, the median margin of error is 30.74 percent of the estimate itself. The median block group reports 1,258 people plus or minus 379. Fully 87.1 percent of block groups carry a margin of error above a fifth of their own estimate and 52.7 percent carry one above 30 percent. At the tract level the same computation gives a median of 13.82 percent, and at the ZIP Code Tabulation Area level 13.77 percent. Relative precision therefore roughly doubles between the block group and the tract, and it is bought with exactly the geographic resolution the site screen was reaching for.

Median household income, the figure most screens lead with, is worse. Table B19013 is published for block groups, but of the 242,297 block groups in the release, 18,638 carry no median household income estimate at all because the sample was insufficient or because the margin of error was larger than the median itself, and another 3,458 carry an estimate whose margin of error cannot be computed because the median falls in an open-ended interval. Of the 220,201 that publish both, the median margin of error is 33.29 percent of the estimate. The median published block group income is $78,920 with a median margin of error of $25,495. One block group in four carries a margin of error above half its own income estimate. A screen that quotes block group median income to the last digit is reporting a number whose 90 percent interval is tens of thousands wide.

The Bureau flags all of this in the file, and the flags are worth learning because they are the difference between a defensible screen and a silent one. A margin of error published as five asterisks, carried in the machine-readable file as the value -555555555, means the estimate is controlled to an independent population or housing estimate: it has no sampling error and the margin of error may be treated as zero. A value of -666666666 means no estimate could be computed. A value of -222222222 or -333333333 means the estimate stands but its margin of error does not (U.S. Census Bureau, Notes on ACS Estimate and Annotation Values, 2026).

Reading those flags across the whole file settles the precision question. Of the 3,222 counties in the 2020 to 2024 release, 3,090 carry a controlled total population, which is to say a figure with no sampling error whatever. All 52 state-level areas are controlled. The national total is controlled. Not one of the 242,297 block groups is. The county number a screen abandons in order to zoom in is the only number in the chain that is exact by construction, and the block groups it zooms into are the sampled ones. Aggregating block groups back up does not recover the county's certainty, and it does not reproduce the county's value either.

Rings do not respect block group lines

The Census Bureau is candid about the gap between its geographies and a lender's. In its April 2026 webinar on creating custom ACS geographies, the Bureau's own statistician told data users that there are no data tools produced by the Census Bureau that output data within a radius of a point, and pointed the audience to an outside university tool for the job. The same demonstration explained the rule that tool applies: it grabs an entire block group that is within the radius, it does not split block groups or grab partial areas, and if any point of the block group falls within the radius it is included (U.S. Census Bureau, Creating Custom American Community Survey Geographies, April 9, 2026).

That is one of three rules in common use, and they give different answers for the same ring. Touch inclusion counts a block group whole if any part of it intersects the ring, which overstates. Centroid assignment counts it whole if a single representative point falls inside, which can overstate or understate and flips on the choice of point, since a geometric internal point and a population-weighted center are not the same place. Areal weighting splits a block group by the share of its land area inside the ring, which is defensible only under an assumption nobody states out loud: that people are spread uniformly across the polygon. In a block group containing a subdivision, a golf course and a gravel pit, they are not.

The size of the resulting error can be measured rather than asserted. Take every county population center published in the 2020 Census, 3,221 points that the Bureau computed as the population-weighted mean center of each county, and draw a three-mile ring at each one. Assign block groups by their own published 2020 population centers, the most defensible version of centroid assignment available, and sum the 2020 to 2024 ACS population estimates for the block groups that fall inside.

At 545 of the 3,221 sites, 16.9 percent, not one block group population center falls within three miles. Centroid assignment returns zero people at the exact point where a county's population is centered. Widen the failure to a one-mile ring and it becomes the normal case: at 1,725 sites, 53.6 percent of the country's county population centers, no block group center falls within a mile. Across the 2,676 sites where a three-mile ring does capture something, the median ring is built from five block groups and returns 6,721 people, 64.7 percent are built from fewer than ten block groups, and at 19.0 percent the ring is a single block group and the trade-area estimate is that one polygon's published number.

