Methodology

A typical-value index for the middle third of the market, not a sale-price average.

One reading per year, always the same month, so year-over-year isn't distorted by seasonality.

Around 3× was the historical norm. Derived rather than stored, so it can't drift from its inputs.

County figures are an effective rate implied by what owners actually pay.

Published for counties with enough rental listings to compute a reliable index — about 1,350 of 3,144. Sparse counties show the state figure. Zillow publishes no state-level ZORI, so state numbers are population-weighted from counties.

The 40th percentile of standard-quality units, used to set housing voucher payments. Published for every county, which is why it appears where the Zillow index does not — but it is a different measure, not a substitute. HUD sets one figure per Fair Market Rent area, so counties inside a metro share a number, and it reflects what a voucher pays rather than what a listing asks.

Dollar levels vary by thousands across sources depending on coverage assumptions. Rank order is more reliable than the absolute figure.

State figures use the 1-year survey; county figures use the 5-year, which covers small counties the 1-year omits.

Published at state level only. County profiles show the state figure and label it as such.

Income restated in national-average dollars. County figures pair county income with the state index.

The metric shows the top marginal rate. The cost calculator computes an effective rate bracket by bracket instead.

Whether pensions, 401(k) and IRA withdrawals are taxed. Three categories only: no income tax at all, retirement income exempt, or taxed as ordinary income. Most states in the last group have age or income exclusions this single label cannot express.

Nine states tax benefits at least partly; almost all exempt lower-income retirees, so “taxed” rarely means taxed in full. The list shrinks most years — West Virginia phased its tax out in January 2026. Verify with your state before acting on it.

Local rates vary within a state; this is an average, not your address.

Scraped from AAA's published page — the only source carrying all 51 regions together.

The full premium, most of which an employer typically pays. The cost calculator lets you exclude it.

State level only.

One-way, and excludes people who work from home.

Straight-line miles from the county's population-weighted centre to the nearest airport the FAA classifies as commercial service, meaning at least 2,500 scheduled passenger boardings a year. Driving distance is always longer and can be far longer where mountains or water intervene — we do not publish a drive time because no routing source we can license commercially would support one, and an invented estimate would be worse than an honest straight line. Whether an airport counts is the FAA's own determination, not ours: a field with a scheduled flight but under 2,500 boardings is categorised general aviation by the agency and appears in that measure instead. The label carries the FAA hub size and the boarding count sits alongside it, so a large hub and a single Essential Air Service route are never presented as equivalent.

Straight-line miles to the nearest public-use airport with no scheduled passenger service. Useful if you fly yourself or charter; otherwise it mostly indicates how rural a county is. Private strips, heliports, seaplane bases, gliderports and balloonports are excluded.

How many airports of each kind sit inside the county or state. Most counties have no airline airport, which is normal and is why the distance measures matter more. Airfields on islands and coastlines can fall just outside simplified boundaries and are snapped to the nearest county in their own state, always within a mile or two. Airports in Puerto Rico, Guam, the Virgin Islands, American Samoa and the Northern Marianas appear in the FAA data but not in any of the 3,144 counties, so they are excluded: fifteen of them have scheduled airline service.

How many Amtrak rail stations sit inside the county or state. Amtrak Thruway bus terminals are excluded, and that distinction does real work: the dataset holds 1,031 stops of which only 564 are rail, and counting a connecting coach from a town with no track would overstate rail access in exactly the rural places the measure is for. A count says nothing about frequency — on some long-distance routes a station sees one train a day in each direction. We publish no distance-to-station measure: Alaska and Hawaii have no Amtrak at all, so their nearest station is on the mainland and a national scale running to 3,845 miles compresses every real difference in the lower 48 into nothing.

Every railway in the country, shaded by use. The four categories are FRA's own published subsets rather than our reading of a code: the network carries NET and PASSNGR fields whose meanings live in a data dictionary we do not hold, and a colour key is a stronger claim than a number because a reader takes ‘freight main line’ as fact. Passenger and Class I are ownership and operator distinctions, not traffic ones — a shortline main line carries real trains and appears under other track. Geometry is generalised to roughly 200 metres, so lines are indicative at county zoom rather than survey-accurate. Yard track also exists in the main network layer and the yards layer carries no shared identifier, so yard segments are drawn twice; the visible result is correct because yards draw on top, but the underlying line is duplicated.

