Feeling Thermometers and Affective
Polarization
Nobody started liking their own side more
Democracy’s Data
Last updated September 15, 2026 at 4:42p EDT
Americans hate the other party more than ever. You have
heard the claim, and it has a number behind it. Since 1978 the American
National Election Studies — ANES, the academic survey of American voters
run by the University of Michigan and Stanford University — has asked
people to rate the two parties on a feeling
thermometer: a hundred is warm, nought is cold, fifty is
neither. Take a partisan, someone who identifies with one of the
parties. Subtract their rating of the other party (the
out-party, in the jargon) from their rating of their own (the
in-party) and you have the standard measure of
affective polarization — dislike across party lines, as
distinct from disagreement about policy.
Here is the number being built, for the first year in the series. In
1978 the average partisan rated their own party 70.6 degrees and the
other party 48.0; 70.6 − 48.0 = 22.6 degrees of polarization. That
number has more than doubled since, to 49.5 in 2024. The doubling is
real. But it is a subtraction, and a subtraction has two halves that the
headline cannot see. The distance could be growing because people warmed
toward their own side, or because they cooled toward the other. This
brief asks the question the number skips: which half moved?
The test
The measure is the pair of party feeling thermometers in the ANES
Time Series Cumulative Data File, 1948–2024, a single table collected by
the University of Michigan and Stanford University for the American
National Election Studies, which stacks every study since 1948 into one
file. It is free, but must be requested through a browser with an
account at electionstudies.org.
The same two questions, in the same words, sit in every study this brief
reports — 18 of them across forty-six years. The ANES chapter introduces the
cumulative file and its codebook, the document that says what every code
in every column means; this brief pulls two columns out of it.
A feeling is the kind of thing only a survey can hold. Nobody
registers a temperature toward a party, and no election ever comes along
to grade the answer the way a vote count eventually grades a horse-race
poll. The question’s construction is therefore not a detail under the
measurement — it is the measurement. A forty-six-year series on
one pair of questions exists only because ANES staff kept asking the
same words, study after study, and kept the codes meaning the same thing
throughout.
The test is one that arithmetic can settle. Undo the subtraction:
compute warmth toward one’s own party and warmth toward the other party
separately, year by year, and see which of the two series produced the
doubling.
The data
The cumulative file is one table, one row per interview: 73,745
respondents from 1948 to 2024. Four columns do the work here.
VCF0218 is the thermometer for the Democratic Party and
VCF0224 the one for the Republican Party.
VCF0301 is the seven-point party identification, used to
decide which party counts as a respondent’s own. VCF0009z
is the weight, which counts some answers more than others so that the
sample’s mix of ages, races, regions and so on matches the country’s,
with the targets taken from the census. Unlike most of this book’s ANES
briefs, every average below applies it.
Two numbers in the thermometer columns are not temperatures, and the
data does not warn you. The scale everybody describes as nought to a
hundred is coded 0 to 97 in this file. It holds 0 ratings of exactly
100, because a hundred is not a code the file uses. What the file does
use is 98 for don’t know and 99 for
not ascertained, the interviewer’s term for an answer that was
never obtained, and 2,183 and 934 entries carry them. A 98 sits inside
the range a reader expects, so it does not look wrong. It looks like
somebody who feels very warmly indeed.
That fact lives in the codebook, not the column, and the cost of
missing it is measurable. Believe the codes and the country warms by 1.4
degrees: the average rating rises from 51.5 to 53.0. Worse, the codes
are not spread evenly across years, so the mistake bends the trend
rather than just shifting it. Everything below drops them, along with 20
negative codes.
One more thing worth knowing about the column before averaging it:
the scale is coarser than it looks. The thermometer offers 98 positions,
and people use about four of them.
| 0 — as cold as possible |
10.1% |
| 50 — the midpoint |
16.1% |
| 85 |
10.5% |
| 97 — the top of the scale |
6.6% |
70.7% of all 97,268 ratings are multiples of ten. The 85s are the
tell that this is habit rather than measurement. Nobody feels exactly 85
degrees about a political party, but it is where the hand goes when the
answer is “warm, not the maximum.” A number this fine-grained, used this
coarsely, still averages perfectly well. It just does not mean what its
precision suggests.
