This paper was generated by AI from the lab's code, data and documentation.
Abstract
The America's Political Pulse survey is a high-frequency study of affective polarization, support for violations of democratic norms and support for partisan violence among American adults. From September 2022 through July 2025 the survey interviewed approximately 1,000 adults from YouGov's online panel each week, and it has since been fielded about once a month; through June 2026 it comprised 157 surveys, 167,000 interviews and 99,000 respondents. The core instrument comprises feeling thermometers toward the two parties, five items on democratic norms and six concrete scenarios of escalating partisan violence, all asked in the same form in every survey. Pooled across surveys, support for arson, assault with a deadly weapon or murder ranges from 1.2 to 2.6 percent of partisans, depending on the act and the party. We also estimate state-level opinion by multilevel regression and poststratification and field the core instrument in five other countries. This paper describes how the survey is designed and why, how to use the public data, and how to reproduce the published estimates.
1 Introduction
The survey tracks three quantities: how much partisans dislike the opposing party, whether they would tolerate their own side's violations of democratic norms, and whether they support violence against the other side. It asks the same core items in the same form in every wave, so that change over time and after political events is not confounded with changes in wording, and it pools waves to estimate opinion in each state.
Two design choices follow from evidence about how these attitudes are mismeasured. Support for violence is measured with concrete, escalating scenarios rather than abstract statements that violence is sometimes justified. Abstract items leave respondents to imagine the act in question, and disengaged respondents who answer arbitrarily register as supporters because most response options on the standard scales indicate support; together these features inflate published estimates, which have ranged from 8 to 40 percent of Americans (Westwood et al. 2022). Each norm and violence item is also paired with a question about how much of the opposing party would agree, because the gap between perceived and actual support is itself a quantity of interest.
Sections 2 through 6 describe fielding, the instrument, the national and state estimates and the international surveys, with the reason for each choice. Section 7 explains how to use the public files, Section 8 the limitations of the design and Section 9 how to reproduce the published estimates.
2 Sample and Fielding
Each survey interviews approximately 1,000 American adults drawn from YouGov's online panel. YouGov selects respondents by sample matching: it draws a target sample from a frame built from the American Community Survey public-use microdata and, for each target case, selects the closest available panelist on age, gender, race and education (Holliday et al. 2024). The frame is built from the most recent ACS release at the time each wave is fielded. Because the panel is opt-in, response rates are not defined. Respondents answer online on a phone, tablet or computer. The first survey was fielded September 16–22, 2022. The ten surveys that followed, fielded between September 23 and December 2, 2022, interviewed 2,000 adults each to cover the midterm election. From December 2022 until July 10, 2025, a new survey was fielded each week, with a single missing week (May 23–29, 2025). Since late July 2025 the survey has been fielded approximately monthly, with each wave in the field for one to two weeks.
The design consists of repeated cross-sections rather than a panel. Respondents become eligible for another of our surveys three weeks after completing one, so some individuals are interviewed more than once: the 167,000 interviews completed through June 2026 were conducted with 99,000 distinct respondents. Analyses that pool surveys should treat repeated interviews with the same respondent as dependent observations, and the public data include a respondent identifier for this purpose.
Inattentive respondents inflate measured support for political violence, because their arbitrary answers fall disproportionately on response options that indicate support (Westwood et al. 2022). From September 2022 through early June 2023, every survey included an attention check consisting of a brief account of a fictional person followed by questions about the person's name, profession, location and experience. Respondents who supplied details the account never contained were classified as disengaged, 31 percent of the 47,950 interviews in that period, and estimates for that period exclude them. Since the survey that closed on June 8, 2023, the check screens respondents rather than flagging them: those who fail it do not complete the survey, so every delivered respondent has passed it. Toplines for surveys through early June 2023 report estimates for engaged respondents together with estimates for all respondents.
Duplicate interviews, identified by their case numbers, are removed when the data are loaded. YouGov removes interviews that fail its checks for speeding, straight-lining and inconsistent answers before delivery and removes repeat offenders from its panel. These checks applied to every wave. The survey is conducted under a protocol approved by Dartmouth College's Committee for the Protection of Human Subjects (STUDY00032564).
