pageok
pageok
pageok
Iowa Poll Results.--

The most important of the local Iowa polls (Des Moines Register) shows substantial leads for Obama and Huckabee.

The composites at Real Clear Politics, on the other hand, show nearly dead heats (with very slight leads for Clinton and Romney).

Although caucuses are notoriously hard to predict, it will be interesting to see which polling agencies (or composites) are closer to the final results.

b.:
Edwards FTW.

I say this not because I support his run, per se, but because I'm clairvoyant.
1.2.2008 8:13am
Lior:
Pollster, an aggregation site, is showing Clinton and Huckabee.
1.2.2008 8:51am
Elliot Reed (mail):
The intrade prediction market puts Huckabee at 53.0 and Romney somewhere between 40 and 44 (the last trade was at 38, but the current bid-ask spread is 40-44). The markets are a pretty good indicator of the conventional wisdom, so it looks like the CW is expecting a tight race between Romney and Huckabee. Makes sense, given that the De Moines Register poll is something of an outlier.

As a Democrat, I say: go Huckabee! Easily the least electable of the Republicans.
1.2.2008 9:35am
HO (mail):
A local newspaper published the results of the last 10 Iowa presidential caucuses. They picked the eventual Democratic nominee correctly all but twice (and one of those times was when a local politician -- Tom Harkin -- was running for president).

The people in Iowa did not do as well with Republicans. They only correctly identified the eventual Republican nominee about half the time.

It looks like it is make or break time for Democrats. Not for Republicans.
1.2.2008 9:42am
AntonK (mail):

...it will be interesting to see which polling agencies (or composites) are closer to the final results"
And why is that?
1.2.2008 9:44am
Sean O'Hara (mail) (www):
Why has compositing polls become so popular? Taking polls with different methodologies and margins of error, and averaging them together doesn't give you any meaningful results.
1.2.2008 10:16am
Dave N (mail):
Two things to remember about the Iowa Caucuses:

1) It is a caucus--which means that people have to go to a specific location at a specific time, not the 12 hours or so that a person has to cast a vote in a regular election. Thus, a voter has to have a high degree of motivation to vote--and there is a premium on organization in getting supporters to the poll. This helps Edwards and Romney, who have the most actual organization in place in Iowa.

2) On the Democratic side, if a candidate does not get 15% of the vote in a local caucus, those candidate's supporters then get to join a group with greater support. My hunch, though I have no proof, is that this helps Obama and Edwards more than Clinton.

My prediction, BTW, is that Huckabee and Edwards win in Iowa; McCain and Obama win in New Hampshire.
1.2.2008 10:22am
merevaudevillian:
Dave N, on the second point, why is that so? It is almost certain that Obama, Edwards, and Clinton will each receive 15% of the vote in every local caucus. The major players outside that are Biden, Dodd, and Richardson, none of whom are likely to get 15% at many, if any, local caucus. Richardson has made clear his support of Clinton for some time now; Biden and Dodd as experienced senators playing up the experience and foreign policy expertise. I would think Clinton would get a substantial portion of their support.

Granted, I think if Edwards or Obama supporters failed to achieve 15% at a local caucus, I think it would help the other. But I'm not so sure the Biden, Dodd, and Richardson supporters would gravitate toward Edwards or Obama.
1.2.2008 10:27am
GV_:
This helps Edwards and Romney, who have the most actual organization in place in Iowa.

I don't believe this is true, although it is "conventional wisdom." Both Clinton and Obama have more money and organizers in Iowa than Edwards. And Clinton and Obama have almost spent as much time this year in Iowa as Edwards.
1.2.2008 10:33am
Cold Warrior:

Why has compositing polls become so popular? Taking polls with different methodologies and margins of error, and averaging them together doesn't give you any meaningful results


Agreed. In my spare time I'm kind of a baseball stats geek, and I see the same thing going on with the various competing measures of defensive performance. What this amounts to is a concession that we lack confidence in the methodology/accuracy of any individual defensive metric. But other than the kind of seat-of-the-pants crude conclusions that may be drawn from referring to several different (flawed) systems -- for example, "all metrics say Manny Ramirez is a very poor left fielder, so I think we can say that it's true" -- "compositing" various systems doesn't necessarily result in a higher level of confidence.


