Too. Many. Daily tracking polls. Update, 22 October.

Politics, Presidential Elections, US Elections, US Politics

The last daily tracking polls update I posted was on 8 October. That’s a lot of daily tracking poll results ago. Yet the bottom line is that nothing much has changed since.

On 8 October, the three-day running average of the daily tracking polls had Obama in the lead by 7%. Today, it has him in the lead by 7.1%. And in the meantime the clock is ticking on, and the window of opportunity for McCain to still close the gap is rapidly closing (or has it already closed?).

One major thing has changed though: there are ever more of the damn things. Daily tracking polls I mean. It seems like no self-respecting pollster can do without one this year (and to think that Gallup didn’t even have one in 2004!).

In early October you already had the long-running Gallup and Rasmussen ones, the Research 2000 one (sponsored by the Daily Kos) and the poll sponsored by Hotline and the alcoholic drinks business Diageo, conducted by FD. Zogby and Reuters started their own on 7 October; Investors Business Daily and the pollster TIPP followed on 13 October, and as of last Monday, ABC and the Washington Post present one as well.

daily tracking polls

Chart 1: All daily tracking polls

If you’re keeping a graph and you were basing any kind of trendline on the average of all daily tracking polls, this – well – sucks. Because every new poll comes with its own house effect, deviating from the others in its own ways. Zogby, for example, started off showing a 2-point lead for Obama when Rasmussen, Gallup and R2000 had it at 8-11 points. Then, just as the Zogby poll had settled nearer to the average, the IBD/TIPP poll came on the market showing Obama in the lead by just 2 points, when the other pollsters had it at 4, 5, 6, 10 and 12 points, respectively. And just when Rasmussen, Zogby, IBD/TIPP and Hotline agreed on a 4-5 point Obama lead last Monday, the ABC/WaPo jumped in with its own tracking poll showing a 9-point lead.

Finally, there was Gallup, on 12 October, providing no longer just data for registered voters, but separate numbers for not just one, but two distinct “likely voter” samples as well, one encompassing a larger universe of voters than the other.

Seriously, you would be put off posting new graphs too.

Which of the three sets of Gallup numbers do you feature? How do you deal with how the ups and downs in your average reflect some new poll having been added into the mix as much as any genuine shift in the numbers? And how meaningful do any of the numbers still seem anyway, when on almost any given day the five-to-nine sets of numbers you’re looking at have Obama’s lead anywhere between 3 and 11 points? (And that’s just the daily tracking polls!)

Chart 1 above, if you click it to enlarge, shows how messy and, by implication, uncertain it all is. The pollsters are all over the map in pinpointing Obama’s lead. But saying that is also identifying the obvious commonality. Every one of the daily tracking polls has Obama in the lead, and that’s been true since 17 September, over a month ago.

How big is that lead though?

During the last week, the IBD/TIPP poll has had it at a relatively cautious 4-7 points. While IBD and TIPP may not be big names, it’s worth pointing out that in 2004, the then-IBD/CSM/TIPP poll came the second closest to the actual results among independent pollsters with its final projection of Bush 50%, Kerry 48%. Rasmussen, which did well in 2004 too (Bush 50.2/Kerry 48.5), has also been cautious, indicating an Obama lead of 4-6 points. Though it can’t be assumed to affect their polling, however, both Rasmussen and IBD have conservative politics.

xkcd cartoon

A propos of nothing in particular...

The other tracking polls have been more boisterous in signalling a comfortable Obama lead. Zogby had Obama’s lead at its lowest in eleven days last Sunday, but since then it’s moved it up from 3% to 10%. The Hotline poll has had it at 5-10 points in the last week. Research 2000 has, as always, been most optimistic, showing Obama 7-11 points up.

Like all polls, though, each of these polls have their raps too. Back when Zogby was showing a comparatively small Obama lead, Nate Silver of fivethirtyeight.com wrote derisively about its methodology:

Zogby is the outlier [..], and that’s because Zogby has the odd practice of fixing his poll’s party identification weights based on what they were in the last presidential election. In Zogby’s world, then, it’s still 2004, when there were roughly as many Republicans as Democrats. Although Zogby’s trendlines may be worth looking at, his topline numbers are basically unusable.

