Princeton Election Consortium

A first draft of electoral history. Since 2004

Just how close is the 2014 Senate race?

July 18th, 2014, 4:06pm by Sam Wang


In my last update, some of you commented that the 2014 Senate race had swung by a lot since May. That is not true! I think perceptions were colored by my emphasis on the snapshot probability. Today, let me take a different tack.

This year, control of the Senate will be closely fought. At the moment, the 2014 Senate race is as close as the 2004 Presidential race (Kerry v. Bush). In an election held today, Democrats/Independents would probably win 48, 49, or 50 seats. There are large distinctions between these outcomes!

We’re building the data pipeline. We’ve now hard-coded our rule for collecting polls (I’ll describe the rule in comments). This allows us to plot how the daily Senate snapshot has evolved. Changes have been quite subtle:

(The gray band shows the 1-sigma confidence interval, about the middle two-thirds of the distribution.)

The decline since May in Democrats’ fortunes comes from a shift of only about 1.5 seats. This shift is driven by changes in the races in Alaska, Colorado, Georgia, Iowa, Kentucky, and Michigan. This is very reminiscent of the 2004 Presidential race, in which the Kerry v. Bush contest was determined mostly by three states: Ohio, Pennsylvania, and Florida.

Median EV estimator from 2004 race

Median EV estimator from 2004 race

In such a circumstance, an overall shift in national opinion by even a few percentage points will swing control to the other side. Let me quantify this fact in terms that some of you are familiar with: the Popular Meta-Margin.

In past years, I have defined the Meta-Margin as the amount by which national popular opinion would have to swing in order to create a perfect tie. In 2012, a Meta-Margin of “Obama +5%” meant that polling margins would have to change by that much toward Mitt Romney in order to create an electoral near-tie.

For the Senate, the Meta-Margin is defined as how much polls would have to shift to create a situation in which exactly 50% of outcomes give 51 or more seats for Republicans (GOP control), and 50% of outcomes give Democratic/Independent control.

For over a month, the Senate Meta-Margin has been within a fraction of 1%:

Today, the Meta-Margin is GOP +0.2%.

This is actually a pretty common situation. Senate control was close in 2006 and in 2010. Then (and probably in 2014), the change in Senate composition was against the President’s party, as is common in midterm election years.

Comparing Senate seats with the Meta-Margin, 1.00 percentage point in Meta-Margin translates to 1.03 Senate seat:

That’s easy to remember: 1 Senate seat per 1 point in national swing.

Thinking in terms of national swing tramples the idea of Senate races as individual contests. It’s not obvious what Begich v. Sullivan (Alaska) has in common with Grimes v. McConnell (Kentucky). However, even if individual factors are at work in different states, they’re all affected by shifts in national mood.

As a measure of national mood, since April President Obama’s approve/disapprove numbers have gone from -7% net disapproval to -10% today. That is a shift of 3 points, which would be enough to account for the 2.0% change in Meta-Margin.

What will happen in November? This is a daunting challenge for any analyst. Any fundamentals-based model (such as the May analysis from The Monkey Cage) has uncertainties equivalent to a few points in popular opinion. Such models are valuable tools for research, but in my view they should not be interpreted as definitive predictions.

Here at PEC, any prediction for Election Day involves modeling how the Meta-Margin will change. That will come later. In the meantime I end with the quotable Yogi Berra (and John von Neumann): “Prediction is hard, especially of the future.”

Tags: 2004 Election · 2014 Election · President · Senate

19 Comments so far ↓

  • bks

    A bit tangential: Surprise upset in primary for Republican candidate for Senate in Georgia. Polling models questioned in this article:

    http://www.peachpundit.com/2014/07/23/five-takeaways-from-georgias-primary-runoff/

    –bks

  • ThePoliticalOmnivore

    Amitabh,
    Exactly–and in that case I think it may have been because of a specific brand of negative campaigning (basically hard-core allegations of racism from another Republican): http://politicalomnivore.blogspot.com/2014/07/going-negative-for-2014.html

  • ThePoliticalOmnivore

    Has anyone done analysis on what the polling looks like if you assume that black turn-out looks like it did in the Mississippi Republican primary (Thad Cochran?).

    What I want to know is this: if we assume that the attempts to turn out African Americans were more successful than polling suggested, what happens if we assume that 2014 might have similar turnout. Now, it’s a different animal: this was an R-primary in Mississippi and I don’t know how to compare that to a state-wide general election.

    Has anyone looked at that scenario (where state-wide turn-out looks proportional to the Mississippi primary for the 2014 elections)?

    • Amitabh Lath

      It would be useful to understand the mechanism by which groups who were previously not motivated to participate in the primary process were mobilized.

      Likely voter models look to previous election cycles to reweight raw poll numbers. If there are new more effective GOTV magic bullets out there the polls could be off by a lot. But the bias could be in either direction.

      So finding the mechanism that boosted GOTV efforts would be useful. Did the MS runoff and VA7 primary voters do it because of the particular candidate or is there a groundswell?

  • Bob Grundfest

    Interesting Upshot article, wherein Nate Cohn wonders where the GOP wave is or whether it will actually appear. Seems a bit early to be saying this.

    http://www.nytimes.com/2014/07/22/upshot/goodbye-to-the-republican-wave.html?ref=todayspaper

    • Ron Pitts

      One thing about that Red Wave thing the GOP keeps pushing, but this Same Period last Off Year Election (not as cool as SPLY) that is to say 2010, Dems were already sunk in several races. The GOP has certain built in advantages that aren’t showing up OR are taking a long time to develop this time. my 2c

  • Joseph

    The shoot-down of Flight MH17 is one of those events that can have implications for the Senate race, mainly by showcasing President Obama as a true leader on the world stage; that is, tough but fair. IMHO, that is more likely to help his standings in the polls than hurt him. And by the same token, those who attack the President using this as a weapon (like John McCain did recently) end up looking like they’re using the tragedy for crass political means, which hurts the chances of Republicans generally. In an election that is this close, it could be a deciding factor.