The series itself is built from partisans, with leaners — people who
call themselves independent but admit to leaning toward a party —
counted alongside: the standard choice, and a choice, which the party identification brief
shows moving a different headline number by a factor of five. Pure
independents, 5,327 of them, have no own party and sit outside every
figure here. A study year must carry at least 200 partisans to be
reported, which leaves 18 study years between 1978 and 2024. Two small
tables come out: thermometers_by_year.csv, with each year’s
warmth toward one’s own party, toward the other, and the gap; and
out_party_zero.csv, the share rating the other party at
exactly nought.
Before you read on, commit to two numbers. In 1978,
partisans rated their own party 70.6 degrees and the other party 48.0.
By 2024, what had each of those become? Write both down.
What it says about democracy
Here is the subtraction, undone.
Figure 1. Warmth toward one’s own party and toward the other,
among partisans including leaners, 1978–2024. Two lines on a
time axis, because the claim is a claim about which of two series moved,
and the shaded band between them is the polarization number itself.
Hovering reads both halves and the gap at any year.
The top line does not move. Partisans rated their
own party 70.6 degrees in 1978 and 70.1 in 2024, a change of 0.5 of a
degree across forty-six years. Through every realignment, every scandal,
and every president in between, the feeling of partisans toward their
own party has been a flat line.
The bottom line falls off the table. Warmth toward the other party
went from 48.0 to 20.6, a drop of 27.4 degrees, and it now sits far
below the fifty the scale calls neither warm nor cold. Redo the
subtraction with the new readings: 70.1 − 20.6 = 49.5. The first number
barely changed; the second did all the work. Put as one number: 98.1% of
the movement in the gap is the out-party half. Nobody started liking
their own side more. They stopped being able to stand the other one. The
gap peaked at 51.9 degrees in 2020, the same year the other party’s
rating touched its floor of 19.2.
The averages also hide how the fall happened, and it is not
everybody sliding down a few degrees.
Figure 2. The share of partisans rating the other party at
absolute zero. One line, because the question is how common the
floor became — the average in Figure 1 fell partly because a growing
group stopped using the scale and went to its bottom.
In 1978, 5.4% of partisans gave the other party the coldest rating
available. In 2024 it was 33.2%. That is a third of partisans choosing
the absolute bottom of a 98-position scale, an answer that reads less
like a temperature than like a refusal to have one.
This matters for how the country gets described. The single number
that gets reported cannot show any of it, because a difference has no
memory of which side moved. Two series drifting apart and one series
collapsing produce the same gap, and they are different stories about
democracy. The first is mutual estrangement; the second is one feeling,
out-party dislike, doing all the work.
The gender gap
brief runs the same arithmetic on a different question and reaches
the same lesson. Ask for the pair. It is one extra number, and it
changes the sentence.
What this brief cannot tell you
A thermometer reading is not a behaviour. It is a
number somebody said to an interviewer. Nothing here shows anyone
refusing a neighbour, a hire, or an in-law over party, and the
literature linking affective polarization to behaviour is contested.
The object being rated changed. “The Republican
Party” and “the Democratic Party” meant different things in 1978 than in
2024. A respondent rating a party colder may be feeling differently
about the same object, or the same about a different one, and the file
cannot separate those.
These are cross-sections, not the same people. Each
study draws a fresh sample, so different respondents answer every year.
A falling series can mean individuals changed their minds, or that the
people replaced by generational turnover felt differently, and those are
different countries. Only a panel — a survey that goes back to the same
people — could tell them apart.
And the series has thin stretches. A year needs 200
partisans to appear, which leaves 18 study years across forty-six years.
The lines connect points that are sometimes several years apart, and the
connecting segments are drawn, not measured.