My prediction, BTW, is that Huckabee and Edwards win in Iowa; McCain and Obama win in New Hampshire.


Funny, I was going to say the same thing. Perhaps a bit of wishful thinking on my part. Not that I necessarily want to see any of these 4 succeed; rather, I'd like to see the whole Iowa-NH thing rendered irrelevant, and the actual nomination thrown to the bigger states.


As a Democrat, I say: go Huckabee! Easily the least electable of the Republicans.


As a libertarian, I disagree. I'm afraid of Huck. He combines the easy, natural b.s. of his predecessor in the Arkansas Gov's Mansion (whenever I see him pick up the electric bass my mind sees Clinton with a saxophone) with the big government "compassionate conservatism" and born-again purity of the current occupant of the White House. Americans have elected some strong mix of these factors in the last 4 presidential elections (o.k., maybe 2000 doesn't count with the popular vote going the other way), and to me the appeal of a folksy professional politician who seems so likeable on the surface is, well, rather scary.
1.2.2008 10:38am
John Neff:
There have been so many polls in Iowa that 40% of those called hang up and some people will not answer their phone (you have to leave a message and it they know you they call back). This means that all polls are worthless and they are averaging as many polls as possible and hoping for the best.

They want us to call people in our precinct and ask them to caucus for our candidate and after talking to some of our neighbors we have concluded we would do more harm than good if we did so. People are also annoyed about the mailings evidently the candidates think the the one who kills the most trees will win.

You will find out how it will turned out late on Jan 3rd or early Jan 4th and if a poll gets it right it will be dumb luck.
1.2.2008 10:42am
warm peaceful guy (mail):
Cold Warrior, you reject compositing polls (and are right to do so), and make a baseball analogy. But then you refer to the "popular vote" in 2000 and say that the election may not count for your purposes. There is no such animal as the "popular vote." It's a meaningless statistic. The national election is a state-by-state contest, and adding up all the votes from all the states to come up with a winner is nonsensical.
To make a baseball analogy (one that I saw here recently), the team that wins the World Series is not necessarily the team that makes the most runs over the series. You don't add up all the runs from all the games and declare a winner. You add up the won games, and the time that won those games obviously had the most runs within those games. But it is not the case where the team that gets the most runs wins the series. So a nominee wins the election by winning the most states, not by winning the most votes from all the states.
1.2.2008 10:48am
Dave N (mail):
merevaudvillian and GV,

Thank you for your responses. In answer to your two critiques--I agree that Clinton, Edwards, and Obama will all get 15% at most caucuses, so the question becomes who Biden, Dodd, Richardson, and Kucinich supporters will gravitate toward. Since Clinton is the "establishment" candidate, my thought is that her support is "hard" but that she is the second choice of fewer people than Edwards and Obama are.

As to why I think Edwards will win overall, he has spent the most time organizing in Iowa, particularly in rural areas that are underrepresented in most polling data.

BTW, Sean O'Hara made a very good point in the post immediately before my previous one--aggregating polls with different methodologies, questions, and sampling error rates is stupid.
1.2.2008 10:50am
Cold Warrior:

But then you refer to the "popular vote" in 2000 and say that the election may not count for your purposes. There is no such animal as the "popular vote."


This is just a weird and incoherent criticism.

I'm trying to point out that Clinton-W "regular guy/feel-goodism" has mass appeal, and to do so I'm noting that Clinton and W have, indeed, won the last 4 national elections. I am then noting that 2000 may suggest that regular guy/feel-goodism is not always the recipe for success, as the glum and robotic Al Gore actually outpolled W nationwide. It has nothing to do with questioning the legitimacy of W's victory.

