Claims about polarization are only as good as the measurements behind them. Some measures look at how politicians vote, some at what citizens say in surveys, and some at how people actually behave when party is quietly in the background. Each has strengths and blind spots, and knowing them protects you from headlines that overstate or understate what is happening.
Measuring politicians: roll-call votes
The best-known measure of elite polarization is DW-NOMINATE, developed by the political scientists Keith Poole and Howard Rosenthal beginning in the 1980s. It places every member of Congress on a scale using their roll-call votes: legislators who vote together end up close together. Applied to the whole history of Congress, it shows that the parties were far apart in the late 1800s and early 1900s, came much closer together in the middle of the twentieth century, and have pulled apart steadily since the 1970s. By the 2000s, overlap between the parties had largely vanished: the most conservative Democrat was usually to the left of the most liberal Republican.
Roll-call measures have limits. Party leaders control which bills come to a vote, so the votes themselves are partly shaped by strategy. A legislator may vote with the party out of discipline rather than conviction. Still, the long-run trend is one of the most robust findings in political science.
Measuring voters' opinions
For ordinary citizens, researchers ask batteries of policy questions and see whether answers cluster. The Pew Research Center found that the share of Americans with consistently liberal or consistently conservative views rose from about 10 percent in 1994 to about 21 percent in 2014. That is a real change, but it also means that most Americans in 2014 still held a mix of views. Scholars disagree about how to read this. Some, such as Alan Abramowitz, see a public that has genuinely polarized, especially among the most engaged. Others, such as Morris Fiorina, see mostly sorting plus a loud, active minority.
Measuring feelings: thermometers
Affective polarization is usually measured with a feeling thermometer, a survey question that asks people to rate a group from 0 (very cold) to 100 (very warm). The American National Election Studies have asked this about the parties for decades. Ratings of one's own party have stayed fairly steady, while ratings of the other party have fallen sharply, from the high 40s around 1980 to the low 20s or below by 2020. The gap between the two, warmth for one's own side minus warmth for the other, is the standard summary measure.
Measuring social distance
Another approach asks how people would feel about close contact with the other side. In 1960 the researchers Gabriel Almond and Sidney Verba asked Americans whether they would be displeased if their child married someone from the other party. Only about 5 percent of Republicans and 4 percent of Democrats said yes. When Iyengar and colleagues asked a similar question in 2010, about half of Republicans and a third of Democrats said they would be unhappy. The wording was not identical, so the exact jump should be treated with caution, but the direction is clear.
Measuring behavior
Survey answers can be cheap talk. So researchers also run behavioral experiments, in which party is a small detail and the researcher watches what people actually do. In a 2015 study, Iyengar and Sean Westwood asked participants to choose between two high school seniors for a scholarship. The résumés differed in grades and in small clues about party, such as membership in a Young Republicans or Young Democrats club. Partisans chose the student from their own party most of the time, even when the other student had better grades. In the same study, party bias was larger than racial bias. Other experiments have tested whether people pay co-partisans more in economic games or prefer them as roommates or employees, with similar results.
Cautions
Several cautions apply to all of these measures. Surveys can attract inattentive respondents who click at random, which can inflate extreme answers. Question wording shifts over decades. And measures built for two-party systems need adjusting for countries with many parties, where hostility may be aimed at one party rather than at "the other side" as a whole. A good rule is to trust trends that show up across several different kinds of measure, and to be skeptical of any single dramatic number.
Worked example: reading a headline
Suppose a news story says, "Americans have never been more divided: 80 percent say the country is split." What does that number actually measure? It is a perception of division, not a measurement of policy views or of personal hostility. To judge the claim you would want to see policy questions over time (ideological), thermometer ratings or social-distance questions (affective), and ideally a behavioral study. You might find that the perception of division has risen faster than real disagreement, which would itself be an important finding: a public that believes it is hopelessly divided may act as if it is, even when it is not. Asking "which kind of polarization does this measure?" is the single most useful habit this level can give you.