By the BestNeighborhood data science team. Last updated July 2026.

This page documents every measure behind the best places to live ranking and the best neighborhoods maps: what it counts, which agency published the underlying data, how the number was built, what it was checked against, and where it stops being reliable. Sections run in default weight order, so the measures that move the ranking most come first.

What this ranking does that other lists do not

Three things.

It measures 39 things instead of about 16. The most popular alternative Best Places to Live list scores cities on roughly 16 metrics grouped into four indexes. This one carries 39, including several that exist nowhere else at small-area grain: modeled traffic and aviation noise, religious adherence, political lean, tree canopy over walkways, estimated heat deaths, and local income tax rates for 3,855 separate jurisdictions.

It scores census block groups, not cities. A block group holds about 1,250 residents. There are 240,160 populated ones covering 338.1 million people, and every measure is computed for each of them. City rankings hide the fact that the range inside one metro area is nearly the range of the entire country: nine out of ten Americans live in a block group scoring between 40.5 and 58.9 on the default weighting, and nine out of ten Chicago-metro residents live between 41.2 and 58.6. Lincoln Park scores better than 86% of American neighborhoods. Englewood, ten miles south and inside the same city, scores better than 4%.

The reader sets the weights. Every measure has a default weight from 0 to 100, and every default can be moved. Places are then re-ranked on the reader’s own weighting across all 53,313 ranked places: 387 metro areas, 445 cities, 2,098 small cities, 50,333 towns, and 50 states. A family weighting schools and commute sees a different list than a retiree weighting weather and crime.

Everything below is built from primary government data, from a documented model whose inputs and validation are stated, or from data BestNeighborhood collected itself. Where a number is an estimate, it says so.

How scoring works

Every measure produces an absolute score from 0 to 100, anchored to a real best case and a real worst case, where higher is better. That absolute score is what drives the ranking. A percentile is carried alongside it for one purpose only: coloring a map of that single measure.

Absolutes rank, percentiles color

Anchoring to a fixed best and worst preserves real spacing. A standout place scores near 95 and a middling place near 70, and the gap between them means something. A percentile would flatten that, because a percentile is uniform by construction: it bunches the top places together and stretches the crowded middle apart. Averaging percentiles across many measures makes it worse, because the average of k roughly independent percentiles compresses toward the middle by about one over the square root of k. Blend 30 of them and almost everything lands between 35 and 65.

So the blend runs on the absolutes, always. A single-measure map is colored by percentile, because most places cluster near the middle on raw values and a raw-value map looks flat. A reader’s combined map is colored by blending the absolutes with the reader’s weights, using the identical arithmetic as the ranked list, then placing each neighborhood’s blended score in the national distribution. The map and the list can never disagree, and a given color means the same thing in every city.

Neutral is almost never 50

Each measure records its real national median and the interface shows it, so an average place reads as average even when its number is not 50. School quality sits near 40. Tree shade near 42. Religiosity near 37. Cell coverage near 90. Pleasant weather near 23. Transit near 2. Ocean proximity is 0 for most of the country. Treating 50 as average would misread every one of those.

Comparing places of different sizes

Places are grouped into tiers: metro areas, cities, small cities, towns, and states. A city can hold more people than a small metro area, so percentiles are computed inside each tier. Metro areas rank against metro areas, cities against cities. The absolute 0 to 100 score stays comparable across every tier, so a neighborhood, a town, a metro, and a state all sit on one ruler when a reader wants to compare across them.

Missing data is dropped, never zeroed

A place is never penalized for data that does not exist. When a measure is missing, it is removed from that place’s blend and the reader’s weights are spread across what remains. Where a defensible stand-in exists, the value is filled from the nearest measured block group within 200 kilometers or from the county, and the fill is flagged. The 200-kilometer cap is load-bearing: without it a Puerto Rico block group would silently inherit a Florida one 1,500 kilometers away.

Zero is used only when zero is the measurement. Four measures have real zeros that must never be treated as gaps: most of the country genuinely is more than 45 minutes from an ocean, 45.1% of block groups genuinely have no transit service, silence genuinely is the absence of transportation noise, and a place with no dense urban land within reach genuinely has none. Puerto Rico and the territories are left blank for measures whose federal source excludes them, rather than filled with a mainland value.

Cost of living and housing cost cannot both be weighted

A cost-of-living index already contains housing. In this one it is just under 28% of the basket by construction. Offering cost of living and housing cost as two independent sliders would count shelter twice, once at full weight and again at 28% of the cost-of-living weight. So affordability ships as one control with four mutually exclusive views, and exactly one is ever active:

Cost of living against local pay (the default). What local wages actually buy after local costs. This is the question for someone taking a job in the area.
Cost of living on its own. The raw price level. This is the question for someone bringing an income with them.
Housing cost against local pay. Home value and rent measured against local income, blended by the share of owners and renters.
Housing cost on its own. The raw price of shelter.

Nothing is divided at render time. All four are precomputed on the same 0 to 100 scale, and picking one selects it.

The shared rollup

Every measure is computed at whatever grain its source supports, pushed down to the 2020 census block, and rolled back up population-weighted through one shared table of 8,174,955 blocks. That is why city numbers survive the fact that a third of block groups and half of census tracts straddle a city line. The rule throughout is to roll the raw value and then score it, never to roll a score and never to average percentiles.

The measures

Crime safety

Default weight 100. National median 53.

Crime carries the heaviest weight because it is the factor most readers rank first, and because national lists tend to underweight it or publish it at a grain too coarse to act on. It is published here at the block group, roughly 1,250 residents, rather than the county.

The underlying figure is not a crime rate. It is weighted by impact to the community, which lets a murder and a stolen bicycle sit in the same number without pretending they are equivalent events. Severity weighting follows the cost-of-crime framework in McCollister, French and Fang (2010), covering direct victim costs, criminal-justice-system costs, and lost economic contribution. Those constants were recovered rather than chosen: measured against production values across one mid-size ZIP code in each of 40 states, they hold to a coefficient of variation of 1.40% and 2.58%, which is rounding.

Correcting for states that report less crime than they experience

Reporting standards are set at the state level, and they differ enormously. On a raw national scale that produces a specific, repeatable artifact: transparent states such as Washington, Colorado and Oregon paint dangerous, and states with thin reporting flood the safest lists. Any ranking that skips this step is partly ranking police paperwork.

The correction measures reporting against two things police cannot under-report. Violent crime cost is fitted against homicide deaths per 100,000 from CDC WONDER, which come from death certificates written by coroners counting bodies. Property crime cost is fitted against motor vehicle thefts per 100,000 from the National Insurance Crime Bureau, which come from stolen-vehicle entries that insurance claims force into the record. Each fit is a log-log regression across the 50 states, and the state residual is winsorized to a maximum correction of 1.82 times in either direction. Fit quality is r = 0.70 for violent and r = 0.80 for property. Washington DC is predicted but excluded from the fit, because a city-state is not a state observation.

Fitting the two components separately is the part that matters. A single free fit of total crime cost against both anchors loads onto the vehicle-theft term and then excuses Louisiana and Arkansas, the two highest-homicide states after Mississippi, as over-reporters. That is exactly backwards. Keeping the anchors separate means a state’s violent-crime under-report cannot be laundered through its car-theft rate, and Denver and Seattle keep their real vehicle-theft problem while shedding only the penalty for reporting it honestly.

The effect is large and it is supposed to be. Mississippi moved from fifth safest to twenty-third worst. Florida’s share of the twenty safest metro areas fell from more than eight to two. Against an independently constructed severity-weighted harm index, the calibrated scores agree on 7 of the 10 most dangerous states and 7 of the 10 safest, up from 6 of 10 before the correction.

Two disclosures follow from this. The dollar figure shown on the site stays the reported cost, so it remains comparable to CrimeGrade, which means the safety score is deliberately not in strict order with the displayed dollar across state lines. And the correction is a modeled estimate, winsorized and disclosed, not a measurement. It is refreshed whenever the source data refreshes.

Coverage reaches 100% of population. Block groups below 100 residents are set aside rather than scored, because a denominator that small produces nonsense: the 2,520 block groups under 100 people have a 90th-percentile cost per capita of $4,959 against a median of $349. Those places, plus 272 airport block groups, become no-data rather than zero, totaling 1.05% of block groups and 37,740 people. Downtown retail districts reading high is real and stays, because the only honest fix would be a cost-per-hour-spent denominator and no such denominator exists.

Full detail on the underlying crime data is at CrimeGrade.

Cost of living and housing affordability

Default weight 90, the second heaviest. National median 40.1 on the default view.

Large data companies sell cost-of-living figures for a few hundred cities. This one is published free for every block group in the country, and building it that way took a significant amount of work.

Why it has to be built in two levels

Published cost-of-living figures cover only a few hundred metro areas, against tens of thousands of candidate features, and a block-group model cannot be fitted to a metro-level figure. So the build separates the two questions it is actually asking. What does this metro area cost, which is learned. And how does this neighborhood differ from its metro, which is measured.

At the metro level, six components are modeled separately: groceries, housing, utilities, transportation, health care, and miscellaneous goods and services. Each is fitted against independent metro price benchmarks on annual panel rows from 2015 forward, cross-validated with five-fold grouping by metro area so that a metro and its neighboring benchmark points never straddle a fold. The model class is chosen per component rather than globally. Three baselines run every round, including a rescale of the federal regional price parities, and a component that cannot beat all three does not ship. The composite lands at R-squared 0.846 with RMSE 7.3 when whole metros are held out, against 0.574 and 12.1 for the price-parity baseline. Predictions are clamped to the observed range, because unbounded extrapolation into sparse rural regions was a real failure mode rather than a hypothetical one.

That produces a level for each of 970 anchor regions: every metro area in the current delineation, plus one non-metro remainder for each state. Where a region carries a recent benchmark observation, the modeled component is calibrated 70% to the observation and 30% to the model.

The within-metro rebuild

The region’s price level is already anchored, so the neighborhood multiplier has to carry the entire distance from a metro average to a specific place. In New York that means carrying the metro’s index of about 227 up to Manhattan’s roughly 505.

The previous multiplier blended a 40th-percentile market rent with the rent households actually report paying. Re-tested on the 32 metro areas that carry more than one independent benchmark point, each demeaned against its own metro, it explained R-squared 0.200 of within-metro housing variation at about half the required amplitude. The reason is structural: a 40th-percentile rent caps exactly where the expensive tail lives. Manhattan’s capped rent is 1.27 times Morristown, New Jersey’s. Its true housing ratio is 3.98 times.

