There is a poll of registered voters in Arizona, Georgia, Nevada and North Carolina

Full Methodology.

Arizonan landline and cellular phone surveys for the New York Times/Siena College were conducted in English and Spanish starting in August. In Georgia and North Carolina, from August 8 to August 15, 2024. 9–14, 2024; additionally, in Nevada from August. December 12 through December 15, 2024. 2,670 voters who were registered to vote in total were interviewed. The margin of sampling error for all registered voters and the likely electorate when all states are combined is plus or minus 2 percentage points for each individual voter.

For every state poll, the margin of sampling error among registered voters is plus or minus 4 percentage points in Arizona, Georgia, and Nevada, and North Carolina, respectively. Arizona, Georgia, Nevada, and North Carolina are within plus or minus 4 percentage points of the likely electorate, while North Carolina is within plus or minus 4 percentage points.

Example.

The survey is a stratified sample of registered voters from the L2 voter file that has been adjusted for response rate. The New York Times took several steps in the selection process to account for nonresponse, varying telephone coverage, and notable state-by-state variations in phone number productivity.

Each state’s L2 voter file was categorized according to factors such as age, homeownership, turnout history, race, gender, marital status, household size, and statehouse district. For each stratum, the percentage of registrants who provided a phone number and the average predicted response rate were determined using data from previous Times/Siena surveys. The reciprocal of a stratum’s mean telephone coverage and modeled response rate served as the initial selection weight. In cases where respondents had more than one phone number listed in the L2 file, the number with the highest modeled response rate was chosen.

fielding.

The Institute of Policy and Opinion Research at Roanoke College, the Public Opinion Research Laboratory at the University of North Florida, ReconMR, and the Siena College Research Institute conducted additional fieldwork on the stratified samples for each state based on political party, race, and region. The interview was terminated if the intended respondent was not present, and the interviewers asked to speak with the person listed on the voter file. Cell phones were used to contact 96% of the total respondents.

ReconMR provided the Spanish translation for the instrument. When conducting a survey, bilingual interviewers were told to take the respondent’s lead and conduct it in either Spanish or English. They started the interview in English. Spanish-speaking interviewers followed up with monolingual Spanish-speaking respondents who had originally been contacted by English-speaking interviewers. In total, 12 percent of weighted interviews and 13 percent of interviews with self-reported Hispanics were done in Spanish.

When an interviewee did not leave the survey before the end of the two self-reported variables used for weighting (age and education) and responded to at least one of the ballot test questions about age, education, race, or the presidential election, the interview was deemed complete for the purposes of being included in the questionnaire.

Voter registration is weighed.

The Times employed the R survey package in a multi-step process to weight the survey.

Initially, stratum-specific unequal probability of selection was taken into account when adjusting the sample.

Secondly, the sample was calibrated to align with voter file-based parameters concerning the attributes of voters who have registered to vote.

The following objectives were applied:.

• Party (based on party registration if available in the state, or based on classification based on involvement in partisan primaries if available in the state, or based on a vote choice model from previous Times/Siena polls) by whether the respondent’s race is modeled as nonwhite or white (L2 model).

• Age (as stated in the voter file if the respondent declines, or as self-reported) by gender (L2).

• Nationality or race (L2 model).

• Education (four categories of self-reported education level, based on census data, Times/Siena polls, and the L2 voter file, and weighted to match NYT-based targets).

• Racial differences between Whites and Non-Whites based on academic attainment, whether attained through college or not (L2 model of race weighted to match NYT-based targets for self-reported education).

• L2 model’s marital status.

• Purchasing a home (L2 model).

State region (as classified by NYT).

The history of turnout, as determined by L2 data classifications by the NYT.

• 2020 election voting procedures (NYT classifications based on L2 data).

• North Carolina’s past voting history in the 2020 presidential primary (L2).

• In Arizona, major party registration or primary participation (NYT classifications based on L2 data).

• The level of education in Arizona and Nevada census tracts.

Ultimately, the weights of the full sample and the outcome of the general election horse race question (including leaners) were identical for the sample of respondents who answered every question in the survey.

Weighting — probable voters.

The Times used the R survey package to weight the survey in a number of steps.

The samples were first corrected for stratum-specific differences in the probability of selection.

Secondly, using a model of 2020 election turnout, the first-stage weight was modified to take into consideration the likelihood that a registrant would cast a ballot in 2024.

Third, the sample was calibrated to align with targets regarding the likely electorate’s composition. By combining the individual-level turnout estimates for registrants on the L2 voter file that were previously detailed, the targets for the make-up of the likely electorate were determined. The weighting criteria were the same as those listed earlier for voters who were registered.

Fourth, self-reported intention to vote was taken into account when adjusting the initial likely electorate weight. Based on past Times/Siena polls, four fifths of the final probability that a registrant would cast a ballot in the 2024 election was determined by their ex ante modeled turnout score, and one fifth was determined by their self-reported intentions, which included a penalty to take into account survey respondents’ propensity to turn out at higher rates than nonrespondents. After multiplying the final turnout probability by the modeled electorate rake weight and dividing by the ex ante modeled turnout probability, the final likely electorate weight was determined.

The sample of respondents who answered every question in the survey was then weighted equally to the outcome of the general election horse race question (including leaners) on the entire sample.

The survey’s design effect, a measurement of the statistical power lost as a result of survey weighting and design, is taken into account by the margin of error. In Arizona, Georgia, Nevada, and North Carolina, the design effect is 1 point2 for registered voters and 1 point37 for the likely electorate, 1 point18 for registered voters and 1 point19 for the likely electorate, and 1 point22 for registered voters and 1 point37 for the likely electorate. These results apply to the entire sample.

The margin of sampling error for the sample of completed interviews is +/- 4 points/6 for registered voters and +/- 4 points8 for the likely electorate in Arizona; +/- 4 points/7 for registered voters and +/- 5 points1 for the likely electorate in Georgia; +/- 4 points6 for registered voters and +/- 4 points8 for the likely electorate in Nevada; +/- 4 points7 for registered voters and +/- 4 points7 points for the likely electorate in North Carolina.

For the sample of completed interviews, the design effect is 1 point 24 for registered voters and 1 point 35 for the likely electorate in Arizona; 1 point 22 for registered voters and 1 point 45 for the likely electorate in Georgia; 1 point 24 for registered voters and 1 point 35 for the likely electorate in Nevada; and 1 point 28 for registered voters and 1 point 28 for the likely electorate in North Carolina.

Surveys conducted over the last three weeks prior to an election have historically shown an error at the 95th percentile of The Times/Siena Poll of plus or minus 5 point1 percentage points. Beyond sample error, real-world error can come from nonresponse bias, coverage error, late shifts among undecided voters, and inaccurate estimation of the electorate’s makeup.