We do not publish a count of commuter rail, subway or light rail stations, and the reason is worth stating rather than leaving as a silent absence. The National Transit Map aggregates agencies' own GTFS feeds, and its coverage of rail operators is partial: it carries 78 rail feeds, and testing it produced 16 stations for New Jersey against roughly 165 NJ Transit rail stations, nothing for Delaware despite SEPTA serving Wilmington, and no sign of Metra in Illinois. It also counts a streetcar stop on a street corner the same as a terminal, which put Philadelphia County above the whole of New York State. A number built on that reports few stations where it means the agency does not publish, and a reader has no way to tell those apart. The railway overlay does show commuter and light rail track, because FRA's network data covers it properly.

Connecticut and Alaska's former Valdez-Cordova area were both blank or short until the boundary handling was corrected; see Geography and county boundaries below. Lewistown, Montana is absent from the FAA boarding file entirely, so it is treated as general aviation. Finally, a renamed airport can lag a year between FAA systems: Palm Beach International became President Donald J. Trump International on 9 July 2026 and its FAA identifier changed from PBI to DJT, while the annual boarding file published in September 2025 still keys it as PBI. We record such changes with their effective date and source rather than letting the airport disappear. On the rail side, Hawaii has no freight network and so no yards, and a small number of waterfront and border stations sit outside the simplified county boundaries and are snapped to the nearest county in their own state.

State level only — no free authoritative county rating exists. School districts don't align to county lines, and some cross them.

State level. Reporting is voluntary and some states have incomplete years — Florida's 2024 data is known to be partial.

Different bases: the state figure is a current monthly rate, the county figure a five-year average. Not directly comparable.

Share below the federal poverty level, which doesn't adjust for local cost of living.

Measures adoption, not advertised availability. Low figures can reflect cost as much as infrastructure.

A proxy for healthcare access, not a direct measure of provider supply.

Measures adoption, not advertised availability. FCC availability data is widely criticised for overstating coverage, so we use what households actually subscribe to. Low figures can reflect cost as much as infrastructure.

Share of the civilian population without health coverage. A proxy for healthcare access, not a direct measure of provider supply.

Physicians per 100,000 residents. Measures supply, not quality. Suppressed where too few providers exist to report — about 240 counties.

Jobs, average weekly wage and average establishments for all industries. The chart indexes each series to its first selected year so differently sized units can share one visual; latest raw values remain visible below it.

Counted from the schools themselves rather than from districts. Every school record carries the county it sits in, so the totals are exact; a district's enrollment cannot be assigned to one county because districts do not follow county lines. Schools reporting no students are excluded as closed or administrative, and virtual schools are excluded because a statewide online school would credit its entire enrollment to whichever county holds its office. The district list for each county is still published alongside these totals.

Enrollment divided by full-time-equivalent teachers, summed across every school in the county. This is a staffing measure and not a quality ranking. A state can simply not report teacher counts in a given year: Tennessee reported them for about 96% of its schools from 2020 through 2023 and for none at all in 2024. Rather than leaving that state blank or mixing this year's students with last year's staff, the ratio falls back to that state's most recent reported year and uses that year's enrollment as well, so it measures a single year. No free authoritative source publishes school quality by county, and we would rather show a number that means exactly what it says than a rating whose method we cannot inspect. We previously carried a proprietary state-level school rating and removed it: it was licensed for other use, and being state-level it shaded every county in a state identically, which reads as a fault in the map.

Births minus deaths per 1,000 residents, and net international migration per 1,000, both from the same file as net domestic migration. They are published separately rather than combined because they routinely point opposite ways: a county can gain domestic movers while deaths outnumber births, or lose domestic movers while growing on international arrivals, and either pattern disappears inside a single total. Birth and death rates are stored alongside for county profiles.

The largest individual county-to-county flows, shown on county profiles. The IRS suppresses any pair below 20 returns for disclosure, so these are the largest flows it will name rather than the largest that occurred — in a small county most of the movement can be suppressed and what remains may not be representative. Counts are households that filed from one county and from another the following year, not people, and they miss anyone who does not file. State-level rollups, foreign flows and unclassified flows are excluded because they are not places and would otherwise rank among them.

