Pollster.com

Articles and Analysis

 

Morning Status Update for 10/2

Topics: Status Update

Marc Ambinder called it the "ObamaPollSplosion." Michael Scherer wasn't sure what to call it, but his headline read "Poll Dance: All Together Now." Whatever the label, yesterday was certainly quite a day for fans of Barack Obama.

In the last 24 hours, we have logged 12 new national polls and 19 new statewide surveys (and that doesn't count 3 new "pre-debate" snapshots also released yesterday by Quinnipiac University) . Ten (10) of the statewide polls and 10 of the national surveys tracked results from prior surveys conducted since September 1. The result?

  • 8 of the 10 statewide tracks showed movement in Obama's direction.
  • 7 of the 10 of the national updates showed movement in Obama's direction.
  • We logged new polls in 7 of the states we currently classified as toss-up or lean, and our estimates in all 7 states moved in Obama's direction.
  • The new polls were enough to nudge Pennsylvania (and its 21 electoral votes) back to lean Obama, and Missouri (11) to toss-up, changing our electoral vote count to 250 for Obama, 163 for McCain and 125 in the toss-up category.
2008-10-02 trend.png

The only consolation for Republicans in yesterday's results were that the daily national tracking polls were relatively stable as compared to the previous day's results. The DailyKos/Research2000 track showed a slight gain for Obama, the Gallup Daily and Diageo/Hotline polls showed a slight decline. The GWU Battleground poll showed a four point net shift to Obama, but it was likely the result of a change in their weighting procedure.

Of all the results yesterday, the most consequential are from Florida. Three new polls, from CNN/Time, Suffolk University and Quinnipiac University (plus a fourth pre-convention snapshot from Quinnipiac) all showed Obama leading by margins of 4 to 8 percentage points. As a result, our usually conservative trend estimate shifted a remarkable 2.6 points in Obama's direction. Although we still classify Florida as a toss-up, Obama currently leads by about a point (47.7% to 46.6%).

Incidentally, before the release of the CNN and Suffolk polls, I received a number of emails asking about the trustworthiness of the Quinnipiac polls. One thing to keep in mind is that our new charts allow you to compare the trend-line for any individual pollster to the overall trend, so you can make your own judgements about whether a particular pollster typically gets different results than the others.

Try this with the chart above: At the end of the trendlines, two of the red dots at the bottom right corner represent the McCain percentage on the two most recent Quinnipiac polls. The blue dot in the upper right corner is the latest CNN/Time percentage for Obama. Click on these dots and you can connect-the-dots for each pollster. How does the pollster trend line compare to our overall trend estimate? And here's one more option. To see the relationship more clearly you might want to click through to the full Florida chart, and use the "choices" tool to display only the Obama or McCain trendline.

Incidentally, Marc Ambinder asked yesterday whether "the polls ever swung this much, this late, and NOT swung back?" Given that so little about the last year has been typical, I'm not sure why we would expect the next five weeks to be any different, however, here are two ways to try to answer that question: First, back before the conventions, we collected poll data for prior presidential elections going back to 1980, and Charles Franklin produced charts showing the trend in the margins for each election. Very few showed much movement in October. It is also worth examining Gallup's remarkable collection of trial-heat trend charts dating back to 1936.

Second, check out Brian Schaffner's post here last night, in which he examines the potential for McCain to come back by winning over the remaining undecided (their ranks are narrowing) or by converting Obama supporters (past panel studies have shown that 95+% of supporters in September stick with that choice on Election Day).

 

Comments
Allen:

It would be hard for me to classify the race as a toss up. In order to McCain to win, he has to prevail in IN, MO, NC, NV, OH, VA, FL and one additional state. Based on the current state probabilities at http://election-projection.net, the chances of this happening are becoming increasingly low.

____________________

Fenders:

Great site! However, I'm concerned that the window (smoothing width) of your lowess curves is too wide. It keeps you from capturing recent trends.

____________________

thoughtful:

Today is even better for Obama!

____________________

Mark Blumenthal:

@fenders:

Have you tried the "smoothing" tool on the charts? You can't access it from the embedded version above, but if you click through it's on every chart on the "tools" menu.

____________________


Last Month the Confederacy was for McCain.

Now the Confederate States Virginia and North Carolina may defect to the Union States.

I do not know if Florida fought during the Civil War. But is is a "weird" state, detaching from the South.

Missouri should have been in the Confederates during the Civil War. I guess. But it is abandoning McCain.

Abraham Lincoln was the Republican Candidate. Reversal of the Earth's Poles ???

Vicente Duque

____________________

mgcsinc:

It would be really nice to see charts of the Obama-McCain difference, since there's a lot of baseline difference (attributable to different amounts of undecided or 'don't know' voters, I would assume) between polls. The behavioral scientist in me is looking for a repeated-measures statistic.

____________________

Allen:

@mgcsinc

The site http://election-projection.net bases its probabilities on the normalized value (Obama - McCain) / (Obama + McCain). The reason is to address the concern you just raised: that different polls push the respondents harder for an answer, and some include "leaners" in their results while others do not. This can result in dramatic differences in the reported number of undecided voters. The normalized value removes these differences, and gives a more accurate measure of which candidate is leading and by how much.

