Methodology

Which counties are included

The 2020 cross-section contains 3,112 released counties across 50 states, representing 155,156,231 two-party votes. All ten former source-geography holds were resolved with official state files and one exact two-component repair; no row was allocated or invented.

Historical model

The separate 2020 descriptive model has a mean absolute error of 4.78 percentage points across 3,112 counties. This is an average absolute error, not a confidence interval. Same-year interpolation is not an election forecast. The forecasting tests below use different specifications and held-out transitions.

Historical forecasting results

V1 trained on 507 certified 2008→2012 transitions and tested on 1,441 counties in 2012→2016. V2 froze a no-state specification and explicit national-swing controls before training on 1,441 2012→2016 transitions and reading errors for 1,445 counties in 2016→2020.

Transition comparison

V2 substantive models / 2016→2020

Analyst views of the 2020 descriptive model are kept out of the public measure picker and remain available here: 2020 model residual map and 2020 fitted Democratic share map. Both are same-year interpolation, not a forecast.

The simple-guess yardstick is our research

Our research A backtest built by this project from cited county presidential returns, with its rules fixed in advance and its method published. It looks back at past elections and is not a forecast. It asks one question of every county in 2016, 2020 and 2024: had you guessed that the county would split its two-party vote exactly as it did four years earlier, how far off would you have been? That guess has no model behind it, which makes it the yardstick: any model of county results, including experiments that simulate voters, has to miss by less than this on the same counties and the same elections before it can claim to add anything.

The two guesses

Same as last time. Each county is guessed to give the same share of its two-party vote to each party as four years earlier. It could have been made before each election.

Same as last time, plus the national shift (hindsight). The same guess moved by the change in the nation's two-party vote. That change is known only after the election, so this guess could not have been made in advance: it is an upper bound on what the simple guess could do.

A miss is the gap, in percentage points, between the guessed and the actual share of the county's two-party vote. Its size is the same whichever party it is counted for, and the size is all this section shows. The rules, the counties and the summaries were fixed in a preregistration dated 2026-10-20, before any error was computed.

How far the simple guess missed

Guessing the same as last time missed by an average of 6.6 points in 2016, 1.9 points in 2020 and 1.9 points in 2024. Adding the national shift, which nobody knew before the votes were counted, gives 5.9 in 2016, 1.9 in 2020 and 1.8 in 2024: a nationwide shift explains little of how far counties moved. The average weighted by votes counts a large county more than a small one, so it is the better comparison for a model of the national vote; the plain average treats every county alike. The root mean square is another average that counts large misses more heavily than small ones. “Same party ahead” is the share of counties where the party ahead in the guess was also ahead in the result; it is given only as a share, and no county is shown with the party either guess favoured.

On the counties whose vote counts we have audited ourselves (the others come straight from the MIT Election Lab's file), the average miss of the same-as-last-time guess was 7.0 points over 1,444 counties in 2016 and 1.9 points over 3,100 counties in 2020; every 2024 county count is a secondary compilation, so 2024 has no audited subset.

Counties, exclusions and checks

Counties are kept the same land in every year: Shannon County, South Dakota, is read as Oglala Lakota; Bedford city, Virginia, is folded into Bedford County; and Clay, Jackson and Platte counties are scored with Kansas City as one composite unit, because Kansas City reports separately in some years. Alaska reports by election district, and the chain of districts across years is not verified, so no county is scored. Connecticut reported 2024 results by planning region, which do not match its eight historical counties, so 2024 is not scored there.

An independent verifier rebuilds the county counts from the raw returns with its own code, re-derives every figure in this section and fails on any difference. An earlier backtest of ours scored the same two guesses on a different scale (the gap between the two parties' shares) and a slightly different set of places; recomputed here, all 6 of its published rows agree exactly, apart from rounding in the computer's arithmetic (the largest difference is 2e-15 points).

Sources: MIT Election Data and Science Lab, County Presidential Election Returns 2000-2024 (v16), Harvard Dataverse (Nonprofit compilation; 2012, 2016 and 2020 (our audited county counts take precedence where they exist)); Tony McGovern, US County Level Election Results 08-24 (Individual compiler; a secondary compilation; 2024). This is a test on elections that have already happened. It is not a forecast and says nothing about who will win any future election.

Net migration for every county, and corridors between places

Net migration is movers in minus movers out over the stated period, per 1,000 residents on July 1 at the end of the period (Census Population Estimates). IRS counts tax returns and the people on them (N2); PopEst estimates all residents.