Fragility follows directly. Among those 2,676 sites, a single block group accounts for a median 17.8 percent of the three-mile ring and for more than a quarter of it at 42.1 percent of them. Moving the radius from 3.0 to 3.1 miles, a change no credit committee would consider material, moves the population answer by more than 5 percent at 24.8 percent of those sites and by more than 10 percent at 14.5 percent. None of that movement is sampling error. It is the boundary rule, and it is invisible in every margin of error the file publishes.

The pattern splits cleanly by density, which is the part most guidance misses. Where the three-mile ring returns more than 25,000 people it is typically built from 39 block groups, its combined margin of error is 5.2 percent, and one block group is 2.3 percent of the total. Where it returns fewer than 5,000 it is typically built from two block groups, its combined margin of error is 17.7 percent, and one block group is half the answer. Ring apportionment is a solved problem in dense urban submarkets and an unsolved one everywhere else, which is exactly backwards from where a ring is most needed. In a dense submarket, published places and tracts already approximate the trade area closely enough to use directly. In a suburban or rural one they do not, and that is where the ring is least trustworthy.

Three worked points make the mechanism concrete. At the 2020 population center of San Francisco County, California, a one-mile ring captures 73 block groups and 81,022 people with a combined margin of error of 3.61 percent; the three-mile ring captures 513 block groups and 617,290 people at 1.51 percent. At the population center of Douglas County, Colorado, a one-mile ring captures exactly one block group, 1,657 people at 33.61 percent, and the three-mile ring reaches eight block groups and 14,889 people at 9.60 percent. At the population center of Sampson County, North Carolina, the one-mile ring is empty, and the two-mile and three-mile rings return the identical figure, 1,609 people plus or minus 707, because the same lone block group is the only one whose center falls inside either. A model that reports a two-mile and a three-mile population for that site is reporting one number twice.

The arithmetic of adding block groups, and the error it hides

When estimates are summed, the Census Bureau's published approximation for the margin of error of the total is to square each component margin of error, add the squares and take the square root. The handbook that carries the formula also carries its warnings: the method ignores covariance between the components, it diverges further from the microdata standard error as more estimates are summed, controlled estimates make the divergence worse because their standard error enters as zero, and users should work with the fewest estimates possible (U.S. Census Bureau, Understanding and Using American Community Survey Data, chapter 8, 2020). For serious aggregation the Bureau publishes Variance Replicate Estimate Tables, 80 replicate weights on selected 5-year detailed tables, which produce a correct margin of error for a combined geography rather than an approximation.

The behavior of that root sum of squares is the trap at the center of this article. Adding block groups outward from the San Francisco County population center, the relative margin of error falls from 23.97 percent on the nearest single block group to 10.22 percent at eight, 3.77 percent at 64 and 1.51 percent at 512. From the Sampson County center it falls from 43.94 percent to 11.25 percent at eight and 1.31 percent at 512, by which point the aggregation has swept up 685,513 people across a 36-mile radius. The absolute margin of error rises the whole way, from 707 people to 8,970. Only the ratio improves, and it improves because the denominator grows faster than the square root of the summed squares.

Read that curve next to the boundary results and the perverse incentive is obvious. A wide ring, a generous drive-time polygon, or a touch-inclusion rule that sweeps in every block group the boundary grazes will all report a smaller relative margin of error than a tight, carefully fitted trade area. The number that looks most precise is the one built with the least care about fit. Statistical error and specification error move in opposite directions, and only one of them is printed.