  • Bill

    I assume that using today’s snapshot as a prediction for November would have considerably larger error than does the snapshot. It is like using the results of a time series analysis to forecast future events: the error bounds kind of balloon out rather quickly.

    • Sam Wang

      Definitely. Today’s snapshot has a nominal uncertainty of +/-1 seat, though in actual fact it is probably smaller (that’s what correcting polls would get us). The added uncertainty from change from now until November is about +/-1.6 seats, which dominates the uncertainty. Which is Amit’s point.

      So, for November: 48-51 Dem/Ind seats. In my view that is the best anyone can do at the moment. Sorry!

      (Addendum added 2:44pm: There is a question about whether the midpoint should be. For now, probably the midpoint of where the Meta-margin has been over time from spring until now. Closer to the election, probably the daily snapshot.)

  • Amitabh Lath

    So 0.2% +- 1.6% is as close to a dead heat as is possible. Amazing.

    As for the polls being biased, I admit I do not have anything definitive. But as a thought experiment it is interesting to think about what happens if the GOTV efforts actually work (as they apparently did with African American voters in the MS runoff).

    • Sam Wang

      Yes, the extent to which we are on the edge is quite underappreciated.

      A commenter on Twitter thought that when today’s snapshot probability crosses 50%, that’s a big deal. This is such a wrong way to think. I must de-emphasize the snapshot.

      One could come up with reasons why the polls are biased. Turnout operations of the type you mention, as well the systematic error that I noted in Election Eve Senate polls in 2012. However, there are reasons for bias in the other direction as well: Obama’s high net disapproval, the off-year GOP turnout advantage, and so on.

  • Amitabh Lath

    So what is the uncertainty on the Senate-2014 Meta Margin? Since the value is effectively zero, the uncertainty matters a lot.

    Give us that gray shaded area!

    Also, should there be some extra systematic uncertainty given how the polling firms are getting the likely voter samples wrong?

    I ask this because this time around I’m getting a funny feeling that poll data is way off. Of course Eric Cantor’s loss and the MS senate runoff election are glaring examples, where likely voter screens were laughably wrong. And your plot from a couple of posts ago (comparing poll margin to election results) shows this bias is widespread.

    Complicating all this is that the pollsters themselves know this, and will be screwing around with their weights to account for it and maybe get it wrong in the other directions.

    I wouldn’t matter, except in the case of a dead heat, such as we have now.

    • Sam Wang

      Yes, yes! You and James A. Young are asking similar, good questions.

      To follow up on my reply to him: I am looking at (a), this year’s generic Congressional preference. Since mid-August 2013, the range has been from R+1.5% (Obamacare rollout) to D+6.4% (GOP-led government shutdown), with today at D+2.1%, almost exactly halfway between. The SD is 1.6%. Since March 1, it is anticorrelated with the Meta-Margin (r=-0.45). That might be noise, though – generic Congressional movements are hard to resolve with current polling methods and abundance.

      (b) Looking at Presidential net disapproval, this has a similar SD, 1.7%. It is highly correlated with the Meta-Margin (r=+0.65). This quantifies what I was saying the other day about linkage between Obama and the Senate campaign.

      (c) I have not examined the 2006 and 2010 Senate campaigns yet.

      Anyway, I think assuming an SD in the Meta-Margin (MMSD) of 1.6% might be in the right range for a November prediction. It would be symmetric around conditions today…which unfortunately does not remove any uncertainty from the November prediction. However, at least now we know what the MMSD is.

      >>>

      I don’t agree with you about polls. Primary races are hard to poll. Me, I keep off that subject. I think general-election polls will be OK.

    • Sam Wang

      P.S. Before doing any of this analysis, I had been leaning toward MMSD=1.25%. Still thinking about this parameter…

  • Billy

    Would it be correct to say that this is the closest “election cycle” since Bush/Gore?

    • Sam Wang

      I stopped short of saying that. The Senate ended up very closely divided in 2000, 2006, and 2010. However, the comparison to the Bush-Gore race did occur to me.

  • James A Young

    Are there typically big swings as Election Day gets closer? What percentage of people track politics all the time, like me, as opposed to those who only care just before election?

    • Sam Wang

      I think swings are usually not large. However, you have put your finger on the exact issue that must be resolved in order to make a polls-only-based prediction.

      The question on my mind is how to estimate those swings. Possibilities include (a) estimating the change in the generic Congressional ballot, (b) relating it to drift in Presidential approval/disapproval, and (c) examination of Senate polls in 2006 and 2010 to see if the meta-margin can be extracted.

  • Sam Wang

    We are tracking 19 races where the final margin is likely to be within about 10 points in either direction. For any given day, the rule is:

    1) Take the last 3 polls or the last N weeks, whichever gives more polls. Currently N=6; that will get smaller over time.

    2) Take only one poll from each polling organization.

    3) If the oldest poll overlaps in time with earlier polls, take those as well.

    Then proceed as for the usual Meta-Analysis (calculate the median and an estimated SEM, turn these into a win probability, and so on).