The map and county profiles show four measures for every county: IRS net migration per 1,000 residents and per 1,000 tax filers, both from each county's complete inflow and outflow totals for IRS filing years 2022→2023, and Census Population Estimates (V2025) net domestic and net international migration per 1,000 residents for July 1, 2024 → July 1, 2025. IRS 2022 series (enhanced address matching from 2022–23); not comparable with earlier years without an overlap design. IRS counts people as exemptions through 2018–19, individuals from 2019–20. Connecticut is reported as its nine planning regions and is never split among its eight former counties. The county population shown is the same Census estimate used as the denominator.

The corridor view is separate. It counts tax returns in the county-to-county corridors IRS publishes for 2022→2023: 6,470,553 returns, 99.82% of comparable published returns. Each county is labeled by the winner of its 2020 presidential vote, and each Connecticut planning region by the summed results of its towns in the state's Statement of Vote. Movers are not labeled by party.

Independent checks compare different measures, so each is cited for what it tests. The map's IRS measure ranks counties much like Census net domestic migration: Spearman 0.850 across 3,087 counties for IRS filing years 2022→2023 (Census V2025), and 0.866 across 2,866 counties against Census Vintage 2023. Net flows from published corridors alone agree less: Spearman 0.694. ACS county-pair movers and IRS corridor exemptions: Spearman 0.613. 26.76% of domestic county-header returns are in IRS residual "other flows" buckets and cannot be assigned to any corridor or context; every ratio here describes published direct corridors only.

Monthly county employment

The explorer’s Current economy panel shows official BLS Local Area Unemployment Statistics for 3,104 of the 3,112 released counties. The map shows the monthly unemployment rate, not seasonally adjusted, compared with the same month one year earlier. A county’s profile adds its resident employment, unemployment and labor-force history and their year-over-year changes; those counts are not mapped because they rank almost exactly with population, so a map of them would show where people live. Coverage is reported separately and never implied to equal electoral coverage.

Status

Published

Latest month

August 2026

Preliminary

Yes, the latest month; BLS revises it the following month

Retrieved

October 8, 2026

Place of residence, not workplace · Not seasonally adjusted · Single current vintage retrieved 2026-10-08 · October 2025 not published (federal shutdown) · Connecticut's eight historical counties have no current LAUS series (replaced by nine planning regions, not coextensive).

Election evidence and district boundaries

Detailed county-to-district relationships are available for the North Carolina and Vermont pilot. National seat and candidate coverage is shown separately on the Elections page, with the source status for each state. Seat identity is plan-scoped: the 2026 North Carolina contest uses enacted SL 2025-95 boundaries, cross-checked block-for-block against the Census 120th-Congress representation, while the current representative’s boundaries are kept distinct. County relationships are many-to-many and versioned on both sides; partial intersections are labeled “district intersects this county,” and no county figure is allocated into a district number. Candidate statuses separate filed, qualified, withdrawn, and unresolved; endorsement records separate the action from its evidence tier and never infer from photos or meetings. District-native ACS appears only where the statistical vintage is genuinely compatible (Vermont’s at-large seat); elsewhere the page says so and shows labeled county and state context instead.

Hot Races: two lists

Statewide, all states. Every 2026 federal contest whose geography is a whole state — 41 in all: the scheduled Senate contests and the at-large House seats — is ranked by how close that state’s 2024 presidential two-party result was, from statewide totals in the compiled 2024 results. A contest is on the list for either of two reasons. Close in 2024: the state is among the 6 closest (15% of the list, rounded down, ties at the cut-off included): Michigan Senate (R+1.4), Georgia Senate (R+2.2), New Hampshire Senate (D+2.8), North Carolina Senate (R+3.3), Minnesota Senate (D+4.3), Virginia Senate (D+5.9). Tightly polled: the current poll average is within 5 points: Michigan (D+3.6), Minnesota (D+1.9), Maine (D+2.0), Ohio (D+3.3), Iowa (D+4.7), Alaska (D+0.8), Texas (D+2.6). The 2024 measure describes how evenly each state voted at the top of the ticket; a state can vote closely for president and lopsidedly for Senate. Neither measure is a forecast or a probability. This list was defined after the 2024 results had been examined, so unlike the North Carolina and Vermont list it was not preregistered.