There is a second trap in the same arithmetic, and it applies to money rather than people. Medians cannot be added and cannot be averaged. Summing block group median household incomes, or taking their mean, produces a quantity with no interpretation. The correct construction uses aggregate household income, table B19025, which is published at block group level, divided by households from table B11001, also published at block group level. That is a ratio of two sums, and the Bureau's derived-estimate guidance carries its own formula for the margin of error of a ratio. The result is a mean household income for the trade area, not a median, and the difference between the two is not cosmetic in a market with a long income tail. The same discipline is what separates a defensible spending model from a decorative one, and it is why BEA regional income data belongs beside ACS tables rather than behind them: the ACS measures who lives in the ring, the BEA series measures the income actually earned and received in the county, adjusted for commuting.

Rings, drive times, and the trade area you actually mean

A ring is a statement about distance. A drive-time polygon is a statement about the road network and the time of day it was computed for. Neither is a statement about where customers come from, and confusing the three is the most common failure in a trade-area section that otherwise cites everything correctly.

Rings have one virtue worth defending: they are reproducible. Another analyst with the same coordinates and the same radius gets the same shape. A drive-time polygon depends on a network dataset, a speed model, a departure time and a routing engine, none of which is usually disclosed, and two routing engines will not return the same fifteen-minute shape for the same address. If a drive-time polygon is used, the file should record the engine, the network vintage, the assumed departure time and whether traffic was modeled. If it cannot record those things, a ring with a stated radius is the more honest instrument, and the reasoning is the same one that governs the provenance standard for every other number in the file.

Both shapes inherit the apportionment problem in full, and the drive-time polygon inherits it worse. A ring is convex and compact, so the block groups it clips are clipped near their edges. A drive-time polygon is a starfish following arterials, with long thin lobes that slice through the middle of block groups along the road rather than around it, which is exactly where the uniform-distribution assumption behind areal weighting fails hardest, since population clusters along those same arterials. Where the polygon follows the road, area weighting will systematically understate the population inside it.

The choice of center point matters as much as the choice of shape. A ring drawn around a rooftop and a ring drawn around a street-segment interpolation of the same address are not the same ring, and the difference is large enough at small radii to flip block groups in and out. That is the subject of geocoding accuracy in CRE analysis, and the short version is that the free Census Geocoder returns a coordinate interpolated along an address range in the MAF/TIGER system rather than a rooftop point, with the ranges covering all possible structure numbers whether or not the structures exist (U.S. Census Bureau, Census Geocoder User Guide, May 2026). At a five-mile radius the difference is noise. At a one-mile radius around a car wash it is the analysis.

Residents are not customers

The ACS counts people where they live. It is a residence-based survey, and the population it reports for a block group at noon on a Tuesday is a fiction convenient for the survey rather than a description of who is there. For most retail, service and hospitality assets, the residential count is half the demand picture at best.

The other half is workplace population, and the public source for it is the LEHD Origin-Destination Employment Statistics program, which tabulates jobs by workplace block, by residence block and by the link between them. Version 8.4 is enumerated on 2020 census blocks, covers data years 2002 through 2023, and is explicitly a partially synthetic product: workplace totals carry noise infusion and residential locations are synthesized before tabulation, so a block-level figure is a protected estimate rather than an administrative count (U.S. Census Bureau LEHD, LODES Dataset Structure Format Version 8.4, revision December 3, 2025). Used carefully, it answers the question the ACS cannot: how many jobs sit inside the ring, and how far their holders travel. That handoff is worked through in reading LODES commuting flows for site selection.

Counting competitors is a third question again, and it belongs to the business register rather than the household survey. Establishment counts by industry inside a trade area come from County Business Patterns and ZIP Code Business Patterns, with their own noise infusion, their own publication thresholds and their own two-year lag, which is the subject of counting competitors and finding market gaps with County Business Patterns. Supply under construction comes from the Building Permits Survey, covered in the Building Permits Survey as a supply pipeline signal. A trade-area section that answers only the residential question has answered a third of the problem and will read that way to a credit committee, as what credit committees expect from market analytics sets out at length.