The measured skill of each candidate on that validation set:

Within-metro housing signal	R-squared	Amplitude
Target		0.177
Reported housing cost alone	0.032	0.083
Capped market rent alone	0.329	0.086
Previous production blend	0.200
Listing price per square foot, county grain	0.647	0.181
Listing price per square foot, ZIP grain	0.723	0.179

Reported housing cost is close to useless inside a metro area, and turns negative when the capped rent is present. Manhattan’s reported rent of $2,679 sits below Morristown’s $3,072, because of rent stabilization and survey top-coding. It survives at a small weight only because it is the one signal that varies below the ZIP code. The current multiplier weights market listing price per square foot at 0.84 and reported cost at 0.12, both as deviations from the region mean, with the exponential renormalized so each region’s population-weighted mean multiplier stays exactly 1. Without that renormalization the curvature of the exponential would push every region off the level it was calibrated to. Listing coverage reaches 99.4% of block groups, 97.9% at true ZIP grain.

Components are combined on a standard household budget allocation: miscellaneous goods and services about a third, housing just under 28%, groceries about 13%, transportation about 12%, utilities about 10%, and health care about 4%. The result is rescaled so the national population-weighted mean is 100, which means 150 is half again as expensive as the national average and 75 is a quarter cheaper.

Validation

Rolled back up to the 301 benchmark areas, the block-group build reproduces those independent benchmarks at r = 0.986 for the composite and r = 0.985 for housing. Isolating the multiplier alone against the one-to-one line gives within-metro R-squared 0.717 at 86% of target amplitude, the deliberate result of a mean-squared-error-optimal shrinkage coefficient. Every one of the 239,780 block groups is populated, 0.02% sit at a clip bound, and the share of Manhattan block groups reading below the national average is 0.0%, down from 0.2% before the rebuild.

The four affordability views are genuinely different lenses, and that is the point. Detroit is top-five on housing cost alone, because its houses are cheap, and bottom-five on cost of living against local pay, because its incomes are lower still. Measured against the census cost-burden tables, the against-income view correlates at Spearman -0.66 for cities over 10,000 residents while the price-only view manages -0.24, which is the direct evidence that the against-income view adds real burden signal rather than restating price.

Sources. The price level is trained against independent metro price benchmarks and market listing data. Public inputs are the U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates for home value, rent, income and tenure, the U.S. Energy Information Administration for electricity, gas and gasoline prices, the Bureau of Economic Analysis regional price parities as a baseline, and the Bureau of Labor Statistics for price series and household budget shares.

Honest limits. Groceries, transportation, health care and miscellaneous costs currently vary between metro areas but not inside them, so within-city variation comes from housing, utilities and income. The very top of the luxury market is still under-shot by roughly 100 housing points at Manhattan, which is a known open problem rather than an artifact. Raw dollar columns are population-weighted means of block-group medians, not true medians, and are published as display approximations: against true census medians they hold rank at Spearman 0.96 to 0.99 for cities with a 3 to 9 percent level bias.

School grade

Default weight 85. National median 40.

Every block group carries an elementary, middle, high and overall grade on an A-plus to F scale, from SchoolGrade, which rates every public school in the country from state assessment results published by the state departments of education, including schools that opt out of annual state testing.

Grades are cut from the achievement percentile into 13 equal-width bands of about 7.7 points each, so the printed grade and the printed percentile can never disagree. Percentiles are what get rolled up, not raw scores, because raw test score and score percentile are not a clean function of one another across states. Rolling the percentile keeps a city’s B meaning the same thing as a block group’s B.

A separate value-added residual is carried alongside the achievement grade. It compares realized outcomes against what a school’s surroundings would predict, using a multilevel expectation model with better than 95% out-of-sample accuracy built on hundreds of covariates. That residual is published as a school’s teaching effectiveness, and it is deliberately not folded into the letter grade, which is cut from achievement only. A reader who wants to know which schools outperform their demographics and a reader who wants to know which schools have the strongest students are asking different questions.

Block groups roll up to places population-weighted through the shared block table. Coverage is 95,053 matched places out of 95,571, with 405 ungraded. The 200-kilometer fill cap isolates those to exactly Puerto Rico. Every mainland place is graded. The United States as a whole lands at a B-minus.

Full detail is at SchoolGrade.

Nice weather

Default weight 70. National median 23.

The score is the share of the year that is genuinely pleasant, and it is strict on purpose. Each of 365 days scores 0 to 3, one point for each condition it meets, all measured on feels-like temperature rather than raw temperature:

Feels-like low above 45F and feels-like high below 85F.
Feels-like average between 55F and 75F.
Below a 14% chance of a significant-rain day, meaning a tenth of an inch or more.

Points are summed across the year and expressed as a share of the possible 1,095. A place has to be both mild and dry to score, which is why the national median is 23 rather than 50. Snow is not scored.

Feels-like uses the National Weather Service bands correctly rather than approximately. Wind chill applies only below 50F with wind above 3 mph. The Rothfusz heat index applies only at or above 80F, using relative humidity derived from dewpoint. Between those bands the raw temperature is used, because both formulas are invalid there.

Station profiles come from ten years of NOAA National Centers for Environmental Information Global Summary of the Day observations, 2016 through 2025. A station qualifies with at least eight of ten years and 2,400 valid temperature days; co-located duplicates within a kilometer are collapsed; all years pool into 365 calendar-day bins smoothed over a 15-day circular window. Rain is only drawn from stations with trustworthy year-round precipitation coverage, which may be a different station than the one supplying temperature. Precipitation flags are honored, so a reading of zero carrying the no-precipitation-reported flag is treated as missing rather than as a dry day.

Roughly 289 marine stations are excluded: buoys, tide gauges and oil platforms report ocean-moderated air temperatures that misrepresent the land. Before the exclusion, Brooklyn read 13 points warmer than the Bronx. Each block draws from its nearest land station through a three-dimensional spatial index, which behaves correctly for Alaska, Hawaii and Puerto Rico, and gets an elevation correction of 6.5C per kilometer against that station’s elevation.

All 8,174,955 census blocks are scored individually and rolled up, rather than one station being assigned to a whole city from its centroid. Against the retired centroid method the rebuild correlates at r = 0.9944 with 94.5% of places within two points, which makes it a refinement rather than a re-scoring. The places that moved are exactly the ones a centroid misrepresents: sprawling or steep. Garland, Texas rose 12.5 points, Denver 7.8, Frisco fell 8.9, Colorado Springs 8.1.

Honest limits. City and metro numbers genuinely differ for metros that span a coast or a mountain range, by 5 to 14 points in Denver, Oxnard, Riverside, Seattle and Los Angeles, so both are published rather than one standing in for the other. Puerto Rico scoring near zero is correct rather than missing, because a place that is warm and wet year-round earns few points on a mild-and-dry measure.

Walkability, transit, and biking

Default weights 70, 35 and 25. National medians 13, 2 and 31.

Walkability here is measured as a walk, not as a distance. The score runs on a pedestrian network graph built from every street, footpath, sidewalk, crossing, driveway and parking aisle in the country, scored on a hexagonal grid of about 66 meters, with routes solved by bounded shortest path rather than by drawing circles.

What that buys is the difference between a place that looks walkable on a map and one that is walkable on foot:

Crossings cost what crossings cost. Crossing a major arterial adds roughly a tenth of a mile of equivalent walking, which is the wait at a typical two-minute signal cycle, plus a danger surcharge. A marked crosswalk keeps the wait and removes most of the danger. Darting across an unmarked arterial roughly doubles the cost. A quiet two-lane street is close to free.
Freeways sever the network. Motorways, trunk roads and their ramps are not walkable edges at all, so getting to the other side costs the real detour to the nearest bridge rather than the straight-line distance.
Walking beside traffic is penalized. An edge running alongside loud, fast traffic costs up to 60% extra distance, scaled by the modeled road noise at that edge, which already encodes traffic volume, speed and lane count.
Curb cuts and parking lots count. Each driveway crossing charges a detour, and ground-level parking aisle length is penalized. Structured and underground parking is not.
Tree canopy is sampled on the road corridor, at each edge midpoint, rather than averaged over the surrounding area. An area average credits a road cut through forest for trees nobody walks under. In Rhode Island the roadside canopy mean is 23% against a 44% area average, and that 21-point gap is the inflation the roadside method removes.

Destinations decay faster than competing scores allow. A destination within a quarter mile counts fully, half a mile counts about 39%, a mile counts 6%, and past 1.25 miles it counts zero. Walk Score still credits about 12% at a mile and counts destinations out to 1.5 miles. Categories are capped so a district of nothing but restaurants cannot score as a complete neighborhood, and a grocery gate caps the amenity score at 35% for any place without a close grocery store. Only groceries open that gate.

Sidewalk presence earns a bonus and sidewalk absence is never penalized. That is a deliberate choice about data quality rather than about urban design: the ratio of mapped sidewalk length to road length runs from 0.19 in Utah to 0.01 in Mississippi, West Virginia and North Dakota, a nineteenfold spread that tracks how thoroughly volunteers have mapped each state rather than how many sidewalks exist. Penalizing absence would publish a mapping artifact as a fact about a neighborhood.

The measured consequence: suburban Sandy, Utah falls from 43 to 10 and a subdivision edge from 35 to 16 once real routes and crossings are counted, while downtown Salt Lake City and Sugar House hold above 90.

Transit and biking

Transit is scored in time rather than in stop counts. The cost of reaching a route is the network walking distance to its nearest stop plus half its headway converted to walk-equivalent distance, decayed on the same curve, so a bus that runs every 30 minutes and stops five minutes away scores close to nothing. Rail, subway and tram carry a 2.0 usefulness multiplier, ferry and cable 1.5, bus 1.0, and evening and weekend service adjusts the result by 0.7 to 1.3. That density term is blended 0.78 with a 0.22 reachability term counting the jobs and destinations reachable within about one frequency-weighted transfer. Where no transit agency publishes a schedule, the assumption is no transit.

Biking blends lane infrastructure at 0.30, destinations and connectivity at 0.30, local bicycle commute share at 0.18, bike-to-transit access at 0.14, and hills at 0.08. Hills are deliberately light where Walk Score gives them 25%. Infrastructure tiers run 3.0 for off-street paths, 1.5 for lanes and 1.0 for shared roadways, with painted lanes discounted by adjacent traffic noise, and a low-stress gate inspired by the Furth and Mekuria level-of-traffic-stress framework.

Sources. Overture Maps Foundation for places, transportation and buildings; the national catalog of General Transit Feed Specification schedules covering 1,115 to 1,145 U.S. agencies parsed to 723,860 stops and 17,876 routes; USGS and MRLC National Land Cover Database tree canopy; USGS 3DEP ten-meter elevation for hills; Census LEHD employment data; Census American Community Survey bicycle commute share; OpenStreetMap for bike lane tiers; and the BestNeighborhood road-noise surface described below.

