« Democracy Corps Post-Debate Panel Survey Gives Kerry Solid Win | Main | A Note on Who Posts What »

A Note on the ABC News/Washington Post Tracking Poll

Alan Abramowitz points out:

The Washington Post tracking poll seems to be suffering from the same ailment that afflicted the Gallup tracking poll four years ago, albeit on a smaller scale so far. In the past few days we have seen Bush's lead among registered voters shrinking while his lead among likely voters has increased. This means that the likely voters and the unlikely voters are moving in the opposite direction, just as they frequently did in the Gallup tracking poll four years ago. This makes no sense, of course. With the WP tracking poll, as with the Gallup tracking poll, the registered voter results are probably a better indicator of the actual standing of the race.

Well said. It's also worth notiing that, in 2000, the ABC/WP tracking poll missed the final vote pretty badly, having Bush up by 3 points at the very end and 3-4 points up on every night of the final week. Looks like they're poised to repeat their fine 2000 performance.


Ruy (or anyone else) I saw 3 or 4 separate posters below ask about Rasmussen's daily tracking polls, now showing bush at 50%. Does anyone know what his internals are? Does he publish them? and what about his trend, assuming his internals are not changing?

Overall, I remain convinced that Nothing has changed in the big picture from something charlie cook wrote mos ago. Bush at 45, kerry at 45, and then 10% undecided; but that 10% skewed democratic, and in addition, historically breaks against incumbents.

Thus, I still see a kerry win, with 52-53% of the vote -

Except for rasmussen...



Dr. Sam Wang's comment of 9/25 is as concise a statistical critique of Gallup and its oversampling problems as I have seen.


ALSO...has anyone in-depth knowledge of Dr. Stimson's (UNC) two party vote multi-poll smoothing methods described here:

"The summary series is Kerry Percent of the Kerry plus Bush total, the right number to know
if one assumes that all electoral votes will be shared between the two major candidates.
The series is based upon commercial results posted to pollingreport.com, supplemented with
the Rasmussen Reports daily tracking poll. It assumes comparability across survey
organizations, but not across sampling methods (adults, registered or likely voters) or
question formats (Bush v. Kerry as opposed to Bush v. Kerry v. Nader). Where
multiple surveys are reported on the same date, aggregation weights by sample size,
i.e., reproducing the grand mean of several studies as if they were one big study.

Dating is by the beginning date of field work, typically about four days earlier than
reporting dates. Thus the series always appears to be a few days old, even though based
on the most recent available data. (One can get a decent approximation of release dates
by forward shifting the time scale about four days.) The series will be updated daily through
election day.

The methodology for rendering mixed date and only partially comparable survey results into
a single underlying series is developed in James A. Stimson, Tides of Consent: Public Opinion
in American Politics, New York: Cambridge University Press, 2004 (July), see particularly
Chapter 4: The Great Horse Race: Finding Meaning in Presidential Campaigns. Cambridge
University Press Listing (excerpt, table of contents, front matter, etc.).

Filtering: The smooth line in the graph is an estimate based upon the Hodrick-Prescott filter,
a means of estimating the smooth underlying movement of a time series while limiting the
effects of unpatterned very short term (one day, one poll) variation. The filtered estimate
should capture real long-term shifts in sentiment better than the raw series. Its final day
estimate is the basis for the electoral college forecast.

Where does one find the ABC registered voter data

I can't even find the RV listed, though last time I saw, Bush led by 3%.

The ABC news web site has a very detailed discussion of the methodology of this poll. This discussion does not really get into detail on their model for identifying likely voters, however.

There are several aspects of this poll (at least its registered voter numbers) that in my opinion are superior to most others, but its failure to adjust for Hispanic voters throws low-income Hispanic voters (who are likely to be very poorly sampled because of multiple jobs and language) in the same categories with low-income whites. This very likely biases the poll a couple of points Republican all by itself.

