If you're the Obama campaign, you have to be discouraged by this just-released SurveyUSA poll showing him down 18 points in Pennsylvania. SurveyUSA has a pretty solid record so far this primary season--accurately predicting the final margins in California and Massachusetts, among other dodgily-polled states. So the fact that SurveyUSA bucks the recent tightening trend makes you wonder if the trend is real. All the more so given the emerging consensus that robo-pollsters like SurveyUSA do a better job capturing public opinion when one of the candidates isn't white. (The idea is that both white and black voters are squeamish about telling live interviewers their true feelings about a black candidate.)

Having said that, I think this may be that rare example of SurveyUSA misfiring. The poll shows Hillary increasing her lead six points since the last SurveyUSA poll, conducted about a week earlier. But if you compare the crosstabs from the two polls, you see that declining black support accounts for almost all of Obama's dropoff. In the last poll, Hillary led Obama 61-33 among whites; this time she leads 61-32. Among African Americans, on the other hand, Obama went from leading 83-17 to leading 74-24. The black share of the electorate also appeared to decline from about 18 percent to about 14 percent. By my rough calculations, Obama's drop among black voters (along with their lower share of the vote) explains more than 5 of the 6 points he lost from the previous poll.

Given SurveyUSA's record, I'm reluctant to dismiss this entirely--maybe there's some below-the-radar reason that Obama's African-American support took a hit last week. (Thoughts welcome in comments.) But my hunch is that it'll more than recover before Election Day.

Update: Oops. Commenter scottlooper correctly notes that the African American share only dropped from 15 to 14 percent, not 18 to 14. I'd backed it out crudely from the crosstabs and didn't realize SurveyUSA actually listed the demographic weights. That means the decline in African-American support and vote-share explains about half the 6-point swing, not 5/6. Still pretty significant though...

Update 2: Mark Blumenthal has some more thoughts about polling trends in Pennsylvania. Nickel version: The outfits that've polled multiple times since early March generally show a modest gain for Obama, and a modest dropoff for Hillary, among white voters. Which is where all the action is likely to be. In an e-mail, Mark also points out, as several commenters have, that the African American result probably isn't statistically significant. It's 15 percent of a sample of fewer than 600 likely voters, which means we're talking about a swing of fewer than ten voters...

--Noam Scheiber