Some people just need to get over it.
1.2.2008 11:03am
warm peaceful guy (mail):
I wasn't suggesting that you were questioning the legitimacy of W's victory. I was saying that you referred to the "popular vote" as if it meant something. It doesn't. I don't care at all about your views on the 2000 election. I'm talking purely about statistics.
1.2.2008 11:17am
Cold Warrior:
The popular vote does mean something. So does that total Runs Scored/Runs Allowed measure in baseball. Neither measure is how we choose winners in presidential elections or in baseball pennant races. But it is silly to say these things don't "mean something." It means what it obviously means: more people went to the polls (or sent in their ballots) to check off "Gore" in 2000.
1.2.2008 11:34am
dew:
Agreed. In my spare time I'm kind of a baseball stats geek, and I see the same thing going on with the various competing measures of defensive performance. What this amounts to is a concession that we lack confidence in the methodology/accuracy of any individual defensive metric.
I don't see how the two are related. In sports, many fans simply seem to have a primal need to come up with a single composite number to measure how "good" a player is, whether or not that one number is actually very meaningful.

These voting polls, on the other hand are a statistical estimates of what might happen, so yes, there is a lack of confidence in the accuracy of any one poll. The reported "margin of error" is usually a sampling error at 95% confidence, or simply that on average, 1 out of 20 polls will have a sampling error even greater than the reported margin of error. But even worse, they sort of lie when they call the sampling error the "margin of error". It presumes that they have eliminated all non-sampling error; that is, they have a purely random, uniform distribution of voters, which seems like a very unrealistic assumption. For example, one of the big unknowns is whether only calling land-line phones skews the results away from younger voters who might only have a cell phone (pollsters do claim it does not have much effect), and there is the continual problem that minor differences in the poll wording, such as the order the candidates are listed, will skew the results of a poll. And each pollster has to estimate what is an "accurate" random selection of potential voters: how many males/females, seniors/students, workers/unemployed, etc.

Averaging multiple polls together smooths out outliers from all these potential error sources, and thus appears to produce results with better accuracy than any one poll. Unless, of course, you already have some special knowledge of which are the "good" polls and which are the less accurate ones; if you can do that then averaging multiple polls is indeed silly. Me, I take them all with a large grain of salt.
1.2.2008 11:48am
Buckland (mail):

BTW, Sean O'Hara made a very good point in the post immediately before my previous one--aggregating polls with different methodologies, questions, and sampling error rates is stupid.


I would tend to disagree with that statement. This is an area of statistics that has become much more interesting in the last few years.

The research equivalent of composite polls is the meta analysis. A meta analysis looks at the results of many studies on a subject (effectivity of a drug, range reduction of an an endangered species, etc.) to see what different studies using different methodologies have found. The results can be more illuminating than a single study of the subject.

I think meta analysis have come into vogue more in the last few years for 2 reasons -- it's less expensive than doing a primary study and the internet gives good access to old data and results from previous studies. Several papers have been written in the last few years on the problems and opportunities in combining such studies (google meta analysis from some work on the subject).

The big problem in all meta analyses, including a poll of polls, is getting a good estimate of the variance (expressed in this case as the margin of error). However it's usually pretty easy to get a reasonable estimate (google 'inverse variance' or 'Peto method' to see some of the ideas in play).

Overall I would say that combining polls with different methodologies is better than trusting a single randomly chosen poll. It tends to rein in outliers and give a feel for how methodology can impact the spread. However it probably won't be better than the best poll of the group. That's where knowledge of a polling org's track record can help. If you know a poll is good than trust it. Else a poll of polls will probably be as good as it gets.
1.2.2008 11:52am
dearieme:
If they want their candidates thoroughly tested, Democrats in the early states should vote - I suggest - for Baghdad Osama. That would keep him in the race. Hellary will stay in the race whatever happens, and Prettyboy will surely stay in at least until chunks of the South have spoken.
1.2.2008 11:53am
John M. Perkins (mail):
From my limited experience of being an Iowa Democratic precinct caucus chair in 1988, I predict that the less than 15% will fairly equally spread out among those receiving less than 15%, with a bonus to Kucinich.
Why?
Because fringe supporters at the precinct caucuses are most likely those who most want to have impact at the county caucuses. [While supporters of the big names often don't look past precint caucus night.] Thus, most of the under 15% will band together under the name of that precinct's plurality. No national delegates are selected till the district caucuses many months away. So there's lots of time for alliances to adjust. Those Iowans seeking advancement to county and then district/state want more time to adjust, and a fringe coalition is the best avenue. A fringe supporter isn't going to county if that supporter joins a big group.