Positive aspects:
1) They get the demographics of everyone they reach on the phone, weight these demographics against Census data, and then throw out the people who aren't registered voters. This has a big advantage over the common practice of getting demographics only from registered voters and weighting those demographics against 2000 exit polls, because ABC's approach should, in principle, pick up new voter registration trends while weighting against 2000 exit poll demographics builds into your results the assumption that the 2004 electorate will be the same as the 2000 electorate.
2) They assign weights to individual cells rather than variables -- e.g., they compare the number of respondents who are college-educated black men over 65 years of age to the census figures for the same category to get the weighting factor, instead of multiplying together 4 weighting factors for black, male, college-educated, and over 65. This avoids two mathematical pitfalls: voting behavior is not always a linear function of the various variables and the estimation problem for variable weights is error-prone because it is likely to be a poorly posed problem.

Both these positives bring negatives with them:
1) Weighting against the population and then throwing out the non-registered may introduce errors.
2) The cell-weighting is what probably motivates them not to weight for Hispanics. You need to have at least one respondent in each cell -- there's too great a risk that your sample of 1000 Americans won't include, for example, any Hispanic woman college graduates over 65. There are ways around this (don't subdivide Hispanics by education, for example) and the poll would be much better if it treated Hispanics as a separate category.

It seems to me that if you subtract the 2000 prediction of this poll from the election results, and apply the same correction to the current predictions of this poll, you will have a reading on the electorate that in many ways is superior to what you can get from any other poll -- since this poll does reflect the changes in composition of the electorate.

I know it has been said, but I think it is a mistake to focus on anyone poll as the gospel. I have seen people come on here to question why one poll or another is showing disparities or a minor Bush increase, and, I believe this is fruitless in such a tight race. Some polls have Kerry up, some have Bush, others have them exactly tied. Here's the thing- even in the WP/ABC poll- Bush's should be elected numbers were below 50 percent (he continues to be in the danger zone) and thats all I needed to know from that poll. At this point, I hope people are focused on three points: 1) Staying on the offensive (that's out of our hands and in Kerry/Edwards hands); 2) Countering any manipulations of the process (ie, preventing suppression of newly registered voters and of FEC law, such as by Sinclair, and, in this arena, I hope everyone has been sending out letters and emails to friends, Congress people, etc to prevent this); and 3) Get out the vote- I hope everyone here is volunteering to make sure that they are phonebanking, doing poll protection (especially if you are a lawyer) and assisting groups such as ACT. For the record, I am phonebanking to swing states over the next three weeks.

Bush is up 4-5 in WaPost/ABC and Rasmussen tracking polls. I haven't seen the bump in Kerry support that I expected after the second debate, which I thought he did well in. I'll wait to see more polls but right now I'm a little concerned.

Interesting thing about Rasmussen. On the day Kerry moved up to within one point he wrote that things have been trending in Kerry's direction, for example Bush's job approval was below 51%, and optimism about the economy was dropping. But the next day, Bush's job approval started going up as well as his poll numbers, all during the Bremer and Rumsfield comments and the WMD report. Doesn't make sense.

Reminder from EDM

1. Comments that directly attack or belittle other participants are not acceptable. Recent comments have been increasingly testing the limits in this area and, as a result, in the future this prohibition will be more strictly enforced.

2. EDM is currently evaluating a variety of options for creating forums or discussion areas specifically for Democrats. As a result, some comments have been temporarily held back in order to introduce them in the context of this upcoming discussion.

3. Please remember that, because EDM is now reviewing all comments, there will usually be a delay before your posts appear.

>Where does one find the ABC registered voter data

RV data can be found at WaPo site, click link for "trend" (although, this link does not appear today for some reason)

Historical RV data for last week can be found at http://2.004k.com/tracking/

check out this article b/c it references something that I have been mentioning about a higher than usual African-American vote this year: http://www.nytimes.com/2004/10/11/politics/campaign/11voters.html

Also, I don't really understand the point of focusing on two polls where other polls are trending differently- are y'all just interested in worrying? Rammussen and WP/ABC say one thing whereas Zogby, newsweek, LA Times, Time Magazine, CBS, Gallup are saying something else. I firmly believe at this point that the polls are like a Rorschach test where people see what they want to see

Don't expect too much of a bump after the debates. All Kerry has to do is explain the differences between him and Bush.

He'll win the anti-Bush vote.

Bush will win the pro-Bush vote.

The next important step then has to do with turnout.

If more Dems vote than reps. vote: we win.

If more Reps vote than dems: we lose.

Turnout is key and history is on our side.