Polls. Senate polls come from the VoteHub Polling API under CC BY 4.0, retrieved October 7, 2026. We keep general-election polls that are not sponsored by a party or campaign, are not internal, and name both candidates of the race’s matchup. The matchup is the one Democrat and one Republican on the certified candidate list where we hold one with exactly one of each; otherwise it is the two leading candidates named in the three most recent polls, and a race is left uncharted when those disagree. Where a pollster published several versions of one poll, we keep likely voters first, then registered voters, then adults, then the fullest ballot. The average is the mean of each pollster’s latest poll ending in the last 30 days, shown only when at least 3 polls from 2 pollsters qualify. There is no pollster weighting, house-effect adjustment or modeling. As a check, 175 of the 189 polls we used when we last read The New York Times poll feed (October 7, 2026) also appear in it, and 172 of those agree within 1 point. We no longer read that feed: its terms allow personal, non-commercial use and ask for written consent before automated collection. Each Senate page charts its polls over time and lists every one with a link to the pollster’s release. House districts outside North Carolina and Vermont are not scored: there are no district-level results here, and several states redrew their maps for 2026.

North Carolina and Vermont. The structural list scores the scheduled North Carolina and Vermont federal contests on four structural features, using rules fixed in advance (preregistered as hot_races_v1): an open seat (4 points when the incumbent is not on the general-election ballot; 2 when the incumbent’s plans are unknown), a first-term incumbent (2 points when the incumbent’s term started on or after January 1, 2025), the size of the general-election field (0.5 per candidacy, up to 3; 1 when no general contest is listed) and primary competition (0.5 per primary candidacy, up to 4). At most 15% of the contests, rounded down, are highlighted: 2 of 16. Scores use no polls, describe how open a contest is, and do not predict a close result.

Known issue under review. The first-term rule reads the start of the incumbent’s current term, so every House incumbent in the pilot currently receives those 2 points. The published ranking will not change until a reviewed rule is published.

County-to-seat relationships: NC/VT pilot only. National race coverage varies by state. Missing names are never reported as verified absence. Refreshes produce new releases; the site shows the release it was verified with, not a fabricated live banner.

How data is checked, joined and modeled

Every analytical row passes a layered release process. The sequence matters: a sophisticated model cannot repair unsupported geography or conflicting official editions.

Places, layers and evidence on the map

When a selected place carries between layers

The map has three geographies: 2024 results areas, audited counties (2020 boundaries) and states. A selected 2024 results area is treated as the same place as a county only when it reuses the county’s code and is not one of the areas below; that holds for 3,103 of 3,160 areas, checked by code and name. Alaska’s 40 election districts, Connecticut’s 9 planning regions and the District of Columbia’s 8 wards are not counties, so a selection there does not carry over to county measures, and the panel says why. A state never carries over to a county measure.

The map reads versioned, fixed files built from checked releases. Every metric exposes its unit, period, geographic grain, adjustment, source tier, release status, coverage and available uncertainty. County demographics come from the 2015–2019 ACS five-year estimates; their public export does not supply margins of error. This is not current demographic measurement. County population is the Census Population Estimates figure for July 1, 2025 (V2025); the ACS population count is used only as an input to the 2020 model.

The Politics map moves through every presidential election since 2000 on one year slider. Each year keeps its own source and geography, and the map says so when the geography changes. 2000 to 2016 are county returns compiled by the MIT Election Data + Science Lab (County Presidential Election Returns 2000–2024, version 16, Harvard Dataverse, doi:10.7910/DVN/VOQCHQ), the same returns as the home page map, with this site’s audited county result in their place wherever one exists; the other counties are as MIT published them and have not been checked here against official canvasses, so these years are labelled a research preview and the map gives no summary or ranking for them. Each year is the Democratic share of the two-party vote, computed from the Democratic and Republican county votes. The 2012–2016 change offered on the 2016 stop is the 2016 share minus the 2012 share, both computed this way from the same returns, for every county with both years, and is a research preview like them. 2020 is the audited county release, and 2024 is as reported, by 2024 results area or state.

State and metro/region benchmarks describe their source geography. Connecticut planning regions and historical counties remain distinct; Alaska election areas are not additive counties. Map geometry only renders shapes. Related seats come from the accepted official relationships, without spatial allocation. Missing, excluded, suppressed and not-covered values are kept distinct.