Honest limits. Grades are absolute, calibrated so 100 means a person can live without a car, and they are not curved. Against a scrape of Walk Score, biking correlates at r = 0.82 in Portland and transit at r = 0.70 in Seattle, both running lower than Walk Score because the wait and crossing penalties are real. Walkability is deliberately no longer validated against Walk Score at all, because Walk Score rewards the car-dependent, arterial-severed places this model penalizes. Transit schedules are static, so real-time service disruption is invisible. Roadside canopy covers the contiguous 48 states only.

Full detail is at Walk Grade.

Short commute

Default weight 65. A 27-minute commute, the national average, scores about 60.

The census publishes a mean commute time at the tract level but only bucket counts at the block group, which is the grain this ranking needs. Combining them takes a step most users of this data skip.

A block-group mean is first computed from the twelve travel-time buckets in table B08303 using bucket midpoints. That estimate is then checked rather than trusted: the block groups inside each tract are aggregated two ways, once by summing their buckets and recomputing the midpoint mean and once against the published tract mean, which isolates midpoint-method error from any geographic mismatch. The error turns out to be almost purely additive, a constant plus 2.0 minutes with a standard deviation of 1.4.

Because the error is additive, the correction is additive. Each tract’s published mean is used as the anchor, and the difference is applied to every block group inside it. The worker-weighted average of the corrected block groups then reproduces the published tract mean exactly, while the minute-level differences between block groups, which the tract file does not contain, are preserved. Block groups with unusable buckets take the tract mean, and tracts absent from the published file take the raw estimate less the global median bias, flagged.

Rollup is worker-weighted rather than population-weighted, using each block’s population scaled by its block group’s worker-to-population ratio, which apportions correctly for block groups straddling a city or metro line.

The score anchors a 5-minute commute at 100 and a 60-minute commute at 0.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates, tables B08303 and S0801. Puerto Rico has no data.

Air quality

Default weight 60. National median 59.

Air quality is interpolated rather than assigned. Assigning each place its nearest monitor draws invisible Voronoi cells across the country, and a hard seam then cuts straight through a metro area wherever the nearest monitor flips. On a block-group map that reads as a data artifact because it is one, an artifact of the tessellation rather than of the air.

Instead, each block group’s population-weighted centroid draws from up to the eight nearest monitors inside a trust radius, weighted by inverse squared distance. The population-weighted centroid matters in large rural block groups where the geometric center can sit miles from any resident. Trust radii come from the measured spatial structure of each pollutant: 106.4 kilometers for fine particulates, ozone and coarse particulates, which is the measured practical range of the particulate variogram, and 25 kilometers for nitrogen dioxide, carbon monoxide and sulfur dioxide, which vary block to block and should not be extrapolated.

The score uses the EPA’s own defining-pollutant rule: the worst sub-index across the pollutants with a candidate reading sets the number. Particulates define 138,599 block groups, ozone 103,408, coarse particulates 322, sulfur dioxide 6. Where regulatory monitors do not reach, a second tier of 524 low-cost particulate sensors fills 2,349 block groups, and beyond that a flagged low-confidence nearest-monitor value fills 823.

Anchoring is theoretical rather than stretched to the observed range: an air quality index of 0, meaning pristine, scores 100, and an index of 100, the federal threshold for unhealthy for sensitive groups, scores 0.

Sources. U.S. Environmental Protection Agency, about 2,397 regulatory monitoring sites, plus a private network of low-cost air sensors that fills gaps where the regulatory network does not reach.

Honest limits, stated plainly. Air quality has almost no block-to-block signal, and that is physics rather than laziness. With about 2,397 monitoring sites nationally and a measured correlation range above 100 kilometers, neighborhoods in the same metro read nearly identical because the air is nearly identical. The standard deviation across block groups is 5.7 points against 23 for noise. No interpolator can invent detail the monitor network never sampled, and stretching the scale to manufacture contrast would be a lie about resolution. The practical consequence is that air quality carries well under its nominal weight in any blend, which is the honest outcome.

Quiet

Default weight 55. National median 52.

No other national places-to-live ranking carries a noise measure at all. This one models sound across the entire country at 100-meter resolution, from five federal agencies and nine separate datasets, and delivers it for all 8.1 million census blocks.

What goes in

Federal Highway Administration, Highway Performance Monitoring System. About 21 gigabytes covering every road in all 50 states plus Puerto Rico, carrying annual average daily traffic, separate single-unit and combination truck counts, posted speed, through lanes, functional class from interstate down to local, surface type, and geometry.
U.S. Department of Transportation, Bureau of Transportation Statistics, National Transportation Noise Map. Three 30-meter raster families for road, aviation and rail, plus separately obtained Alaska aviation, Alaska rail and Hawaii aviation coverage.
Federal Aviation Administration. The Terminal Area Forecast for itinerant and local operations by category and based-aircraft counts, runway geometry for orientation, and the Airport Master Record to reconcile identifiers across the three files.
BTS T-100 segment data for supplementary carrier departures, seats and aircraft type.
USGS and MRLC, National Land Cover Database 2021. Land cover, percent impervious surface and percent tree canopy, the canopy layer produced by the USDA Forest Service.
U.S. Census Bureau TIGER/Line boundaries and 2020 tabulation blocks.
EPA’s 1974 Levels Document, which sets 55 A-weighted decibels as the threshold for outdoor activity interference and is the anchor for the published share of area above 55 decibels.

How it is modeled

A source sound level is predicted at 15 meters from each road centerline by a 16-feature model discovered by machine learning against the federal rasters and then deployed in a transparent linear form. It was fitted on 970,319 road-noise samples across Rhode Island, Texas, Florida and California, and tests at R-squared 0.6809 with RMSE 5.51 decibels. A gradient-boosted model reaches 0.7473, which means the transparent linear form captures 91% of the predictable variance while remaining inspectable. Mean bias by road class runs between -0.01 and +0.03 decibels.

Two results from that model are worth quoting because they are the intuition behind the map. Every tenfold increase in traffic volume adds about 4.5 decibels. A fifth of traffic being heavy trucks adds about 1.2. A quiet residential street sits near 55 decibels at the centerline and an interstate near 80.

Sound then decays at 25 decibels per tenfold increase in distance, a coefficient fixed nationally between the roughly 18 of open country and the roughly 30 of dense urban fabric. Propagation stops at a 35-decibel floor, below which road noise is inaudible against ambient sound, and at a hard 8-kilometer cap. That 80-decibel interstate reads about 59 decibels at 100 meters, 48 at 300, and 39 at a kilometer, with an audible glow reaching about 950 meters.

Overlapping sources are summed in the energy domain rather than in decibels, because decibels are logarithmic and adding them arithmetically is meaningless. The same rule governs aggregation up to the block group. Doing it the naive way produces an error of about 9 decibels, which is about 31 score points.

What it is checked against

The model is validated against the federal government’s own noise model rather than against field readings, and that distinction is worth stating clearly: the Bureau of Transportation Statistics rasters are the Department of Transportation’s modeled output, not measured sound-level meter data. There are no field measurements anywhere in this pipeline.

The R-squared of 0.68 also understates the accuracy of the delivered map, because the model is not the only source. Wherever the federal raster is louder than the prediction, the federal value wins. Aviation and rail are taken straight from the federal rasters and carry no model error at all. The synthetic aviation model that fills airports the federal map misses is deliberately conservative, with peaks reduced, extents trimmed 15%, a hard 75-decibel cap, and airports already covered federally skipped entirely. That 0.68 governs only the minor, local and rural roads where this model is the sole source.

The score anchors 40 decibels at 100 and 70 decibels at 0. Coverage is 100% of population with no fallback and no missing data across all 242,335 block groups. The loudest large cities are Manhattan at 61.7 decibels, Boston, Philadelphia and Minneapolis; the quietest are Naples, Lincoln, Bradenton and Anchorage.

Honest limits. Industrial, construction and military aviation noise are not modeled. Silence is a true zero and is never renormalized away.

Full detail is at the transit noise map.

Disaster safety

Default weight 55. National median risk 26.

The headline number FEMA publishes is expected annual loss in dollars, which mechanically overstates dense cities: more buildings means a bigger number even at identical risk. This score uses the rate instead, expected annual building loss divided by building value, which is what a resident actually faces.

The scale is pegged to the highest building-loss rate among stateside metro areas, currently Jacksonville, North Carolina at $609 per $100,000 of building value, driven by coastal hurricane and flood exposure at Camp Lejeune. Pegging to the worst metro rather than the worst census tract avoids letting one coastal sliver define the scale while still driving genuinely disaster-prone places toward 100. A square-root curve rather than a linear one puts the typical place at a moderate 20 to 26 instead of near 3.

FEMA publishes at the census tract, so each tract’s building loss and building value are allocated to its blocks by population share and re-divided after summing, giving a building-value-weighted rate for each place. Nine separate hazard percentiles ride alongside the headline for readers who fear one hazard in particular: hurricane, tornado, riverine flood, coastal flood, wildfire, earthquake, hail, strong wind and winter weather. Drought is excluded because FEMA models it as an agricultural loss with no building component.

Heat wave was tested for double-counting against the separate heat measure and removed from consideration: it is 0.02% of national building loss, and dropping it changes the score at r = 1.0000. The two measures correlate at only 0.45 to 0.53, confirming they measure different things, property damage against heat mortality.

Source. FEMA National Risk Index, expected annual loss by census tract across 18 hazards.

Honest limits. The National Risk Index is a modeled multi-year expected annual loss, not an observed year. Puerto Rico and the territories are scored and mapped but excluded from setting the scale, because they are the most disaster-prone places in the system and would otherwise compress everywhere else; their metro areas clamp at 100.

Household income

Default weight 55. National median 47.3, at a household income of $85,700.

Median household income, log-anchored so that $43,700 scores 0 and $181,400 scores 100. Log spacing rather than linear, because a linear scale bunched cities into the 75 to 95 range and hid the differences that matter.

The weight is moderate rather than heavy on purpose. Income already appears inside the affordability measure and correlates with the local economy measure, and a heavy weight here would count earnings three times.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates. Vintage was proven rather than assumed: all six measures in this family match the live census API at Spearman 1.0000 across four states.

Local economy

Default weight 55. National median 57.

This measures the labor market rather than pay, deliberately in the opposite lane from income and affordability. It blends three factors, each scaled between winsorized fifth and ninety-fifth percentiles:

Job growth, 0.40. Compound annual growth in jobs physically located in the block group, from 2019 to the latest complete year, from Census LEHD workplace data.
Current unemployment, 0.35. The latest trailing-twelve-month county rate from the Bureau of Labor Statistics, plus each block group’s own deviation from its county rate, so neighborhoods differ while the level stays current.
Participation change, 0.25. Change in employed residents per working-age adult from 2019 to the latest year.