This uncertainty does not prevent Stephanopolos from presenting the results as an established reality or fact, as he did on his Sunday program. The news about the polls is their inherent inaccuracy, and it is intellectually dishonest to pretend otherwise.

I know that rasmuessen does weight for party ID.

They did'nt in the last election, and they had bush winning by 9. This election they are weighing it.

The Kerry-mentum is continuing:

Reuters/Zogby has Kerry 47:44 ahead in the latest (monday) three-day tracking poll, up two points from Sunday. 6 percent undecideds; 16 percent of independents are undecided.

"The close race turns up the pressure for Wednesday's final debate", they write. That is so true.


ABC and Rasmussen are going one direction, Zobgy (phone daily tracking) the other. Some of the divergence might be the method used to identify LV's.

Ben R.-

Thanks. That was a very informative post.

As one of those who cited Rasmussen for comfort back when other polls showed >10% Bush leads, of course I am concerned also. Objectively speaking, though, the interesting thing about Rasmussen is that to my recollection he's consistently had the smallest leads for either side and the smallest swings in either direction of all the pollsters. I first noticed his polls back in the early spring when Kerry first became the likely nominee and surged to a large lead over Bush in just about every poll _except_ Rasmussen's. I can't help wondering just what about his methods leads to this result, and whether it's closer to reality than the others, or farther from it.

We all know to be wary of polls. Still, I can tell you from being a political junky for many years and following this stuff that Zogby has an excellent reputation for accuracy in his polling. Look for Zogby polling data for a more accurate snapshot of the electorate.

Go Kerry!

St. Louis, MO

This just occured to me:
Since my youth I have heard or read over and over the hilarious story about the start of Presidential polling, in 1936, when the _Literary Digest_ predicted a solid Landon victory, because they polled by telephone and in those days a disproportionate number of Democrats couldn't afford one yet. (For those who aren't up on their history, Landon carried Maine and Vermont, for 8 electoral votes.)
Wouldn't it be ironically appropriate if this year the whole polling enterprise is finally revealed as futile because a disproportionate number of Democrats don't have land lines, or are never home, or screen their calls?

Today's Wall Street Journal has a very interesting article about the difficulty of getting low-income Hispanics to participate in market research. The same must apply, at least to some degree and maybe even with greater force, to their participation in political polls.

I hope I won't bore everyone if I give a simplified example to explain the significance of "cell weighting" and also show the importance of weighting low-income Hispanics properly.

Suppose you are surveying a population that contains equal numbers (25% each) of college-educated Anglos, non-college Anglos, college-educated Hispanics, and non-college Hispanics. Suppose further that college Anglos vote 52-48 Kerry, non-college Anglos and college Hispanics vote 60-40 Bush, and non-college Hispanics vote 76-24 Kerry.

After completing 1300 phone inteviews, your sample consists of 400 college Anglos, 400 non-college Anglos, 400 college Hispanics, and 100 non-college Hispanics. If you use cell-weighting like ABC/Wash Post, you will count each non-college Hispanic 4 times and your poll will give Kerry a 52-48 advantage.

If you use parameter weighting like CBS/NY Times, and I think most other polls, you will count Hispanics double and count non-college respondents double. (In this simple example, you can solve exactly for the weights so unlike the real case you don't have a numerical error in your estimation of the weights.) Your weighted population will consist of one-third non-college Anglos, one-sixth college Anglos, one-third college Hispanics, and one-sixth non-college Hispanics. (But looking at the variables one at a time, you have half college-educated and half Hispanic, so you might think you have accurately weighted the sample.) With parameter weighting, your poll will give Bush a 52-48 majority.

Note the importance of non-linearity here -- the reason this example works out the way it does is that college education has a different effect on the voting patterns of different ethnic groups. This seems likely to me -- certainly here in the Washington suburbs where a non-college Hispanic is usually a Salvadoran immigrant while there is a sizable population of college-educated Hispanics with occupations like economist at the IMF.

Does anybody have an explanation of the generally more sympathetic reading of the Kerry-Bush race right now that we're getting from the Zobgy/WSJ battleground polls?

The one from a couple weeks ago had Kerry up 0.1% in Arkansas, even, and something like a 10% bulge in NM and Minnesota. Does anyone know anything about their methodology? I, of course, want to believe all of those polls, but considering a few others (like some of those showing Bush with a lead in NM) I'm wary, of course.

