Princeton Election Consortium

A first draft of electoral history. Since 2004

A technical correction to the Meta-Analysis

September 18th, 2012, 8:10am by Sam Wang


We must make a technical correction to the Meta-Analysis. This correction does not change the general history of the campaign. However, an explanation is due to those of you who watch the twists and turns of the Meta-Analysis closely.

Because of a change in the format of our data feed from the Huffington Post, it no longer returns polls in the same order as earlier this summer.  The change affects us when multiple polls appear on the same day, which is why this problem is only beginning to appear. The problem was noticed by reader Froggy, who noticed that recent median margins in FL, OH, and WI did not fully reflect fresh polling data.

The corrected output is shown as the black curve here (old output in green):

(click on the graph for a larger version)

The general turning points of the campaign have not changed. The Ryan-VP bounce for Romney looks smaller, as does the post-Republican-convention “negative bounce.” But both events are still present.

The corrected graphs for the EV estimator and the Meta-margin history go live today. Importantly for today’s news, the post-Democratic-convention bounce is now more apparent – and more reflective of national polls. Also, the November prediction is unchanged.

I apologize profusely for this error. It was fortunate that the corrections were as small as they were. Andrew Ferguson and I thank Froggy, whose close attention to the data allowed us to track down this problem.

Let the flogging begin.

Tags: 2012 Election · Meta-analysis · President

18 Comments so far ↓

  • Billy

    Sam, on the state-by-state map a bunch of states (ie. Michigan, Wisconsin, Virginia, Ohio) jumped from “toss up” (white) to “97.5% democrat”. Is this a result of the tweak or was there some kind of weird polling result this week?

    • Andrew Ferguson

      hi Billy, the jump from “toss up” to “97.5% democrat” reflects the fact that, now that we are using all of the polling data, those states are “97.5% democrat.” unfortunately, I can’t tell you at this time whether this change is simply because we are now using a complete dataset, or because there has been a shift in the polling.

      I’m happy to share the complete dataset with you if you’d like to dig in, but perhaps these links will suffice to study the Wisconsin, Virginia, and Ohio polls:
      http://elections.huffingtonpost.com/pollster/2012-wisconsin-president-romney-vs-obama
      http://elections.huffingtonpost.com/pollster/2012-virginia-president-romney-vs-obama
      http://elections.huffingtonpost.com/pollster/2012-ohio-president-romney-vs-obama

      best,
      Andrew

    • Froggy

      As I recall, Michigan was unaffected by the correction. Wisconsin was wrong a little bit just before the correction (it was Obama +1 when it should have been Obama +1.5 — it actually should have been Obama +2 instead of Obama +0.5 between late August and last week). Ohio was way wrong since about last Thursday (only Obama +1 when it should have been Obama +4). Virginia’s shift to dark blue is a combination of the correction and the good polls Obama had there yesterday and today (and last week).

      Colorado was a little off in the other direction (the correction pulled it back toward Romney, from Obama +3 to Obama +2), and today’s Rasmussen poll moved it even further toward Romney.

  • Amitabh Lath

    What caused the deviation in July? Right by the arrow labelled “Bain”? Were there days of multiple polls then, or some other bugfix?

    • Andrew Ferguson

      hi Amitabh, we were missing some polls from those days. now, we are processing all of the polls from that period, and thus the graph had to be adjusted. there have been no changes to the calculation, only to the input data.

      thanks,
      Andrew

  • MarkS

    Um, I didn’t understand that at all. “The Huffington Post … no longer returns polls in the same order as earlier this summer. The change affects us when multiple polls appear on the same day.” Affects you *how*? Were you multiply counting polls that should have been counted only once? Or what, exactly?

  • Alfred G. Cuzán

    It appears that the DNC convention bump was two points or less, whereas the VP Ryan announcement depressed Obama by about three. So the net effect of the Ryan announcement and the conventions has been a point in favor of Romney?

  • badni

    Under the corrected graph it is less obvious that Bain is a key moment in the timeline.

    It looks like something started a couple of weeks before that that was negative for Obama, and something a week or two after that started helping Obama.

    Any ideas on what started pushing the metamargin down?

    And what started pushing it back up? Akin? Or maybe the Bain ads just took that much time to really have an effect?

  • Joseph Marshall

    I certainly applaud your honesty and scientific rigor. One things strike me as interesting, the appearance in the new EV estimator of a week of extreme volatility in early September, that appears to my amateur eye to be larger than anything comparable in the rest of the chart. Is this a polling artifact? Or some reflection of actual events a week or two ago?

  • Pat

    Thanks for the correction. While the Meta-margin was not changed significantly, it is impressive how much this correction has affected the EV distribution: now there is about 33% chance of Obama getting 332 EV, while earlier this number was at around 15% with many other peaks. The general aspect of the histogram shows a change which is pretty spectacular!

  • MAT

    Bravo for not sticking your head in the sand, instead making the changes and then publically throwing yourself onto the mercy of the court of public opinion. As noted above, this is how it’s supposed to work.

  • Matt McIrvin

    …only with higher time resolution, of course.

  • Matt McIrvin

    With this change, the EV chart now bears an uncanny resemblance to the one on electoral-vote.com, which should be a reassuring check.

  • Olav Grinde

    Excellent! You’re crystal clear, as always.
    No reason for flogging — only complements for the transparency of your methods. That’s why the Princeton Election Consortium and the Meta-Analysis is exemplary; I just wish other poll aggregators were in your league.

  • Eric Fisher

    Real scientists take data seriously. This correction is the way the Internet ought to work. Thank you, Sam, Andrew, and Froggy.