OTOH, Kucinich has instructed his supporters to go to Obama. But Kucinich supporters will try to reorganize as Kucinich first. Add that Edwards and Clinton precinct leaders can count also, then a few of their of their unnessary* suporters may join Kucinich to take the reported count away from Obama. (Richardson/Dodd groups may also gain with the in-precinct threat of "slide some of your supporters to us or we will join your opponents." Heck, deals like this are some of the fun group dynamics that attract some to the caucuses. And it only takes one with the easy math skills and a loud mouth.

*Unnessary means, if Clinton has more than enough supporters to get 4 county delegates, but not 5 delegates, then those extra supporters may slide to Kucinich to give Kucinich 1 delegate, taking one from Obama.
1.2.2008 11:55am
Qwinn:
"I am then noting that 2000 may suggest that regular guy/feel-goodism is not always the recipe for success, as the glum and robotic Al Gore actually outpolled W nationwide."

Well, if you're going to note -that-, then you should also note that in 2000, Bush got way more of the popular vote (47.9%) than Clinton did in his first election (43.3%) and only slightly less than his second election (49%).

Qwinn
1.2.2008 12:02pm
AK (mail):
warm peaceful guy:

If you're really a baseball stats geek, you might remember that a couple of years back the SF Giants won the NL West despite scoring fewer runs than their opponents. They'd get blown out 12-1 in a game, then win the next two or three 2-1 or 1-0. I don't know how often something like that has happened, but it's great parallel to losing the popular vote and winning the election.
1.2.2008 12:22pm
Duh (mail):
Why has compositing polls become so popular? Taking polls with different methodologies and margins of error, and averaging them together doesn't give you any meaningful results.


You will continue to be confused so long as you imagine news programming and MSM news reporting is about the news. It isn't. It's about finding any drivel at all to fill up all that dead air between the ads.
1.2.2008 12:23pm
Cold Warrior:
Having echoed the prediction of Huck and Edwards winning in Iowa, I think I'm changing my mind.

Huckabee Rep '08 = Dean Dem '04: late-breaking frontrunner who craters when caucus goers decide that they need a candidate who can win in November.

Thompson Rep '08 = General Wesley Clark Dem '04: on paper, a heckuva candidate. Late entrant with poor organization, and additional lack of any kind of emotional spark will finish him off early.

McCain Rep '08 = Kerry Dem '04: The "electable" candidate that desperate out-of-power party faithful turn to; written off for dead as late as September, he blasts into the lead after NH on pure "electability" compromise grounds.

Romney Rep '08 = Lieberman Dem '04: Party core just not sold on him.

Giuliani Rep '08 = ? Dem '04: I can't find an analogy here. He's still the wildcard, but I see his potential voters moving (particularly in NH) toward McCain as safer/less weird.

Thus, my history repeats analysis leads to a prediction: McCain will emerge as the clear frontrunner by this time next week.
1.2.2008 12:23pm
Soccer Dad (mail) (www):
Say what you want about the composite polls. I have no idea how it works. However look at the results of RCP's final tally based on composite polls in 2004. Notice something?