EIA publishes electricity prices by utility, not by county, and separately lists the counties where each utility has distribution equipment. The county figures are written into the same three metrics as the state figures rather than parallel ones, so each concept appears once and resolves county-first. They are the median across the utilities serving that county, with the number of them published as its own metric so the reader knows whether a figure describes one price or summarises several. The two levels are derived differently — the state figure is EIA's own state series — so a county can sit above or below its state number for reasons other than local prices. A customer-weighted average is not offered: customer counts are reported per utility per state, not per county, so the weighting would be invented rather than measured. Only bundled service is counted — in states with retail choice a delivery-only utility bills for part of the supply, and averaging it against bundled utilities would print a rate no household pays, so counties served only by such utilities keep the state figure instead. Natural gas has no county equivalent and is not approximated: EIA publishes gas utility prices but explicitly does not publish gas utility service territories, so there is no basis for saying which county a gas utility serves.

The map and the statically generated state and county pages used to keep separate lists of which metrics exist. They drifted, and nineteen metrics with data were missing from the page list — so measures that had been live for months, including county frost dates and three health measures, never appeared on a profile. Nothing reported it, because a metric absent from a list is not an error, it is simply absent. The definitions now live in one file and the page list is derived from them, so a new metric appears everywhere or nowhere.

Neither has an honest county equivalent. The FBI stopped publishing county-level crime and the remaining route — summing individual police agencies into counties — would measure reporting compliance rather than crime: agency participation collapsed when the FBI moved to incident-based reporting, with roughly a third submitting nothing and another quarter submitting partial years, so a county whose sheriff files while its municipal forces do not would appear safer than one where everybody reports. Cost of living is published for states and urban areas by construction, with no county series to substitute. Both are shown as the statewide figure applied to every county, which is what they are.

Industry mix, class of worker and labour force participation, pulled on the same call that already supplied the occupation breakdown. Labour force participation is the share of adults working or looking for work, which separates a retired county from an unemployed one — a jobless rate counts only people actively looking, so both read low there. Self-employment covers unincorporated businesses only, so it understates owners who incorporated. Largest industry share measures concentration rather than which industry: a county where one industry employs a third of workers rises and falls with it, whichever it is. The thirteen industry groups are listed on profiles rather than stacked into a bar, because thirteen segments in one bar cannot be told apart.

Global horizontal irradiance — the solar energy reaching a horizontal surface, in kilowatt-hours per square metre per day, averaged over years. Sunny days are not published for every county; this is, so it is what the site shows, labelled as what it is rather than converted into hours of sunshine it cannot support. It tracks cloud cover closely, which is what the question usually means, and it doubles as an indication of what rooftop solar would produce. Modelled from satellite cloud observations at roughly four kilometres and read at each county's population-weighted centre, so a large county with varied terrain is described by where its people live. Humidity is not shown: the same source publishes it only in an hourly download per location, which is not workable for 3,144 counties, and the gridded dew point normals that would support it are a separate build.

The standard proxy for poverty among the students a district serves, since eligibility is set against household income. One caveat is worth stating plainly: districts increasingly use community eligibility, which feeds every child free and reports the whole school as eligible, so a high figure can mean a genuinely poor area or a district that opted into universal meals, and the published data does not distinguish them. Shown only where schools covering at least 60% of a county's enrolment reported it — below that the share describes the schools that answered rather than the county.

The share of a county's colleges that are publicly owned, which matters mostly for price: in-state tuition exists only at public institutions. Shown only where a county has at least one college, since a share of nothing is not zero percent. Student enrolment is not shown — the IPEDS directory endpoint this cron reads carries ownership and location but not enrolment, which lives on a separate endpoint, and a metric that would always be blank is worse than one that is absent.

