Methodology
The New York Times Upshot/Siena College poll of 985 voters in Virginia was conducted from Oct. 29 to Nov. 2 on landline and cellular telephones with live interviewers.
The sample was selected from an L2 voter file of active Virginia voters stratified by age, gender, partisan primary history, region, L2 modeled race, modeled turnout probability, census-tract level density and educational attainment, and precinct-level presidential vote.
The probability that a voter was selected for the sample was in inverse proportion to the number of voters with a telephone number in each strata and the estimated probability that a voter would respond to the survey, multiplied by their estimated probability of voting in the gubernatorial election.
The initial survey weight adjusted the sample to match the estimated composition of the likely electorate by age, party primary history, L2 modeled race, gender, region, education and a modeled turnout score. Voter file data was used for weighting, except for education.
Self-reported education was weighted to match estimates for the educational composition of the likely electorate, based on data from Nov. 2014 Voting and Registration Supplement to the CPS and the Jan.-Sept. 2017 CPS.
The voter file based targets were estimated with a model of turnout in the 2013 gubernatorial election, based on vote history, age and date of registration, and applied to active registered voters on the L2 file in Virginia.
The standard deviation of the weights was .35. The maximum weight was 2.33 and the minimum weight was .37; 95 percent of weights were between .45 and 1.84.
The probability that a respondent would turn out was based on a combination of poll responses and a modeled turnout score. The modeled turnout score represented 60 percent of the estimated turnout probability; the self-reported likely voter score represented 40 percent.
The self-reported likely voter score was based on responses to five questions about interest in politics, the gubernatorial campaign and intention to vote.
The final stage of weighting adjusted the likely-voter-weighted sample to match the turnout profile of an electorate with a turnout of 2.4 million.