It's my nature to remain as cautiously optimistic as possible...rather than cocky or confident. Three more weeks of this tension...


That was an excellent and very edifying example you gave us. Can you possibly tell us something about your education/training, just so we understand where these insights are coming from?

I am not a professional pollster or statistician, but as a physician who does some clinical research I know a tiny bit about statistical & sampling methods, and I would wonder if the process of weighting by cell, as you described, might produce some volatility & inaccuracy in a poll with regard to narrow population subgroups. For example: this country being the way it is, the number of college educated black men over the age of 65 is pretty low. Let's say you surveyed 1000 people, and there was only 1 voter who fit that category. Let's say it just so happened that he's a neurosurgeon, and a one-issue voter for this election (i.e. tort reform), so he's backing Bush. His response will now be weighted to represent his whole slice of the electorate, even though we all know darn well that the vast majority of blacks at every age and educational level overwhelmingly support Kerry.

In other words, if your cell values get small enough, then those values become potshots rather than good statistical samples. This is obviously the trade-off for trying to get some greater accuracy by subdividing the population as much as possible. Do you think this is a serious consideration, or is it likely to just wash out in the end?

Another good article :


Basically, it says what anyone who is not worrier already suspects- that this is a tight race, but right now Kerry has the Big Mo- even in the swing states. And its showing what we can see from several of the electoral college sites- that Kerry is making headway in swing states like Ohio and Penn and its clear that it is possible that he w ill not even need Fl (and to my surprise- my state Va maybe a swing state this year- as for one reason why I go back to the tech crowd in the northern part of the state and angry AA (we black folk got a longer memory than most Americans - if you will allow my insider view of why this is the case) who will turn out in higher number (as I mentioned there is an excellent article on NY Times about black turn out this year where it ends with one saying at age 80 plus that if it is six feet of snow out and she has to walk- she's going to vote and this has been true across the political divide- in part b/c we represent a disproportionate number of soldiers onthe ground among other more traditional reasons.)

> As one of those who cited Rasmussen for comfort
> back when other polls showed >10% Bush leads, of
> course I am concerned also.

According to CNN columnist Mark Shields, the thing to look out for is "Shrub's" current share of the vote since it likely estimates his "ceiling". Undecided voters typically favor the challenger, so they almost always receive a bigger share of the vote than e.g. the Labor Day polls suggested.


My feelings are a mix of cautious pessimism and cautious optimism. I do feel Kerry has a good chance *if* he can focus consistently on issues that favor him, and *if* he gets a couple of lucky breaks during the final three weeks. This month has been fairly good to him so far. As long as polls indicate a majority of voters believe the country is on the wrong track, anything can happen.


Ben Ross...I did an extensive analysis of Asian nnd Hispanic voting behavior in SF using the Roper Social Capital Benchmark survey and local elections and polling data.

The long and short of it which I am fairly certain is not simply a left coast city phenomenon is that Hispanics (broadly defined) have extremely low political efficacy/participation rates..

Of the four broad ethnic groupings they rank lowest and are the most difficult to get to the polls...

Only the Southeast Asian recent immigrant is more difficult.

I think this may go a long way towards explaining why Texas, despite an ever growing Hispanic community, remains solidly Republican

The latest Zogby poll has Kerry 47, Bush 44.
Even though those numbers are very encouraging, there was one troubling element in the poll - he's showing Bush now leading with 18-29yo's 47-38.

Before getting overly concerned about these numbers with young voters, i think it needs to be put in perspective. Zogby's poll 2 days prior showed Kerry at 53 in the 18-29yo sub (I don't see what Bush was polling at in the subgroup). Also, virtually every other poll I have seen broken down by age has kerry blowing bush out of the water with the 18-29yo. Further, what has happened in the last 2 days to produce a 15pt shift away from Kerry among this group, while at the same time Kerry's overall support seems to be increasing.

If the shift is real, then we should start worrying. However, at this point I'm willing to chalk it up to being a bad sample within the subgroup. I think 18-29yo's are particularily difficult to get a good reading on because the cell phone/caller ID issue and the fact that this polling was conducted entirely during the weekend: young urban kerry voters are out painting the town red while young rural bush voters are watching watching football in between trips to the local wallmart for entertainment - not to sound like david brooks.

