Princeton Election Consortium

Innovations in democracy since 2004

Outcome: Biden 306 EV (D+1.2% from toss-up), Senate 50 D (D+1.0%)
Nov 3 polls: Biden 342 EV (D+5.3%), Senate 50-55 D (D+3.9%), House control D+4.6%
Moneyball states: President AZ NE-2 NV, Senate MT ME AK, Legislatures KS TX NC

What the Gerrymander giveth with one hand: House control in 2014 now a toss-up

October 10th, 2013, 10:30am by Sam Wang

Andrew Sullivan has collected some commentary that argues against my suggestion that the House could potentially flip in 2014. There were some good points made. However, it should be noted that some of these analysts’ comments have been OBE (Overtaken By Events) – the shutdown. The incredible numbers in yesterday’s NBC/WSJ poll pretty much demonstrate that point: approval of the Republican party is at an all-time low…and 60% of voters would vote out every single member of Congress at once including their own if given the chance.

All of the arguments on this subject can and should be quantified. It’s just like last year’s Presidential race: is any of this punditry data-based? Today I start to outline a true prediction. I give a “toy model,” i.e. step 1 toward something more realistic. The toy model relies on a prediction of popular vote only. At the end, I start to add a little bit of complication. I invite you to add more complication in comments.

Provisionally, it looks like the following: In a little over a week, the shutdown has increased the probability of a Democratic House takeover in 2014 from 13% to as high as 50%.

The first number (13%) is is exactly as expected if House trends were following expectations from analysts’ conventional wisdom and political science research. But the shutdown, combined with the fluidity of gerrymandered districts, has added a highly unexpected twist. The 50% figure could swing back toward the Republicans, or it could go further toward the Democrats. For certain, it gives a measure of an unusually fast change in the national mood. Let’s dive in.

Predicting the nationwide vote. Kevin Drum suggests that Congressional candidate preference polls a year out are not all that predictive. But they might have just enough value to be useful. Quantifying that value is a key step in turning my presentation yesterday (which was a snapshot) into a true prediction.

To project current Congressional opinion forward in time, we have to quantify Delta, the difference between opinion in October of the previous year and the actual national election vote. Using generic Congressional ballot data from RealClearPolitics, I get the following:

Year 1 year before Outcome Delta
2002* D+0.0% R+4.6% Toward R 4.6%
2004* D+1.5% R+2.6% Toward R 4.1%
2006 D+8.0% D+7.9% Toward R 0.1%
2008 D+12.0% D+10.7% Toward R 1.3%
2010 D+1.0% R+9.4% Toward R 10.4%
2012 D+1.0% D+1.3% Toward D 0.3%
Average D+3.9% D+0.5% Toward R 3.4 +/- 4.0%

The correlation between polls now and eventual outcome is r=+0.89, which is pretty good. Basically, polls now are fairly predictive of Election Day, but Democrats are likely to lose ground by then**. The average movement is Delta=3.4% toward Republicans, with a standard deviation of 4.0%. To account for further uncertainty, I will bump that standard deviation up to 5.0% for prediction purposes.

Let’s make a prediction of the national popular vote, based on pre-shutdown conditions. The last data points before the shutdown gave D+5.0% (Quinnipiac, Rasmussen, PPP, Sept. 23-29). Adding in the expected drift, that translates to an average predicted popular vote in November 2014 of 5.0-3.4=D+1.6%. This is similar to the 2012 election, when Democrats won 1.3% more votes  (but not more seats).

Now let’s make a second prediction, this time based on post-shutdown data. One piece of information is my previous analysis of the MoveOn data indicating a swing of 8 points away from Republicans, and 14 points in gerrymandered districts. That number comes from a surveys for specific Republican members of Congress, paired with a generic Democratic challenger. In short, the MoveOn data is equivalent to current conditions of D+9% nationally.

Harry Enten and Nate Cohn speculate that the MoveOn/PPP polls might not be accurate, either because (1) asking “Congressman vs. Generic” gives an unfair advantage to Generic, perhaps because (2) Democrats might run a bunch of lame candidates. But the generic Congressional preference question has the same issues and has a rather good track record. Recent generic-Congressional surveys (D+4%, D+5%, D+8%) are not so far from the MoveOn prediction for non-gerrymandered districts. However, I agree with Enten that it would have been good to have data for some Democratic districts.

Combining the MoveOn data for non-gerrymandered districts with the other post-shutdown polls gives a median margin of D+6.5%. Projected forward to November, that’s a prediction of D+3.1+/-5.0%, which means that Democrats have a 73% probability of winning the national popular vote in 2014.

Converting votes to seats. If shifts in opinion occur nationwide by the same amount in every district, the probability of a Democratic takeover is the probability that their popular margin will exceed D+7%, the threshold number in the 2012 election. Then the probability of Republican retention of the House would be 78% post-shutdown, or odds of 4-1 in their favor. That is not so far from most analysts’ views. (To put this in perspective, applying this method in October 2009 would have predicted the 2010 GOP takeover with a probability of about 70%.)

