New Marquette poll has Mary Burke 2 points ahead of Scott Walker with likely voters, and 4 points behind with registered voters.

"Poll director Charles Franklin called the results 'the epitome of what a horse race looks like.'"

Also in the poll: "A solid majority, 54 percent, say the state is heading in the right direction, while 42 percent say it's on the wrong track...."

ADDED: The new poll has Walker with 47.5% and Burke at 44.1% among registered voters, while last month it was 45.8% percent support, with Burke at 44.8%. Likely voters are identified by Marquette as those who say they are "certain to vote in November’s election," and it's Burke 48.6%, Walker 46.5%. In July it was Burke 46.8%, Walker’s 46.3%.  The questionnaire (PDF) asks people whether they are "Absolutely certain" or "Very Likely" to vote in November, and I wonder what the spread would be if the "Very Likely"s were included. Who can be "Absolutely certain" that they will even be alive in November? It would seem that they are screening for fuzzy thinkers and liars.

In the comments to this post, sane_voter said: "The first poll in history where the Dem does better with likely voters. I call BS." I found this WaPo article that addresses the question why Republicans always seem to do better with "likely voters":

This isn’t always true. The likely voter population is different in midterm versus presidential elections. Midterm elections, in which there is no national candidate and less of a focus on a broad, national campaign attracts different types of voters. Specifically, non-white and younger voters have been less apt to turn out for midterm elections. 

That article also looks at how pollsters screen for likely voters: How do pollsters identify likely voters?

1. Screening questions: Pollsters ask respondents screening questions about past voting and/or vote intention and set a threshold of who is considered a likely voter (this is the Post’s standard method). The Post begins by excluding those who are not registered to vote and uses questions to predict voting such as past voting behavior, self-reported vote intention, interest in the election and knowledge of one’s voting place to further thin the herd. A related method uses similar questions to assign each respondent a likely voter score based on how many questions they answer “correctly” (e.g. Did you vote in the most recent presidential election? Yes = 1 point). Gallup and the Pew Research Center employ methods like this.

2. Voter list sampling: Campaign pollsters typically use voter lists to recruit respondents, targeting those who have voted in a particular set of recent elections according to voter registration files. This method is often combined with the screening questions above to establish voter likelihood.

3. Probability-weighted models: Less commonly used is a “probability voting” model for identifying likely voters. Probability voting models assign probability of voting on a 0 to 1 scale for each respondent and weight results by this. In other words, if a respondent is considered 66 percent likely to vote and another 33 percent, the first respondents’ vote choice will receive twice the weight in data. CBS News, the New York Times and Marist College polls have employed such models in the past.

This seems so much more sophisticated than what Marquette seems to be doing!