The tables a trade-area screen actually needs

A working screen needs five things about the people in the shape: how many there are, how many households they form, what those households earn, whether they own or rent, and how the mix is changing. Four of the five are available at block group in the 2020 to 2024 file, and the fifth is not.

Population comes from B01003 and age structure from B01001, both published at block group. Households come from B11001, housing units from B25001, tenure from B25003 and average household size by tenure from B25010, all at block group. Income is available in three forms at block group: the median in B19013, the bracketed distribution in B19001 and the aggregate in B19025, with per capita income in B19301. Housing value and rent come from B25077 and B25064, structure type from B25024, education from B15003, commute time from B08303, employment status from B23025, race and Hispanic origin from B03002, and the income-to-poverty ratio from C17002. Each of those was confirmed present at block group level in the published 2020 to 2024 Summary File.

Two absences are worth planning around. B07003, geographic mobility by residence one year ago, is not published at block group, so the mover rate that drives self-storage and multifamily demand models has to be taken at tract level or above; the per-capita saturation logic in self-storage demand analysis is built on that constraint rather than around it. B17001, poverty status by age and sex, is likewise not published at block group, although the coarser C17002 ratio table is, which is usually enough for a screen.

The precision of the individual cells is the reason to prefer the coarser table wherever the question allows it. In one ordinary Alabama block group taken from the published file, the median home value is $244,600 with a margin of error of $158,784, and per capita income is $34,754 plus or minus $11,167. Those are not unusual values; they are what a block group cell looks like. Reporting either to a credit committee without its margin is not a rounding decision, it is a misstatement of what the file says.

What counts as current in August 2026

A trade-area model is dated by its slowest input, and for demographics that input is the 5-year file. The 2024 cycle ran as follows. The 2024 1-year estimates were released on September 11, 2025, 254 days after the reference year closed, for areas of 65,000 or more. The 2024 1-year supplemental estimates and the 1-year microdata followed on December 4, 2025, 338 days out, for areas of 20,000 or more. The 2020 to 2024 5-year estimates and the matching Variance Replicate Estimate Tables, the only products that carry tracts and block groups, arrived on January 29, 2026, 394 days after the period closed. The 5-year microdata came on March 5, 2026, at 429 days.

Two facts about that calendar belong in the memo rather than in a footnote. The 5-year release slipped: on April 2, 2025 the Bureau scheduled it for December 11, 2025, and on November 20, 2025 it moved to January 29, 2026, a 49-day slip announced three weeks before the original date. And the oldest observations in that file were already 6.1 years old on the day it was published, because the period opens on January 1, 2020. Read on August 23, 2026, they are 6.6 years old.

The forward calendar is worse than a lag. As of the Bureau's update of August 6, 2026, the release date for the 2025 ACS 1-year estimates is being determined while the Bureau assesses the impact of the new Department of Commerce administrative order on Disclosure Avoidance for Statistical Products and seeks a compliant path to releasing the 1-year tables later in the year. That is 235 days past the close of the 2025 reference year with no date set. No screen written in August 2026 may assume a September 2026 ACS 1-year release, and any model whose refresh logic assumes an annual September drop needs a manual check rather than a scheduler.

The practical consequence for a fast-growing suburb is the opposite of the usual advice. Analysts are told to prefer the 5-year file because it is more reliable, and at block group level there is no choice in the matter. But in a submarket that added a thousand rooftops in 2023 and 2024, the 2020 to 2024 file weights those years alongside 2020 and 2021, when the subdivision was a field. The five-year average is not a lagged estimate of today; it is a correct estimate of a period that includes a market that no longer exists. The honest treatment is to name the period in the sentence that carries the number, cross-check the direction of travel against building permits, and never present the 5-year figure as a current-year count.

A method that survives a credit file

The following sequence produces a trade-area demographic exhibit that holds up under examination, and it is the demographic leg of the 30-minute pre-term-sheet site screen.