3 Instrument
The core items appear in identical form in every survey, so that their trends are not confounded by changes in wording. Many items refer to the respondent's own party and the opposing party, which are inserted for each respondent. Partisanship is measured with the standard seven-point party identification question, and independents who lean toward a party are classified as partisans, since leaners resemble partisans in their attitudes and behavior (Petrocik 2009). Pure independents are assigned one of the parties as the opposing party for items that require one.
3.1 Affective polarization
Respondents rate Democrats and Republicans on a feeling thermometer: “We'd like you to rate how you feel towards some groups on a scale of 0 to 100. Zero means very unfavorable and 100 means very favorable. Fifty means you do not feel favorable or unfavorable.” For each partisan, affective polarization is the rating of the respondent's own party minus the rating of the opposing party. The thermometer is the most widely fielded measure of partisan affect, which allows our series to be compared with other surveys. Since 2025, each respondent also rates two of six social groups selected at random: Black people, white people, Jewish people, Hispanic people, illegal immigrants and transgender people.
3.2 Support for partisan violence
Abstract questions about whether violence is ever justified leave respondents to imagine the act in question. We instead describe specific acts committed by a named man against members of the opposing party and ask, “Do you support or oppose [name]'s actions?” Responses are recorded on a five-point scale from strongly support to strongly oppose. The six acts, which escalate in severity, are summarized in Table 1.
| Item | Act | Scenario |
|---|---|---|
| 1 | Protest without a permit | Led a protest against the opposing party on the grounds of the county courthouse without a permit and refused to leave when asked by police |
| 2 | Vandalism | Vandalized several large signs supporting the opposing party's candidates |
| 3 | Assault | Threw rocks at peaceful protesters from the opposing party; no one was seriously injured, but paramedics bandaged a man with a head wound |
| 4 | Arson | Set fire to the opposing party's local headquarters after it had closed for the night, while adjacent buildings were still occupied |
| 5 | Assault with a deadly weapon | Drove his car into a crowd of protesters from the opposing party, seriously injuring several people, one of whom spent a month in the hospital |
| 6 | Murder | Stabbed a prominent member of the opposing party to death, believing the victim had helped prevent him from voting as part of a conspiracy against voters of his own party |
Each scenario states that the man was convicted of the crime, and his name is drawn at random from six alternatives so that no single name's associations drive the response. The items form a ladder: a respondent is shown each act only after supporting or strongly supporting the act before it. The ladder spares the large majority who oppose minor acts from reading descriptions of murder, and it shortens the instrument, at the cost that support for each act must be computed as a share of all respondents rather than of those shown it (Section 4). In practice the ladder is nearly strict; of 111,466 respondents who did not support the first act, 257 were nonetheless shown the second. Two further items ask respondents to estimate the percentage of the opposing party that would support the third and sixth acts, the assault on protesters and the murder.
3.3 Democratic norms
Five items ask whether the respondent's own party should violate a democratic norm to gain an advantage (Table 2; Holliday et al. 2024). Each is answered on a five-point scale from strongly agree to strongly disagree, with the respondent's own and opposing parties inserted where indicated.
| Norm | Wording |
|---|---|
| Courts | “[Own party] elected officials should sometimes consider ignoring court decisions when the judges who issued those decisions were appointed by [opposing party] presidents.” |
| Polling places | “[Own party] should reduce the number of polling stations in areas that typically support [opposing party].” |
| Executive orders | “If a [own party] president can't get cooperation from [opposing party] members of congress to pass new laws, the [own party] president should circumvent Congress and issue executive orders on their own to accomplish their priorities.” |
| Censorship | “The government should be able to censor media sources that spend more time attacking [own party] than [opposing party].” |
| Loyalty | “When a [own party] candidate questions the outcome of an election other [own party] should be more loyal to the [own party] party than to election rules and the constitution.” |
Each item is paired with a question asking what percentage of the opposing party's voters the respondent believes would agree. Respondents who rate democracy as important but endorse at least one violation are asked to justify, in their own words, one of the violations they endorsed.
