How Pull works

What Pull measures

How many points one county moved its state’s margin. Take the county’s Democratic votes minus its Republican votes and divide by every vote cast in the state.

Fulton, Georgia, 2024: (384,752 − 144,655) ÷ 5,250,905 = D+4.57

Add up every county and you get the state’s margin exactly, which is what the two bars in each card show: each block is one county, as wide as its pull, and the longer bar wins by the difference. National Pull does the same against the national vote.

Pull and the vote margin

Pull is the vote margin, rescaled. Within one election every county’s net votes are divided by the same number, the votes cast in its state, so a State Pull map of a state looks just like a vote-margin map of it: the same counties count for the most, in the same order. The idea is the same too. A place matters for the votes it nets, not for its percentage.

The difference shows when two elections are set side by side. Raw vote margins grow whenever more people vote, so a bigger number can mean a place mattered more or only that turnout rose. Pull is in points of the final margin, which mean the same thing every year.

Fulton, Georgia: a margin of 141,864 votes in 2008 and 240,097 in 2024, up 69%. Georgia cast 1.3 million more votes in 2024, so Fulton’s pull went from D+3.61 to D+4.57, up 26%.

For the same reason it lets a small state be compared with a large one, and the pulls of a state’s counties always add up to its margin in points.

One figure is a hybrid. In a state’s list of elections, a Senate or governor’s race shows a pull worked out against that year’s national vote for the House: the race’s net votes in the state, divided by every vote cast for the House in the country. It says how big that margin was on a national scale. It is not part of any national margin, since only some states hold those races, so these pulls do not add up to anything. Races in odd years have no House vote to set them against, so they have no pull.

Why the House vote and not the presidential vote, even in a presidential year? For decades the country voted one way for president and another way for everything else. In 1984 it was R+18.2 for president and D+5.3 for the House; in 1972, R+23.2 and D+5.2. Measured against the presidential vote, every Senate and governor’s race in those years looks 20 points more Democratic than it was. Since 2000 the two votes have stayed within about 3 points of each other, so recent years barely change.

The House vote has a flaw of its own. A seat with only one candidate adds votes to one side alone, or none at all, and before the 1960s much of the South sent Democrats to the House unopposed. So its national margin can lean a few points Democratic in those years.

Lean versus turnout

Pull is a place’s share of the vote times its margin, so a change in Pull has two sources. Lean is the margin moving while the share is held steady. Turnout is the share moving while the margin is held steady. The two add up to the full change.

Read both loosely. A margin can move because voters switched, because one side stayed home, or because different people live there now. Turnout is relative: a county that grew still loses weight if the rest of the state grew faster.

A map shaded by turnout alone does not show how places voted. A Democratic county is red there if its share of the state’s vote shrank, even when its own margin moved left, and a county whose share held steady is pale however far its margin moved.

Where the name comes from

Pull started life as Mikert, the name its inventor gave it in high school. Elsewhere there is no settled name. Analysts usually quote the raw ingredient, net votes, and sometimes call the normalized figure a county’s contribution to the statewide margin.

