The Methodology Behind Times Voter Trends New York Times (Online) Nate Cohn Times Voter Trends reports are based on individual-level estimates of the likelihood that voters would have supported the Democrat, Republican or a minor-party candidate. The estimates incorporate poll results from The Times/Siena Poll; voter file data from L2; precinct-level election results collected by The New York Times and the Voting and Elections Science Team; county-level election results from The Associated Press; and U.S. Census Bureau data including the Current Population Survey and the American Community Survey. The L2 voter file is used as the foundation for a synthetic population of registered voters representing the joint distribution of the self-reported demographic characteristics of registered voters in the United States. The probability that each registrant supported the Democratic, Republican or minor party candidate for president in 2024 or 2020 is based on a model of vote choice in Times/Siena polls, fit as a function of respondent's self-reported demographics, data on the L2 voter file, and geographic data. The resulting individual-level probabilities are adjusted to match the result by precinct. The estimates do not report a margin of sampling error because they are not based on a sample of the population. The estimates are derived from multiple models, which are each subject to different kinds of error. The results for small subgroups, states without a Times/Siena pre-election poll and states without reliable data on partisanship, like party registration or primary vote history, should be treated with particular caution. The first post-election L2 voter file with comprehensive vote history in each state is used as the base frame for a synthetic population of registered voters, which represents the joint distribution of the demographic characteristics of registered voters in the United States. The synthetic population is built iteratively and probabilistically. For each successive trait, the self-reported demographic characteristic of Times/Siena poll respondents is modeled as a function of voter file traits like party registration and vote history, geographic-level predictors, and the respondent's self-reported characteristics for any previously modeled trait. These models are used to estimate the probability that each registered voter belongs to a given demographic group; the demographic trait assigned to each registrant is simulated probabilistically. The demographic characteristics were assigned or modeled in this order: Gender: The L2 voter file contains a likely gender for voters, usually based either on how voters identified on their voter registration form or the likely gender of their first name. The L2 classification is accurate for 99 percent of respondents in recent Times/Siena polls; it is used without revision. Age: The L2 voter file contains either a date or year of birth for nearly all voters in most states, based on the information provided by registrants when they submit their voter registration form. When age is provided on the L2 voter file, it is used without modification. However, the voter file does not have age for a meaningful share of the population in some states, such as Alaska, New Hampshire and Wisconsin. For voters without a date or year of birth on the voter file, the age of the registrant is probabilistically simulated based on a model of self-reported age in Times/Siena polls. Race: The racial classification available on the L2 voter file is never assigned to the registrant — not even in the seven states where state voter files contain the race provided by registrants on their voter registration form. Race is probabilistically simulated for all registered voters based on multiple models of self-reported race in prior Times/Siena polls. The polls were joined to the given year's voter file and reweighted to match the voter file by L2's modeled race, party, age and interactions thereof, with the original survey weight as the base weight. The adjusted survey weight was used in all subsequent race models. The first model is a Times/Siena polling-based implementation of Bayesian Improved Surname and Geocoding, which estimates the probability of someone's race based on their surname and geographic information. In traditional BISG, the surname-derived probability of someone's race is based on the racial composition of people with their surname in U.S. Census data. In this version, the name-based probability is based on a multinomial logistic regression model with cross-validation of self-reported race in previous Times/Siena polls fit as a function of multiple measures of the racial makeup of the respondent's name: the 2010 decennial Census surname file and the racial makeup of voters' first, middle and last names on the L2 voter file in the states where voters can provide their race when they register to vote. The geo-level data is Census tract data from the American Community Survey 5-year of the racial composition of the citizen voting age population. The second model uses a machine learning algorithm, Gradient Boosting, to model self-reported race among previous Times/Siena poll respondents as a function of the aforementioned BISG-like estimate, voter file characteristics, geographic predictors and simulated self-reported traits like age and gender. Three different versions of the model are fit for states with different data on race and party: one for states with party on the voter file, one for states without party on the voter file (trained on voters from all states, but without partisanship as a predictor) and one for states where self-reported race is available on the voter file (trained only on the states with race on the voter file, with voter file race as an additional predictor variable). Education: Self-reported education was modeled for all registered voters using three models of self-reported education in Times/Siena polls: a model of whether the respondent had a college degree; a model of whether college graduates had a postgraduate degree; and a model of whether people without a four-year college degree had attended some college. All three models were trained using a machine learning algorithm, Gradient Boosting, which modeled education as a function of self-reported race, age, gender, census-tract educational attainment, census-tract educational attainment by race, turnout, party and salient socioeconomic variables on the voter file, including home value, education and occupation. Separate models were fit for states with and without party registration. The models were adjusted to match New York Times