3.4 Other content
A battery on values asks whether a stranger would return a lost wallet and asks about bribery, the responsiveness of public services, the importance of voting, pride in being American and whether people like the respondent are treated fairly. Policy items rotate: each respondent answers five of ten policy questions selected at random, and since January 2025 the survey has included items on immigration, the economy, tariffs and free speech. Election-year items cover primary preferences, vote choice and turnout. The survey also carries modules commissioned by partners, such as a 2026 module on higher education, which are analyzed separately and withheld from the public file. Demographic, partisan and religious profiles are supplied by YouGov.
4 Weighting and National Estimates
YouGov supplies a weight for each respondent, constructed with propensity scores: the matched cases are weighted to the frame and then post-stratified to national benchmarks for age, gender, race, education, region, party identification and income (Holliday et al. 2024). Large weights are trimmed; no weight in the public file exceeds 7.1. Within each survey the weights have a mean of one. Our national series are weighted means and weighted proportions computed for each week of fielding; since monthly fielding began, each wave's interviews are assigned to the wave's most common week, so that each wave yields one estimate. We report these estimates without smoothing, so that a sudden change after an event is not spread across neighboring weeks; the cost is that differences between consecutive surveys include sampling error. For a proportion near one half, a sample of 1,000 implies a margin of error of approximately three percentage points before accounting for the design effect of weighting, so a difference between two consecutive surveys is weak evidence of change unless it persists in later surveys.
Affective polarization is the mean, among partisans, of the difference between the thermometer ratings of the respondent's own party and the opposing party. Negative values, from partisans who rate the opposing party more warmly than their own, are kept. Support for each violent act is the proportion of partisans who support or strongly support it, where respondents who were not shown an act because they did not support the preceding act are counted as not supporting it. Each estimate is therefore a proportion of all partisans rather than of those who reached that item. Endorsement of a norm violation is the proportion who agree or strongly agree, and our summary measure, the proportion endorsing at least one violation, excludes the executive-orders item, which Holliday et al. (2024) treat as a lesser violation.
The treatment of the ladder has a large effect on the estimates. Among respondents shown them, support for the more severe acts appears high, because most of the few respondents who support assault also support arson. As a proportion of all partisans, pooled across surveys, support for arson, assault with a deadly weapon or murder ranges from 1.2 to 2.6 percent, depending on the act and the party. We report the latter quantity because it corresponds to the population proportion of interest.
Each survey also produces a topline report giving the distribution of responses to every core item. The norm and violence items are tabulated among partisans.
5 State Estimates
A single survey contains too few respondents in most states to support direct estimates; a typical wave includes only a handful of respondents from Wyoming or Vermont. We therefore estimate state opinion by multilevel regression and poststratification (Park, Gelman and Bafumi 2004), pooling every survey since September 2022. Pooling trades timeliness for precision: the estimates describe the entire period rather than the most recent month.
For each binary measure we fit a multilevel logistic regression to engaged respondents,
\Pr(y_i = 1) = \operatorname{logit}^{-1}\big(\beta_0 + \beta_{\text{age}[i]} + \beta_{\text{race}[i]} + \beta_{\text{educ}[i]} + \beta_{\text{gender}[i]} + \gamma\, v_{s[i]} + \alpha_{s[i]}\big),\qquad \alpha_s \sim N(0, \sigma^2_{\text{state}})where age has four categories (18–29, 30–44, 45–64, 65 and older), race four (white, Black, Hispanic and other, with Hispanic ethnicity taking precedence), education four (high school or less, some college, a four-year degree and a postgraduate degree) and gender two. Poststratification requires the joint distribution of these variables in each state, which the American Community Survey provides. The state-level predictor v is the Democratic share of the 2024 two-party presidential vote. A state predictor of this kind is what makes MRP estimates reliable in states with few respondents, because it lets those states borrow strength from states with similar politics rather than from the national mean (Lax and Phillips 2009; Buttice and Highton 2013). The term α is a state random intercept. We fit the model by variational Bayes as implemented in statsmodels, with its default priors and without survey weights; poststratification adjusts for the composition of the sample. If the variational fit fails, a linear probability model with the empirical-Bayes state intercepts described below, clipped to the unit interval, is used instead.