Costs and housing are a research preview

The cost and housing measures are a research preview: they have not had the full checks that published layers get. Housing, grocery CPI, modeled energy burden, permits, and state gasoline/electricity keep their source periods and geographic basis. Grocery CPI is a metro or region benchmark, never an individual county price. State values are never relabeled as county observations. The housing measure compares Aug 2026 vs Aug 2025; the Zillow file was downloaded again for the economic score and reproduces the published value for every county that has one; its hash is pinned in that package. Electricity is displayed as the published state residential price in cents/kWh, rather than confusing a growth rate with the level unit. Employer and additional real-estate data are not published.

The county economic score is our research

Our research The score, its two lenses, every weight and every rule here are this project's own research design, not an official statistic. Every input is a published figure from a named source and vintage. It reads every county through two lenses: For people already here (homeowners and longer-standing, older residents, including those on fixed incomes) and For people getting started (households headed by people under 35, and newcomers). The same rise in home values counts as a gain in the first and a barrier in the second.

How it is built

1. Each figure becomes a percentile among the counties that have it, flipped where lower is better for that lens.
2. A lens score is the weighted mean of those percentiles, from 0 to 100. A lens is computed only when the indicators present carry at least 75 of its 100 weight points; otherwise no lens score, and the reason lists the missing indicators.
3. Economic conditions: the simple mean of the two lens scores; computed only when both lenses are computed. Neither group is privileged: a place scores high only if it works for both those already there and those trying to get in.
4. The four groups split each lens at 50, the middle of the scale, fixed in advance: Works for both (966 counties); Works for owners, hard to get into (560 counties); Open door (554 counties); Struggling for both (1,012 counties).

State electricity prices are shown on each county card as context and never scored: they are the same for every county in a state. State gasoline prices were shown the same way until October 15, 2026, when they were withdrawn because AAA’s terms do not allow republishing them; they were never scored. The Energy Information Administration’s weekly gasoline prices for a state or its region now appear under Faster signals, outside the score. Grocery prices are published for metro areas and regions only, so they carry a small weight and are left out of one sensitivity test.

Uncertainty: ranges, never one exact number

Each survey input is redrawn 1,000 times from its published margin of error, and every percentile, lens score, headline and rank is recomputed each time. Cards show the middle 90% of those draws. The typical county's headline range is 10 points wide, and its rank range spans 849 places, which is why no single rank is published. LAUS, Zillow, LEAD and CPI publish no county sampling error; they are held fixed, so the bands understate total uncertainty. 1,285 counties land in another group in more than one draw in five and are marked as near the line.

Coverage

3,092 of 3,112 counties are scored. The 8 Connecticut counties are not: Census and BLS publish Connecticut by planning regions, which are never mapped onto the historical counties. 12 other counties have too few inputs, and their cards name the missing figures.

Sensitivity to the weights

• Every indicator of a lens weighted equally: 895 of 3,092 counties (28.9%) change group; the headline's rank correlation with the published weights is 0.90.
• Grocery price change removed: 122 of 3,048 counties (4.0%) change group; the headline's rank correlation with the published weights is 1.00.
• Home value change weight doubled in both lenses: 575 of 3,048 counties (18.9%) change group; the headline's rank correlation with the published weights is 0.99.
• Employment change and unemployment rate weights doubled in both lenses: 540 of 3,092 counties (17.5%) change group; the headline's rank correlation with the published weights is 0.95.

The group names are sensitive to the weights. The preregistered rule (if any variant moves more than 25% of counties to another quadrant, the methods page says the quadrant names are sensitive to the weights) applies: weighting every figure equally moves more than a quarter of counties to another group. Read a county's group together with its two lens ranges.

The outside check

Does the 'getting started' lens line up with later net in-migration of 25-34-year-olds? The check was fixed in advance and uses nothing the score uses: the “For people getting started” lens, rebuilt from 2019 figures with the same rules, against what happened next. The outcome is the change in the number of people aged 20 to 29 in July 2020 as they became 25 to 34 by July 2025 (Census population estimates, Vintage 2025). At these ages deaths remove well under one person in a hundred over five years, so the change is close to the group's net migration.

Result: rank correlation 0.118 across 3,085 counties (95% interval 0.083 to 0.154). The rule was: Spearman >= 0.10 AND the lower end of the 95% interval > 0. It passed, so the headline score is published.

Read it modestly. The relation is weak, and the “For people already here” lens built the same way lines up more strongly with the same outcome (0.192; the headline, 0.191): places that were doing well overall drew young adults, and the check does not show that the “For people getting started” lens captures something separate. Against the 2020–2024 share of 25-to-34-year-olds who moved in from elsewhere (people moving in, not net), the current lens shows no relation (-0.017 across 1,837 counties).

