The Sun-Times covers the campaigns’ view of early voting:
The Giannoulias campaign has composed a model that includes every voter in Illinois with a grade from zero to 100, a score of 100 meaning they are most likely to vote for Giannoulias.
So far, of the 150,000 Illinois voters who voted early as of Thursday morning, 54 percent of them scored 50 or above on that predictor, Rendina said.
Among those more intensely for Giannoulias or Kirk, the most-likely Giannoulias voters — those scoring 70-100, accounted for 47 percent of the early votes.
Those judged to be most likely to be Kirk supporters — scoring 0-30 — accounted for only 38 percent of early votes.
There is a lot of guessing involved there, Rendina admits. Not everyone will vote as predicted. The projections are based on past tallies of whether they have generally requested Democratic or Republican ballots.
On the Republican side, Illinois’ Republican Party has been out-performing every other state in America for making phone and in-person contacts with voters, said state GOP Chairman Pat Brady — 3.5 million since June; 110,000 on Saturday alone.
This is going to be very close. The modeling may seem complicated and such, but in reality it isn’t. It essentially asks people in surveys who they are supporting and then a series of voting behavioral questions. The data is then crunched–I imagine using similar methodology to Nate Silver by running simulations based on the likelihood of support for Alexi based on matching past behaviors to the survey data. The probability at the end then tells you the likelihood of a vote for the candidate.
It is far better than precinct level results because it focuses on individual voter behavior and not a geographic proxy. Where the potential weakness exists, it’s very similar to most statistical models. If the models assumptions and underlying data from surveys is not consistent with reality then you’ll get bad predictions. Overall though, the basic technique is relatively simple and should improve upon likely voter models in surveys since it takes behavioral cues as the most important predictor instead of respondent opinions of the moment.