What the housing in a county actually is, rather than what it costs. Median year built comes from B25035; the decade breakdown and the pre-1980 share from B25034, whose buckets run newest to oldest, so 1980 is the boundary between the sixth bucket and the fifth. Manufactured homes come from B25024, combining mobile homes with the boats, RVs and vans the Census also counts as housing, since both are financed the same way. Vacancy comes from B25002 and second homes from B25004. Second homes are given as a share of all housing units rather than of vacant ones: a third of all homes being second homes describes a resort town, while the same figure as a share of vacancies describes nothing. The line numbers within each table cannot be resolved by name through the API, so each group is checked against its own published total and skipped rather than written if the parts do not sum to the whole. State figures are weighted by housing units, not population, so a county with many homes and few residents — which is what a resort county is — carries its full weight.

Housing units authorised for new privately-owned residential construction. Permits are the first step in building, so these lead every other measure on the site by a year or more — population change records where people went, permits record where housing is being made for them. Counted as units rather than buildings, since a forty-unit block is one building and forty homes. The headline figure is per 1,000 existing homes, because a raw count would only re-rank counties by population. The multifamily share is withheld below 25 authorised units, where two duplexes would otherwise read as a wholesale shift to flats. State figures sum the counties and re-divide by the state's own housing stock rather than averaging county rates, which would let a county authorising four homes weigh as heavily as one authorising four thousand. The survey covers permit-issuing places, which the Census puts at more than 99% of privately-owned residential building.

Every Medicare-certified hospital with an address in the county, how many run an emergency department, and the average CMS star rating of those rated. What this measure cannot do is give a distance: the CMS file publishes address, city, ZIP and county but no coordinates, so a nearest-hospital distance cannot be derived from it, and a county reading zero may still be twenty minutes from a hospital across the line. Zero is shown as zero rather than left blank, because most rural counties genuinely have none and a blank would read as missing data. Hospitals too small or too new to be rated are excluded from the average rather than counted as zero stars, so a county average can rest on a single hospital and should be read beside the count. State ratings are weighted by the number of hospitals behind each county figure, so a county with one hospital does not pull as hard as one with thirty. The dataset URL is read from the CMS catalog on each run, because CMS re-cuts the file quarterly under a new address. Connecticut is placed by town. CMS files its hospitals under the eight counties abolished in 2022, so no planning region matches by name — but every hospital carries a city or town, and the town-to-region map built to fix the Connecticut map resolves them. Where neither the county name nor the town can place a state's hospitals, it is left blank rather than written as having none. County names in this file are hand-entered and contain both spacing variants and outright typos, so matching strips every non-letter and names the known misspellings explicitly.

State and ranking pages are generated ahead of time. County and comparison pages — 4,400 of them — are rendered when first requested and cached from then on, which takes a deployment from 4,646 pages to about 235. Data reaches the site through one scheduled rebuild a day rather than through a rebuild per data commit: thirty-four automated jobs write throughout the day, and rebuilding the whole site for each one was consuming almost all of this project's infrastructure budget while changing nothing about the site's code. Rendering pages on request does not make their figures any fresher — the data is read from the deployment's own files, so an on-demand page shows exactly what the last build shipped. Fetching data at request time instead would make every page view depend on an external service and was rejected as worse than the problem it solves.

County GDP divided by population. The dollar total alone would only sort counties by size, so this is per resident, which measures how intense an economy is rather than how large. The caveat is structural and cannot be corrected: output is counted where the work happens and population where people sleep, so a county of offices that empties each evening reads very high while the suburb its workers drive home to reads low, with neither being richer. It describes the local economy, not local incomes. BEA withholds some county figures for disclosure and marks others unavailable; those are left blank rather than read as zero. The latest year present in the feed is used rather than a fixed one, since county GDP is published each December for the year before last. State figures divide total output by total population rather than averaging county figures.

Land area as the Census Bureau publishes it, taken from the same service that supplies county boundaries rather than computed from the simplified polygons drawn for the map — a shape smoothed for display is the wrong thing to measure an area from. Land only, excluding water, because a county that is half lake would otherwise read as far emptier than it is. Density is population divided by that land area, recomputed at state level from state totals rather than averaged across counties, which would let one small dense county lift a mostly empty state. Shown to one decimal below 100 people per square mile and as whole numbers above, since 0.4 is meaningful in the Alaskan interior and a decimal place in Manhattan is false precision.