Anyhow, I'm generally encouraged otherwise by Zogby's latest poll

Another good site for poll analysis is:


They have a good discussion today about how Kerry handily won the second debate among independents.

BTW I think many people are greatly overlooking the negative effects of Bush's winks, shouts, and jabs in the second debate. I think it left a lasting negative taste in the mouths of many undecides.

Ane earlier post mentioned the Literary Digest poll of 1936. I too have been thinking about that poll for weeks now.

For those not familiar with it, one of the greatest textbook cases of botched polling came in the 1936 presidential race, when the Literary Digest's poll (the most respected poll at the time) said FDR would lose in a landslide. He, in fact, WON in a landslide. What was the flaw? The people surveyed (by postcard) came from their subscriber list, the phone book, and from lists of car owners...and in those days, at the height of the Depression, only the more affluent subscribed to the Literary Digest, had phones, or owned cars. So a great deal of FDR's supporters were not even included in the sample. That's called coverage error.

I think the polling biz might be about to pull another Literary Digest. I think there could be big time coverage error going on, seriously undercounting Kerry supporters. Huge numbers of new voters are being registered (and are thus not included in the polls we see), more and more younger voters have only cell phones (which are not called in polls), and lower SES (socio-economic status) voters are more difficult to reach on the phone. All of these groups favor Kerry. Furthermore, it seems that new Democratic registrations are far outstripping new Republican registrations in many of the key battleground states.

I also think the "likely voter" models being used may be out of date or just plain irrelevant: Ever since the primaries, there seems to be evidence of a larger than usual turnout motivated by intense and widespread desire to defeat Bush. Demonstrations of approximately 500,000 people each in Washington and in NY also suggest vast discontent that hopefully will lead to high turnout. And Kerry is now attracting huge crowds even in swing states (60,000 in Portland; 20,000 in Seattle; 12,000 in Albuquerque; 40,000 in Philadelphia; about 20,000 in Elyria, Ohio just a few days ago; etc etc) that belie a flagging campaign. In other words, this year's likely voter may be quite different from 2000's likely voter.

I'm not smarter than the major pollsters; I mean, they too are aware of all this. But they are in denial that their techniques may be antiquated. They think all their data weighting, modeling, and other statistical sleight of hand will pull their chestnuts out of the metaphorical fire. I think they could be wrong.

Also, the internals of the polls seem to indicate that, compared to 2000, Bush has less support now among Hispanics, Blacks, Arabs, Jews, and other demographic groups -- most notably, independents. The internals also show (I think) that more Bush 2000 voters support Kerry than Gore voters support Bush. Only Fear And Smear could explain a Bush lead in key states. I just hope Kerry can somehow overcome it.

To answer scottso, I have a Ph.D. in statistical physics, and haven't forgotten all of it. A part of my work also involves critiquing computer models (not related to polls), trying to look at the inputs and outputs and figure out if there are any mistaken assumptions or approximations, or numerical artifacts that lead to misleading results.

It's true that cell weighting increases the sampling error when you heavily weight a cell that has only one respondent. But if your cell had only one respondent and you did parameter weighting, you are effectively estimating the responses of the "missing" respondents from the responses of other people who share only some of their demographic characteristics. This risks introducing both systematic and random errors. As a reader of polls, I am more concerned about systematic than random errors because I can greatly reduce the random error by looking at several polls.

Aside from the nonlinearity issue, my problem with parameter weighting, which I haven't seen discussed elsewhere, is the numerical errors that you can get in estimating the parameters. How big they are I don't know -- it would be extremely interesting to look at the total contribution of individual cells (number of respondents multiplied by weight) in successive polls by the same organization, and see how variable that number is. If the number jumps around, it would be an additional source of random error in addition to sampling error.

To John McC - As far as Hispanic vote goes, it's certainly less than the Hispanic percentage of the population, but it's not negligible. Tarrance group estimates 6% nationally. See:
This useful article also will tell you that the Tarrance Lake Battleground poll weights to equal numbers of Republicans and Democrats -- so you need to raise the Kerry percentage several points to get a reasonable prediction out of that poll.

