How Much of 2024 Was Already Decided in
2020
One election against the next, one office against
another, and the sliver of the map that was genuinely in play
Democracy’s Data
Last updated September 15, 2026 at 4:42p EDT
An election year runs for months: primaries, conventions, debates, a
billion dollars of advertising, and a night of maps changing colour on
television. All of that spectacle carries an assumption — that the
results are genuinely up for grabs. Do election results actually change
from one election to the next?
The question has a citizen’s stake, not just an analyst’s. If nearly
every place’s result can be written down four years in advance, then
campaigns and their attention are aimed at a sliver of the country. Most
voters are being watched rather than courted. This brief measures how
much of 2024 was already fixed in 2020, at the county level and then
across offices on the same ballot.
The test
Two comparisons need two files. The first sets every county’s 2020
presidential vote against its 2024 vote, using the county returns
compiled by the researcher Tony McGovern from news organizations’
results pages and downloadable at https://github.com/tonmcg/US_County_Level_Election_Results_08-24
— a file whose own README calls its figures not authoritative, and which
this book’s county-returns
chapter checks against the fifty-one official state publications. No
agency publishes county returns nationally, so a private compilation is
the only kind of file there is. The second comparison sets offices
against each other in 2024: president against Senate. It uses the Clerk
of the House’s Statistics of the Presidential and Congressional
Election, the official tally the Clerk compiles from the state
canvasses — each state’s formal count of its own returns — and published
in March 2025.
The Clerk’s document fits the second test for one specific reason: it
prints the presidential, Senate and House results for a state on the
same pages, from one canvass, on one date. The measure below is a
subtraction, how a state voted for the Senate minus how it voted for
president, and a subtraction is exactly where two differently built
files leak their differences into the answer. Taking both numbers from
one document closes that door. What a certified return can bear at all
is this cluster’s
chapter on election returns.
The data
One row of the county file is one county, with both elections beside
each other:
| Maricopa County, Arizona |
48.9% |
51.8% |
+2.9 |
| Allegheny County, Pennsylvania |
39.6% |
39.8% |
+0.1 |
| Starr County, Texas |
47.5% |
58.0% |
+10.6 |
| county_fips |
the five-digit federal code for the county |
categorical |
| r20 |
the Republican share of the votes cast for the two
major parties in 2020 |
continuous |
| r24 |
the same quantity in 2024 |
continuous |
| swing |
the second minus the first, in points |
continuous |
| flipped |
whether the county backed a different party in the two
elections |
yes or no |
The share is of the two major parties, not of every
vote cast. Third-party votes were not the same size in the two
elections, and a share-of-everything comparison would read that
difference as movement between the parties, which it is not.
3,104 counties appear in both files and are drawn below. 48 reporting
units from 2020 and 56 from 2024 have no partner in the other year, and
the reason is geography rather than a broken key. Alaska reports by
state legislative district and renumbered them; Connecticut replaced its
eight counties with nine planning regions in 2022; the District of
Columbia is one row in 2020 and eight in 2024. The ground was recut, so
there is nothing to join. The Senate file has one row per contest, with
each contest’s Republican share beside the state’s presidential one — 35
contests in 33 states, since Nebraska and California each held two.
Before you read on, write down two numbers. Knowing
only how a county voted in 2020, how close do you think you could come
to its 2024 result — within twenty points, ten, five? And how many of
the 3,104 counties do you think switched parties between the two
elections? Commit to both.
What it says about democracy
Figure 1. Every county, twice. Each dot is one
county: left to right is how Republican it voted in 2020, height is how
Republican it voted in 2024. A scatter of one election against the next
fits the question because persistence has a shape, the diagonal line —
any county that escaped its past sits visibly off it. The shaded boxes
are the only places a county that switched parties can land, and the
ringed dots are the ones that did.
Persistence wins, and it wins hard. The correlation between the two
years, meaning a measure of how far the counties stay in the same order
where 1 is a perfect line, is 0.9950. Draw the best
straight line through this cloud and it misses a typical county by 1.6
points. Knowing only 2020, you could have come that close to almost
every county’s 2024 result before a single ballot was cast, and 96.7% of
counties finished within five points of where they started.
Only 86 of the 3,104 counties changed sides, and every one
moved the same way: all 86 went from Democratic to Republican,
and the count in the other direction is 0. That is why one shaded box
holds every ring and the other is empty. A country whose places were
trading with each other would have filled both.