Land area excludes water, because counting a lake as living space would make lakeside and coastal counties read as emptier than they are. Area comes from the Census Bureau's own published figure rather than being computed from the simplified boundaries the map draws — a shape drawn for display is the wrong thing to measure from. Density is population over land area, and it is the figure that actually separates rural from suburban from urban: a county of two thousand square miles is a different place at four people per square mile than at four hundred. County averages hide what happens inside them, so a county holding one town and a great deal of empty land shows a middling number describing neither. State density divides the state's people by the state's land rather than averaging its counties, which would let a small dense county weigh as heavily as a vast empty one. Density, population, land area, school enrolment, airport traffic and distance to a train station are all shaded on a logarithmic scale, because each spans an order of magnitude or more — density runs from 1.3 people per square mile in Alaska to 11,169 in the District of Columbia — and on a straight scale one outlier compresses every other place into the palest tenth of the ramp, producing a map that looks empty rather than informative. The six were chosen by measuring where each measure’s median falls on its own ramp, not by eye, and only measures with no zero values qualify, since a logarithm has nothing to say about zero.

Age structure, tenure, one-person households, place of birth and language. Each variable is resolved by the label the Census publishes for it rather than by a line number: a table renumbered between vintages is still found, where a hardcoded line would quietly begin reporting something else. The age bands are classified by reading the label — a band is counted as under 18 only if its whole range falls below 18, so the band spanning 18 and 19 is counted in neither figure rather than being split by assumption. Renters are a share of occupied homes, so the figure is unaffected by second homes and vacancies. Living alone counts households, not people: a place where a third of households are one person is not a place where a third of residents live alone. Language covers residents five and over who speak something other than English at home, and says nothing about how well they speak English, which the Census publishes separately. State figures weight each county by the population or the households the measure describes, not by county count.

Any measure can be embedded in another site as a single-metric map, with the code on that measure's ranking page. The chart carries its own source line and a link back, so a copy of it on someone else's page still says where the figures came from and what shading was used — a chart that travels without its provenance is the thing this project exists against. Embeds are not indexed by search engines, so they do not compete with the page hosting them. They are rendered on request rather than built ahead: there are 135 and most will never be used.

Adherents reported by the 372 religious bodies that took part, as a share of population, plus congregations per 100,000 residents. What this counts is narrower than it looks and the difference is the whole story. Adherents means members, their children and others who participate: broader than membership, narrower than belief. Those 372 bodies account for 48.6% of the US population nationally, so the rest are not counted as irreligious — they are people whose tradition did not take part, who attend somewhere that keeps no rolls, or who believe without belonging. A county at 35% has 35% on a participating body’s rolls, and nothing here describes the other 65%. Some denominational totals are estimated where a body could not supply county detail. A few counties report more adherents than residents, because bodies record where a congregation sits rather than where its members live; those are capped at 100%. The share is recomputed from the published counts rather than copied from the published percentage, so the figure and its inputs cannot disagree. State figures divide the state’s adherents by its population rather than averaging county shares.

The share of adults who could not reliably afford enough food in the past year, and the share who received SNAP. Both are modelled estimates rather than counts: CDC projects the national BRFSS survey onto counties using local demographics, so a county’s figure is what a population like it typically reports, not what that county was asked. Close to a measurement in a large county; leaning on the model in a small rural one. Food insecurity measures affordability and not distance to a shop — a place can have groceries on every corner and still rank high. Food assistance is shown as neutral, because a high figure reflects both how much need exists and how well a state enrols those who qualify, and two counties with identical hardship can differ purely because one state makes applying easier. Coverage is partial and unevenly so: these come from an optional survey module that not every state runs, so about 2,300 of 3,143 counties have a figure and the gaps are whole states rather than scattered counties — a blank state means its survey did not ask, not that nobody there is affected. Only these two measures are taken from PLACES: obesity, diabetes, physical inactivity and mental distress already come from County Health Rankings, and taking them twice would duplicate rather than add.