Three measures are continuous: affective polarization, in thermometer points with negative values kept; the share of the six violent acts a respondent supports, with the ladder coded as in Section 4; and the share of the five norm violations a respondent endorses, among the items answered. For these we estimate the fixed effects of equation (1) by least squares and the state intercepts by empirical Bayes. Each state's mean residual is shrunk toward zero by the factor τ²/(τ² + σ²/nₛ), where σ² is the residual variance, nₛ the number of respondents in the state and τ² the variance between states, estimated by the method of moments (DerSimonian and Laird 1986) and floored at zero. When τ² is zero, differences between states come only from their demographic composition and the 2024 vote. We use this closed form because the iterative mixed-model estimator stopped at the boundary of zero variance for these measures and returned no state effects.
The state estimate is the population-weighted average of the model's predictions across demographic cells,
\hat\theta_s = \frac{\sum_{c \in s} N_c\,\hat\pi_c}{\sum_{c \in s} N_c}where c indexes the 6,528 combinations of state, age, race, education and gender in the 50 states and the District of Columbia, Nc is the adult population of the cell in the Census Bureau's 2023 American Community Survey five-year sample, and π̂c is the predicted probability from equation (1). Intervals are the 2.5th and 97.5th percentiles over 200 simulations that draw the fixed and state effects jointly from a normal approximation to the posterior, centred at the variational estimates, with covariance given by the inverse curvature of the log posterior there; for the continuous measures the draws combine a robust covariance for the fixed effects with each state's posterior variance. Each published estimate is accompanied by its interval and by the number of respondents on which it rests, which ranges from 281 in the state with the fewest respondents to more than 16,000 in the state with the most.
We publish state estimates for 25 measures: affective polarization, the violence and norm indexes, support for each of the six violent acts, agreement with each of the five norm violations, extreme ideological identification, identification with the MAGA movement (among Republicans and Republican leaners, the only respondents asked) and nine items on immigration, the economy, tariffs and free speech. The violence measures use the ladder coding described in Section 4. The violence and norm measures are estimated among all engaged respondents and are therefore proportions of all adults in a state, whereas the national series in Section 4 are proportions of partisans. A measure is not estimated if fewer than 200 respondents or fewer than ten states answered it. An earlier version of the file, replaced on October 20, 2026, measured support for the more severe acts only among respondents who had supported the preceding acts and should not be used.
6 International Surveys
Since January 2024 we have fielded the core instrument quarterly in Brazil, Germany, India, Israel and Poland, five countries with competitive elections and different party systems. Each round interviews approximately 1,000 adults per country, and the eleven rounds completed through July 2026 comprise approximately 11,000 interviews in each country. YouGov fields these surveys on its online panels. In each country and round it interviews somewhat more respondents than needed and matches 1,000 of them to a sampling frame on gender, age and education. According to YouGov's methodology statements for the April 2025 round, the frames were drawn by stratified sampling from the 2021 Latin American Public Opinion Project for Brazil, the 2022 Eurobarometer for Germany and Poland, and the 2019 Pew Global Attitudes survey for India and Israel, and the matched samples were weighted by iterative proportional fitting to gender, age and education, and also to region in Brazil and Germany. The Indian sample represents India's online population rather than all adults.
The items follow the American instrument, adapted to each country's parties and institutions. Respondents rate each major party on a feeling thermometer, and affective polarization is defined as the rating of the respondent's own party minus the average rating of the competing parties, one of several measures used for multiparty systems (Wagner 2021); we use it because it reduces to the American measure when there are two parties. The Israeli survey asks directly for ratings of the respondent's own party and its opponents. The violence scenarios are retained, with the country's parties as targets. Norm items are translated where the relevant institution exists and replaced where it does not; the German and Polish surveys carry their own sets, which address, among other matters, the European Union and the dissolution of parliament.
These series support comparisons within a country over time rather than comparisons of levels across countries. Because the opposing party is defined differently in multiparty systems, the level of affective polarization in Germany or Israel cannot be compared directly with the level in the United States. Support is coded as in the United States, as support or strong support. The international estimates are weighted with the weight supplied for each respondent and include all respondents, not only partisans.