However, the shutdown has upended things, and the shift is not equal across districts, as indicated by the MoveOn data. To repeat yesterday’s point, gerrymandered districts have swung by an average of 14 16 points against Republican incumbents. (Note, 10:47am: 12 new PPP polls reinforce the picture. Gerrymandered R’s have slipped an average of 15.6%, compared with 8.6% for non-gerrymandered R’s. It’s highly significant, p=0.01 [spreadsheet].)

The gerrymandered districts are shown in white.

If gerrymandered districts swing harder, that puts dozens more seats into play. If the finding holds up (needs more data!), that brings the probability of Republican retention down to around 50%.

Whether or not this holds up precisely, let’s look at the big picture: The usual midterm outcome would lead to a gain of seats and votes for Republicans. But the change in national mood has changed the narrative. The GOP is likely to lose the popular vote, and even lose seats. I believe that Republicans are still slightly favored to retain control…but that depends on what they do in the coming weeks and months.

From an analytical standpoint, this is somewhat mind-blowing. Think about it: the shutdown has turned a likely Republican lock in 2014 into a situation where their political survival hangs in the balance. That is quite an accomplishment in one week.


*Data are from October in the year before the election, except for 2002-2004, which are data for January in the year of the election.

**Bafumi, Erikson, and Wlezien have a different interpretation based on data going back to 1946: over that period, that midterm opinion during the year tends to move against the incumbent president. For Democratic presidents, the shift from January to November is about 6.7 +/- 3.4%. Applying that rule would predict a near tie: R+0.2+/-3.4%. However, note that the rule has been seriously violated twice in the table above. Politics is in a disrupted period, so I’ll take my value above for now.

***It is also consistent with the finding (Bafumi et al., op. cit.) that a party’s midterm popular vote tends to lag the previous Presidential vote by 3.3+/-2.6%. This predicts D+0.7+/-2.6% in 2014. However, there’s a long tail on that distribution; Republicans gained 3 points in 2002, totally contrary to Bafumi-driven expectations.

In other words, based on Bafumi et al. the pre-shutdown generic Congressional vote and the 2012 election result lead to the same prediction, and match analyst conventional wisdom. But Enten, Cohn, Sullivan, and others, think of it like this: Republicans losing the House would only come to pass if something strange and big happened to alter the national mood. What do you think is happening right now?

Tags: 2014 Election · House

17 Comments so far ↓

  • mediaglyphic

    Dr. Wang,
    your election model used statewide polls. Can they be used here to get a better prediction? Or are there not enough statewide polls that are performed regularly enough. Of course district by district polls would be the best, but i am pretty sure we don’t have a lot of these.

    • Sam Wang

      For now, I think a combination of the generic Congressional ballot is a pretty good approach, supplemented by district-by-district swings. What I’m doing with the MoveOn/PPP data can be done with any district polls – and will be more credible once it’s not just one liberal-messaging (but actually pretty accurate) polling organization.

      Statewide polls could also be used, but we are starting to get into Nate levels of complexity there. At that point it might be best to just ask Charlie Cook or Larry Sabato what they think, since their teams are good at keeping all the districts in their heads at once.

  • mediaglyphic

    Larry Sabato might do a quant analysis, but i found Charlie Cook’s analysis to be much less rigorous.
    btw i found out that Dems won 7 Romney districts (where romney won more than 50%) while repubs won 17 seats in districts where the President got more than 50%.

    • Sam Wang

      Cook may be less rigorous, but if you read closely he is very good. A totally different approach. I have learned over time that when my calculations do not match Cook, it is worthwhile to check the calculations.

    • mediaglyphic

      I have read cook for years, but in this last cycle i did not find him insightful at all. In fact he seemed to just echo the prevailing wisdom.

      However your endorsement counts for a lot so i will pay more attention.

  • Michael S

    You’re going to need district by district polls here, and not just for the general election. What you’ve not looked at yet is districts where sitting Republicans will be primaried by Tea Party types. In the last election cycle they took over the gerrymandered seats. In the 2014 cycle it’s becoming ever more clear that more seats will be primaried, even districts that are somewhat safe or in play.

    What you’re going to see is a backlash from the Tea Party base displacing incumbents with name and fundraising advantage with more radical candidates that will not be so palatable in the general.

    This suicidal scenario is going to play out all over the country, and there might be enough primary challengers in seats that would be safe with the current incumbent to shift things even more in the general.

  • Ebenezer Scrooge

    Gerrymanders decay over their decade, as the composition of Congressional districts changes. This is particularly true for party-packing gerrymanders which–unlike incumbent-protecting gerrymanders–rely on the smallest possible edge that will create a noncompetitive district.