estimates for the educational composition of registered voters, based on the Census Current Population Survey and the American Community Survey. The probability that each registrant supported the Democratic, Republican or minor party candidate for president is based on multiple models of vote choice in pre-election New York Times/Siena polls taken from September to November in 2020 and 2024. Major party and minor party vote choice were modeled separately. The major party models use a machine learning algorithm, Gradient Boosting, to train a model of vote choice in Times/Siena polls as a function of the following covariates: Age Race Education (four categories and two categories) Gender Party registration and partisan primary participation Region (Northeast, South, West, Midwest) State, if the respondent is from a state poll Vote history in the last two general elections Census tract race and ethnicity Census tract educational attainment Census tract educational attainment by race and ethnicity Precinct-level election results in 2024 (or imputed based on the demographic makeup of the precinct) Precinct-level ecological inference by race and ethnicity (eipack::bayesei) Two separate models are fit for states with different voter file data on partisanship. One model is fit for states with party registration or primary vote history. It is trained only on data from states with such data. A second model is fit for states without partisanship data on the voter file. The model is trained on all survey respondents from all states, but without party registration or partisan primary participation as a predictor. The minor party model was fit using a logistic regression model, with minor party support modeled as a function of the standard self-reported demographic traits, vote history, party registration, precinct-level election results in 2024, and whether the respondent's ethnicity is classified as likely to be Arab by L2. The resulting individual-level probabilities were adjusted in multiple steps to match the actual election result by precinct. First, the probability that a voter would support a minor party candidate was adjusted to match the minor party proportion of the vote. Second, the probability that a voter would support a major party candidate was adjusted to match a candidate's major party vote share in the precinct. Finally, the probability that a voter would support the Democratic or Republican candidate was calculated by multiplying the adjusted Democratic and Republican share of the major party vote by the adjusted major party vote share of the precinct. The adjustment is calculated by finding a value on the logit scale that minimizes the difference between the actual result by precinct and the mean of the individual-level probabilities among validated voters. The racial makeup of the electorate is sensitive to question-wording and whether to use the race and ethnicity provided by registrants in the seven states where the voter registration form asks about race. Across Times/Siena polls from 2020 to 2026, only 90 percent of Times/Siena respondents' self-reported race aligns with the race they reported when they registered to vote in Alabama, South Carolina, Louisiana, Florida, Tennessee, North Carolina and Georgia. The number is similar, at 92 percent, among those who provided their cellphone number when they registered to vote, suggesting that the difference is mostly due differences in measurement. The effect varies considerably by race. Among self-reported white voters to the Times/Siena poll, 96 percent reported the same classification when they registered to vote, compared with 93 percent of Black respondents, 74 percent of Hispanics and only 32 percent of other nonwhite voters. On net, Times/Siena respondents were less likely to identify as white in response to the survey than when they registered to vote. Overall, 63 percent of Times/Siena respondents from the states where self-reported race is available on the voter file identified as non-Hispanic whites in the survey, compared with 65 percent when they registered to vote. The difference between the survey and voter file-based measures of a respondent's self-reported race is at least partly a result of differences in question wording, which varies across state voter registration forms and between those forms and The Times/Siena Poll. Beginning in early 2024, Times/Siena polls have asked the combined race and ethnicity question planned for the 2030 census, not the traditional two-part race and ethnicity question used in the 2020 census. For the Times Voter Trends, responses to the Times/Siena race question have been recoded to more closely align with the 2020 Census definition by reassigning multiracial Hispanics to Hispanic and Middle Eastern and North African respondents to non-Hispanic white. Native Hawaiian or Pacific Islander, American Indian or Native American, multiracial, and “some other race” have been combined into “other.” To be classified as “some other race,” respondents had to provide an answer that's about race, ethnicity or national origin and clearly interpretable to the person responsible for recoding the open-ended answer recorded by the interviewer. For example, “human” is treated as Don't know/Refused, not “some other race,” and excluded from our modeling because it does not offer a specific race, ethnicity or national origin. Conversely, a respondent recorded as responding “Castanian” is recoded as Don't know/Refused because the person responsible for coding the race of the respondent doesn't know whether this is a misheard or misspelled version of Caucasian (White), Kazak (Asian), or something else, not because the respondent themselves doesn't know or is refusing to answer. The change in the race and ethnicity question has particularly significant consequences for the composition of the “other” category, making it challenging to compare the 2020 and 2024 results for this group. Consequently, the Times Voter Trends pages do not display the shift between 2020 and 2024 for the “other” race category, as well as the shift between 2020 and 2024 by race in Hawaii, where “other” race represents nearly one-third of the vote. Source: Times Voter Trends, Current Population Survey, Catalist Voter Behavior Study, Pew Research Center study of validated voters; exit poll conducted by the National Election Pool and Edison Research. By The New York Times Source: Times Voter Trends, Current Population Survey, Catalist Voter Behavior Study, Pew Research Center study of validated voters; exit poll conducted by the National Election Pool and Edison Research. By The New York Times