Every state was called correctly (in terms of electoral winner) except for Wisconsin. My best guess is that there are many competing biases built into different polls and that by compositing them, you correct for those biases.
1.2.2008 12:30pm
Civic Lessons Needed:
I love how Liberals scream about the popular vote mattering when referencing the 2000 election but go batshit crazy when you point out that Slick Willy never won the popular vote in either 1992 or 1996. I tend to think that this is the result of substandard schooling in the public education system; I doubt the teachers know what the electoral college is, how do we expect the taught to know?!
1.2.2008 12:35pm
Cold Warrior:
Qwinn, my point is that the "regular guy" or "feel good" candidate has broad appeal, whether Republican or Democrat. The last clear "glum" winner I can think of: Richard Nixon over Hubert Humphrey/George McGovern.

Think about it:

1976: Smiling Jimmy Carter over monotone Gerald Ford
1980: Smiling Ronald Reagan over (by then) "America is under a great malaise" Jimmy Carter
1984: Smiling Ronald Reagan's Morning in America trounces monotone Mondale preaching "sacrifice"
1988: Kind of a tie. Bush 41 wasn't exactly a glad-hander, but Dukakis was ridiculed as a robotic technocrat
1992: Clinton feels your pain and understands middle class Americans as Bush fiddles while America goes into a minor economic downturn
1996: Clinton the triumphant conciliator glad-hands his way to a rout over droll Dole
2000: Compassionate Conservative folksy-talking W plays to a statistical tie in against robotic Algore (but emerges victorious in the end ... and I don't care to get into that long story again).
2004: Vanquishing hero W over bitter war hero (or not?) and serial questioner of authority Kerry

So don't underestimate Huck's national appeal. I understand he's the lead guest on Leno's return to live TV tonight. Watch him play bass with the band (Clinton '92) and engage in effortless repartee with the clearly Dem-leaning host, and watch him pull it off perfectly. We are talking world-class (or at least U.S.-class) politician here.
1.2.2008 12:37pm
Cold Warrior:
Civics Lesson Needed:

Get over it!

Clinton got more votes than Bush in 1992. He got more votes than Dole in 1996. Gore got more votes than W in 2000. That's my point. Nothing more, nothing less.

Are you some kind of web troll who finds pounces on perceived liberal bias whenever/wherever someone mentions a plain fact?

The fact is not a liberal or conservative fact. It is a fact.
1.2.2008 12:44pm
Bender (mail):
Buckland beat me to it by mentioning meta analysis. Assuming that those who are combining polls did get all the data they needed to do a reasonable meta-analysis, the combined results will be more accurate.

For polls to be useful they need to produce precise (reasonably small standard errors) and accurate (no bias) estimates. Appropriate sampling designs and a large enough sample size can ensure the former. It is my suspicion that sampling in this country has become unintentionally more and more biased (and towards ending up with a more "liberal" than the general population sample) over the past few decades.

I am using "bias" here in its purely technical/statistical sense. The generally "liberal" orientation of pollsters and their interviewers leads to unconscious and unintentional pressure to give "liberal" answers. Also, "conservatives" have become more and more suspicious of "mainstream" organizations like polling companies and more and more likely to opt out of responding to polls.

Based on my suspicions, I generally give "conservative" candidates one or more more points and "liberal" candidates one or more less points than most polls suggest. (For Gallop and other more biased polls, I add a couple of points in the appropriate direction.)

Based on the above arguments, I predict that Romney will win in Iowa with Clinton/Obama a dead heat.
1.2.2008 12:48pm
dew:
I don't see how the two are related. In sports,

I will actually disagree with myself -- the two do seem closely related, just not in the same way "Cold Warrior" claims. If taking aggregating multiple polls from different pollsters is bad because they are not exactly the same polls, then it also seems pretty silly to take a player's stats from one game against one team and aggregate it with the results of his performance against different team, and then (horrors) comparing it to the same alleged statistic for a player on another team who didn't even play against the same two teams as the first. Seriously, aggregating multiple polls together is not necessarily bad; you just have to understand what you are aggregating, and what you are gaining and losing by doing that aggregating -- just like with sports stats.

I have been suspicious of single poll results for a while; I think it was in 1984 (maybe 88) that I noticed two poll results that came out a day apart, one with a "2% margin of error" and a second with a "3% margin of error", yet they were 6% apart. If you pay attention, that seems to happen at least once each presidential election cycle.

