1. First, as Emory's Alan Abramowitz correctly pointed out in an email to me, "Seat exposure and the midterm dummy variable predict substantial Democratic losses regardless of what happens to either the generic ballot or Obama approval." See, for instance, Abramowitz's statistical model of House seat swings, which predicts a 38 seat loss for Democrats on the basis of those two factors alone.
2. Second, Rahm Emanuel seems to have bought into the hype about the president's approval rating as the overriding factor in midterms. In Bai's article, he is paraphrased as follows:
For every point that Obama’s approval rating dips below 50 percent, Emanuel said, there are probably four or five more House districts that will swing into the Republican column, and vice versa.
These results are not corroborated by the statistics. Controlling for other factors, Abramowitz's model predicts that a one point decrease in Obama's net approval (approval-disapproval) is associated with a .22 seat shift toward Republicans. (He indicates by email that the coefficient for raw approval is less than 0.5.) Even the simple slope in a bivariate plot is far less than 4-5 seats per point of approval. Unless there's some massive non-linearity around 50% approval, Emanuel's estimate is off by an order of magnitude.