7 Using the Data
7.1 Public files
The interview-level data are public. The file all-data.zip contains one row per interview, with every core and rotating item in the public instrument, the survey weight, the demographic and partisan profile, the survey number and field dates, and a pseudonymous respondent identifier that links repeat interviews with the same respondent. A codebook documents every variable. We also publish the topline report for each survey, the state estimates with their own codebook, and a file for each country in the international surveys. Files are rebuilt as new surveys are added, so users should record the date on which they downloaded a file. The data are released under a Creative Commons Attribution 4.0 license and should be cited as Westwood and Lelkes (2024).
| File | Contents |
|---|---|
| /data/all-data.zip | One row per interview (all-data.csv), including respondents classified as disengaged |
| /data/codebook | Definitions of every variable in all-data.csv |
| /toplines/ | One topline report per survey, listed in /toplines/index.json |
| /data/state-opinion-estimates.csv, .json | State estimates: state, measure, estimate, ci_low, ci_high, n_respondents and the date of the data (asof) |
| /data/state-opinion-estimates-codebook.md | Definitions of the state measures |
| /data/international/{BR,DE,IL,IN,PL}-all.csv | One file per country, with response labels |
| /toplines/international/ | One topline report per country and round |
| /data/state-opinion-frame.csv, state-opinion-predictors.csv | The poststratification frame and state predictor used for the state estimates |
The variables needed for the core measures are the survey weight (weight); the survey number, year, week and interview times (survey, year, week, starttime, endtime); the engagement flag (engaged); party (party, coded dems or reps, and pid7); the two thermometers (democrat_therm_1, republican_therm_1); the six violence items (violence1 to violence6) and the two perception items (violence3_perception, violence6_perception); the five norm items (norm_judges, norm_polling, norm_executive, norm_censorship, norm_loyalty), each with a _perception counterpart; the demographic profile (among others, birthyr, gender, race, hispanic, educ and inputstate); and the pseudonymous respondent identifier (uid_public).
We withhold information that would identify respondents or that belongs to partners. YouGov's respondent and case identifiers never appear in public files. The pseudonymous identifier is derived from YouGov's identifier by a fixed one-way function; it allows a respondent's interviews to be linked but reveals nothing to anyone who does not already hold YouGov's identifier. ZIP codes, responses to the attention check and modules commissioned by partners are also withheld. County, congressional district, birth year, exact interview times and the open-ended justifications of norm violations remain in the file.
7.2 Reproducing the national series
The national series can be reproduced from all-data.zip. Keep interviews with engaged equal to 1 and party equal to dems or reps (party is derived from pid7, with independents who lean toward a party counted as partisans), and group them by year and week. Affective polarization is democrat_therm_1 minus republican_therm_1 for Democrats and the reverse for Republicans, averaged with weight. Support for act k is the weighted share of interviews in which violencek is Support or Strongly support; a missing response means the respondent was not shown the act and counts as not supporting it. Agreement with a norm is the weighted share answering Agree or Strongly agree among respondents who gave one of the five substantive responses. The perception items are weighted means on a scale from 0 to 100. Interviews by the same respondent share a value of uid_public, which should be used to cluster standard errors in pooled analyses. The headline figures on the website, among them the share endorsing at least one norm violation (excluding the executive-orders item), are weighted in the same way and pooled across surveys.
8 Limitations
Respondents are people who have joined YouGov's panel, and the national estimates are only as representative as YouGov's sampling and weighting make them. Online panels also contain insincere respondents, who can inflate estimates of rare attitudes such as support for murder. Before mid-2023, respondents who failed the attention check were delivered and excluded from estimates; since then they are screened out before delivery, so YouGov's weights from that point are computed for passing respondents only. The move from weekly to monthly fielding in July 2025 reduced the density of the series, and two longer gaps followed, from September 23 to November 14, 2025, and from December 17, 2025 to January 20, 2026.
Repeated cross-sections describe change in the population rather than in individuals; the repeat interviews in the data are incidental and do not constitute a panel. The state estimates pool four years of surveys and therefore cannot describe change within a state. They depend on demographic composition and the 2024 presidential vote, and in states with few respondents they lie close to the values those predictors imply.