Data and caveats

  • County president 2000–2020, state president 1976–2024 and Senate 1976–2024: MIT Election Data and Science Lab.
  • President by state, 1868–1972: each state’s counties added up, from the same Amlani and Algara file (Alaska 1960–1972 from its certified totals), with the electors from the state table in their replication files. That county file counts only the votes on each party’s own line, so a nominee who also ran on another line comes up short (New York’s fusion tickets of 1936–1952, Bryan in 1896, Greeley in 1872), and in Alabama it counts Thurmond and Wallace as the Democrat. It also put Baltimore city’s votes into Baltimore County a second time in most years before 1950; the city is taken back out. Where a state’s sums were still more than 1.5 points from the published shares (58 state results), the county table on Wikipedia’s page for that state’s election is used instead, if it adds up to within a point of the published margin: 42 of them. The other 16 are rebuilt from the published shares, their vote counts are close but not exact, and their counties are left blank (hatched) in that election. Every county from 1868 to 1996 was then checked against Wikipedia’s table: 96% agree to within half a percent. In 73 counties where the file’s count is plainly wrong (a dropped digit, two counties’ rows exchanged) and Wikipedia’s fits the same county in the elections either side, Wikipedia’s figures are used; about 300 other large disagreements, where neither source is clearly the right one, are left as the file has them.
  • County maps before 1976 are drawn on the county lines of the nearest census, from the NHGIS boundary files that come with the same dataset (IPUMS NHGIS, University of Minnesota, www.nhgis.org). Each is fetched when one of its elections is first opened. Alaska is one unit, and territories that could not vote are blank.
  • Gold marks a place a third candidate led, and its card reads I+ (his lead over the better of the two nominees). For a state before 1976 that means most of its electors went to a third candidate. For a county, the third candidate’s own votes (Weaver 1892, Roosevelt 1912, La Follette 1924, Thurmond 1948, Wallace 1968, Anderson 1980, Perot 1992 and 1996) are read from the county table on Wikipedia’s page for each state’s election, about 20,700 counties in all, wherever that table’s statewide share for him matches the published one. They are added to the county file’s own figures and never exceed the votes it has left over after the two nominees. Where Wikipedia has no county table (Texas, Mississippi and South Carolina in 1912; Connecticut and Rhode Island in most years) a county is still marked only where everyone else together out-polled both nominees by more than the minor candidates could explain, and its card shows everyone but the two nominees. On a state’s card his figure is his own share from the state table, to a tenth of a point. Comparisons between two elections show no gold.
  • Before 1976 electors are counted as pledged, including those won by Weaver (1892), Roosevelt (1912), La Follette (1924), Thurmond (1948), unpledged slates (1960) and Wallace (1968). A state a third candidate carried is gold on every map of one election; its Democrat-against-Republican margin still counts toward pull. Florida in 1868 and Colorado in 1876 chose electors by legislature, with no popular vote.
  • County president 1976–1996, county Senate 1976–2020 and every governor’s race 1976–2020: Amlani and Algara, Partisanship & Nationalization in American Elections (Electoral Studies, 2021). That file gives each county’s Democratic and Republican nominees and its total vote, and has no Alaska. Governor statewide figures are its county sums.
  • A Senate race has a county map where its counties come within 2% of the statewide count for both parties. Four do not: Louisiana 1998, 2004 and 2020 (the statewide figure adds up every candidate of each party in the all-party primary), and Colorado 2020 (counties incomplete).
  • Where the two sources disagreed and the counties matched the certified result, the statewide record now follows the counties: president in Vermont 1976 and South Carolina 1992; Senate in Utah 1976 and 2010, Wyoming 1978 (the parties were swapped), North Carolina 1986, West Virginia 2006 and the Louisiana runoffs of 2002, 2014 and 2016. North Dakota’s 1980 Senate race, missing before, is added. Two specials with no party on the ballot (Georgia 2000, Mississippi 2008) now count their Democrat and Republican as such.
  • Governors 1976–2020 are set against Wikipedia the same way as the earlier races below. 575 are their counties added up. Seventeen had an independent or third-party candidate, or two from one party, in the top two; they are named from Wikipedia and hatched: Alaska 1978, 1990, 1998 and 2014, Hawaii 1982 and 1994, Illinois 1986, Louisiana 1987, Connecticut 1990, Utah 1992, Maine 1994, 1998 and 2010, Minnesota 1998, Vermont 2008, Colorado 2010 and Rhode Island 2010. Alaska’s other races from 1978 to 2018 come from Wikipedia’s page for each race and are one statewide unit. Missouri 2004–2016 has Kansas City, which the county file leaves out, put back into Jackson County. Three elections the file lacks are drawn from the county table on their Wikipedia page: Utah’s 2010 special, West Virginia’s 2011 special and Wisconsin’s 2012 recall.