A block group missing one factor spreads that weight across the others, but a place with only one factor is left as no-data rather than scored on a single generous input. Data windows are detected per state, which caught Michigan filing employment files that were about 97% incomplete for two years. Alaska has no workplace data at all and reweights onto the other two factors.

The design goal was independence from pay, and it was measured: correlation with the income percentile is 0.21 across 242,000 block groups. The result reads as momentum rather than prestige. San Francisco, New York and San Jose land in the mid-50s despite very low unemployment, because their job and resident-employment growth from 2019 to 2023 was weak. The top metros are the real low-unemployment boomtowns: St George, Bozeman, Provo, Idaho Falls, Fayetteville, Huntsville, Raleigh.

Sources. Census LEHD Origin-Destination Employment Statistics, Bureau of Labor Statistics Local Area Unemployment Statistics, and the American Community Survey table B23025.

Honest limits. Two of the three factors carry block-group texture, but the current unemployment level is county-resolution; within-county differences come from the survey pattern rather than from current filings. Puerto Rico and one remote Alaska place have no data.

Healthcare access and health outlook

Default weights 50 and 25. National medians 53 and 51.

These are two separate measures kept apart on purpose, and the separation is the interesting part. Health outlook covers the current health burden of residents plus their preventable risk factors. Healthcare access covers barriers to getting care, inverted so that more access reads as better. They share zero inputs.

That matters because the obvious composite would double-count. The underlying community health index blends current burden at 0.45, preventable risk at 0.30, and access at 0.25 into one number, and weighting that composite alongside an access measure would count access twice. It is not shipped here at all. As built, a reader can weight both at full strength and an area with less-healthy residents is not penalized a second time for also lacking care. The two genuinely diverge: Seattle scores 48.9 on access and 72.7 on outlook.

Thirty-nine measures feed the three sub-indexes. Current burden covers 18, including diabetes, coronary heart disease, stroke, cancer, chronic obstructive pulmonary disease, asthma, arthritis, depression, self-reported general health, and six functional limitations. Preventable risk covers 7, led by high blood pressure and adult smoking. Access barriers covers 14, including uninsured share, food insecurity, housing insecurity, four cancer and cholesterol screening rates, medication adherence, loneliness, lack of reliable transportation, and utility shut-off threat. Individual weights are justified per measure against global burden-of-disease disability shares, CDC attributable-death counts, and published hazard ratios rather than assigned by feel.

Each measure becomes a directional z-score against the national tract distribution, mapped through a bounded transform so that 50 is the national average. That is a monotone transform of a real magnitude rather than a rank, which is why the resulting distribution is visibly not uniform.

Sources. CDC PLACES as the primary source, with the CDC and ATSDR Environmental Justice Index filling individual missing values, mainly in Pennsylvania and Kentucky where the current PLACES release ships only 5 of 40 measures.

Honest limits. The source is native to the census tract, so every block group inside a tract carries an identical value: 99.0% of multi-block-group tracts have exactly one distinct access value. Health contributes no within-neighborhood discrimination and the map is flat inside a tract. That is the resolution of the data. Where the source ships a blank rather than a number, the block group is left as no-data rather than cast to zero, because a zero would appear on the map as a perfect score; the true minimums are 8.95, 2.73 and 12.84, so zero is never real.

Diversity

Default weight 30. National median 53.1 for cities, 40 for block groups.

The probability, times 100, that two randomly chosen residents belong to different racial or ethnic groups, computed over six groups. The theoretical ceiling is 83.3, an even six-way split; the highest block group observed is 81.4.

This replaced an older index rather than refreshing it. The legacy measure was an effective-number-of-groups index whose percentile encoding was generous enough to make American cities look more diverse than they are. The current score is the raw interaction probability with no stretch applied. Ranks are nearly identical, at Spearman 0.966, but the levels are honest: Queens and Oakland land near 77 rather than at 100.

This is a preference-neutral measure. A reader who prefers a more homogeneous area moves the control the other way and the same stored number is inverted. No mirrored field is published.

Because a place’s score is computed from that place’s own group shares rolled up and then measured, a city-wide score can legitimately exceed the score of its typical neighborhood. A city of homogeneous but different neighborhoods is diverse at the city scale and not at the street scale, and both readings are correct.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates.

Cheap car insurance

Default weight 25. National median 65.6, at a premium of about $1,623.

Annual full-coverage premium for a good adult driver, built from a state base rate and a block-group variance model, then scored 0 to 100 on the logarithm of the premium so that about $1,082 scores 100 and about $3,515 scores 0. It is a cost-of-living factor that national lists leave out, and it varies more than most readers expect: the spread across cities has an interquartile range of nearly 25 score points.

The rollup reproduces the source’s own independently computed place-level dollars at Pearson 1.000 with no level bias across 52,877 cities, 33,110 ZIP codes and 3,131 counties.

Source. CoverageWatch, which collects and publishes the underlying state and neighborhood premium data.

Honest limits. The state base rate is the dominant term, so this separates states strongly and adds a neighborhood gradient within a metro, with urban cores rating worse, which is consistent with how insurers actually price. Puerto Rico and the territories have no data.

Low property tax

Default weight 25. National median 73.9, at an effective rate of 0.751%.

This scores the effective rate, real estate taxes paid divided by home value, rather than the dollar bill. That choice is what keeps it from restating housing cost. Honolulu is the proof: high dollars at $2,564 a year, but a bottom-decile rate of 0.310%, scoring 89.9.

The rate is computed from aggregate taxes paid over aggregate home value, not from the published median-dollar table, because that table tops out at $10,001 a year and hits that ceiling in 4,502 census tracts. Aggregates are not top-coded. The median dollar figure still ships as a display value and is never used to rank.

Scoring is linear in the rate, since rates run roughly symmetrically between 0.2% and 2.5% and need no log transform, anchored between population-weighted city percentiles at 0.033% and 2.786%. The national population-weighted mean rate is 0.989%. Highest rates are New Jersey at 2.00%, Illinois 1.87%, Connecticut 1.76%; lowest are Hawaii 0.29%, Alabama 0.35%, Nevada 0.48%.

When the source moved from the 2023 to the 2024 survey vintage, county-level rates correlated at r = 0.99 while falling 2 to 3 percent nationally, because home values outran tax levies. That is a re-leveling rather than a re-scoring, and every published constant moved with it.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates, tables B25090, B25082 and B25103.

Honest limits. 4,034 of 85,382 tracts lack a published rate and inherit their county’s, then their state’s, flagged; 940 of 95,571 places are marked as mostly imputed. The score is a population-weighted mean of tract rates, which is the rate the typical resident faces rather than the dollar-weighted aggregate rate.

Low income, sales and gas tax

Default weight 25. National median 43.2, at an effective rate of 6.64%.

This is the most source-intensive measure on the site: roughly 40 separate issuing agencies, and 3,855 hand-verified local income tax rate rows covering 15 of the 16 states that levy one. It answers a question no ranking answers well, which is what a household actually hands over in income, sales and gasoline tax before property tax.

Ten scenarios per state, not one

The same national household is run through every jurisdiction’s current law, deliberately, so the result measures tax policy rather than local incomes or local prices. But a single household was measurably blind in both directions. Scoring one median single filer put Portland’s two local income surtaxes at zero, because neither begins below $125,000, and never saw that Tennessee’s roughly 9.5% combined sales tax is a 9.1% effective rate on a bottom-quintile household.

So the law is evaluated at five national household-income quintiles, each through both the single and the married-filing-jointly schedules. Ten runs per state, 520 state rows and 290 local rows. Income at each quintile comes from the census, and the consumption base at each quintile is built line by line from the Consumer Expenditure Survey, covering food away from home, alcohol, utilities, household operations and furnishings, apparel, vehicle purchase and repair, public transportation, entertainment, personal care, reading, tobacco and miscellaneous. Groceries and gasoline are excluded from that base because they carry their own rates, and shelter, health care, education, insurance and non-consumption items are excluded outright.

Income tax applies marginal brackets after the standard deduction and personal exemption, with credits applied after tax. Nine states levy no wage income tax. Alabama and Oregon get federal deductibility computed against the real current federal schedule rather than a hardcoded constant. Sales tax applies the combined state and average-local rate to the consumption base plus a separate grocery rate where a state taxes food. Gasoline tax applies the state cents-per-gallon rate to a modeled gallon count.

Both survey inputs are two years older than the tax law being modeled, and the brackets are inflation-indexed, so running old nominal dollars through current brackets would make every household look about 5% poorer than the law assumes. That is a one-directional bias rather than noise, so income, consumption base and grocery spending are lifted by the measured consumer price change. Gallons are deliberately not lifted. Gallons are a physical quantity, not a nominal one, and gasoline demand is inelastic enough that a 10% price move changes consumption by about 2%. The pump price never enters the tax at all, because gasoline excise is levied per gallon; the price is only a deflator, and the correct deflator is the one matching the year of the spending figure.

Local income tax, and how it is joined to geography

Three states levy at a grain finer than a county and not expressible as a city: Ohio taxes at the incorporated place and again at the unified school district, Pennsylvania at the municipality, and Iowa at the school district. Those are resolved by reading the census Block Assignment Files directly, producing place, minor-civil-division and school-district identifiers for 788,612 Ohio, Pennsylvania and Iowa blocks. Every other state rides a county code or a city.

The join was the hard part and each piece is deterministic rather than approximate. Ohio’s municipal file keys directly to census place codes for 673 of 678 rows. Ohio school districts are never matched by name, because the state has three separate districts called Buckeye Local and both a Canton City and a Canton Local; instead the state education identifier routes through the federal education directory to the census school-district code, matching 214 of 214. Iowa uses the same route. Pennsylvania is the only source with no census key at all, so it needs name matching on county and municipality type, hitting 2,626 of 2,627, with the single miss a genuine spelling divergence handled by an override.

Combination follows two different rules, and getting them backward would be a real error. Within a level of municipal taxation the most specific key wins, so a Philadelphia block owes the municipal rate rather than that plus a county rate. Across tax types the rule is addition, because an Ohio household genuinely owes both a municipal and a school-district tax and an Oregon household genuinely owes two separate county-level surtaxes. An unrecognized schedule now raises an error rather than resolving to zero, because a named schedule with a null rate had previously evaluated silently to nothing, which is exactly the failure mode by which Oregon’s two surtaxes could have been modeled as free.