Every state was called correctly (in terms of electoral winner) except for Wisconsin. My best guess is that there are many competing biases built into different polls and that by compositing them, you correct for those biases.

Bingo. Except that you are probably not correcting the biases, just smoothing them out a bit.
1.2.2008 12:52pm
Cornellian (mail):
Edwards FTW.

I'll bet that Obama knows what "FTW" means, and that Hilary doesn't.
1.2.2008 12:56pm
Civic Lessons Needed:
And see Cold Warrior go batshit crazy at my post! Gosh, so predictable...lmao
1.2.2008 1:04pm
Wugong:
I love how Liberals scream about the popular vote mattering when referencing the 2000 election but go batshit crazy when you point out that Slick Willy never won the popular vote in either 1992 or 1996.

Um, the obvious response is that Clinton did win the popular vote. That is, he received more votes overall than did any of his opponents, just as Gore did in 2000. That's different from winning the majority, of course, but you didn't make that point. Perhaps you could do with a civics lesson yourself.
1.2.2008 1:07pm
Cold Warrior:
Civics Lesson Needed:

Yes, you do need one. A civics lesson, that is. As I just informed you, I am not a liberal. I am a libertarian. And I voted for W in 2000, not because I agreed with everything he said but because he appeared, overall, to be somewhat more in line with my political philosophy than Gore.
1.2.2008 1:11pm
Cold Warrior:
dew:

The point is not that a meta-analysis is no better than the individual studies.

The point is that we don't know when it comes to polling. We'd like to believe that sampling error cancels out, but maybe it doesn't.

You are correct that in my baseball example I'm not talking about sample error; in fact, the defensive metrics I'm talking about take into account every single batted ball. It's very common for baseball analysts to "average together" the findings of several different metrics, and to simply assume that the average is better because the errors tend to cancel out. But we don't really know that, do we?

And there's another problem in polling. In the baseball stat example, we're concerned with what really happened in the past. How good a fielder was Jeter compared to other shortstops in 2007? The polling phenonmenon more closely resembles projection systems in baseball. We'd like to think that the "composite" approach is better because it smooths out spread from the mean and eliminates the more extreme projections. And that's probably true, but we don't know that it's true. We also don't know the extent to which a hypothetically perfect poll/projection (one that records the preferences of the voting pool with 100% accuracy) will correspond with the actual behavioral dynamics of the Iowa caucus system. The way to figure that out would be to look at "composite" approaches in the last few IA caucus cycles and see if they predicted the results any better than individual polls did, or any better than some other rule ("the candidate who spent the most money in Iowa will win") did.
1.2.2008 1:20pm
Ralph Phelan (mail):
Overall I would say that combining polls with different methodologies is better than trusting a single randomly chosen poll.
It's good for reining in outliers due to poor sample choice, question variation, etc. But there are problems I don't see how to solve:

(1) What if there's a correlation between response rate and response? For example, I've heard it claimed that conservatives are far more likely to hang up pollsters than liberals (forgot the source - sorry). How do you even check if that's true, let alone measure it and adjust for it? Do Kucinich supporters respond at the same rate as Clinton supporters?

(2) "Preference falsification." Election after election, liberal New Yorkers were unwilling to admit to a pollster that they really were going to vote for that man the NYT said such nasty things about, Rudy Giuliani. But in the privacy of the voting booth, they did. Does Iowahave any social taboos that will tend to cause a given candidate to be over or under represented?

(3) As is pointed out above, this race involves a fairly small number of political junkies who are familiar with arithmetic and enjoy political gamesmanship. So how much strategic lying to pollsters is going on?
1.2.2008 1:38pm
Baseballhead (mail):
If you're really a baseball stats geek, you might remember that a couple of years back the SF Giants won the NL West despite scoring fewer runs than their opponents.
That was back in 1997. The Diamondbacks did the same thing this season, but a team outperforming their Pythagorean by such a margin as to win a division is pretty rare. I'm not about to find out just how rare, though. Who has that kind of time! *grins*
1.2.2008 2:17pm
dew:
Cold Warrior: The point is that we don't know when it comes to polling. We'd like to believe that sampling error cancels out, but maybe it doesn't.