  • Senate 1914–1974 and governors 1868–1975 use the same county files, which count each party’s own line only and name only the Democrat and the Republican. So every race was set against Wikipedia twice: the page for that year’s elections (each candidate’s party and share) and the race’s own page (each candidate’s votes). Where the file’s Democrat-Republican margin is within 2 points of Wikipedia’s (888 Senate races and 1,506 governor’s), the state is its counties added up and the counties are drawn. Where it is further off, or the file has no such race (106 Senate, 97 governor’s), the state is Wikipedia’s count and its counties are left blank: the race page’s own count where it has one, otherwise one rebuilt from the year page’s shares. For 20 more governor’s races the race page has a county table that adds up to its result, and the counties are drawn from that. Eight races the file lacks have no vote count on Wikipedia either: their shares are hung on the state’s vote for president that year, and the card says the counts are estimates. Races whose top two were not one Democrat and one Republican, or that had one name on the ballot (142 and 237), are named from Wikipedia and hatched. Nine are left out because Wikipedia gives no usable result, six Senate specials among them; Maryland’s governor in 1969 was chosen by the legislature. Twelve could not be checked and stand as the file has them. In six governor’s races the leader in the count did not take office, and the card says so. The Senate starts in 1914, the first elections after the Seventeenth Amendment.
  • Where a race’s Wikipedia page has a county table that adds up to its statewide result (871 of the 3,808 Senate and governor’s races drawn by county), every county was compared with the map’s: 40,341 of 40,920 match within half a percent. The other 579 stand as the county file has them for now. 64 of those are Louisiana’s parishes in the 2022 Senate race, where the map adds up every candidate of each party and Wikipedia’s table has only the leading two.
  • The county map is today’s. A county that did not exist yet is blank in earlier elections: La Paz, Arizona before 1984; Cibola, New Mexico before 1982; Broomfield, Colorado before 2002. Counties since merged away are added to the county that took them in.
  • County president 2024: the MIT Election Lab’s certified county returns (County Presidential Election Returns 2000–2024, CC0). Virginia, where that file’s county sums miss the certified count, uses tonmcg’s county compilation instead. In Maine and Rhode Island the shortfall is ballots counted only statewide (Maine’s overseas ballots, for one), shown as ballots not counted by county. Connecticut 2024 is by its eight legacy counties from Wikipedia’s town-based table.
  • Pull divides by the certified statewide total. County files leave out some write-ins and late counts, so county sums can miss the state margin: worst are New York 2012 (1.8 points), Mississippi 2004 (1.4) and Maine 2004 (1.0).
  • Alaska counts election-day votes by precinct but absentee, early and questioned ballots only by state House district, so it has no official borough results. For the 2024 presidential and 2022 governor’s races its boroughs are estimates: each district’s absentee, early and questioned vote is shared among its boroughs in proportion to each party’s election-day vote there (precincts placed in boroughs by MIT’s 2024 precinct file; Chugach and Copper River shown together as Valdez-Cordova). They add up to the certified statewide count. In every other election Alaska is one statewide unit.
  • Kansas City is folded into Jackson County, Bedford city into Bedford County, Kalawao into Maui. In Missouri’s Senate races of 2004–2018 the county file leaves Kansas City out, so Jackson County is given what the statewide count has left over.
  • An independent who sat with one party and had no challenger of note from it is counted as that party: Bernie Sanders in Vermont (2006, 2012, 2018, 2024), who won the Democratic primary each time and declined the nomination. His counties are from the county tables on the races’ Wikipedia pages. Independents who faced a real Democrat or Republican (Angus King, Joe Lieberman in 2006, Harry Byrd Jr.) stay independents, and so do Farmer-Labor and Nonpartisan League senators, who belonged to another party.
  • Senate maps show each state’s regular race. A special election joins its year’s map when it was the state’s only Senate race that year (Massachusetts in January 2010); one that shared a year with another race in its state, or fell in an odd year, is under “Senate special elections”. Off-cycle specials are from each race’s Wikipedia page. Where a runoff settled the seat, the runoff is the race on the map, and the state’s card has a switch to the first round where that is on file (Georgia 1992, 2008, 2020 twice and 2022; Louisiana 2002). County results exist for only one of the two rounds in each. Louisiana 2014 and 2016, Mississippi’s 2018 special and Texas’s 1993 special show the runoff alone. Every race is tied to one of the state’s two seats, so “the election before” means that seat’s last election. States are hatched when the top two finishers were not one Democrat and one Republican.
  • County results for the 2022 and 2024 Senate races and the 2022 governor races are added up from MIT’s precinct returns, and kept only where both parties land within 2% of the certified statewide count. That leaves out Indiana (both years) and California’s 2022 Senate race; Tennessee’s 2022 governor race is not in the file. Governor statewide figures are the county sums.
  • Governor races of 2021, 2023 and 2025 (New Jersey and Virginia; Kentucky, Louisiana and Mississippi): the county table on each race’s Wikipedia page, taken from the state’s certified canvass. Louisiana’s 2023 race was settled in the October open primary, so the map shows that primary’s leading Democrat and Republican. The Senate specials of Oklahoma (2022) and Nebraska (2024) are by county from the MIT Election Lab’s precinct returns, which add up exactly to the certified results.