Scoring, and what it turned up

The score is linear in the mean effective rate across the five quintiles, equal-weighted because each quintile is a fifth of households, anchored between 0.81% in Alaska and 11.09% in New York City. Averaging dollars instead would be dominated by the top quintile’s bill, and progressive states would be punished by an arithmetic artifact rather than by their policy. The headline dollar figure published alongside it is a single filer at the middle quintile, so the number a reader sees stays interpretable.

The per-quintile shapes behave the way tax theory predicts, which is its own check: Tennessee runs from 33.0 at the bottom quintile to 84.3 at the top, steeply regressive; Oregon runs the other way, 75.3 down to 14.6; New Hampshire scores 100 at every level. California at the fourth quintile owes $12,191 filing single against $8,076 filing jointly, which is why both filing tracks ship.

Assembling this from primary sources turned up several errors in widely republished figures, which is the strongest available evidence the work is original rather than copied:

Indiana’s maximum local income tax is 3.00% in Randolph County, not the 3.38% circulating on calculator sites, which is Pulaski County’s 2017 rate. Pulaski is now 2.85%.
Saginaw, Michigan is 1.5%, not the 1.0% in a widely copied municipal reference table. Michigan has 24 income-tax cities, and one city that appears on several third-party lists levies none.
Jefferson County, Alabama has no occupational tax; it was struck down by the Alabama Supreme Court in 2011. A U.S. Treasury appendix still lists one.
Aurora, Colorado repealed its occupational privilege tax effective January 2025, so Colorado has four such cities rather than five.
Opelika, Alabama cut from 1.5% to 1.0% in April 2025, and both the state municipal league table and the city’s own business page still show the old rate.

Kentucky’s roughly 200 uncentralized occupational licenses are out of scope, and that was verified to evaluate to exactly zero across all 1,886 Kentucky places rather than assumed.

The national population-weighted average bill for a single middle-quintile filer is $5,318 a year. Lowest: New Hampshire $169, Alaska $553, Wyoming $1,619, South Dakota $2,115, Florida $2,162. Highest: Maryland $8,186, New York $7,817, Illinois $7,095, California $6,612.

Sources. Rates and scenarios come from the Tax Foundation, the U.S. Energy Information Administration, the U.S. Census Bureau, the Bureau of Labor Statistics, the Federal Reserve Bank of St. Louis, the Internal Revenue Service, and the National Center for Education Statistics. Local rates come from the revenue and taxation agencies of each levying jurisdiction, all of which are listed in the source table at the bottom of this page.

Honest limits. Household-income quintiles are run through individual filing schedules, so the single track at the top quintile is somewhat artificial and the joint track overstates tax for the many one-person households at the bottom. Both ship. The two surveys define a household slightly differently, about 10% apart at the middle quintile. A place footprint is not a city hall boundary: Pittsburgh’s true municipal rate is pinned at 3.0%, but only 62.5% of the place’s residents live inside the city, so the place-wide rate reads 2.34%. Employer-side payroll taxes are out of scope. This measure is scored and ranked but is not currently published as a map layer.

Educational attainment

Default weight 25. National median 44.8.

The educational attainment of adults 25 and over, as an equal-weight blend of three separately anchored shares: high school completion or better, bachelor’s degree or better, and graduate degree or better. Neutral sits at 72.5%, 35.6% and 26.4% respectively.

This is a measure of the adult population and is not a substitute for the school grade. A college town with weak public schools scores high here and low there, and both readings are correct.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates, table B15003.

Internet speed

Default weight 25. National median 75 for cities, 82 for block groups.

The fastest advertised download speed available at each block, rolled up population-weighted, scored on a logarithmic curve where no service scores 0 and a 10 gigabit advertised tier scores 100. The federal broadband benchmark of 100 megabits lands at 50.

Anchoring is theoretical rather than stretched to the observed distribution, and that fixed a real problem. The previous percentile-based floor pinned 7,091 cities and 3.37 million people at exactly zero. Under the theoretical anchor no city, block group or metro scores zero, and those 7,081 recovered cities now span 369 distinct scores from 26.0 to 65.4.

Source. FCC National Broadband Map and the Broadband Data Collection.

Honest limits, and this one is significant. This measures advertised marketing tiers, not delivered speed, because the federal collection is advertised-only by statute. The practical consequence shows up in face validity: Enid, Oklahoma reads as the second-fastest metro in America. Dispersion is genuinely low, at a standard deviation of 8.8 points across block groups, because 315.7 million of 334.7 million Americans live in blocks advertising gigabit service or better, so the measure carries roughly half its nominal weight in any blend. Treat it as a floor check for underserved areas rather than as a quality ranking. Measuring real throughput requires a different source and is a known open item.

Cell coverage

Default weight 25. National median 90.

Advertised coverage is near-universal 5G, so a score built on advertised technology alone sits at about 97 for most of the populated country and produces a flat, useless map. Two things fix that here.

First, each carrier’s advertised tier is scaled by modeled signal strength at that location, running from full credit at -85 dBm down to 40% credit at -115 dBm, because real throughput tracks signal strength rather than the label on the tier. That single change took the block-group standard deviation across the Los Angeles basin from effectively zero to 6.6.

Second, carrier choice counts. The score is 60% the best available carrier of any kind, 30% the second-best national carrier, and 10% the third. One carrier at top service scores 60, a second national carrier takes it to 90, and all three reach 100. The primary slot accepts any carrier, so a rural area on a strong regional network is not zeroed out, while the redundancy bonus is restricted to the three national carriers, which is the choice set almost everyone can actually buy.

Each block is placed at its population-weighted centroid and matched to its coverage hexagon, about a tenth of a square kilometer. No coverage from anyone is a true zero and covers 2.5% of blocks.

Source. FCC National Broadband Map, Mobile Broadband hexagon coverage.

Honest limits. Coverage is what carriers claim they can deliver. This measure mainly separates the underserved tail and is a weak differentiator among large cities by design, most of which land between 90 and 96. The signal scaling thresholds are an editorial choice, documented and tunable. A block inherits the coverage of the single hexagon containing its population centroid.

Big-city access

Default weight 20. National median 57.

How much genuine urban fabric is within reach, which is a proxy for access to the jobs, airports, specialty hospitals and cultural institutions that only cluster at density. It exists because a ranking without it puts isolated small metros above places with real cities nearby, on the strength of cheap housing and short commutes.

The first version measured reach to metro population, and that turned out to be the wrong quantity. A metro area’s population count includes whatever farmland the delineation happened to fold in. The current version measures reach to dense land: square miles of blocks at 15,000 or more people per square mile, decayed exponentially with a 50-kilometer scale and summed across all 393 metro areas.

Two construction choices were tested rather than assumed. The 15,000 threshold, roughly 23 people per acre or rowhouse-to-midrise fabric, came from a sweep: a mid-size real city holds its rank across the entire threshold range because it has real urban fabric and just less of it, while a place whose density tops out at suburban falls off a cliff between 5,000 and 20,000. It is not an East Coast artifact either, since Austin holds 9.3 square miles above the threshold, Denver 20.9, Seattle 26.8 and San Jose 28.9. And blocks dominated by group quarters are excluded, because at a high density threshold a dormitory, prison or barracks looks exactly like an urban core. Nationally that is 4.73% of dense land, but locally it is decisive: 57.9% of Ames, Iowa’s dense land, 57.0% of Pensacola’s, 47.8% of Blacksburg’s. The tell was college towns rising as the density threshold rose.

Adjacent cores stack on purpose. Manchester, New Hampshire borrows 89.9% of its score from Boston, Appleton 91.9% from the Fox Valley and Milwaukee corridor, and Cheyenne 93.3% from the Front Range. Borrowing access from a nearby city is the property being measured. No dense land within reach is a true zero and means it.

Source. U.S. Census Bureau 2020 Decennial Census block population, land area and group-quarters counts, with TIGER/Line geometry. Entirely self-contained, with no external feed that can go stale.

Heat safety

Default weight 20.

This is not a temperature score. It is a modeled estimate of excess heat deaths per 100,000 residents, and it is built that way because the thing a reader cares about is whether the heat where they live actually kills people.

Heat exposure is five measurements, not one

Peak temperature alone is a poor predictor of heat mortality. What kills is intensity, duration and the absence of overnight recovery, so exposure is a fixed-weight blend of five components computed from ten years of daily station observations, 2016 through 2025:

Component	Weight	What it counts
Heat intensity	0.45	Annual sum of degrees above a 90F afternoon heat index, the National Weather Service extreme caution threshold
Unusual for here	0.15	Days above this station’s own ten-year 95th percentile and at least 80F, which is the acclimatization effect
Runs of consecutive hot days	0.15	Days that are the third or later day of a 90F-plus run, when the body’s buffer is gone
Nights that do not cool down	0.15	Nights with a minimum at or above 75F, meaning no overnight relief
Extreme intensity	0.10	Days at or above 103F, the National Weather Service danger threshold

The heat index itself is recomputed rather than taken from a general weather file. The usual approach evaluates the heat index at the day’s average relative humidity, which inflates humid places because humid nights drag the daily mean humidity up. Here it is evaluated at the daily maximum temperature with dewpoint held at its daily mean, since dewpoint is roughly conserved through the day and that produces realistic low afternoon humidity. The result represents dry heat in Phoenix and humid heat on the Gulf Coast correctly, and the two are genuinely different hazards.

Two populations, because the deaths are not evenly distributed

The model estimates deaths separately for the housed population, scaled by exposure and by a vulnerability multiplier, and for the unsheltered population, scaled by exposure alone at a far higher rate. That split exists because 49% of Maricopa County’s heat decedents were unhoused, from under 1% of the population. A model without it would misplace both the level and the geography.

The vulnerability multiplier is built from residents aged 65 and over, cardiovascular and chronic disease prevalence, disability, substance use, and urban heat island intensity from low tree canopy and high impervious surface, with exponents matched to the actual decedent profile in the one county that documents it thoroughly: 60% aged 50 and over, 47% with a cardiovascular comorbidity.

What it is calibrated against

The model has three free parameters, and they are fitted against the four places in the country that count heat deaths carefully enough to be treated as ground truth, all of them local or state health departments rather than the national death-certificate file:

Place	Model	Reported	Reporting body
Maricopa County, Arizona	696	620	Maricopa County Department of Public Health heat surveillance
Clark County, Nevada	354	400	Southern Nevada Health District
Arizona statewide	1,093	980	Arizona Department of Health Services
New York City	316	350	NYC Department of Health and Mental Hygiene

The fit runs about 10 to 12% above the attributed benchmarks by design, because attributed counts are themselves a floor. Miami-Dade and the New York boroughs are suppressed in the national file precisely because they under-report, which is why their benchmarks come from local health departments instead.