Fix the point first and record how it was fixed. Rooftop, parcel centroid or address-range interpolation, with the geocoder and vintage named. Then choose the shape and state it: a radius in miles, or a drive time with its engine, network vintage and departure assumption. Then choose the geography the shape will be built from, and choose the largest one the question tolerates. If the ring is wide enough to contain 25 or more block groups, block groups are reasonable. If it contains fewer than ten, drop to tracts and say why, or report the block group figure with the count of block groups printed beside it so the reader can see how few decisions the number rests on.

Apportion explicitly. Name the rule, whether centroid assignment, touch inclusion or areal weighting, and state the assumption it carries. Where the ring cuts a block group that holds more than a tenth of the total, run the estimate both ways, with the block group in and with it out, and report the range. That two-line sensitivity is worth more to a credit committee than any single point estimate, and it costs one extra query.

Aggregate correctly. Sum estimates, combine margins of error as a root sum of squares, and use the Variance Replicate Estimate Tables where the combined figure carries weight in the decision. Build income from aggregate income over households, never from medians. Carry the margin of error into the exhibit rather than dropping it at the last formatting step, and quote it in the same units as the estimate.

Then cross-check against something the ACS does not know. Jobs inside the ring from LODES, competitor establishments from County Business Patterns, units under construction from the Building Permits Survey, county income and its residence adjustment from the BEA. Where a vendor product is used instead of the source files, know what it did with the apportionment, because demographic vendors against Census-direct data differ less in their inputs than in their undocumented allocation rules and their proprietary current-year updates, which are models rather than measurements.

MMCG Analytics is a map-first commercial real estate analytics platform built on federal, state and public data with source and vintage provenance carried on displayed values, and its demographic layer is assembled from these public records. It supplies data and analytics. The credit decision rests with the lender, and so does the choice of trade area.

What to write about uncertainty

A trade-area exhibit that hides its uncertainty invites the reader to discount all of it. One that states uncertainty in the same breath as the estimate earns the estimate. Four sentences do the work.

State the shape and the rule: a three-mile radius around the geocoded site, built from 2020 census block groups assigned by their population centers. State the vintage and the period: 2020 to 2024 ACS 5-year estimates, released January 29, 2026, describing the five-year period as a whole and not any single year. State the estimate with its combined margin of error at 90 percent confidence, computed as a root sum of squares over the component block groups per Census Bureau guidance. State the boundary sensitivity: the number of block groups in the ring, and the effect of the largest single one being included or excluded.

Then say what the number does not measure. It does not measure daytime population. It does not measure spending, only the income of residents. Its group quarters component carries characteristics estimated at state level. And in a small ring, its stated margin of error understates total error because the apportionment decision carries no margin of error at all. A memo that says those things is not weaker than one that does not. It is the only version a reviewer can check, which is the standard the rest of the quantitative core of feasibility analysis is held to as well.

Frequently asked questions

Can I get ACS data for a three-mile radius?

Not directly. The Census Bureau stated in its April 2026 webinar on custom geographies that no Census Bureau tool outputs data within a radius of a point. Radius figures are always built by assigning published geographies, usually block groups, to the ring under a stated rule: touch inclusion, centroid assignment or areal weighting. Each rule gives a different answer, and the difference is largest where the ring holds few block groups.

What is the smallest geography ACS data is available for?

The block group, and only in the 5-year estimates. The 2020 to 2024 release publishes 242,297 block groups, generally holding 600 to 3,000 people each, and only for a selected subset of detailed tables. Census blocks carry decennial counts but no ACS characteristics, and the 1-year and 1-year supplemental files stop at areas of 65,000 and 20,000 people respectively.

Why are block group margins of error so large?

Because a block group of about 1,258 people receives a very small share of a sample that has been shrinking. The ACS has selected close to 3.54 million addresses a year since 2012, but completed housing unit interviews fell from 2,375,715 in 2012 to 1,944,002 in 2024. In the 2020 to 2024 release the median block group total population margin of error is 30.74 percent of the estimate, and the median block group median household income margin of error is 33.29 percent of the estimate.