  • 2024 governor races: county results added up from the MIT Election Lab’s official 2024 precinct returns. Where that file is incomplete or counts a county twice, the county comes from the county table on Wikipedia’s page for the race: in Indiana, Carroll, Hendricks, Monroe and Tippecanoe; in North Dakota, Burleigh and Divide. Each state then matches its certified total to within 0.3%.
  • 2026 Senate and governor races (preliminary, not yet certified): county results from the state election offices of AK, AR, AZ, CA, CO, FL, GA, IA, ID, LA, MI, MN, MT, NC, NE, NM, OK, OR, PA, RI, SC, SD and VA; for AL, CT, DE, HI, IL, KS, KY, MA, MD, ME, MS, NH, NJ, NV, NY, OH, TN, TX, VT, WI, WV and WY, AP’s count as published by NBC News until each state certifies. Towns are added up to counties in New England. The 2026 national House vote is the Cook Political Report tracker as of November 17.
  • New York’s county results for 2004, 2008, 2012, 2016 and 2024 count every line a nominee ran on (Democratic and Working Families, Republican and Conservative, and so on), from the county tables on Wikipedia’s page for each election; the county file this map otherwise uses counts only the party’s own line in those years.
  • Alaska reports its votes statewide, not by borough, so it is left out of the totals by kind of place.
  • A selected county’s place type (big city, suburb, medium metro, small metro, small town or rural) is the National Center for Health Statistics’ 2013 urban-rural class (large central metro, large fringe metro, medium, small, micropolitan, noncore). It is shown for every election, so for older ones it describes the county as it is now, not as it was then.
  • City names and locations are from GeoNames (CC BY 4.0). A selected county shows its largest place, plus any others of at least 100,000 people or a third of the largest. A selected market shows its largest place and up to four more (five on a whole market) of at least 150,000 people, or 40,000 and a quarter of the largest, leaving out any that sit too close to one already named to label.
  • Electoral votes are counted as the results pledged them, so electors who voted for someone else are ignored. The tipping-point state is the one holding the winner’s 270th electoral vote, counting from the safest state down; the Electoral College edge is its margin minus the national popular vote. Maine’s and Nebraska’s stray districts are counted separately, using approximate district margins.
  • Two-party drops every vote not cast for the Democrat or the Republican, so the two shares add up to 100 and a strong third-party year (1980, 1992, 1996) is measured like any other.
  • Tap a candidate on a card to shade the map by that candidate’s share of every vote cast (one election at a time; tap again to go back). A third candidate’s county shares are his own votes where Wikipedia’s county tables have them, and otherwise every vote not cast for the two nominees, so outside his strongholds they run high.
  • Markets are Nielsen’s television markets (DMAs) as of 2016: 210 groups of whole counties, the areas campaigns buy local TV ads by. The county lists come from a public transcription of Nielsen’s 2016 list (Nielsen sells the official file), with its typing slips put right, and were checked against a second public source, an older set of market shapes: 2,992 of 3,113 counties agree, and where they differ (mostly counties on a market’s edge) the 2016 list is used. A few counties Nielsen splits between two markets are put whole in the market with the larger part (Kern, Riverside, El Dorado and Solano, California; Apache, Arizona; Lea, New Mexico; Oneida, New York). Alaska’s three markets are one unit, as Alaska is on most maps. The same 2016 markets are used for every election, back to 1868, long before there was television: a county on an old map goes to the market of the county that holds its middle today. On the national map of a presidential election a market is whole, across state lines; elsewhere, and whenever the map is shaded by State Pull or each state is set to 50-50, it is cut at state lines into its part in each state (Kansas City (KS) and Kansas City (MO)), so each part pulls on its own state.
  • US margin moves every place by the same amount, a uniform swing, so the map shows how places would fall with the country somewhere else. It has three settings that hold from year to year: 50-50 (each election with its national vote tied), each year’s own national House vote (a presidential election set to how the country voted for the House that day; for a Senate or governor’s race this is the result as counted), and a custom margin. A custom margin stays fixed as you move between elections, and when it was taken from a particular year’s vote it says so, as in D+13.6 (1976 House vote). Open the panel from the dotted words or the small bar under a card’s big number, or from the US margin setting, then drag the bar under the pointer, tap a state to move the line just past it, type a margin, or pick another year’s. On the small bar, the arrow marked US is the nation and the dot is the place on the card. The other setting, each state at 50-50 takes each state’s own margin off its counties instead. Real elections do not swing uniformly, so read it as a rough guide; vote counts shown in italics are estimates from the adjusted shares. A presidential election is measured against the national vote for president, a Senate or governor’s race against the national vote for the House in the same year, taken from Wikipedia’s page for each year’s House elections (1870–2024), because the presidential vote and the rest of the ballot were far apart for much of the twentieth century. The House vote is rougher: a seat with one candidate counts for one side only. Races in odd years have no House vote.