Sources

• Federal·Census Bureau Census American Community Survey 5-year estimates, tables B19049, B25007, B25064, B25072, B25077, B25093, 2020-2024. Survey estimates with 90% margins of error, carried into the score's uncertainty.
• Federal·EIA EIA Electric Power Monthly, Table 5.6.A, residential price, July 2026. State level: the same for every county in the state.
• Federal·BLS BLS Consumer Price Index, food at home, not seasonally adjusted, August 2025 to August 2026. Published for metro areas and Census regions, not counties.
• Federal·Census Bureau Census ACS 5-year median household income (B19013), the earlier window restated in 2024 dollars with BLS CPI-U annual averages, 2015-2019 to 2020-2024. CPI-U, U.S. city average (CUUR0000SA0), 2019 and 2024 annual averages.
• Federal·BLS BLS Local Area Unemployment Statistics, county, not seasonally adjusted, by place of residence, June-August 2026 against June-August 2025 (jobs); September 2025-August 2026 (unemployment rate). October 2025 was never published. August 2026 is preliminary.
• Federal·DOE DOE Low-Income Energy Affordability Data (LEAD) Tool, ACS 2022 PUMS (2018-2022). A modeled estimate, not observed bills.
• Private·Zillow Zillow Home Value Index (ZHVI), county, typical home value, smoothed and seasonally adjusted, August 2025 to August 2026. Zillow data, used with attribution.

Package: county_economic_score_v1 (preregistration, build, an independent recomputation of every score, and a clean-room check of the release archive).

Young-adult labor-market context is our research

Our research Built by this project from published Census figures, with its method published. Context beside the county economic score, not an input to it. The county economic score uses one all-ages unemployment rate and an all-ages employment change, so on its own it says nothing about how young adults are doing in the job market. Two age-specific measures are shown beside the score on each county's card, the explore panel and the Counties page. They are context: neither feeds the score, whose figures, weights and preregistration are unchanged.

Unemployment rate, ages 20 to 34

From the American Community Survey 5-year estimates for 2020-2024, table B23001: civilian unemployed people divided by the civilian labor force, ages 20 to 34, with the all-ages (16 and over) rate from the same table and the ratio of the two. Ages 16 to 19 are left out by design, because the score's young group is householders under 35; they remain in the all-ages rate. The young-adult rate is compared only with the all-ages rate from this same survey, never with the monthly BLS rate, because the periods and sources differ.

Each figure carries a 90% margin of error, derived with the Census Bureau's published formulas for sums, proportions and ratios. It is published only where its coefficient of variation (the standard error as a share of the estimate) is 0.30 or less, a rule fixed before any figure was computed; up to 0.15 is marked reliable and above that “use with caution”. Any other county is shown as “too uncertain to publish”, never blank and never zero. The young-adult rate is published for 1,475 of 3,104 counties with data; 1,629 are too uncertain, mostly the smaller counties, where the survey's sample is too small for a narrow age band. The all-ages rate is published for 2,132 counties and the ratio for 1,161.

Jobs held by workers aged 25 to 34

From the Census Bureau's Quarterly Workforce Indicators: the number of jobs on the first day of each quarter for workers aged 25 to 34 and for all ages, at private employers, by the county where the job is located. These are jobs located in the county, not its residents, so a county that people commute into can differ from its residents' jobs; each job counts once, and public-sector jobs are not included. The change compares the average of the four quarters ending at each state's latest quarter (2025 Q4 in this release) with the average of the four before; a four-quarter average avoids seasonal swings. The difference between the two changes is shown in percentage points.

QWI protects confidentiality by adding noise to every released count, and publishes status flags rather than sampling errors. A county is left blank, with the reason shown, unless every count behind its figure carries the “OK” flag, and also when the average number of jobs held by workers aged 25 to 34 is under 100 in either window, because a change in a handful of jobs is dominated by single employers. The figure is published for 2,830 counties. The Census has no current county figures for Michigan (its last quarter there is 2021 Q4), so those counties show no jobs figure rather than an old one. Connecticut's eight historical counties have neither measure: Census publishes Connecticut by planning regions, which are never mapped onto them.

The sentence

Where a county's home-value sentence says values rose, and its young-adult and all-ages rates are both published, one more sentence states the two rates with their margins of error and how they compare (1,160 counties). It is generated from the county's own figures, in neutral wording, and makes no claim about cause.