Supermarkets and grocery stores, counted as establishments rather than chains, with convenience stores counted separately — a county with a filling station and no supermarket is a different place from one with three supermarkets, and combining them would hide exactly that. The Census omits a county from a code’s results rather than returning a zero, so an absent county has either no such shop or was withheld and the two look identical. A broader code settles it: all retail covers 3,226 counties, essentially everywhere, so absence at the supermarket code is a real absence. A zero is written only where the county appears under all retail, which proves the Census reached it; where it does not, the field is left blank rather than assumed. Around 850 counties have no supermarket. The count covers shops inside the county line, so a county with none may still be a short drive from a large store across the border — this measures what is in the county, not what is reachable from it.

The share of residents more than a mile by road from a SNAP-authorized food store in an urban tract, or more than ten miles in a rural one, and the share living in a tract that is both low income and low access — USDA’s own definition of a food desert, which requires both conditions rather than distance alone. Driving distance along the road network rather than straight-line, since how far a shop is as the crow flies is not a question anyone has. Published by census tract and aggregated here to counties by population, so a county figure is the share of its residents rather than the share of its tracts. This is the one source on the site that cannot update itself: USDA builds its download links with JavaScript, so nothing can fetch the file automatically, and it is downloaded by hand and aggregated by a script in the repository. USDA republishes roughly annually, which is close to the rhythm the data changes at, but the step is manual and the date above says when it last ran. Read it beside the store counts: a county can hold several supermarkets clustered in one town and still leave most of its residents far from one.

The districts based in each county, with enrolment, school count, students per full-time teacher and whether the district is a city, suburb, town or rural one. They are listed rather than ranked, and that is not a gap we intend to fill: no district test score is comparable across state lines, because every state runs its own assessment, so there is nothing honest to rank by. Educational service agencies are excluded — Puget Sound Educational Service District 121 has no students and teaches nobody, and counting such bodies would inflate every county. Charter districts are included and marked, since a single-school charter is a different proposition from a district of twenty schools. A district is filed under the county where its offices sit, so one serving two counties appears under one of them and a county’s list is near-complete rather than certainly complete. The ratio is shown only for districts of at least a hundred students, below which it is noise. District finance is deliberately absent: the only year the portal returns is 2020, six years old and dominated by pandemic relief lines.

Thirty-four counties have fewer than 1,000 residents. Loving County, Texas has 33, so its two fatal crashes read as 6,060 deaths per 100,000 against a national median of 32; a county of 647 with a single physician reads as 272 doctors per 100,000; and Loving County's real oil production divided by its 33 residents reads as $348 million of output each. The arithmetic is right and the figure is meaningless, and on a shared colour scale a handful of these flatten every other county in the country. Rates — traffic deaths, the road-safety trend, net migration, primary care physicians and economic output per resident — are therefore withheld below 1,000 residents rather than shown as outliers. Counts, distances, dollar figures and shares are unaffected, because they do not divide a few events by a tiny denominator.

Connecticut abolished its counties in 2022 and replaced them with nine planning regions, so it cannot be handled once and forgotten — each publisher migrated on its own schedule. Census, NCES, IPEDS, HUD, FEMA and the Population Estimates Program all publish the planning regions. Zillow, County Health Rankings and NHTSA FARS are still keyed to the abolished counties, and where that is so the figure is left blank rather than borrowed from a county that covered different ground. Seven of the nine regions draw 90 to 100% of their towns from a single old county, but Capitol splits 77/23 between Hartford and Tolland and Naugatuck Valley splits 43/43 between New Haven and Litchfield, which is a coin flip rather than a number.

Net domestic migration per 1,000 residents: arrivals from elsewhere in the United States minus departures to it. International arrivals are deliberately excluded, because the question this answers is where people already in the country are choosing to move, and mixing the two hides it — most large coastal metros are growing in total while losing domestic movers. Published as a rate so counties of different sizes can be compared; the raw yearly counts are on each county profile.

The share of adults 25 and over at each level of schooling completed, summed from the table's detail lines. This describes the people who live in a county rather than the schools in it, which are different questions: a county with a university can have modest attainment if its graduates leave, and a commuter county can have high attainment with few schools of its own. County figures come from the 5-year survey, which covers every county at the cost of being a rolling average.

Active, currently reporting institutions, placed in the county of their campus. Distance is measured from the county's population-weighted centre to the nearest campus granting bachelor's degrees or above, so it measures from where people actually live rather than the county's geometric middle. The bachelor's-or-above category is named for what the source code means rather than as "four-year": community colleges that grant applied bachelor's degrees fall inside it, and calling those four-year institutions would read as an error.

