Princeton Election Consortium

A first draft of electoral history. Since 2004

How will the last holdouts break?

November 3rd, 2008, 2:45pm by Sam Wang


The national two-candidate margin is currently Obama leading by 7.0 +/- 0.9% (median+/-SEM, n=7, 10/31-11/2). There are between 1 and 6% “undecideds.” Can they wreak havoc anywhere?

Today, Charles Franklin at Pollster.com writes about how undecideds have broken since 1948. These data can give us outer bounds on what may happen this year. Bottom line: they don’t matter.

In these data from the National Election Study, the largest break ever occurred in 2000, when undecided voters broke 3:1 against the incumbent party. Overall, the average swing has been a net loss of 7 +/- 21 (mean+/-SD) % of undecided voters for the incumbent party. There’s no statistically significant bias in either direction.

We can put an upper limit on how much effect these unaware souls will have. The most-extreme-case scenario is 6% undeclared (NBC/WSJ) with the same size break as in 2000, for a maximum net swing of 2.6% for either Obama or McCain. Assuming an average of 3.1% undecideds across surveys, the net swing has a 95% confidence interval of -1.4% to +1.4%. These values are fairly close to the “Obama +2%” and “McCain +2%” scenarios that you can see by clicking on the right sidebar.

Using current median margins (second entry on each line), the two-candidate margin is smaller than 2.6% in four states: Indiana (tie), Missouri (tie), North Carolina (Obama +1%), and North Dakota (tie). In all other cases, the win probabilities would not flip (i.e. change between >50% and <50%). In the case of a break toward McCain, what’s left is 278 safe EV for Obama, and a median outcome of approximately Obama 338 EV, McCain 200 EV.

This uncertainty does not affect the overall Median EV Estimate, as I wrote back in August. My topic at the time was the effect of correlated fluctuations in polls, i.e. “what if all the polls are off in the same direction?” My conclusion at the time was that unless we know in advance what the error is (for instance, with the cell-phone effect), there’s not much change to the analysis.

I will probably include this in my final estimate tonight. You’ll still see what the hypothetical effect would be in my bias analysis, where I consider the possible effect of biased polls due to missed cell-phone users, the [inverse] Bradley effect, and so on.

Tags: 2008 Election

12 Comments so far ↓

  • Eddie

    NICE!

  • Brad

    I’ve been following this site for a few weeks now and I really appreciate the statistical analysis of the polls you’re providing. I find this analysis to be extremely interesting and it is exactly what I’ve been looking for.

    What I’m finding kind of odd however, is the disconnect between what, intellectually I believe will happen tomorrow, and emotionally what I fear is going to happen. It is irrational, but I can’t help but worry that somehow Obama won’t win. This sort of analysis goes a long way though to assuaging my concerns.

    Cheers,
    Brad

  • Bruce W

    I agree the effect of undecideds will be minor, especially when compared to the effect of total turnout, the disparity of GOTV operations, and general cell-phone underpolling, but I think you are pushing your luck assuming a normal distribution and giving SD and 95% CI for a collection of 15 points.

  • Collin

    Interesting. It’s certainly hard to see any pattern.

    On a related point, perhaps already hashed out here, is there any reason why we should expect political races to tighten at the close? Do they generally? If so, one (unsupported) explanation comes to mind: maybe low-information voters who previously expressed a preference for the frontrunner, based mostly on a bandwagon effect, eventually break more evenly after the onslaught of advertising and coverage at the end of the campaign.

  • Frank

    Sam, You solicited pre-election requests, but how about post-election? I’d like to see a coattail graph of Senate races that shows the Senate Democratic candidate’s observed margin minus his or her expected margin by the same for Obama.

  • Evans

    wow, expert analysis. Now about that (Hays) Alaska poll today showing the race was within the margin of error there… I’m assuming that’s an outlier? Maybe the Stevens thing really matters that much though, is this the first survey since then?

    Either way, I’m sticking by my guns and predicting a 390+ EV landslide win for Obama because everyone is still underestimating the importance of his ground game advantage. Want one crazier prediction? Republicans will actually have an election-day weakness when their supporters decide they aren’t enthusiastic enough to wait 2 hours for McCain-Palin.

    I’m saying it every day: Enthusiasm matters this time.

  • Jack Rems

    It appears to me the “share of undecided” chart shows a reliable break AGAINST reelection of any president. (So it should break against Obama in 2012….)

  • Richard

    There have been six sites I have followed closely since the conventions. RCP and Pollster.com for the raw data; fivethirtyeight because Nate Silver is fun to read and quirky and Sean Quinn does great reports from the field; Gott and Colley because they keep a win/loss/tie record on polls in each state for the past 30 days, which is another nice way to think about how close a state is (for example Obama is 31-0-0 in PA, 30-6-1 in OH, 25-4-1 in FL, and 26-0-0 in VA, which gives a pretty good indication he is going to win these states); the Computer Science group at University of Illinois (Roll to the White House) since their model offers a check on this site–they only update every one to two days and currently have it 357.5 to 180.5, remarkably close to the meta-analysis; and of course this site, which is unquestionably my favorite. The median EV estimator is simply the best graphical scorecard for keeping track of the race that I have seen. I hope you are back in 2012. Thank you.

  • The Numbers Guy : Where the Poll Numbers Stand on Election Eve

    […] Undecided voters could break sharply for McCain — though history suggests this wouldn’t gain him more than a few percentage […]

  • Wall Street Journal » Blog Archive » Where the Poll Numbers Stand on Election Eve

    […] Undecided voters could break sharply for McCain — though history suggests this wouldn’t gain him more than a few percentage […]

  • JohnL

    […] an average of 3.1% undecideds across surveys, the net swing has a 95% confidence interval of -1.4% to +1.4%. These values are fairly close to the “Obama +2%” and “McCain +2%” scenarios that you can see by clicking on the right sidebar.

    Also, how about the already decided? If 2 or 3 in 10 have voted already (hooray for them) and they gave [40 or 50 or 60%] of their votes for one of the major party candidates, what percentage of the remaining LV (or EV!) does each candidate need to obtain to win?

  • MTGAP

    You say that Obama is leading by 7%. You also say the median outcome is Obama 338 EV, McCain 200 EV. But that’s more than a 7% difference. How does that work?

Leave a Comment