Well, yes. But the point I was making (perhaps badly) was simply that aggregating polls is actually useful - and I apologize, I may have been conflating you in my responses with other posters who were more flippantly dismissing composite poll results. I would never say that aggregating multiple polls is certain to give better results, or even that anything would cancel anything out. I would say that aggregating more data points will smooth out many bad outliers and will be statistically more likely to produce a result with more confidence than any individual polls. As I said, if we could separate the more accurate polls from the less accurate polls, we would do that rather than aggregating polls. But since it seems we cannot, aggregating polls is a useful tool.

And there's another problem in polling. In the baseball stat example, we're concerned with what really happened in the past.

I deleted a paragraph on exactly this in my first post because I thought I was getting too wordy. This is a big problem will election polls that keeps pollsters up at night; they are taking imperfect snapshots of the current state of things to project to the future, unlike sports, where we can know the actual current and past stats to do our projections (that are still often bad projections). Another difference is that aggregating stats over several months for a team or player tells us something useful, while for election polls, aggregating any polls more than a week or so apart probably is pretty silly.

... (comments on Iowa caucuses) ...
I would rather look at primaries; predicting the Iowa caucus results seem to be better suited for reading sheep entrails, in part because of the 15% rule. Seriously, I think the Mystery Pollster (now aggregated into pollster.com) had some pretty good discussions about how accurate (or inaccurate) certain polling organizations are; maybe choosing the better pollsters would be better than using aggregate results. I have not had time to read pollster.com in a long while, they may have some good discussions on this topic.
1.2.2008 2:27pm
Cold Warrior:
I can't find a good link providing the poll results in the run-up to the 2004 Iowa Caucus (Dem side). I'd be very interested to see whether the "composite" was any better at predicting the results than the individual polls were.

Here's why the composite isn't necessarily better.

Poll A: oversamples Christian conservatives because it oversamples rural voters, who are more likely to be Christian conservatives (perhaps because they are more polite and more likely to be at home in the evening when called)

Poll B: overestimates the Christian conservative vote because it strongly weights the answer to the question, "Did you participate in the last Iowa Caucus" in determining who is a "Likely Voter" (and rural/small town populations tend to be less transient than urban ones)

Poll C: overestimates the Christian conservative vote based on some combination of A and B.

Poll D: draws a truly representative sample, and more accurately predicts who is a likely caucus participant.

Under these circumstances, Poll D standing alone is far better than the average of Polls A, B, C, and D.
1.2.2008 2:43pm
Cold Warrior:
Here's a link (look at the comments, too) about the perceived accuracy as well as the real-world success in 2004 of various IA Caucus polls:

pollster.com/blogs/poll_of_posters_rating_the_i_.php

One comment:



Everybody hates Zogby but his final poll in the 2004 Iowa caucuses was nearly identical to the Des Moines Register/Seltzer poll, which everyone loves.

Zogby, 1/16-18/2004
n = 502 LVs

Dean 22%
Edwards 21%
Gephardt 18%
Kerry 25%


Des Moines Register/Seltzer
1/14-1/16/2004, n = 606 LVs

Dean 20%
Edwards 23%
Gephardt 18%
Kerry 26%

In addition to those two, Democracy Corps, Pew, Research 2000, the LA Times, the Quad-City Times (not sure if those were Research 2000 polls too or not) and Survey USA all polled the caucus race during the 2004 season and they all generally agreed with one another fairly well except for Survey USA, which was completly out to lunch. The Edwards and Gephardt campaigns also made polls public that were well off the mean in some respect. But there was really nothing particularly wrong with Zogby last time.