Presidential results only. Local and down-ballot races often differ sharply. Shown as a fact, and excluded from Find your fit.

The average coldest winter reading over 15 years, cut into the standard 10°F bands. Not taken from the USDA map, whose data is not published for reuse — it should agree in most places, but a county spanning a mountain range varies inside itself more than one number can show.

Average last spring and first fall freeze at 32°F, averaged across the stations in each county, with growing season as the gap between them. Present for 2,908 of 3,144 counties; the rest have no station reporting daily minimum temperatures.

Price is revenue ÷ sales and consumption is sales ÷ customers, the same derivation EIA uses in its own tables, averaged over twelve months. The bill is the product of the two and is not predictable from price alone: Louisiana’s rate is 40% of California’s, yet the bills are within a few dollars.

Dollars per thousand cubic feet, averaged over 24 months with the monthly series kept for the trend chart. Per-unit price rises in summer, when fixed monthly charges spread across almost no gas, and falls in winter as volume climbs. Read it alongside how many homes actually burn gas.

Share of occupied homes by heating fuel. The one metric in this group with genuine county detail. It explains the rest: electric-heat states are exactly the high-kilowatt-hour ones, and propane and fuel oil are bought by the tankful at market prices rather than billed monthly, so they appear on no utility statement.

National percentiles from 0 to 100. Each blends how likely the hazard is with how much population and property sit in its path, so a populated county near wildland scores higher than empty land that burns as often. State figures are population-weighted from counties. Flood risk is documented by FEMA but rejected by their live data service, so it is omitted rather than estimated.

Average daily fine particulate matter (PM2.5) in micrograms per cubic metre. The EPA annual standard is 9. Western counties read higher in wildfire years.

Counts major federal declarations affecting the county, not individual storms, fires or damages. A declaration depends on federal assistance criteria, so it is useful context alongside risk scores, not a complete count of every hazardous event.

The only hand-maintained data on the site, and labelled as such wherever it appears. We tried to derive it automatically from Wikidata, which is CC0 and self-updating, but the query returned individual players and defunct franchises alongside clubs, and ten of the thirty-two NFL teams — the Bears, Seahawks, Packers and Giants among them — had no usable location at all. A source that silently drops a third of a league is worse than a list refreshed deliberately. Rosters come from the league feeds; the state each club plays in is checked by hand, because it is not the state in the name for every club: both New York NFL teams play in New Jersey and Washington plays in Maryland. Teams relocate roughly once every year or two and it is widely reported when they do.

Every league except college football is placed in the county its venue is in, which is often not the county in its name. The Cowboys play in Arlington, the Patriots in Foxborough, the Braves in the Cumberland area of Cobb County, the Gwinnett Stripers in Lawrenceville, and the Augusta GreenJackets across the state line in North Augusta, South Carolina. County comes from a Census place lookup rather than by hand, and anything it cannot resolve gets an entry recording why. Two cases are worth naming because they resolve cleanly and wrongly if left alone: the Census place called New York spans five counties, so the Knicks and Rangers are entered as Manhattan rather than letting the lookup answer, and Boise resolved to Boise County, a rural county that merely shares the name, until the ordering was corrected to find the place Census calls Boise City in Ada County.

The per-league counts shade the map at both state and county level. Only college football remains state-level, shown the same for every county in the state, because those clubs have not yet been given venue cities.

Currently the NFL and top-division college football. A state showing nothing has no team in those competitions, which is not the same as having no teams: minor league baseball, lower college divisions and other leagues are not included, and in many places those are what people actually attend. The other major leagues will be added the same way.

All 3,144 counties and county equivalents in the 50 states and the District of Columbia. Puerto Rico, Guam, the U.S. Virgin Islands, American Samoa and the Northern Mariana Islands are outside that set, so an airport or station there appears in a federal source but in none of our counties. Fifteen airports with scheduled airline service fall into that category and are excluded rather than misfiled.