Two cautions keep the finding honest. A correlation measures
order, not sameness: press the button on Figure 1 and
every county moves ten points Republican, a national landslide, while
the correlation stays 0.9950. The typical county did move, by +1.5
points toward the Republicans, and Maverick County, Texas moved +14.1.
Persistence rules out places changing places; it says nothing against
the whole field moving at once. And counties are not people: weighting
by votes cast, which means counting some counties more than others,
drops the correlation only to 0.9918.
So one election predicts the next. Does one office predict another,
on the same day, on the same ballot?
Figure 2. The two offices, on one pair of axes. This
is Figure 1 with one axis replaced: the state’s Senate vote instead of
the next presidential election. Colour is the party of the senator
elected, an asterisk marks the two contests for the last weeks of a
term, and the magnified panel shows the crowded middle. The dashed line
is a Senate candidate finishing exactly level with their party’s
presidential ticket.
Almost every point sits on or just under the line. The two offices
correlate at 0.952, and the typical Senate candidate came within 2.2
points of their ticket. The gaps that do open are small and one-sided:
Larry Hogan ran 8.7 points ahead of Trump in Maryland and still lost,
and Deb Fischer ran 7.1 points behind him in Nebraska and still won.
The exceptions are where the sliver decides everything. 4 contests,
in AZ, MI, NV, WI, produced a senator whose party did not carry the
state for president — a crossover, in this brief’s
vocabulary: a place that voted one way for president and the other way
for another office on the same ballot. All four went the same way: Trump
carried the state, and the voters sent a Democrat to the Senate. In
each, the Republican Senate candidate ran an ordinary 0.9 to 4.0 points
behind Trump. That ordinary gap changed the outcome only because Trump’s
own margin there was smaller than the gap. Take Arizona. Trump took
52.8% of the state’s two-party presidential vote; Kari Lake, the
Republican for Senate, finished 4.0 points behind him, and 52.8 − 4.0 =
48.8, under half, so Ruben Gallego took the seat. Those four seats did
not decide which party controlled the Senate — Republicans won a
majority without them — but they are the difference between a narrow
majority and a commanding one.
Even that count of 4 is a decision wearing the costume of an
observation. Two points in Figure 2 are grey, and grey is not a
party:
| Vermont |
Bernard Sanders |
Independent |
Harris |
32.6 pts |
| Maine |
Angus S. King, Jr. |
Independent |
Harris |
20.1 pts |
Maine and Vermont elected independents who were not Democratic
nominees, sit with the Democrats in Washington, and represent states
Harris carried. Read “a different party” as the ballot line and the
count is 6; read it as the two sides in Washington and it is 4. Neither
reading is wrong — they are different questions wearing the same
sentence, and the decision between them does not travel with the number.
When you are handed a count of a category, ask what it would take for a
case to sit on the other side of the line.
The House tells the same story at a larger scale. 16 districts out of
435 voted one way for president and the other for the House, which is
3.7% of the chamber. The widest of those splits was PA-01: the
Republican share of its two-party presidential vote was 49.9%, and of
its House vote 56.4%, a gap of 6.5 points that put the two offices on
opposite sides of fifty — so a district Harris carried sent Brian K.
Fitzpatrick, a Republican, to the House. It was not always so:
Move across the chart for each election’s share.
Figure 3. Districts splitting their vote between president
and House, 1952–2024. Only presidential years are drawn, so
every point compares a House election with the presidential election
held the same day.
Split districts were once ordinary: 44.4% of contested districts
split in 1972, and the line sits above thirty per cent in 6 of the 19
elections drawn. It falls through the 1990s and keeps falling, and the
last two elections are the two lowest readings in the series. One
caution about the denominator, the bottom number the percentage is
computed over: the line counts only districts where both parties ran, so
2024 reads 4.0% there and 3.7% over the whole chamber. Neither is wrong,
and quoting one where the other belongs is the commonest way this
statistic is misused.
What this brief cannot tell you
No voter appears anywhere in this brief. Every
number is a total for a place. A state that voted for Trump and elected
a Democrat could be one where many people split their ballots, or one
where different sets of people turned out for the two contests. Guessing
what individuals did from group totals is called ecological inference, a
problem with a seventy-year literature of its own, and nothing here is
evidence about any person’s ballot.