What you should have learned
- Affective polarization in the ANES is 49.5 degrees, up from 22.6 in
1978, with a peak of 51.9 in 2020.
- The gap is a subtraction, and essentially one half moved: own-party
warmth changed 0.5 of a degree while other-party warmth fell 27.4 —
98.1% of the movement is the out-party half.
- Part of the fall is a floor: the share of partisans rating the other
party at exactly nought went from 5.4% to 33.2%.
- Two codes in the thermometer columns, 98 and 99, are not
temperatures; believing them warms the country by 1.4 degrees and bends
the trend.
- The scale is used far more coarsely than it looks: 70.7% of ratings
are multiples of ten.
Extensions
None of these is answered above, and every one can be answered from
the tables the Sources section hands over.
thermometers_by_year.csv has in_party and
out_party. Divide one by the other, year by year. When does
the other party first earn less than half the warmth of one’s own?thermometers_by_year.csv has gap for every
reported year. Work out how much of the total rise arrived in each
decade. Is the doubling steady, or does one stretch carry most of
it?out_party_zero.csv has pct_zero. Find the
year it first passes 20%, and decide whether the climb to a third is a
ramp or a staircase.thermometers_by_year.csv carries n for
every year. Find the thinnest years, and decide whether the shape of the
trend survives without them.
Two stretch ideas point past the spreadsheet. The same ANES file
carries thermometers for dozens of other objects: unions, big business,
liberals, conservatives. A reader with the cumulative file can test
whether coldness rose toward out-groups in general or toward the
out-party alone. And the Pew Research Center publishes its own series on
partisans holding “very unfavorable” views of the other party, for
instance in its 2022 report on partisan hostility (https://www.pewresearch.org/politics/2022/08/09/as-partisan-hostility-grows-signs-of-frustration-with-the-two-party-system/);
check whether its timing matches Figure 2’s climb.
Sources
Data
- American National Election Studies. ANES Time Series Cumulative
Data File (1948–2024), version of 5 February 2026, CSV distribution
anes_timeseries_cdf_csv_20260205.csv. Ann Arbor, MI and
Palo Alto, CA: University of Michigan and Stanford University. https://electionstudies.org/data-center/anes-time-series-cumulative-data-file/
- ANES. Time Series Cumulative Data File Codebook, same
release. The meaning of codes 98 and 99 was read out of it rather than
assumed.
- The file must be requested through a browser with a free account,
and is not redistributed with this book; the ANES chapter describes what it costs
to get. Every number above is computed from the two tables below at the
moment this document is rendered, and checked against these
assertions:
| the whole cumulative file arrived |
yes |
| there is no value of 100 in a column described as
running to 100 |
yes |
| the codes that masquerade as warmth are present and
worth removing |
yes |
| believing the codes warms the country |
yes |
| the scale is used far more coarsely than its 98 points
allow |
yes |
| the series spans at least four decades |
yes |
| the gap is wider at the end than at the start |
yes |
| out-party warmth moved further than in-party
warmth |
yes |
| rating the other party at absolute zero became far more
common |
yes |
Academic literature
- Iyengar, Shanto, Gaurav Sood and Yphtach Lelkes (2012). “Affect, Not
Ideology: A Social Identity Perspective on Polarization.” Public
Opinion Quarterly 76(3): 405–431. The paper that made this
subtraction the standard measure. https://www.jstor.org/stable/41684577
- Iyengar, Shanto and Sean J. Westwood (2015). “Fear and Loathing
across Party Lines: New Evidence on Group Polarization.” American
Journal of Political Science 59(3): 690–707. Whether the feeling
reaches behaviour. https://www.jstor.org/stable/24583091
- Iyengar, Shanto, Yphtach Lelkes, Matthew Levendusky, Neil Malhotra
and Sean J. Westwood (2019). “The Origins and Consequences of Affective
Polarization in the United States.” Annual Review of Political
Science 22: 129–146, doi:10.1146/annurev-polisci-051117-073034.
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.
out_party_zero.csv,
thermometers_by_year.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.