    What was an efficient gerrymander in 2010 will not be as good in 2012, and worse in 2014 . . . . I’m wondering if there is some way to quantify this–decay times and magnitudes.

  • Amitabh Lath

    Midterm turnout skews conservative. Is there any indication that the current unpleasantness is going to make the 2014 D/R ratio any higher?

    Either by energizing Democratic constituencies or suppressing Republican ones?

    • Some Body

      Adding my doubt to Amit’s. The 3.4% figure lumps together Presidential and mid-term years, where these are two rather different sets of data. And while we do have the outstanding case of 2006, on the whole, especially with a Dem President, the prospects are for the swing to be considerably higher than 3.4%. At least other things being equal.

    • Sam Wang

      That is a good point. See Bafumi et al. for a clear separation for on-year and off-year changes. I didn’t use that but it does lead to a larger range of outcomes. How does this bias cut?

  • Sophie

    I’m a bit skeptical of the finding that gerrymandered districts are swinging much harder than nongerrymandered districts. There isn’t an obvious reason why that would occur, other than the possibility that gerrymandered districts might contain more swing voters – but the effect you’ve found is so large that this is unlikely to explain all of the disproportionate swing. It is also true that gerrymandering spreads party loyalists thinly – but that should affect the baseline result rather than the swing amount. I’d like the result you’ve found to be true, but am slightly doubtful of it.

    Let me offer instead a counter-hypothesis, which I’d be interested in testing. The hypothesis is as such: Gerrymandered districts are more advantageous for incumbents than their PVI numbers would predict, because they often have distorted and contorted shapes, which makes it far more difficult for a challenger to effectively advertise themselves (in TV/etc.) and gain name recognition. As a result, the incumbent finds it easier to keep an advantage in name recognition and become re-elected. Notably, this effect would not affect the generic ballot – voters may be willing to vote for a generic Democrat, but without the ability to effectively advertise, they aren’t willing to vote for a complete unknown. For instance, most of the gerrymandered seats held by Republican incumbents in Michigan were won by Obama in 2008, and lost by <10% in 2012. Yet far fewer of them were close in Congressional vote, presumably because of effects of this sort. Even in MI-11, which featured Kerry Bentivolio as the Republican candidate (perhaps most famous for stating in a court deposition "I have a problem figuring out which one I really am, Santa Claus or Kerry Bentivolio"), the Democratic challenger was not able to equal Obama's performance (losing by 6.4%, compared to a 5.4% loss for the President.)

    The hypothesis, then, would be that Republican gerrymandered districts find a stronger than normal dropoff for Democratic candidates between the generic ballot and the actual election. If true, part of the heavy swing you've found for gerrymandered districts would represent this effect. I'd be very interested in trying to check if that is indeed the case.

  • Rob

    I suspect that there is a lot of mean reversion in the time-series of congressional candidate polls.

    If so, this would suggest that numbers will swing back towards the norm and the most recent polls will be an unusually poor predictor of 1-year-ahead results.

    • Sam Wang

      Well, erm, that is basically the main topic of this post!

      Future movement is included in the analysis. That is why I did not cite current conditions, which are a snapshot indicating that in an election today, Democrats would control the chamber with >90% probability and a 50-seat margin.

    • Rob

      Err, yeah sorry I kind of skimmed that a little quick!

      Let me try to spell out the idea that was at the back of my head. Suppose the series of democratic results in generic ballots (we’ll call this x) follows the form:
      x(t+1) = x(t) + (avg – X(t)) + i
      Where i is some error term (that doesn’t appear to be very normally distributed!), and avg is the average of the past year of poll results.

      Now you’ve said, “We’re trying to predict x(t+365) using x(t). Let’s look at some past relationships between x(a) and x(a+365).”

      On average you will find that the expected x(a+365) is x(a). However, if at time t you are above the average results for the past year, your expected x(t+365) will actually be a lot less than x(t).

      Now let’s say that the i here is very very fat tailed, as it appears to be. Then, if you have a small sample size (as you do), you may even find that not only is x(a+365) equal to x(a) in expectation, but the std deviation is very small as well.

      That is my concern. Of course I don’t know the polls follow this form, but it doesn’t seem unrealistic to me. And I think this is the intuition behind pundits dismissing your idea, although I doubt most would articulate it this way.

  • Bill Miller

    Hi Sam, I enjoy your site as always, but disagree. I recommend American Nations: A History of the Eleven Rival Regional Cultures of North America [Kindle Edition] I hope you convince your liberal buds how safe they are in 2014! :) and beyond…

    • Bill Miller

      We believe your models haven’t factored in all those canceled insurance plans, which will awaken those who ignore politics at this interim season.

  • Don Derham

    Crystal ball cloudy that day.

Leave a Reply to Amitabh Lath (Cancel)