How do I combine margins of error when I add block groups together?

Square each component margin of error, add the squares and take the square root. That is the approximation the Census Bureau publishes for aggregated counts, and it ignores covariance, degrades as the number of summed estimates rises, and is distorted by controlled estimates whose standard error is zero. For a figure that carries weight in a decision, use the Variance Replicate Estimate Tables, which supply 80 replicate weights on selected 5-year detailed tables and produce a proper margin of error for a combined geography.

Should a trade-area screen use 1-year or 5-year ACS data?

At tract or block group level there is no choice: only the 5-year file publishes them. Above that threshold the trade is recency against reliability, and for a fast-growing suburb the usual advice inverts. A 2020 to 2024 estimate is a correct description of a five-year period that includes years before the subdivision was built, not a lagged snapshot of today. Name the period next to the number and cross-check direction with building permits.

Can two consecutive ACS 5-year releases be compared to measure growth?

No. The 2019 to 2023 and 2020 to 2024 releases share four years of data. The Census Bureau's comparison guidance treats overlapping multiyear periods as an improper comparison and non-overlapping ones, such as 2015 to 2019 against 2020 to 2024, as proper. Differencing consecutive releases measures mostly the same households against themselves.

Does the ACS measure daytime population?

No. The ACS counts people at their place of residence. Workplace population comes from the LEHD Origin-Destination Employment Statistics program, which tabulates jobs by workplace and residence block and is a partially synthetic product with noise infusion on workplace totals. For retail, service and hospitality assets, the residential count answers only part of the demand question.