At least with respect to the 2004 Dem IA Caucuses, aggregating the various polls certainly wouldn't have acheived greater accuracy. That may, of course, have a lot to do with the special nature of the IA Caucuses, since I agree with earlier commenters that the state-by-state general election "composites" (of reliable polls) were pretty darn good.
1.2.2008 3:09pm
dew:
Under these circumstances, Poll D standing alone is far better than the average of Polls A, B, C, and D.
Yes, of course. But as I have suggested a couple of times, unless you can pick the "good" poll" from the "bad" polls, what is the usefulness of this fact? You are likely to dismiss D as an outlier (which would be worse than the composite poll result), or some talking head on TV will pronounce that a candidate is trending up or down or something like that based on the multiple polls, all of which would also be wrong. If multiple polling organizations have heavily skewed poll results on the same non-sampling bias (that is, they are claiming 2% margin of error when they are really off by something like 10% because of the christian conservative bias), then the polling is pretty much just useless news &talk show filler.
1.2.2008 3:14pm
Buckland (mail):
Cold Warrior sez....


Under these circumstances, Poll D standing alone is far better than the average of Polls A, B, C, and D.


Absolutely right. That's where knowledge of methodology comes in, and especially as it relates to how the poll will be used. That's why I said earlier "However it probably won't be better than the best poll of the group. That's where knowledge of a polling org's track record can help. If you know a poll is good than trust it."

A meta analysis is useful when you don't know what is good and what is not. In your example if you know that 3/4 polls has some bad methodology by all means eliminate them. Unfortunately that's not usually obvious. In that case a meta analysis is likely to give better results than a random choice of 1 poll.
1.2.2008 3:31pm
Cold Warrior:
The IA Caucus on the Dem side is even weirder than I thought. I'm inclined to say that polling is pretty useless, or at least that IA Dem polling has the ability to make only the coarsest distinctions (e.g., Edwards/Clinton/Obama are Tier 1 candidates, Richardson is Tier 2, everyone else is Tier 3).

From Jeff Greenfield's recent article on slate.com:


But the drama doesn't answer one basic question: Why can't the public learn how many of the participants actually voted for the different candidates? The Democratic Party has these voter-preference numbers—but will not release them to the public, as this New York Times op-ed pointed out. Instead, the percentages reported on election night reflect only the share of state delegates each candidate has won. Why? No one seems exactly sure. But it means that a candidate who turned out more total supporters than anyone else, across the state, could wind up in second or third place—and no one will know.
1.2.2008 4:43pm
John Neff:
I am baffled by the statement that the Iowa Democratic party leaders have the voter preference numbers but will not release them. There is no ballot the only information the precinct chair has is the initial number of qualified registered attenders and the number of viable groups. All they have time to verify is if the viable group had enough members to make the cut. It is not practical to write down and verify the initial distribution. Some people will leave before the final count (I know that sounds strange but if you have more than enough to hold your delegates it works).

If your group is not viable you can leave, become uncommitted, join another group or become a spectator. If my candidate is not viable I am not inclined to join another group but I am providing transportation for others so I cannot leave. In that case my best option is uncommitted if they are viable.
1.2.2008 6:23pm
Syd Henderson (mail):
Civic Lessons Needed:
I love how Liberals scream about the popular vote mattering when referencing the 2000 election but go batshit crazy when you point out that Slick Willy never won the popular vote in either 1992 or 1996.
Clinton did win the popular vote in 1992 and 1996. He didn't get a majority, but that's not winning the popular vote did.
1.2.2008 7:51pm
Ralph Phelan (mail):
I love how Liberals scream about the popular vote mattering when referencing the 2000 election

I love how both sides are so focused on refighting that old war that neither noticed that Cold Warrior was equating Bush and Clinton! He considers both of them examples of "feel-good" candidates, and a vote for either to be a potential vote for Huckabee. So which of them did better is irrelevant to the fact that they both did pretty good.

Just to mess up everybody's analyses, including Cold Warrior's: I voted for both Clinton and Bush, and was reasonably satisfied with the policy results in both cases, and I loathe Huckabee and would vote for almost anyone the Democrats ran against him.
1.3.2008 8:01am