The national number, and why it is so far above the official count

Summed nationally the model produces about 17,700 heat deaths a year. Official death certificates record roughly 1,500 to 2,400. That gap is not a modeling error, and two independent checks say so.

The first is seasonality. Month-by-month, the model and the national death file agree closely: both peak in July, at 34.1% of annual deaths against 40.5%, and both place 96 to 98% of deaths between May and September. A model that was merely too large would not reproduce the shape of the year.

The second is that the model was calibrated bottom-up on four local benchmarks and was never tuned to hit a national figure, yet it lands between the two leading published estimates built by an entirely different method. Shindell and colleagues put current-decade heat mortality at 12,000 deaths a year across the contiguous United States, with a confidence interval of 7,400 to 16,500, extrapolating exposure-response relationships measured in ten cities (Shindell et al., 2020, GeoHealth). The 2023 Lancet Countdown U.S. policy brief estimates 23,200 heat-related deaths in 2022 among adults 65 and older alone. A third study, restricted to 297 counties covering 62% of the population between 1997 and 2006, estimates 5,608 a year within that footprint (Weinberger et al., 2020, Environmental Epidemiology).

Those studies count all-cause excess mortality above each location’s optimal temperature, which is why they all run roughly an order of magnitude above the death-certificate count, and they define their populations differently, so they are a bracket rather than a single target. Two approaches sharing no inputs, built for different purposes, landing in the same range is convergent evidence rather than a coincidence.

The ten-year back-cast is consistent with the gap closing rather than the model drifting. Modeled deaths rose from about 14,600 in 2016 to about 20,000 in 2024, roughly a third each from hotter weather, from a 51% rise in unsheltered homelessness, and from the national share of residents 65 and over rising from 14.5% to 17.3%. Official counts grew about 2.4 times over the same period against the model’s 1.27, which means a quarter to a third of the official increase is real additional deaths and the rest is the death-certificate system counting better. The implied undercount narrowed from roughly 16-fold to roughly 8-fold.

Sources. NOAA NCEI Global Summary of the Day, CDC and ATSDR Environmental Justice Index 2024, CDC PLACES, HUD Point-in-Time homeless counts, CDC WONDER heat-related mortality, Census age structure, and USGS and MRLC canopy and impervious surface for the heat island term.

Honest limits. The estimate is a typical recent year rather than any specific one; 2023 and 2024 both run higher. A score of zero means what it says, no estimated heat deaths at all, and only 83 block groups reach it, all in Alaska. Cold places are low-risk rather than zero-risk, and the difference is shown rather than rounded away. The heat index formula is unreliable above 130F and is capped there. The homeless-count crosswalk is several years old and drops about 4% of counted unsheltered people. The national death file suppresses county cells below ten deaths.

Housing crash safety

Default weight 20.

Anyone can assert that a price-to-income ratio looks high. This measure instead asks which pre-crash traits actually predicted the last decline, learns them from what happened, and applies them to current data.

The outcome trained against is the real peak-to-trough fall in home values for every ZIP code between 2006 and 2012, mapped to census tracts. Ten factors survived a 22-candidate screen and were weighted by non-negative least squares: price run-up at 0.21, owner cost burden 0.15, renter cost burden 0.14, vacancy 0.12, speculative vacancy 0.09, junior-lien share 0.09, price-to-income 0.06, price-to-rent 0.06, boom-era construction share 0.05, and mortgaged share 0.03.

Run-up carrying the most weight is what makes this a bubble score rather than a decline score. Detroit fell about as hard as Las Vegas in 2008, but Las Vegas had run up 120% first and Detroit only 55%. Only one of those was a bubble.

Factors are converted to percentiles within their own cross-section, so the 2008 fit ranks tracts against 2008 peers and the current score ranks block groups against current peers, which is what lets the weights transfer as a statement about being frothy relative to your era.

The surroundings dominate the block group itself. A place’s roughly one-mile ring predicts its crash better than its own numbers do, so each factor blends 35% own value with 65% ring. Vacancy leans harder on the ring still, at 15% own, because block-group vacancy is the noisiest factor and often reflects new-construction lease-up rather than overhang. Each block group’s own share also scales down with its population, so thin or suppressed block groups lean on their surroundings rather than on a small sample.

Validated by spatial cross-validation holding out whole metro areas, the transparent index predicts the 2008 decline at Spearman 0.63, against a 0.74 gradient-boosted ceiling and 0.56 for run-up alone. It puts the canonical 2008 bubble metros on top and the steady ones at the bottom without being told which were which.

The scale is pegged so that 100 is as stretched as the worst 2006 metro, meaning a predicted 47% decline. Higher scores mean safer.

Sources. U.S. Census Bureau American Community Survey, Zillow Home Value Index for the crash outcome, Federal Reserve Bank of St. Louis for the national price, mortgage and inflation series, and the S&P CoreLogic Case-Shiller national index for dating the crash window.

Honest limits. The national median block group sits near 62, and that is real rather than a scaling bug: current price-to-income sits at or above 2006 levels, the recent run-up rivals 2000 to 2006, and cost burden is high. The percentile column is the better tool for comparing places against each other. The outcome is ZIP-resolution mapped down to tracts, and run-up is uniform within a ZIP code.

Low vacancy

Default weight 15. National median 77.9, at 7.0% vacant.

The share of housing units that are vacant, inverted, anchored so 1.7% scores 100 and 25.5% scores 0. High vacancy can mean blight or it can mean a seasonal resort town, so the weight is light and the reader is left to interpret it. Naples, Florida scores near the bottom for seasonal reasons rather than distressed ones.

It is computed here from the source table rather than taken from an existing database field, because that field silently fills block groups whose vacant count is zero, undocumented, affecting about a fifth of block groups.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates, table B25002.

Voter turnout

Default weight 15. National median 69.

Votes cast as a share of voting-age population, modeled to the block group from certified precinct returns. It reads as civic engagement, and higher is treated as better. It already is a percentage with natural anchors at 0 and 100, so no re-anchoring is applied. Madison, Wisconsin is highest at 80.4.

A non-voter rate is deliberately not published, because it is exactly 100 minus this number, and shipping both would let a reader weight both and silently cancel them out.

The modeling approach is described under political lean below, since both fields come from the same model.

State infrastructure

Default weight 10.

A state-level measure, which means it paints hard borders at state lines. That is honest rather than a defect: the things it measures genuinely are set at the state level. The weight is light because local factors matter more to most readers.

Six factors combine into one score: grid reliability at 26%, drinking water safety at 24%, road quality at 20%, bridge quality at 12%, waterway quality at 12%, and clean energy at 6%. Each is scaled between winsorized fifth and ninety-fifth percentiles so one outlier cannot compress everyone else, and a state missing a factor reweights across the rest rather than taking a zero. The scale leaves headroom by design and nothing scores 100.

Grid reliability uses outage minutes per year excluding major event days, which isolates baseline grid quality from storm exposure that the disaster measure already covers.

Drinking water safety is the most detailed of the six. It counts only active community water systems, the ones serving residents year-round, excluding transient systems like highway rest stops. A system is affected if it recorded a health-based violation beginning in 2021 or later, a recent window so that large states are not punished for decades of history. Each affected system is then classified by its worst violation using the EPA’s legally defined public notification tiers: acute violations requiring 24-hour notice, such as E. coli or acute nitrate, are about 13% of health-based violations, and non-acute violations requiring 30-day notice, such as arsenic, lead or disinfection byproducts, are about 86%. Systems are rolled to the state by population served, and acute exposure counts at full weight against 0.4 for non-acute. Severity genuinely changes the answer: Wisconsin and Pennsylvania score well despite a high affected share because nearly all of it is non-acute, while Arizona and Maryland score worse than their raw share implies.

Waterway quality counts the share of active dischargers in significant noncompliance under the EPA’s own definition, covering effluent limit exceedances, schedule violations more than 30 or 90 days late, and failure to report.

The two water measures are not redundant, which is checkable because they disagree. Arizona scores 99.6 on waterways and 69 on drinking water. Nevada scores 99 on drinking water and 85 on waterways. Puerto Rico scores 35 on drinking water and 92 on waterways, because its problem is home distribution rather than industrial discharge.

Sources. U.S. Energy Information Administration Form EIA-861 for grid reliability and generation, EPA Enforcement and Compliance History Online for both Safe Drinking Water Act and Clean Water Act discharge records, Federal Highway Administration highway statistics, and the National Bridge Inventory.

Honest limits. Puerto Rico has no road or clean-energy data and Washington DC has no road data, so both lean harder on the factors they do have and should be read as approximate. The 1.0 against 0.4 severity weighting for drinking water is the one judgment call in the build, and the underlying acute and non-acute population counts are published alongside the score so it can be checked.

State fiscal health

Default weight 10.

Whether a state can pay its bills through the next downturn without a tax rise or a service cut. Seven factors, again winsorized and reweighted rather than zero-filled: pension funded ratio at 23%, budget balance at 20%, reserves at 20%, debt burden at 20%, pension contribution discipline at 7%, long-run structural balance at 5%, and unemployment insurance trust fund solvency at 5%. Near-term factors carry 40% because they are what hits residents in the next recession; pensions carry 30% as the largest long-term claim on future budgets.

Two construction choices are worth stating because they change the answer. Budget balance counts positive revenue adjustments as resources, since Alaska’s Permanent Fund draw genuinely is its recurring funding, but does not subtract negative ones, since a surplus swept into reserves is surplus disposition rather than reduced revenue. A plain revenue-minus-expenditure calculation would score Alaska at minus 52% a year, which describes nothing real. Pensions aggregate from plan to state by liability weight rather than by simple average, so a small well-funded plan cannot offset a large underfunded one.

Validation is the strongest external check on this site. Against Moody’s January 2026 state ratings the score correlates at Spearman 0.54, and the mean score by rating tier is perfectly ordered: Aaa 58.6, Aa1 49.2, Aa2 38.1, Aa3 and A2 around 14. The two lowest-scoring states here are Moody’s two lowest. That agreement is reached without ever reading a credit rating.

The divergence is by design and is also informative. Moody’s scorecard is roughly half economy and governance, which this measure deliberately excludes because the composite covers economy elsewhere, and ratings lag published actuals. Maryland is the clean example: rated above where this score puts it, then downgraded in 2025.

Sources. National Association of State Budget Officers Fiscal Survey of the States, the Public Plans Database at the Center for Retirement Research at Boston College, Equable Institute for a fresher pension vintage, Census Annual Survey of State Government Finances, Bureau of Economic Analysis state personal income, and the U.S. Department of Labor unemployment insurance trust fund solvency report.