A high correlation across places is not stability over
time. Figure 1 rules out counties changing places with each
other. It does not rule out the whole country moving, and the whole
country moved +1.5 points in the typical county. Both sentences describe
the same picture.
Two elections are not a trend. The county comparison
covers one pair of years. Figure 3 is a long series for the House, but
there is no comparable series here for counties or the Senate, so
whether 2020 predicted 2024 unusually well is a question this
brief cannot answer.
The vertical axis of Figure 2 is not always Republican
against Democrat. It is the Republican against whoever finished
closest, which in three contests was an independent. A strict
Republican-against-Democrat axis would report Maine as three-to-one
Republican and Nebraska as unopposed, so this brief does not use one;
the cost is three points whose presidential axis has no true
counterpart.
Running ahead of the ticket is not a measure of a
candidate. The gap from the dashed line has two sides: a strong
candidate, a weak opponent, or a state full of willing ticket-splitters
all produce it, and this brief cannot tell you which did.
The county file is a compilation, not a public
record. No agency publishes it, and the 2020 file is the weaker
of the two used here; the county-returns
chapter measures how far such compilations sit from what the states
published.
What you should have learned
- County results persist: 2020 and 2024 correlate at 0.9950, a
best-fit line misses the typical county by 1.6 points, and 96.7% of
counties finished within five points of where they started.
- Change was real but uniform: the typical county moved +1.5 points,
and all 86 counties that switched parties switched in the same
direction.
- On one ballot, offices move together: the Senate vote correlates
with the presidential vote at 0.952, and the typical Senate candidate
ran within 2.2 points of the ticket.
- The mobile sliver decides control: 4 Senate contests crossed over,
each because an ordinary gap from the ticket met a presidential margin
smaller than the gap.
- A published count hides a definition: counting senators by ballot
line instead of by caucus moves the crossover count from 4 to 6.
Extensions
None of these is answered above, and every one can be answered from
the tables the Sources section hands over.
counties.csv carries swing for every
county. Sort by it. Do the counties that moved most have anything in
common — size, state, how lopsided they already were in
r20?house.csv has ran_ahead for every
contested district and a crossover flag. Count the
crossover districts, then look at which party’s candidates hold them and
how far ahead of the ticket they ran.split_trend.csv gives pct_split for every
presidential year. Find the last year it sat above twenty per cent. What
was the first election after it where the figure had halved?senate.csv compares each Senate result against the
presidential result in the same state, in ran_ahead. Find
the candidates who ran furthest ahead of their party’s ticket. Did any
of them win?states.csv holds each state’s presidential result in
pres_r_two. Compare its closest states with the
margin column of senate.csv: did the closest
presidential states also hold the closest Senate races?
One stretch idea points past the spreadsheet: when the 2026 midterm
returns are certified, download the Senate results and see whether
crossover wins survive without a presidential race on the ballot.
Another: pick one flipped county from Figure 1 and pull its precinct
returns from the state’s election site, to see whether the county moved
evenly or in patches.
Sources
Data
- U.S. House of Representatives, Office of the Clerk. Statistics
of the Presidential and Congressional Election of November 5, 2024,
published 10 March 2025. Read through this book’s house-competition
chapter, which holds a copy. Every presidential, Senate and House
result in this brief comes from that document. https://history.house.gov/Institution/Election-Statistics/Election-Statistics/,
every volume from 1920 on. The Clerk’s own address for the series, https://clerk.house.gov/Votes/ElectionStatistics, no
longer resolved when this was last checked.
- County presidential returns for 2020 and 2024: Tony McGovern, US
County Level Election Results 08-24, compiled from news
organizations’ results pages and, by its own README, not authoritative.
The data-sources
chapter describes it and the county-returns
chapter checks it against the fifty-one official state publications.
https://github.com/tonmcg/US_County_Level_Election_Results_08-24
- The long House series in Figure 3 is the house-competition
chapter’s, built from Gary Jacobson’s House elections file and the
Clerk’s own documents.
Academic literature
The data itself
Every figure above is drawn from these tables. They are plain CSV,
they open in a spreadsheet, and they are yours to take further.
counties.csv, house.csv, senate.csv, split_trend.csv, states.csv
Two more files here are not data about the subject at all – they are
how this chapter checked its own work: facts.csv, the individual
numbers the text above quotes, and checks.csv, the
arithmetic verified before anything was drawn.