Sources

• Federal·Census Bureau Census American Community Survey 5-year estimates, table B23001: sex by age by employment status for the population 16 years and over, 2020-2024. Residents, a survey estimate with 90% margins of error. A figure is published only where its coefficient of variation is 0.30 or less. Compared only with the all-ages rate from the same table, never with the monthly BLS rate.
• Federal·Census Bureau Census Quarterly Workforce Indicators (QWI), beginning-of-quarter employment, private employers, by the county of the workplace, the four quarters to 2025 Q4 against the four before. Jobs located in the county, not its residents. QWI counts are noise-infused (“fuzzed”) for confidentiality and carry status flags; flagged counts and averages under 100 jobs aged 25 to 34 are not published.

Documents read for the method:

• Census LEHD, Quarterly Workforce Indicators 101 (QWI20260310)
• Census LEHD, public-use data schema (QWI variables, age groups, ownership and status flags)
• Census Bureau, ACS General Handbook, chapter 8: calculating measures of error for derived estimates (2020 edition)

Package: young_adult_labor_v1 (a preregistration fixed before any figure was acquired, with 2 dated amendments that each record when and why they were made; an independent recomputation of every published figure from the raw responses; and a clean-room check of the release archive).

Faster signals: recent figures beside the county data

The county data the site is built on arrive months or years after the period they describe. The faster signals add the newest figures that federal agencies and private providers publish, each as its provider’s own series for one geography, with its chip, a County, State, Region or National tag, the last day it describes and its age. They are context only: not part of the economic score, never averaged or blended with another figure, and never used to fill a county’s gap. Which series are shown, the comparisons we compute and the rules below were fixed in a preregistration on October 15, 2026, before any file was downloaded. The gasoline, electricity and personal income series were added in a dated amendment (October 15, 2026), likewise fixed before their data were downloaded.

Federal·Census Bureau Housing units authorized by building permits in the county, January to August 2026, as the Census Bureau estimates them, with the share actually reported by permit offices (the rest is the Census Bureau’s imputation). Published for 2,981 counties; the change from the same months a year earlier is shown for 1,763 counties where both years have at least 20 units (1,218 have fewer).

Federal·FHFA FHFA’s own annual change in its developmental county house price index for 2025 (2,595 counties), and its seasonally adjusted purchase-only index for each state and the nation, with the change from the same quarter a year earlier.

Private·Indeed Hiring Lab Indeed Hiring Lab’s job postings index for each state and the nation (February 1, 2020 = 100), with the change from the same date a year earlier. Postings on one job site, not openings or hires.

Federal·Labor Department Initial unemployment-insurance claims in each state, by state of residence, for the latest week in the Labor Department’s revised state file; the change compares the average of the four latest weeks with the same four weeks a year earlier, because single weeks swing with holidays and layoffs.

Federal·Census Bureau Business applications in each state and the nation for the latest month, seasonally adjusted; the change compares the unadjusted month with the same month a year earlier. These are applications for an employer identification number: not new businesses and not jobs.

Federal·EIA The Energy Information Administration’s weekly price of regular gasoline (all formulations; the cash pump price including taxes at 8:00 a.m. on Monday), estimated from a sample of about 1,000 stations. EIA publishes the nation, seven regions and 9 states, and says it has no county prices. A state’s page shows its own EIA series where EIA publishes one and otherwise the smallest EIA region that contains the state, tagged Region and naming its states; the two are never shown together, averaged, or presented as a county’s. The change compares the latest Monday with the Monday 52 weeks earlier. Where EIA publishes a state or its region this replaces the withdrawn AAA state averages.

Federal·EIA The average price residential customers paid for electricity in each state and the nation for July 2026 (revenue divided by sales; EIA’s preliminary estimate), with the change from the same month a year earlier. We checked that these figures equal the ones EIA prints in its Electric Power Monthly table for the same months.

Federal·BEA The personal income of each state’s residents from the Bureau of Economic Analysis, seasonally adjusted at annual rates and in current dollars (not adjusted for inflation), with income per person beside it and the change from the same quarter a year earlier.