Connecticut replaced its eight counties with nine planning regions in 2022, after the 2020 county centres of population were tabulated. That left the state with no measurable centre and therefore no distance-based figure at all: airports, rail, frost dates and garden zone were blank statewide. We now derive the centres the way Census derives its own, population-weighted from the tracts inside each region. Both halves are Census data — tract centres from the 2020 release, boundaries from TIGERweb — and a derived centre that falls outside its own region aborts the update rather than publishing.

Three metrics are blank for the nine planning regions, for three different reasons. County Health Rankings began publishing planning-region figures with its 2025 release, but fine particulate matter was not among them — it is still issued only for the eight former counties, which covered different ground, so there is nothing to join to. Primary care physicians appears on neither of that release's Connecticut lists and was not updated for the state at all. Zillow's rent index has not migrated and still reports the eight abolished counties. We leave all three empty rather than borrowing former-county values. Rent is the one case where the site is not silent: HUD Fair Market Rent publishes on the planning regions and is shown for all nine.

NOAA's nClimGrid county product covers the contiguous states only and is keyed to the counties NOAA holds on file, which left 46 counties with no temperature, precipitation, snowfall or day counts at all: every Alaska borough, every Hawaii county, the nine Connecticut planning regions, the District of Columbia and one Virginia city. Because that product is already aggregated to counties it carries no coordinates and cannot be re-placed, so these counties are filled instead from weather station observations assigned to the county the station physically sits in, against current Census boundaries. ACIS reports Connecticut stations under the abolished county codes, so its own county field is not used. This is a different estimator from the gridded average — it is the mean of the stations inside a county, which leans toward valleys, airports and populated ground — so it is used only where the gridded figure does not exist, never in place of one, and the number of stations behind each such county is recorded.

Any file built before 2022 still describes Connecticut's old counties, and files built before 2019 still describe Alaska's Valdez-Cordova census area. Joining on someone else's county code therefore carries a hidden dependency on when their file was made, and the failure is quiet: counts read zero rather than erroring. Where a source's own geography can drift, we assign by coordinates against current boundaries instead, because coordinates do not go stale.

County figures reflect local annual averages. Alaska, Hawaii, Connecticut, the District of Columbia and one Virginia city fall outside that product and are measured from weather stations placed by coordinate instead; see the geography section for what that changes.

FEMA publishes eighteen hazard scores per county; twelve are shown. Each combines how often the hazard occurs with what it would cost when it does, so a score reflects exposure and consequence together rather than frequency alone. Riverine and coastal flooding are kept separate because they threaten different places for different reasons, and coastal flood is zero inland by definition. Flooding was absent from this site until August 2026 because the field was requested under the wrong name and the rejection was discarded on successful runs; the cron now reports every hazard field the service offers and names any it does not.

Blends likelihood with what is exposed to it, so populated land near wildland scores higher than empty land that burns as often. State figures are population-weighted from counties.

Current five-year estimate.

People who moved in minus those who moved out, published per year so any window from one to five years is a real total. Deliberately excludes births and deaths: a county can grow without anyone moving in, and presenting net population change as migration would be wrong. Values available from 2021.

Origin-destination pairs — how many people moved between two specific states. Survey estimates carrying margins of error, so large flows are more reliable than small ones and a difference of a few hundred between two states may not be meaningful. A different survey and vintage from net migration above, so the two will not reconcile exactly.

Households that filed a tax return from one county and from another the following year. This does not reconcile with the Census net migration above and is not meant to: Census counts people, the IRS counts filed returns, so this misses anyone who does not file — the very poor, and some retirees — and counts a household once rather than per person. Totals are taken from the IRS's own summary rows rather than by summing county pairs, because flows under 20 returns are suppressed for disclosure and summing pairs would undercount, by more in small counties than large ones. The 2022–2023 release also changed how returns are matched between years, adding about five percent more of them, so it should not be trended against earlier files.

Average adjusted gross income of the households that moved in, and of those that moved away. Worth reading as a pair: a county can gain population while the people arriving earn less than the people leaving, which population change alone will not show. Neither direction is straightforwardly good — leavers earning more can mean people cashing out or being priced out — and the figures say nothing about why anyone moved. Adjusted gross income is not take-home pay and excludes income that is not taxed.

Neutral: low often means families and college towns, high means retirees or out-migration.