Young Kerry voters "painting the town red?" I think "painting the town blue" would be a better choice of words.
But more to the point, age and other subgroups often break down into too few voters for statistical reliability, and I sometimes wonder how they screen for age. Zogby' s poll is not in and of itself anything to worry about, but young voters (of which I am one) are often unreliable for either party, frequently changing their minds and with lower registration and turnout rates. Hopefully Democratic efforts will change that this year.

Gallup now shows Kerry up 1% in a 3 way, LV, and tied in a 3-way RV. That's a slight improvement from their last poll, which puts their trend more in the Zogby camp and away from the Rasmussen/WaPo camp.

"Also, the internals of the polls seem to indicate that, compared to 2000, Bush has less support now among Hispanics, Blacks, Arabs, Jews, and other demographic groups -- most notably, independents."

That's probably not true among Jews (of which I'm one, so I follow this trend pretty closely), but that's largely because he's starting from such a low number: Bush got about 20% of the Jewish vote in 2000. Republicans have been bragging for a long time now about getting a big boost in that number because of Bush's support for Israel, but some of their predictions have been crazy, like 35-40%, or even more, of the Jewish vote going to Bush this time. Nothing even remotely close to that is likely. The handful of polls done this year with data on Jewish voters seem to indicate low-to-mid 20s as the most likely number for Bush, probably no higher than 25%. Republicans then sometimes brag about how they'll do well enough with Jews to make the difference in some swing states (like Ohio and Florida) even if they don't improve that much from 2000 nationwide, but this also doesn't seem very likely to me.

In reference to one of the posts above...why do newly registered voters not show up in the polling numbers?

A few comments about Rasmussen and other polls. I've been watching the Rasmussen polls for a few months. Generally, 3 things go together: Bush-Kerry difference, dem-republican congressonal preference and Bush approval. I suspect that the dem-republican congressional preference is actually more stable than the Bush-Kerry choice. Therefore, I tend to use it as a measure of sampling preference. Right now, the dem-republican difference is at a low ebb for the dems (still, dems are preferred), suggesting that the Bush preference is due to republican oversampling. I've read that Rasmussen accounts for this; nonetheless, watch the corresponding drift of democratic-republican congressional preference and Bush-Kerry preference.

Second note; one thing I really hate about the 3-day rolling average polls is that you can't determine the source of a one-day shift. If a poll shifts in one day it could be because of a sharp shift the evening before or a shift in the opposite direction 3 days prior. This is extremely annoying. I like seeing the rolling average. But I would also like to see the underlying daily polls.

Finally, there were several polls that indicated that Kerry won on Friday. One is the zogby poll, which indicated that he won by about 15%. Second was a Democracy Corps poll of a sample taken before and after the debate, again giving Kerry a strong win, especially among uncommitted voters. In general, the sample-resample approach (even with relatively small numbers) seems much better suited to determining whether a particular event, such as a debate, leads to a change.

I think they do. I too have been surprised by many posters saying that they aren't being counted in the polls. I think they are, if they have a phone.

Why did Bush do so much better among African Americans in Virginia (21%) in 2000 than any other state? Was it the military vote? Just curious. Because if the black vote in VA mirrors the rest of the country - 8% for Bush - then I think Kerry can win it.

What Alan neglected to mention regarding ABC News' fine performance in 2000 was that most of the other polling organizations ended their tracking with Bush ahead by 2. So I guess they're all suspect. Someone has an axe to grind.
G B N Und Oth
Gallup/CNN/USA Today 46% 48% 4% 0% 2%
Pew Research 47% 49% 4% 0% 0%
IBD/CSM/TIPP 46% 48% 4% 0% 2%
ICR/Politics Now 44% 46% 7% 1% 2%
NBC/WSJ 44% 47% 3% 4% 2%
ABC/WashPost 45% 48% 3% 3% 1%
Battleground 45% 50% 4% 0% 1%

*source: National Council on Public Polls

There are a number of reasons new registrants may be undercounted.
One is that they are disproportionally young, and the Democrats among them probably urban; therefore more likely to rely entirely on cell phones, or not be home when a pollster calls their land line.
Also, I wonder: do any pollsters pick their victims from the voter rwegistration rolls? They may not have the latest updates.
In any case new registrants won't pass the "likely voter" screen because they didn't vote the last time!