Federal·BEA The personal income of the county’s residents, wherever they work, for 2024, in current dollars, with income per person and the change from the year before. BEA publishes small Virginia independent cities and Kalawao County, Hawaii combined with a neighbor, and Connecticut by planning region, so 60 counties say “not available” with the reason. These BEA figures are read from BEA’s public data files, which stay our source of record. We also asked BEA’s Data API for the same tables, lines and periods (the latest quarter and the same quarter a year earlier for the states and the nation; the latest year and the year before for the counties) and checked that each value equals the one in the files, as text. The API supplies no figure of its own, and the check stops the build if the two ever differ.

This product uses the Bureau of Economic Analysis (BEA) Data API but is not endorsed or certified by BEA.

Comparisons always use the same week, month, quarter or date one year earlier, so seasonal patterns cancel without an adjustment of ours, and they are marked “computed by us”. A line is not shown as current once its last day is older than its source’s schedule allows: 28 days for claims, 21 for job postings, 75 for business applications, 100 for permits, 200 for state house prices, 550 for the county index, 21 for gasoline, 120 for electricity, 200 for state personal income and 770 for county personal income. Every file is stored with its hash; a test rebuilds every displayed number from those bytes, and a second program recomputes them independently.

Not shown, and why: ADP’s terms forbid public reuse; Zillow’s rent and listings files, Realtor.com and Redfin do not clearly allow republishing; the BLS payroll and job-openings data are available, but the BLS method pages refuse automated readers, so we could not read them first; Indeed’s metro file and the Census business trends survey are pending decisions. Each is listed on the sources page with its reason.

Who publishes each number

Every source the site cites is listed once on the sources page, with the kind of organization that publishes it, the kind of number it is, the places and dates it covers, its license, how old its newest figure is, and what we read before using it. Of the 73 sources listed, 52 are shown on the site: 27 federal agency sources, one state or local government, seven university or nonprofit sources, five private company sources, two individual compiler sources and ten our research releases. The other 21 were read and are linked or listed with the reason they are not shown.

Source types

Every figure carries a chip that names its publisher and the kind of organization it is, and a tag that names the geography it describes (County State National). A federal agency is a U.S. government statistical or program agency; the Federal Reserve System is shown separately because the regional Reserve Banks are not executive agencies; state and local governments publish election records; universities and nonprofits publish research datasets; private companies publish figures from their own platforms or customers; an individual compiler is one person’s public dataset; our research is work we built, with its method published. The chip is a label of who made the number, not a grade: a federal source is not better and a private one is not suspect.

The kind of number is a second, separate label: a survey sample (estimates with sampling error), administrative records (counts kept to run a program, such as tax returns or unemployment-insurance claims), a model estimate (built from several inputs by a stated method), platform or transaction data (what one company sees among its own users or customers), a compilation (records gathered from other publishers), or derived by us.

Independence

Agreement between two sources proves less than it looks when they share inputs. Private series are often benchmarked to federal ones (ADP, which we do not show, re-weights its payroll sample to federal benchmarks every year), and federal series share inputs too: the Census Bureau’s business applications are applications for an employer identification number filed with the IRS. Where two figures on a page move together, the site does not present that as confirmation.

Never blended, never in the score

Figures from different providers are never averaged, summed or combined into one number, and a state or national figure is never used to fill a county’s gap: a county without a figure says so, and the state line beside it carries its own State tag. Nothing from a private provider enters the county economic score; a future score that used one would be a new version with its own preregistration.

Who is in each private dataset

A private provider sees its own users or customers. Indeed’s index counts postings on one job site: openings that are not advertised online are missing, one posting can cover several openings, and occupations with high turnover post more often. Each private source’s entry on the sources page says who is in it and who is not, in the provider’s own words where we could read them.

Licenses, revisions and discontinuation

We read each provider’s terms before showing a figure. Where the terms do not clearly allow republishing, we link to the provider and show no figures: ten sources are linked this way. Providers revise their history (FHFA revises its whole index at every release; Indeed revised its whole series when it changed its seasonal adjustment in November 2024; the Census Bureau revises its seasonally adjusted business applications every month), so each figure is dated by the file it came from, and a refresh replaces it rather than mixing vintages. When a provider revises history in a way that changes what the site said, the change goes in the corrections log. A source that stops publishing or changes its file shows “not available”, never its last number as if it were current.

Federal interruptions

Federal series have gaps around the 2025 federal funding lapse: the BLS never published county unemployment for October 2025; the Census Bureau’s business trends survey has no estimates for its periods from October 6 to November 16, 2025; and the September and October 2025 county permit files were posted together in January 2026. Gaps are shown as gaps, never filled.