Deliberate exclusions. Credit ratings, because they are a redundant composite of the factors above expressed in letters. Tax burden, because that is cost of living rather than fiscal responsibility. Advocacy-organization taxpayer-burden composites, because they double-count debt and pensions that are held separately here. Federal funds dependency, because it mechanically penalizes poor states. Retiree health care obligations were dropped in July 2026 because no current machine-readable 50-state source exists.

Honest limits. Washington DC has five of seven factors and Puerto Rico three, both reweighted and both approximate. Nevada reports no pension contribution percentage, so that weight redistributes onto its stronger factors and its rank is probably a touch generous. Metro areas crossing state lines are population-weighted.

Tree shade

Default weight 5. National median 42.

Canopy cover measured three ways and blended, because area-average canopy misses the thing people actually experience. General canopy is a population-weighted mean of block canopy. Walkway canopy is a length-weighted mean over every walkable edge in the country, sampled at edge midpoints. Sidewalk canopy is the same restricted to sidewalks and footpaths. The walkway blend weights sidewalks at 0.6, and the final score weights general canopy at 0.6 against that walkway blend at 0.4, so trees over the routes people walk count for more than trees in general.

The scale is a ratio to a fixed anchor rather than a rank: 100 is as shaded as a top-1% American city, currently 67.4% canopy. Using the 99th percentile rather than the maximum stops one small forest-edge town from compressing everyone else. Because it is a ratio rather than a percentile, a bare desert city scores near 0 and stays there. Across roughly 52,000 cities the score runs from 0 through a median of 42 to 100, with 31% of cities at or below 20.

Cross-checked against the canopy computed independently for the noise model, the mean absolute difference is 0.07 points with 93% within 0.15.

The weight is light and the measure is off by default in some views because readers genuinely differ on whether they want trees.

Sources. USGS and MRLC National Land Cover Database Tree Canopy Cover 2021, produced by the USDA Forest Service, sampled along the Overture Maps pedestrian network.

Honest limits. The canopy raster covers the contiguous 48 states and Washington DC only. Alaska, Hawaii, Puerto Rico and the territories are published as no-data rather than as a fabricated match, and are excluded from the national anchor.

Measures that ship turned off

Six measures carry a default weight of zero. They are fully built and fully documented; they are off because they encode a personal preference rather than a general statement about a place being better. A reader who wants one turns it on and the ranking and the map redraw.

Close to the ocean

Default weight 0. National median 0.

Drive time to salt water, not straight-line distance. The coastline is converted to about 1.4 million points at 300-meter spacing, then every block within 130 kilometers is routed over the federal road network at posted speed limits. Every road node within 800 meters of the coast is treated as ocean access, tied together by a zero-cost virtual source, and a single shortest-path pass times the whole coastal network out to 75 minutes.

The road graph needed one specific fix worth mentioning, because it is the kind of thing that quietly ruins a result. The federal road file is route-based rather than topological, so crossing routes do not automatically share a node. Vertices are quantized onto a roughly 92-meter grid to force intersections to connect. At a finer 37-meter grid the network shattered into fragments and coastal cities came out averaging bogus 75-minute drives to water they can see.

The score decays logarithmically: touching the water scores 100, five minutes scores 76, ten minutes 59, twenty minutes 35, and 45 minutes or more scores 0. Tidal bays and sounds count, so Seattle sits 0.4 miles from counted coast and Annapolis 1.1. The Great Lakes do not count, because this measures the ocean rather than large water.

It defaults to zero weight because proximity to an ocean is a preference rather than an editorial claim that a place is better. Left on, it was quietly promoting coastal metros over comparable inland ones.

Sources. U.S. Census Bureau TIGER/Line 2020 coastline and the Federal Highway Administration Highway Performance Monitoring System road network.

Honest limits. Most of the country is a true 0 and the national median is 0. That is a measurement, not missing data. Speed limits ignore congestion, so these are free-flow estimates suitable for a decay score rather than for a commute forecast.

Built-up surroundings

Default weight 0. National median 65.

A three-way control reading in town, no preference, or out in the country. The measurement is population within one mile of the block group’s center divided by the land area inside that circle, on a logarithmic ramp.

Using a one-mile circle rather than the block group’s own density is measured rather than assumed. Checked against the 2020 census urban and rural classification as ground truth, the share of block groups that a threshold would misclassify falls from 8.6% using the block group’s own density to 0.9% using the one-mile circle. A block group that is one apartment complex surrounded by fields reads dense on its own numbers and correctly not-dense on the circle.

It is a rate rather than an amount, and that was tested too. The version parallel to big-city access, counting square miles of dense land within a mile, zeroes out 38.3% of all block groups, which ties the entire rural half of the country and destroys any ordering within it.

The zero default weight is also measured rather than assumed. Holding big-city access fixed and sweeping this weight from 0 to 30 moves no metro area more than one rank, because 68.8% of its variance sits inside metro areas: every metro rolls up to between 48 and 62, since a metro’s residents live in its built-up parts wherever the metro is. It does not even add map texture, since the median within-metro spread of the blended score falls slightly as the weight rises. It earns its place only when a reader turns it up, which is what a control is for.

It is not a restatement of walkability either. Correlations at the block group are 0.66 against walkability, 0.64 against transit, 0.73 against biking, and 0.60 against big-city access. Density predicting walkability is the expected relationship rather than a duplicate.

Source. U.S. Census Bureau 2020 Decennial Census block population and land area.

Cheap childcare

Default weight 0. National median 59.3, at $817 a month.

Monthly price of center-based childcare for one child, averaged across infant, toddler and preschool care and adjusted to current dollars. Scored on the logarithm of the price between $382 and $2,473 a month. The national population-weighted average is $1,075. Arlington, Virginia at $2,495 and San Francisco at $2,439 sit at the expensive end; El Paso at $564 sits near the cheap end.

236 counties have no published price, including all of Indiana and New Mexico and most of Missouri. Those are imputed from county median household income through a national regression, plus a state-level offset where the state has at least three priced counties, reaching an R-squared of about 0.76. The imputed share rolls up population-weighted, and a place is flagged when more than half its population sits on imputed values, which is 6,127 of 95,571 places.

Source. U.S. Department of Labor, Women’s Bureau National Database of Childcare Prices.

Honest limits. The source is county-grain, so every block group in a county carries one value. That is fine for choosing a city and honestly weak as a map layer. Values are per child rather than a family budget.

Newer or older homes

Default weight 0. National median 54.5, at a median year built of 1980.

Median year of construction, anchored so 1951 scores 0 and 2004.6 scores 100. This is a preference-neutral control, so a reader who wants prewar housing moves it the other way and the same stored number is inverted. No mirrored field is published. The poles are Frisco, Texas at 2008 against Cleveland and Detroit at 1948.

Source. U.S. Census Bureau American Community Survey 2020 to 2024 five-year estimates.

Political lean

Default weight 0.

Some people want neighbors who vote the way they do, some want the opposite, and plenty do not care. This is published as a signed value running from about -100 for strongly Democratic to about +100 for strongly Republican, and the reader picks the direction.

Precincts are the finest geography at which results are certified, and precincts do not line up with census block groups: a single precinct typically spans four or five of them. So the returns are treated as a modeling problem rather than as a lookup. Certified results are matched to census geography, and the matched rows become training data for an ensemble model over roughly 14,900 candidate demographic, health, environmental and land-cover features per block group.

The honest test is predicting states the model has never seen, so validation holds out whole states. On that test lean reaches R-squared 0.867 with a mean absolute error of 11.6 points, and turnout reaches R-squared 0.390. Within-state ordering is close to perfect nearly everywhere: excluding Washington DC, the nine weakest states still rank their own areas at Spearman 0.947 to 0.969.

Political lean deliberately ships no percentile and no single-color map ramp. A percentile is defined on a good-to-bad scale, and lean has no better end. Publishing one would smuggle in a direction the reader is supposed to choose. The map uses a diverging blue-to-red ramp on the raw value instead.

Sources. Certified precinct-level results from the Secretaries of State and Divisions of Elections of all 50 states and Washington DC, with predictors from the U.S. Census Bureau, CDC PLACES, the CDC and ATSDR Environmental Justice Index, USGS and MRLC land cover, the PRISM Climate Group and NOAA NCEI for election-day weather, the MIT Election Data and Science Lab for vote-mode shares, and the National Conference of State Legislatures for election administration rules.

Honest limits, and this one is disclosed on the pages themselves. Six states are excluded from training: Alaska has no precinct-level data published, and Connecticut, Maine, New Hampshire, Rhode Island and Vermont report by town rather than by precinct. Their numbers come from demographic and health features rather than from local ground truth, and the per-place pages say so. Hawaii carries an additional caveat because two of its predictor sets are imputed. Because most of the country’s 53,000 places are small rural towns that lean Republican while the median person’s block group is close to neutral, read the block-group number for where people live and the city number for what most towns look like.

Full detail is at the political spectrum map.

Religiosity

Default weight 0. National median 37.

Whether a community is churchgoing or secular changes daily life, and no other site maps it to small areas. This runs from 0 for fully secular to 100 for about as religious as anywhere in America is on every axis at once. It is preference-neutral: a reader who wants a secular area moves the control the other way and the same stored number is inverted.

The inputs are a proprietary religious signal measurement. The construction is described as far as possible.

Places of worship are verified against building footprints

A registered congregation can be a post office box or a basement. To count only real places of worship, every candidate has to survive a footprint test. Points of interest are kept only where the taxonomy resolves to a place of worship, which correctly drops religious schools and religious goods stores. Each surviving point is matched to the largest building footprint within 25 meters, which recovers offset map pins and picks the sanctuary rather than an adjacent shed, with one place of worship per footprint so nothing is double counted.

The minimum footprint of 250 square meters is calibrated rather than chosen. Sweeping the cut from 100 to 400 square meters moves the national surviving count from 435,502 down to 290,852, and 250 is the point at which that count lands within a few percent of an independent national benchmark. Every surviving building keeps its footprint area, so the threshold can be retuned later without rescanning the country.

That footprint signal validates against observed truth at Pearson 0.953 and Spearman 0.966, which is a near-perfect independent geographic check on a signal built from an entirely different kind of data.

Model-based refinement

After signal measurement, 21 predictors were found to be valid in a sensible direction. Every demographic predictor enters as a blend of the block group’s own value at 0.45 and its five-mile ring at 0.55, so the model never leans on a single noisy block group. Validation again holds out whole states, and the ordering is informative: the footprint signal alone reaches Spearman 0.448, demographics alone reach 0.306, the combined constrained model reaches 0.484, and an unconstrained model reaches 0.541. Verified places of worship carry most of the accuracy, demographics add texture, and the observed anchor carries what neither can.

