Note added, September 2014: This analysis did not take into account the effects of redistricting, incumbency, and other factors. For more up-to-date views see an October analysis and this post-election piece, The Great Gerrymander Of 2012 (NYT).
Conditions through August showed a 2% lead on the generic Congressional ballot for Democrats. As of September 20th, in the wake of the Democratic convention, the lead has widened to 4.0 +/- 2.0%. Although it has yet to be appreciated by pundits, this could well translate to a November loss of the House of Representatives by Republicans. Based on the generic Congressional ballot, the probability of a Democratic takeover is 74% with a median 16-seat majority. Whichever party is in control, the seat margin is headed for being narrower than the current Congress. Like any probability in the 20-80% range, this is a knife-edge situation. This picture may change over the coming six weeks as more information, especially district-level polls, becomes available.
[Update: Based on later discussion (here and here), some other factors (i.e. incumbency) do not affect this prediction. Two remaining unknown parameters do: (1) the quantitative effect of redistricting measured in units of the generic Congressional preference poll, and (2) whether the generic poll is more likely to move toward or away from the Presidential winner's party in coming weeks. These factors are likely to be small...but this poll is very close at the moment. More on this later.]
As seen in recent articles in Politico and U.S. News, few pundits think the Democrats will re-take the House. However, analysis of a leading indicator suggests to me that transfer of control is a distinct possibility.
Predicting the House outcome is challenging. First, there is the basic problem that we have to estimate how far opinion will move between now and November. On top of that, there is uncertainty in knowing how the polling measurement – generic Congressional ballot preference – translates to a seat outcome.
Another approach would be to use district-by-district polls and ratings. An estimate like that can be seen from our data partner, Pollster.com. Their House outlook shows retained GOP control, and RealClearPolitics implies the same. However, many of those polls are weeks or months old. My estimate today suggests that in the coming weeks, we might look for district polls to move in the Democrats’ direction. This is also an opportunity for a detailed analytical approach, as taken elsewhere, to shine.
In 2010, the national Congressional vote was a big 6.6% margin of popular vote win for the Republicans. That outcome was very close to the pre-election R+7% polling median (but 8 points less than the final Gallup poll of R+15%). So the generic Congressional preference poll, aggregated across pollsters, can give some sense of where the vote will go.
In 2012, the picture looks very different from 2010. Congressional voter sentiment before Labor Day is often movable (see 2008 and 2004 history). But we are now entering the high season, so some sense of the outcome is starting to emerge.
In summer polls leading up to the 2012 conventions, Republicans were behind Democrats by a median of 2%, a 9-point swing from 2010. Consequently, many seats won in that Republican wave are now at risk, as I pointed out in my previous House outlook. To put it another way, midterm election conditions (as in 2010) work against incumbent Presidents; the pendulum has now swung back.
In post-convention polls, Democrats got a big bounce that peaked at D+6% and now appears to be subsiding. It is a good thing for the Republicans that the election was not last week. The most recent polls (Sept. 7-17) indicate a median lead of D+4.0+/-2.0% (+/- estimated SEM, n=7 different pollsters). (Note that the HuffPost smoothing software uses a different algorithm.) This 11-point swing from 2010, if it were to hold, would lead to big Democratic gains in the House.
As I pointed out two weeks ago, the eventual national D-minus-R House popular vote share is strongly predictive of the corresponding margin of House seats. Here is a graph based on data from 2000-2010 House elections.
It shows that each 1.0% of popular-vote margin translates to a 6.0-seat advantage. This plot shows no long-term advantage* for either side: a nearly-tied popular vote would translate to a nearly-tied House (x-intercept = -0.3 +/- 1.1%). Individual data points deviate from the fit line by 7 to 17 seats in either direction.
However, there is a known advantage to incumbency, which I estimate as being worth an equivalent of 1.3 +/-
1.01.8% of popular vote. This is smaller than a recent estimate**. In other words, on average a national win of D+1.3% is required for the House to change hands.
Next, let us translate the generic Congressional ballot to estimate House seat outcomes. Applying the Wisdom-of-Pollster-Crowds principle that has served so well in Presidential and Senate races, we will use the poll median to approximate the actual popular preference.
The HuffPost graph above includes telephone polls (i.e. not robopoll or Internet), but that is strictly for purposes of clear display. Using all polls and median-based statistics to address issues of outlier data gives the median of D+4.0% that I gave. That translates to a narrow 16-seat Democratic majority in an election held today.
This would be an unusual outcome. It would involve a Democratic net gain of over 30 seats, much more than the typical gain for a re-elected president’s party. But 2010 was also an exceptional wave year for the Republicans. Again, think of the pendulum. In any event, this is what the numbers are currently telling us.
The principal caveat. The main issue with this analysis is that it does not use district-level data. In the coming weeks, those surveys will become more abundant. In 2008, district polls did a very good job of estimating the outcome – on Election Eve. Six weeks out, the generic ballot preference is the week-to-week indicator that is available.
Where things could go in the next seven weeks. Assume a +/-4% opinion shift between now and November, and this leads to a popular vote prediction of D+0% to D+8% (1 sigma). This gives a Democratic takeover probability of 74%, approximately three out of four.***
It should be noted that current conditions emphasize the post-convention bounce, which could be transient. Conversely, if the Democratic lead increases, that would take House control out of the knife-edge territory that I defined previously. For now, a smart use of campaign donations is to donate to the DCCC through this ActBlue page, or the GOP through Crossroads GPS.
Are you interested in a specific race? For tracking specific districts, watch the aggregators such as our data source (HuffPost/Pollster.com) and RealClearPolitics. Considering the current 11-point swing, any race that was won in 2010 by a Republican by less than 15% should be considered competitive, and worth getting involved if you are a Democrat or Republican currently on the sidelines. I myself find this surprising, but that’s what the arithmetic is saying.
*We can ask whether redistricting should affect the x-intercept, i.e. give one side a structural advantage. Often the goal of redistricting is to protect specific incumbents. This is different from the goal of altering the total number of seats expected by either side. Indeed, maximizing one goal can work against the other by collecting partisan advantage into a few incumbents’ districts, at which point even a little swing can flip the outcome in the other districts. This is a complex subject, and to my knowledge it is not proved that gerrymandering leads to a global structural advantage.
**A four-factor model at FiveThirtyEight has led to a meme propagating through the political blogosphere that there is a 3-4 point structural advantage for the Republicans. However, that estimate is less precise than it appears.
When doing multiparameter regression, it is important to quantify uncertainty, as well as avoid fitting to noise. Here, two quality checks are to (i) make sure the the uncertainty does not blow up when all the extra factors are added, and (ii) make sure a large fraction of the 7-17-seat residual can be accounted for by each added factor. After playing around with the four-factor analysis, I have concluded that it is likely to fail these criteria. The consequence is low confidence in estimating the structural advantage. A simpler model using only this-year’s-national-vote and last-Congress’s-seats (i.e. two factors) applied to 1996-2010 data has less of this problem, and indicates a structural advantage to Republicans in 2012 of 1.3 +/-
1.01.8%. Note: when fitting to data going back to 1946, this becomes 0.6 +/- 1.0%. It is not possible to rule out the possibility that in the aggregate over the whole chamber, this parameter is zero.
***This is in marked contrast to the InTrade price of 0.17, which corresponds to an imputed market opinion-driven probability of 17%. As I have previously pointed out, InTrade market opinion is often quantitatively wrong. In this case it might even be getting the direction (i.e. <50% or >50%) wrong.