How current each number is

Each figure shows the last day it describes and how old that is, counted to the reader’s today, not from the day we retrieved the file. A five-year survey average describes all five years, and its age is counted from the last of them. A quarterly jobs count that measures jobs on the first day of the quarter is dated to that day. Each source has a stated schedule; a figure older than its schedule allows is not shown as current. The county profile draws every source’s age on one ruler.

Our research and theirs

Figures we built carry the Our research badge and link to their method here. Everything else is the provider’s figure, cited to them; where we compute something from it (such as a change from a year earlier), the figure says “computed by us”. The faster signals were last refreshed from files retrieved on October 15, 2026.

Political Volatility Pressure

A proposed way to examine rapid change, divergence from history and comparable places, and several pressures occurring together. No validated measure is published. It would describe unusual conditions, not which party benefits or the probability of an election outcome.

Level, rate of change, historical divergence and comparisons with the state, nation and defensible peers would be separate inputs. Before any measure is published, its definitions, geographic coverage, reference periods, uncertainty and validation must be specified. The current site does not combine these inputs or classify places by volatility.

What kind of claim is this?

Observed data includes published measurements and survey estimates. Derived measures are calculations from those inputs. Comparisons state a benchmark and period. Signals require a defined detection rule; models estimate or forecast a quantity. Interpretation explains possible meaning and must be separated from the measured facts. Neither a large change nor a high rank, by itself, establishes causation.

A second axis is who made the number. Every figure carries a chip naming its publisher and the kind of organization it is (a federal agency, the Federal Reserve System, a state or local government, a university or nonprofit, a private company, an individual compiler, or our research) and a tag for the geography it describes. The two axes are independent: a federal model estimate and a private platform count are different kinds of claim from different kinds of publisher, and each is read as what it is. Who publishes each number · every source.

Counties in focus: what the leverage probabilities mean

How the simulation works

The Counties in focus page picks its counties with a county-leverage simulation that ranks 3,102 county-level units by the chance that replacing a county’s own result with a fresh draw changes the Electoral College winner. Each county’s margin is the 2024 result plus its response to a national swing and coalition-cluster, division, state and county-specific shocks, with a mix of calm and turbulent cycles; states are exact sums of counties. The national environment is a stated assumption with no polls, fundamentals or registration data, so every probability is a comparison between counties under that assumption, not a forecast. The ± figures beside each probability are simulation error, which is small; the larger uncertainty is the modeling assumptions, which the page tests with alternative assumption sets.

Leverage, the chance of flipping a state, information about the winner, swing probability, net-vote movement and trend are kept as separate measures. Rolling-origin backtests show leverage concentrating where past races were close, with imperfect calibration that the page reports. Elasticity is shown with its interval and prior-dominated flag; information value is model-implied; Maine and Nebraska districts are county proxies; Alaska and Connecticut are state-level units; economic labels are descriptive.

How the simulation was tested

We backtested the simulation on 9,312 county results and 153 state results from the 2016, 2020 and 2024 elections (earlier elections lack the history to test); its 80% county-margin intervals covered 60% of results in 2016 and 97⁠–⁠98% in 2020 and 2024. Each test rebuilds the model from the information available before that election, then scores it against what happened: 3,104 county-level units in each election, of which Kansas City is one composite and Alaska and Connecticut are one state-level unit each (9 of the 9,312 results are not a single county), and each state and the District of Columbia (153 state results). Earlier elections are not tested because the model needs at least 3 earlier election-to-election changes to estimate how places move together. An interval is calibrated when it covers as many results as its label promises; these covered fewer in 2016 and more in 2020 and 2024, and three testable elections cannot show how well the extremes are calibrated. A single county’s chance of being decisive cannot be tested directly: in each tested election no one state alone would have changed the winner, so no county was decisive. These tests compare the simulation with past results; they do not make its probabilities for 2028 a forecast.

Package county_consequentiality_v1 · model run 2026-10-06 · publication 35ba9b37fa41… · limits and file hashes

What is still unresolved

Two distinct held-out transitions are enough to reject durability, not enough to characterize every future election regime. National-cycle forces remain unforecasted; uncertainty is under-calibrated; feature missingness is patterned; and the strict test samples do not cover every US county. Migration histories cross IRS series breaks and cannot identify mover ideology or motive. The eight state-years below remain withheld from strict county analysis.

Corrections log

Errors we have found or been told about, what we changed, and the data affected. Newest first.