Sources

1. U.S. Census Bureau, When to Use 1-Year or 5-Year Estimates, American Community Survey guidance, 2026. https://www.census.gov/programs-surveys/acs/guidance/estimates.html
2. U.S. Census Bureau, Comparing ACS Data, American Community Survey guidance on overlapping and non-overlapping multiyear periods, 2026. https://www.census.gov/programs-surveys/acs/guidance/comparing-acs-data.html
3. U.S. Census Bureau, American Community Survey Accuracy of the Data (2024), 90 percent confidence convention, standard error and controlled estimates, 2026. https://www2.census.gov/programs-surveys/acs/tech_docs/accuracy/ACS_Accuracy_of_Data_2024.pdf
4. U.S. Census Bureau, Multiyear Accuracy of the Data, 5-year estimates for 2020 to 2024, 2026. https://www2.census.gov/programs-surveys/acs/tech_docs/accuracy/MultiyearACSAccuracyofData2024.pdf
5. U.S. Census Bureau, Understanding and Using American Community Survey Data: What All Data Users Need to Know, chapter 8, Calculating Measures of Error for Derived Estimates, 2020. https://www.census.gov/content/dam/Census/library/publications/2020/acs/acs_general_handbook_2020.pdf
6. U.S. Census Bureau, American Community Survey Data Suppression, data quality filtering and the block group table restrictions, September 27, 2016. https://www2.census.gov/programs-surveys/acs/tech_docs/data_suppression/ACSO_Data_Suppression.pdf
7. U.S. Census Bureau, Notes on ACS Estimate and Annotation Values, jam values including the controlled-estimate flag, page last revised July 20, 2026. https://www.census.gov/data/developers/data-sets/acs-1year/notes-on-acs-estimate-and-annotation-values.html
8. U.S. Census Bureau, 2024 Data Release New and Notable, release dates of September 11, 2025, December 4, 2025, January 29, 2026 and March 5, 2026, the December 11, 2025 schedule and the block group table statement, 2026. https://www.census.gov/programs-surveys/acs/news/data-releases/2024/release.html
9. U.S. Census Bureau, 2024 ACS Release Schedule, 2026. https://www.census.gov/programs-surveys/acs/news/data-releases/2024/release-schedule.html
10. U.S. Census Bureau, American Community Survey News and Updates, entry of August 6, 2026 on the undetermined release date for the 2025 1-year estimates. https://www.census.gov/programs-surveys/acs/news/updates/2026.html
11. U.S. Census Bureau, Creating Custom American Community Survey Geographies, webinar transcript, April 9, 2026. https://www2.census.gov/about/training-workshops/2026/2026-04-09-creating-custom-acs-geographies-transcript.pdf
12. U.S. Census Bureau, 2020 to 2024 American Community Survey 5-Year Summary File, table-based format, released January 29, 2026. https://www.census.gov/programs-surveys/acs/data/summary-file.html
13. U.S. Census Bureau, Variance Replicate Estimate Tables, 80 replicate weights on selected 5-year detailed tables, 2020 to 2024 release, 2026. https://www.census.gov/programs-surveys/acs/data/variance-tables.html
14. U.S. Census Bureau, Centers of Population, 2020 Census, block group and county files with population and the latitude and longitude of the mean center, 2021. https://www.census.gov/geographies/reference-files/time-series/geo/centers-population.html
15. U.S. Census Bureau, Geography Program glossary, definitions of census block, block group and census tract, 2026. https://www.census.gov/programs-surveys/geography/about/glossary.html
16. U.S. Census Bureau, Areas Published, American Community Survey geography, counties published by product, 2026. https://www.census.gov/programs-surveys/acs/geography-acs/areas-published.html
17. U.S. Census Bureau, 2020 Census tract and block group relationship files, intersections of geography without population counts, 2021. https://www.census.gov/geographies/reference-files/time-series/geo/relationship-files.2020.html
18. U.S. Census Bureau, American Community Survey Sample Size, United States, initial addresses selected, final interviews and group quarters interviews for survey years 2012 to 2024, table read August 23, 2026. https://www.census.gov/programs-surveys/acs/methodology/sample-size-and-data-quality/sample-size.html
19. U.S. Census Bureau, user note, Changes to ACS Group Quarters Small Area Estimation, September 2012. https://www.census.gov/programs-surveys/acs/technical-documentation/user-notes/2011-01.html
20. U.S. Census Bureau, 2024 TIGER/Line Shapefiles Technical Documentation, positional accuracy disclaimer and the statement that boundaries are not a legal land description, 2024. https://www2.census.gov/geo/pdfs/maps-data/data/tiger/tgrshp2024/TGRSHP2024_TechDoc.pdf
21. U.S. Census Bureau, Census Geocoder User Guide, address range interpolation and match indicators, May 2026. https://www2.census.gov/geo/pdfs/maps-data/data/Census_Geocoder_User_Guide.pdf
22. U.S. Census Bureau, Center for Economic Studies, LODES Dataset Structure Format Version 8.4, revision of December 3, 2025. https://lehd.ces.census.gov/data/lodes/LODES8/LODESTechDoc8.4.pdf
23. U.S. Bureau of Economic Analysis, County Personal Income: Concepts and Methods, residence adjustment for intercounty commuting, April 2026. https://www.bea.gov/system/files/methodologies/BEA-County-Personal-Income-Concepts-and-Methods.pdf
24. 13 U.S.C. 141(a), Population and other census information, Cornell Legal Information Institute, 2026. https://www.law.cornell.edu/uscode/text/13/141
25. 13 U.S.C. 9(a), Information as confidential, Cornell Legal Information Institute, 2026. https://www.law.cornell.edu/uscode/text/13/9
26. MMCG Research, ACS Trade-Area Precision Series: geography counts, margin of error distributions, the three-mile ring at 3,221 county population centers and the root sum of squares curves, computed from the 2020 to 2024 ACS 5-Year Summary File and the 2020 Census Centers of Population files; MMCG database, 2026. https://mmcganalytics.com/methodology/