9 Replication
9.1 Code and software
The survey pipeline is part of the lab's code for America's Political Pulse. Loading of waves, the national series and the international series are in lambdas/survey-processor/ (processing/csv_ingestion.py, us_processor.py and international_processor.py); the topline reports are in elite/surveys/toplines/; the public file is built by elite/surveys/regenerate_all_data.py; and the state estimates are in elite/surveys/mrp/. The code is not public; the lab shares it with researchers on request. The state estimates, toplines and public file are produced in Python 3.11 with statsmodels 0.14.6, pandas 2.3.3, NumPy 1.26.4, patsy 1.0.3 and SciPy 1.17.1, pinned in docker/requirements-batch-light.lock. The loading function pins only major versions (pandas below 3, NumPy below 2).
9.2 State estimates
The state estimates are produced by a single command:
python -m elite.surveys.mrp.model [--measures ...] [--n-sims 200]It reads every engaged interview from the survey database, the same interviews that appear in all-data.zip with engaged equal to 1, and two auxiliary files. The poststratification frame (elite/surveys/mrp/data/frame.csv) sums the person weights of adults in the 2023 American Community Survey five-year public-use microdata into the 6,528 cells of equation (2); it is rebuilt by python -m elite.surveys.mrp.build_frame, which requires a Census API key. The state predictor file (elite/surveys/mrp/data/state_predictors.csv) gives each state's Democratic share of the 2024 two-party presidential vote. The file does not record where its vote shares were obtained, but in every state its two-party share is within 0.0001 of the share computed from the certified state results compiled by the Federal Election Commission (2025). Each measure's simulations use their own NumPy generator, seeded from 20240101 and a hash of the measure's name, so a measure's intervals do not depend on which other measures are estimated. The command writes the CSV, JSON and codebook files listed in Table 3, together with copies of the frame and predictor files it used, which are published with the estimates.
9.3 Limits on exact reproduction
The national series and topline reports involve no randomness and can be reproduced exactly from all-data.zip. Three limits remain. The state estimates read the survey database at the time they are run, recorded in the field asof, whereas all-data.zip is rebuilt separately, so the public file can differ from the interviews behind a given set of estimates. YouGov's construction of the weights is not public. And the attention-check responses are withheld, so the engagement flag for surveys before July 2023 cannot be recomputed from the public file.
References
Buttice, Matthew K., and Benjamin Highton. 2013. “How Does Multilevel Regression and Poststratification Perform with Conventional National Surveys?” Political Analysis 21(4): 449–467.
DerSimonian, Rebecca, and Nan Laird. 1986. “Meta-Analysis in Clinical Trials.” Controlled Clinical Trials 7(3): 177–188.
Federal Election Commission. 2025. Official 2024 Presidential General Election Results [data set]. https://www.fec.gov/documents/5645/2024presgeresults.xlsx. Accessed October 2026.
Holliday, Derek E., Shanto Iyengar, Yphtach Lelkes, and Sean J. Westwood. 2024. “Uncommon and Nonpartisan: Antidemocratic Attitudes in the American Public.” Proceedings of the National Academy of Sciences 121(13): e2313013121.
Lax, Jeffrey R., and Justin H. Phillips. 2009. “How Should We Estimate Public Opinion in the States?” American Journal of Political Science 53(1): 107–121.
Park, David K., Andrew Gelman, and Joseph Bafumi. 2004. “Bayesian Multilevel Estimation with Poststratification: State-Level Estimates from National Polls.” Political Analysis 12(4): 375–385.
Petrocik, John Richard. 2009. “Measuring Party Support: Leaners Are Not Independents.” Electoral Studies 28(4): 562–572.
Wagner, Markus. 2021. “Affective Polarization in Multiparty Systems.” Electoral Studies 69: 102199.
Westwood, Sean J., Justin Grimmer, Matthew Tyler, and Clayton Nall. 2022. “Current Research Overstates American Support for Political Violence.” Proceedings of the National Academy of Sciences 119(12): e2116870119.
Westwood, Sean J., and Yphtach Lelkes. 2024. America's Political Pulse [data set]. Polarization Research Lab. https://americaspoliticalpulse.com/data. Accessed October 2026.