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

Stat-urday: Lombardo’s Quadratic Molasses

September 7th, 2012, 11:00am by Sam Wang

You are so impatient. Today is still all pre-DNC data. But I predict you’ll recognize the bounce when it hits.

Another Stat-urday. Let’s look at Steve Lombardo’s claim that “you didn’t build that” was a quadratic turning point on July 13th. Aside from the fact that this “trend” extrapolates to a 20-point win by Romney on November 6th, which is wrong, there is a small grain of truth here. It’s a story I have already told, that of the Ryan-VP bounce.

First, Lombardo’s original plot:

(Note the axis convention that a Romney lead is plotted as a positive quantity, converse to the style of plots here at PEC.)

Fitting to a quadratic doesn’t make sense. There is no Iron Law of Opinion that says the national margin has to follow a trajectory like that. Extrapolate it out 60 more days, and Romney wins by a Johnson-Goldwater-level landslide. Partisan fantasies are one thing, but this is obviously wrong.

But the core claim might have value: that since July 13th, the Obama-Romney margin has been moving in one direction. Could that be true?

One test would be to run a simple correlation on post-July-13 data to see if there’s an upward trend. The data look like this:

In this dataset, the correlation coefficient is about r=+0.27, which works out to a p-value of 0.11 (one-tailed). As I’ve written before, p<0.05 is when a scientist starts thinking something is there.

So it’s close to significance. Perhaps Lombardo has formed a hunch, an intuition. In that context, the quadratic fit is a way of asking is there was a minimum point somewhere in this series of data. But such a fit is imprecise at best. If we were to look at the details of the fit, we would probably find that this method only allows the minimum point to be localized to within about a month, given the quality of the data.

However, we have state polls, which provide higher time resolution. As it turns out, all the Romney-ward movement since July 13th happened in a single week in August.
Ryan bounce: about 3.0-3.5% in Meta-margin units
To match this to the Lombardo plot, August 13th is 85 days before the election. Maybe now you can where the shift is likely to have occurred in the national-poll data.

As I wrote Wednesday, the Ryan bounce has partly subsided, as seen in the Meta-analysis history. So I wouldn’t go investing in Lombardo Quadratics quite yet. It’s a bit molasses-like in its ability to detect quick changes.

Tags: 2012 Election · Meta-analysis · President

19 Comments so far ↓

  • BCC

    Looks like climate “skeptic”-style analysis has come to the polling world

  • Iseeurfuture

    Could you explain how today’s job report will impact your analysis. Personally, and it’s just me I believe the numbers are being crunched purposefully. Why? Think about it. Thanks.

  • Matt McIrvin

    A jobs report shouldn’t directly affect his analysis at all, because it depends only on polls. If the economic situation has some effect on polls, that should show up.

  • Ralph Reinhold

    As I have stated before, I come into statistics from a different world. Lombardo’s data is a combination of point process and shot noise (in my world). There is a definite 1/f shape to the data that is common with both of them. At this stage, the best that can be said about the data is that it is very noisy. We are looking for a signal that is very low in frequency. The lower you go in frequency, the more the data is affected by the noise.

  • Ralph Reinhold

    It is important to note that a moving average is not the ideal filter for a point process or for shot noise. If I had time, I would love to do some temporal analyses on the data. Since each source (Pew, PPP, Rassmussen) have prefiltered, it would be best to look only at their data. I have no idea where Lombardo’s data came from. If it is from multiple sources (which I suspect) the margin for error will run from close to 2 to close to 10 which means that the noise is different for each sample point…that is, it is not stationary. Therefore, it cannot be used as an ensemble.

  • Chris R

    Sam: Ot, but I can’t resist:

    Well, it looks like we have an answer to “can you get a bump in the polls from a convention”, based on Gallup and Rasmussen’s shift.

    You can, provided your convention doesn’t feature an elderly man berating furniture.

  • Iseeurfuture

    Gallop has BO up 48-45 7 day tracking.

  • Hans

    Why Quadratic? why not Cubic? Linear? A sine wave?

    The very premise of the analysis seems faulty, without some explanation as to why a quadratic trend line was chosen, and without a measure of goodness of fit.

  • Ralph Reinhold

    Hans: It is intuitively obvious to the most casual observer that he got a hold of Excel and when you only have a hammer, everything looks like a nail. It appears to me that he used data from multiple sources. As such, it is not stationary because the variation changes from sample to sample and most statistical methods fail. GIGO

  • Matt

    Chris: it’s still a bit early but Gallup and Rasmussen are consistent with a bump so far. Certainly that will be the narrative if the bump does materialize, especially if it is significant relative to the static nature of the polls this election.

    Hans: I suspect he chose his narrative first and then tried to illustrate it statistically. FWIW, I thought Clinton’s choice of dates for his job numbers falls in the same category of abusing math to make a point. Not that he didn’t have a lot of valid points to make, such as reducing a deficit by cutting taxes doesn’t add up, but politics and abuse of statistics go hand in hand.

  • BillSct

    “politics and abuse of statistics go hand in hand.” True enough, but Bill makes it fun…

  • pechmerle

    “There are three kinds of lies: lies, damned lies — and statistics.”

    Often attributed to Disraeli, or Twain. Most likely neither is correct. See the extensive collection of possible first sightings here:

  • Matt McIrvin

    There’s an ARG poll out taken Tuesday through Thursday (a period covering the Michelle Obama and Bill Clinton speeches) that shows Romney up by 3. I think what we’ve seen so far is still consistent with no Democratic bounce.

  • wheelers cat

    Matt, I disagree. Even Silver is seeing signs.
    As hypercautious as he is this year I think that post means there will definitely be a DNC bounce.
    Silver was infinitely better in 2008, not so risk averse.
    I can’t decide if hes being paid to pimp the horserace or just terrified of falling off the top of the mountain (he is the most influential statistician in known blogspace).
    Like always, we get the statistics we deserve.
    Nate and the other big name poll aggregators are serving up junk-food statistics– cheap, tasty, and filling whether you are a red consumer or a blue consumer.
    Artificial uncertainty is the poll aggregator equivalent of pink slime.

  • baw1064

    Why not fit it to the Gamma function? You could extrapolate it to Romney capturing 187% of the vote in November…

  • wheelers cat

    Dr. Wang:
    the cubic might actually have predictive value.
    if Obama’s convention bounce steps on Romney’s non-bounce, will it drive the “negative bounce” further into south-of-zero territory?

  • Matt McIrvin

    Even in 2008, Silver seemed to be adding a lot of poorly motivated uncertainty to his numbers. But the nice thing about his site then was the series of road reports about the campaign; I miss that.

Leave a Comment