Of course, right now Zogby has Obama up by 10, so if that’s still based on data weighted by 2004 party ID numbers, all the better. On the other end of the scale, the Kos/Research2000 poll has raised some eyebrows by systematically polling better results for Obama than any other tracking poll. Seriously – in the first 37 days of its existence, it had the best or shared best result for Obama among the tracking polls on all but 5 days. It’s only the last five days that it’s fallen in line.

The Hotline poll, meanwhile, is criticized for using significantly smaller sample sizes than the standards of the genre, Gallup and Rasmussen. The latter two’s numbers are based on rolling three- and four-day samples comprising a total of over 2000 respondents. The ABC/WaPo, IBD/TIPP and Zogby numbers are based on totals of 1000 respondents or more. Today’s Hotline poll numbers, in contrast, are based on 782 interviews over three days. This also explains why the Hotline numbers tend to be more volatile than the others, while Rasmussen’s numbers (based on a rolling set of 3000 interviews) are extremely stable. The result is that Rasmussen’s numbers reduce the role of statistical noise, but are also rather boring to report. As they write today:

It is difficult to overstate the stability of this campaign [..]. For the past eleven days, Obama has been at the 50% or 51% level of support every day [..]. During those same eleven days, McCain has been at 45% or 46% every day and the gap between the candidates has stayed between four and six percentage points.

If you go back a couple of weeks further, the results are pretty much the same. It’s now been 27 days since Obama’s support moved below 50% or above 52%. During that period, the number voting for McCain has stayed in the 44% to 46% range and the gap between the candidates has ranged from four to eight percentage points.

Finally, there’s Gallup, which has shown generous leads for Obama among all registered voters (6-11 points) and its expanded likely voter model (4-10 points), but its traditional likely voter model has had it as low as +2 and now has it at +5. FWIW, both pollster.com and fivethirtyeight.com are opting to use Gallup’s expanded likely voter model for their calculations.

daily tracking polls, successive averages

Chart 2

Clear as mud, right? An average sure would help, but there’s this problem with new polls being added into the mix and jerking the average up or down in the process. Chart 2 constitutes an attempt to solve that dilemma.

In this chart, every time a new poll is added into the mix, the “Average all” line takes it into account, but an alternate line continues on its way reflecting only the polls that existed before, for apple-to-apple comparisons. Of course, the more polls are integrated, the less volatile the fluctuations become, so it’s no surprise that the lines converge pretty closely near the end.

All in all it gives one a good feel for the overall trend. There’s a faint suggestion that the race tightened slightly in between the 11th and the 18th, but that this tightening has not continued since. (Possibly relevant fact: reactions to the last presidential debate will have started showing up in the polls published on the 17th and become fully integrated by the 19th or 20th.) Note, however, the analysis of Alan Abramowitz on pollster.com of the variation in the daily tracking poll numbers between the 10th and the 20th and his conclusion that it most likely amounted to nothing but random statistical variation.

daily tracking polls, respective averages

Chart 3

Chart 3 offers another alternative look aimed at charting an average, but still doing some justice to the variety between the polls. It plots how the daily tracking polls average varies when you include all but one of the polls, each line representing the exclusion of one of the respective polls. Again, even leaving out one poll, the averages still each represent six polls come yesterday’s numbers, hence why the difference between the options becomes smaller over time.

The most striking thing here, apart from the same overall trend, is how excluding the Kos/Research 2000 poll impacts the average. Excluding it instantly decreased the average Obama lead by up to a percentage point – that’s the yellow line.

If you’re interested in the house effects of the different tracking polls compared to each other, click on through to this post from 12 October by Charles Franklin at pollster.com as well. It has a good graph clearly illustrating the house effect of the Research 2000 poll, as well as the greater volatility of the Gallup poll when compared to the Rasmussen one.

A final graphical nod to pollster.com. The above charts cover only the daily tracking polls. Pollster.com charts all national polls, tracking and otherwise. Plus, as user you can yourself set the parameters of the graph you get to see. Below is the current chart if you exclude internet polls, limit the date range to start at 24 August, and set the percentage rate at 35%-55%:

2 Comments

2 Comments

  1. sozobe  •  Oct 23, 2008 @5:57 am

    Ha, I sympathize! It’s really been a tsunami of polls lately. Nice analysis though, thanks!

  2. nimh  •  Oct 23, 2008 @2:07 pm

    Just saw that Nate Silver published a useful run-down of the daily tracking polls, with their respective strengths and quirks, last Tuesday: Tracking Poll Primer

Leave a Reply

Allowed tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>