____________________

Jotu:

That Gallup stuff is very interesting, but a bit scary for an Obama fan for a couple of reasons: There seem to be several instances of fairly dramatic changes in October, even late October; and in several cases, the final poll numbers don't seem to be very close to the actual results.

____________________

ARingMD2B:

Mark-

I love the new features on the site. A feature that I would suggest would be the ability to apply the same transformation of the data to all states at the same time and then recalculate the map. I imagine this would take quite a bit of processing power, but it would be truly awesome.

____________________

Nindid:

I am sorry to focus on minutia here, but the last sentence of the first paragraph is a problem.

"Whatever the label, yesterday was certainly quite a day for fans of Barack Obama."

I am very happy about these poll numbers but I am certainly not a "fan". This is not some sporting event and to suggest that people simply have a rooting interest in the most important political position in the world is somehow offensive this morning.

It is not too big of a deal, but it was unnecessary as well and detracts from the seriousness of your point.

____________________

jme:

@Allen

Can you explain further the rationale behind normalizing the differences?

I agree completely that examining the differences (Obama-McCain) gives a somewhat clearer picture of who may be leading at the moment. But when you normalize by (Obama+McCain) aren't you saying that (38,42) is a bigger lead than (44,48)?

I'm not saying that's necessarily crazy, I just don't follow why that would be reasonable. Do we have empirical evidence that when pollsters push leaners/undecideds that the observed margin tends to increase? Because that's what you would seem to be implying by normalizing in this fashion. My off the cuff intuition would be the opposite; that on average polls with smaller undecided rates would tend to have smaller margins.

____________________

jme:

Just as another quick note, I would second the mentions above; namely I'd love to be able to see plots of differences (Obama-McCain) with trendlines, and maybe even CI's...

____________________

Allen:

@jme:

I should add that another difference is that some pollsters include 3rd party candidates and some do not.

Dividing by Obama + McCain basically recomputes the percentages based on only the respondents who favor either Obama or McCain. It excludes respondents who: favor a third party candidate, are undecided, or refuse to answer.

In other words, the pollster reports the following values:

Obama% = N-Obama / N-total

McCain% = N-McCain / N-total

where N-total = N-Obama + N-McCain + N-3rdparty + N-undecided + N-refused

The normalization excludes 3rdparty, undecided and refused and computes:

Obama%' = Obama% * N-total / (N-Obama + N-McCain) = Obama% / (N-Obama/N-total + N-McCain/N-total) = Obama% / (Obama% + McCain%)

McCain%' = McCain% * N-total / (N-Obama + N-McCain) = McCain% / (N-Obama/N-total + N-McCain/N-total) = McCain% / (Obama% + McCain%)

The difference in these normalized quantities is:

Obama%' - McCain%' = (Obama% - McCain%) / (Obama% + McCain%)

That normalized difference is used to compute a weighted average margin across all polls.

The example you gave is correct:

38 to 42 = (0.42-0.38)/(0.42+0.38) = 5%

44 to 48 = (0.48-0.44)/(0.48+0.44) = 4.35%

Again, the point is to put all of the polls on the "same scale". The theory is that if the first poll had excluded 3rd party candidates or pushed the respondents harder, it would have come up with numbers like 41.3 to 45.65, a margin of 4.35 vs the second polls margin of 4. It tacitly assumes that the undecided and refused votes would break in the same percentage as the decided vote if the respondents were "pushed". It also assumes that votes for 3rd party candidates can be ignored (which is a true statement because they have no effect on the outcome) and that if the polls that do not report 3rd parties candidates either ignored respondents who would vote for them, or lumped them into other/refused, but in any case, the methodology did not change the votes for the major party candidates (if that assumption is incorrect, then the pollster should not be ignoring third parties).

The above assumptions might not be true in all cases, but they do provide an easy and consistent way of normalizing the data that comes from a number of different sources so that it can be combined into a single statistic.

____________________

jme:

@Allen

Huh. Well, I see your point, I'm just unconvinced. On the one hand, I worry that all the assumptions that you listed that are implicit in this re-scaling are speculative; we just don't know if they are reasonable or not. (It would be really cool if it were possible to justify some of them if there are pollsters that provide results to sequential questioning; i.e. numbers for who initially said one thing but when pressed, etc)

On the other hand, it seems to me that the practical difference between simply modeling the differences directly and re-scaling them would be very minimal. (With the obvious exception of a state with a small number of polls where one of them has a _huge_ number of undecideds/3rd party).

But I'd be interested to know: how much do electoral-projection.net's results change if you use just the differences?

____________________



Post a comment




Please be patient while your comment posts - sometimes it takes a minute or two. To check your comment, please wait 60 seconds and click your browser's refresh button. Note that comments with three or more hyperlinks will be held for approval.

MAP - US, AL, AK, AZ, AR, CA, CO, CT, DE, FL, GA, HI, ID, IL, IN, IA, KS, KY, LA, ME, MD, MA, MI, MN, MS, MO, MT, NE, NV, NH, NJ, NM, NY, NC, ND, OH, OK, OR, PA, RI, SC, SD, TN, TX, UT, VT, VA, WA, WV, WI, WY, PR