The composite

Four components, each anchored to its population-weighted first and ninety-ninth percentile: adherent share at 55%, verified places of worship per 1,000 residents within five miles at 15%, the same within one mile at 10%, and observance intensity at 20%.

Observance intensity exists because a once-a-year adherent and a weekly attender count identically in any headline adherence figure. It reweights the local mix of traditions by typical weekly attendance rates, running roughly 0.18 to 0.85 across traditions.

Face validity runs the way it should: Provo 74, Ogden 64, Salt Lake City 62, then Deep South and Catholic border metros, with Madison, Albany, Denver, San Francisco and Portland at the bottom.

Honest limits. This is an allocation model, not a survey of individuals: observed totals are ground truth and further placement is demographic inference. Building footprint coverage is thinner in rural areas, so a few real rural places of worship lack a footprint and are dropped; national calibration offsets the level and the correlation shows the geography holds. Observance intensity is roughly a fifth of the score and is an editorial construction, which is disclosed rather than hidden. Block groups under 10 residents are not scored at all, because airports, water, parks and industrial land carry no residents and would otherwise render as spuriously secular. Puerto Rico and the territories are honest no-data rather than a fabricated zero, and Puerto Rico is in fact highly religious.

Coverage, limits, and how this is kept current

The things a knowledgeable reader should know before relying on any of this:

Puerto Rico is ranked as a place but not in every measure. Four measures have no data there, so it is excluded from the ranked lists rather than shown with a partial blend. Its maps and its per-place pages still work.
Washington DC is excluded from the states tier only. It still ranks as a city and as a metro area. A city-state is not comparable to a state on state-level measures.
Three measures are state-resolution and paint hard borders at state lines: state infrastructure, state fiscal health, and the state component of car insurance.
Two measures are county-resolution: childcare cost, and the current unemployment level inside local economy.
Health is tract-resolution, so it is flat inside a census tract and contributes no within-neighborhood signal.
Air quality is metro-resolution in practice, because the monitor network is.
Tree shade covers the contiguous 48 states only. Alaska, Hawaii, Puerto Rico and the territories are blank rather than filled.
Internet speed measures advertised tiers, not delivered speed. Read it as a floor check for underserved areas.
Other taxes is ranked but not currently mapped.

Census-based measures update when the Census Bureau releases new five-year estimates, on a roughly yearly cycle. Federal agency measures update on each agency’s own release schedule. Models are refit when their inputs refresh, and the reporting calibration in the crime measure is refit alongside its two anchor series. Where a measure is an estimate, its method and inputs are stated above so a reader can weigh it rather than take it on trust.

Frequently asked questions

Why does this ranking differ from another site’s?

Three reasons, in order of how much they matter. The weights are the reader’s rather than the publisher’s, so moving one changes the list. The grain is the neighborhood rather than the city, and a city ranking averages together places that share nothing but a mayor. And the measures themselves differ: crime here is corrected for state reporting differences, walkability is routed rather than measured in straight lines, and cost of living is computed for the block group rather than for the metro.

Why is a measure blank for my town?

Either the federal source for that measure excludes your state or territory, or your area lacks a local filing. Blank means unknown. A blank measure is dropped from the blend and the remaining weights are rebalanced, so it never counts against a place as a zero.

Why is the neutral point not 50?

Because 50 is rarely the real median. Scores are anchored to real best and worst cases rather than curved, so the median lands wherever the country actually is. Pleasant weather sits near 23 because most of the country is not mild and dry. Cell coverage sits near 90 because most of the country has it. The interface shows each measure’s real median so an average place reads as average.

Why does a score of 0 appear for real places?

For four measures a zero is the measurement rather than a gap: ocean proximity, transit service, transportation noise, and access to dense urban land. Most of the country genuinely is more than 45 minutes from salt water and genuinely has no scheduled transit. Everywhere else, a missing value is published as blank rather than as a zero.

Can these numbers be used or cited?

Yes, with attribution to BestNeighborhood and a link back to this page. For bulk data or research use, use the contact form.

Sources

Every agency whose published data appears in the ranking, linked to the agency rather than to a specific file. Sources marked as BestNeighborhood data are collected or modeled in house and are described in the relevant section above.

Measure	Source	Publisher
Crime safety	Uniform Crime Reporting and NIBRS incident data	Federal Bureau of Investigation
Crime safety	Homicide mortality, reporting calibration anchor	CDC WONDER
Crime safety	Motor vehicle theft, reporting calibration anchor	National Insurance Crime Bureau
Crime safety	Block-group crime model	CrimeGrade
School grade	State assessment results	State departments of education, via SchoolGrade
Cost of living, affordability	Metro price benchmarks and market listing data, modeled to the block group	BestNeighborhood
Cost of living, income, education, diversity, vacancy, home age, commute, property tax	American Community Survey five-year estimates	U.S. Census Bureau
Cost of living, taxes, infrastructure	Energy price and utility statistics, Form EIA-861	U.S. Energy Information Administration
Cost of living, fiscal health	Regional price parities, state personal income	Bureau of Economic Analysis
Cost of living, taxes, local economy	Consumer price and expenditure series, Local Area Unemployment Statistics	Bureau of Labor Statistics
Local economy	LEHD Origin-Destination Employment Statistics	Census LEHD
Weather, heat, politics	Global Summary of the Day, GHCN-Daily station observations	NOAA National Centers for Environmental Information
Weather	Wind chill and heat index formulations	National Weather Service
Air quality	Regulatory air monitoring, defining-pollutant rule	U.S. Environmental Protection Agency
Air quality	Private low-cost particulate sensor network	Private, not named
Noise, ocean, infrastructure	Highway Performance Monitoring System, highway statistics, National Bridge Inventory	Federal Highway Administration
Noise	National Transportation Noise Map, T-100 segment data	Bureau of Transportation Statistics
Noise	Terminal Area Forecast, runway data, Airport Master Record	Federal Aviation Administration
Noise, shade, walkability, heat, politics	National Land Cover Database land cover, impervious surface, tree canopy	USGS and MRLC, canopy by the USDA Forest Service
Walkability, transit, biking, shade, religiosity	Places, transportation and buildings	Overture Maps Foundation
Transit	General Transit Feed Specification schedules, about 1,140 U.S. agencies	Transit agencies, via the Mobility Database
Biking	Bike lane classification	OpenStreetMap
Biking	Elevation, 3D Elevation Program	U.S. Geological Survey
Disaster safety	National Risk Index, expected annual loss	FEMA
Healthcare access, health outlook, heat, politics	PLACES local health estimates	Centers for Disease Control and Prevention
Healthcare access, heat, politics	Environmental Justice Index	CDC and ATSDR
Heat safety	Point-in-Time homeless counts	U.S. Department of Housing and Urban Development
Heat safety	County heat mortality surveillance	Maricopa County, Southern Nevada Health District, Arizona DHS, NYC DOHMH
Housing crash safety	Home Value Index	Zillow
Housing crash safety, taxes	Economic and price series	Federal Reserve Bank of St. Louis
Housing crash safety	Case-Shiller national home price index	S&P Dow Jones Indices
Internet speed, cell coverage	National Broadband Map, Broadband Data Collection	Federal Communications Commission
Childcare cost	National Database of Childcare Prices	U.S. Department of Labor, Women’s Bureau
Car insurance	Premium data by state and block group	CoverageWatch
Religiosity	Proprietary religious signal measurement	BestNeighborhood
Taxes	State income, sales and gasoline tax rates	Tax Foundation
Taxes	Federal brackets and deductions	Internal Revenue Service
Taxes	School district identifiers, Common Core of Data	National Center for Education Statistics
Taxes, local income	Municipal and school district income tax rates	Ohio Department of Taxation
Taxes, local income	Act 32 earned income tax register	Pennsylvania Department of Community and Economic Development
Taxes, local income	Wage tax	City of Philadelphia
Taxes, local income	School district and emergency services surtaxes	Iowa Department of Revenue
Taxes, local income	County local income tax rates	Indiana Department of Revenue
Taxes, local income	County and Baltimore City income tax rates	Comptroller of Maryland
Taxes, local income	New York City and Yonkers resident schedules	New York State Department of Taxation and Finance
Taxes, local income	City income tax authority and rate list	Michigan Department of Treasury
Taxes, local income	City income tax	City of Detroit
Taxes, local income	Earnings taxes	Kansas City, St. Louis
Taxes, local income	Wage tax	City of Wilmington, Delaware
Taxes, local income	Preschool for All and Supportive Housing Services income taxes	Multnomah County, Oregon Metro
Taxes, local income	Arts Education and Access income tax	City of Portland, Oregon
Taxes, local income	Federal tax subtraction and standard deduction tables	Oregon Department of Revenue
Taxes, local income	Municipal occupational tax survey	Alabama League of Municipalities
Taxes, local income	Occupational taxes and license fees	Birmingham, Auburn, Opelika, Alabama
Taxes, local income	Occupational privilege taxes	Denver, Glendale, Greenwood Village, Sheridan, Colorado
Taxes, local income	City service and street maintenance fees	Charleston, Huntington, Morgantown, Parkersburg, Wheeling, Fairmont, Weirton, Romney, West Virginia
Taxes, local income	Income tax schedule	Puerto Rico Departamento de Hacienda
State infrastructure	Safe Drinking Water Act and Clean Water Act compliance records	EPA Enforcement and Compliance History Online
State fiscal health	Fiscal Survey of the States	National Association of State Budget Officers
State fiscal health	Public Plans Database	Center for Retirement Research at Boston College
State fiscal health	State of Pensions	Equable Institute
State fiscal health	Unemployment insurance trust fund solvency	U.S. Department of Labor
Political lean, voter turnout	Certified precinct-level election results	Secretaries of State and Divisions of Elections, all 50 states and DC
Voter turnout	Vote mode shares, Survey of the Performance of American Elections	MIT Election Data and Science Lab
Voter turnout	Election administration rules by state	National Conference of State Legislatures
Political lean	Election-day weather grids	PRISM Climate Group, Oregon State University
Big-city access, built-up surroundings, all rollups	2020 Decennial Census block population, land area, group quarters, TIGER/Line geography	U.S. Census Bureau

Cost-of-crime severity weighting follows McCollister, K. E., French, M. T., and Fang, H. (2010), “The cost of crime to society: New crime-specific estimates for policy and program evaluation,” Drug and Alcohol Dependence 108(1-2), 98-109.

Data and methodology by the BestNeighborhood data science team. Last updated July 2026.

Best Places and Best Neighborhoods to Live in the U.S.: Methodology for 2026-2027

On this page