Pollster.com

Articles and Analysis

 

Lunchtime Status Update for 9/23

Topics: Status Update

Seventeen new polls logged since yesterday's update help shift five states either away from John McCain or toward Barack Obama:

  • Michigan and Wisconsin (and their combined 27 electoral votes) shift from toss-up back to lean Obama

Thus, our current classification of the Electoral College vote stands at 229 for Obama, 166 for McCain with 143 electoral votes classified as toss-up.

Two minor shifts that do not affect the electoral totals: Oregon moves from lean to strong Obama and Arkansas shifts from lean to strong McCain.

[More to follow shortly]

 

Comments
OGLiberal:

PPP just released a new CO poll showing Obama leading in that state, 51-44. He was only up 1 a week ago. Big shift in indies and Palins faves are down and her unfaves are way up. Will this +7 Obama result in CO turn that state light blue?

____________________

serpounce:

The way things are going I really think that Mr. Toss-up has an outside chance of pulling off a victory. He's made some serious gains in the past few weeks.

____________________

javakah:

I'm not sure where best to post this, but this was a rather interesting article that in many ways is poll related:

http://www.msnbc.msn.com/id/26833397/

It's looking at the donation numbers in some of the counties around the country that were very close/battleground counties.

I'm rather curious as to how likely the number of donors (and the amounts donated) in such counties might relate to actual votes. Unfortunately while it does provide Excel sheets for the current numbers for this election, it doesn't provide such numbers for the 2004 election.

____________________

bmcavan:

Nat. Polls are starting to tighten again. The state polls are always quick to follow. Let's not get excited yet.

____________________

s.b.:

Do either of the Marks know why Gallup is not using likely voter screens this time around as they have in every other presidential election they have polled after the conventions. Likely voter polling started in August last time and we are almost in Oct with no likely voter numbers fromm Gallup. Does anyone know why Gallup has made this siginificant change?

____________________

jme:

I'm going to try to renew our discussion from the 9-19 lunchtime update...let's see if anyone joins me...

I still think that this state coloring exercise is simply a can of worms. Here are some things I thought about over the weekend:

1.) Mark and Charles intentionally chose their coloring system to be "conservative", but it is important to note that there are different _ways_ for a statistical procedure to be conservative. Their working definition of conservative means that they will tend to err on the side of calling states toss-ups. That's fine, but one could also say that it would be conservative to resist the _switching_ of the colors of states. In this sense, a conservative procedure would tend to avoid _switching_ the classification of a state form its present color. Which is better? Who knows? My gut tells me that erring towards coloring states yellow might increase the variability of (some) states' switching colors repeatedly, whereas erring towards keeping the color the same might make the color overly dependent on the current state (and hence on older polling). Altogether now: bias-variance tradeoff!

It will be interesting to go back after the election and look to see if the procedure they are using now _needlessly_ switched the classification of a lot of states, compared to the sign of (Obama - McCain) in the trend line. In other words, take, say, MI. If the margin there is straddling the cutoff point between yellow and light blue, I could envision it switching back and forth 1-3 more times before the election, even though Obama's trend line remains at nearly the same margin above McCain's for the whole time, and then Obama eventually wins MI. It will be interesting to see how much of this (if any) takes place.

2.) I'm still bothered by a procedure that calls MT and WV toss-ups based on such a small number of polls. I sort of wish there were a 6th color for abstention due to lack of polling.

3.) Is anyone at pollster willing to comment (speculate) about the variability of the variability of polling across states? (Did that make sense?) I mean, compare the polling for NC and, say, PA. Just eyeballing it, it seems like the variation in PA has settled down over the last 2 months, whereas the variability in NC has exploded over the last 2 months. But better plots looking at this specifically would make it clearer...

____________________

s.b.:

Of the four tracking polls released today on RCP Obama is up by an average of 1.2% Two of these polls have no likely voter screen. You know what that means? Obama is not winning as of today. The likely voter polls were closer last time than the RV's. In The LV polls McCain is ahead by 1% today. Things change. Its a very close election. No one is running away with this thing as of yet.

____________________

Allen:

@jme

"Coloring" states as safe/lean/tossup is somewhat arbitrary. Other sites such as election-projection.net compute a probability for each state and then run simulated elections. You might prefer that approach. It also has an Interactive Presidential Election Probability Calculator if you want to run your own numbers.

____________________

judas_priest:

Regarding s.b.'s comment, given the size of the samples, the true average of the two LV polls is .7%. (RR has 3000, Battleground has 800, which means that Rasmussen should have roughly twice the weighting of Battleground. Before anybody jumps on my case, please remember this is a quick eyeball assessment.) In any event, this is well within sampling error.

Also, LV screens are only as good as the underlying model. Since there is considerable evidence that we can expect higher turnout from certain demographic groups that has been typically, these models may themselves be skewed.

____________________

s.b.:

I just went back to RCP from NOv 2, 2004

There are only three organizations that run RV polls in Sept, the last one being Newsweek from 9/30-10/02

The other two are CBS news with two polls, the last on Sept 16th and Pew with two polls the last on Sept 26.

So at this point in the race in 2004 only two more polls came out with RV's One Newsweek and one PEW.

EVERY OTHER POLL is using LV's by this time in 2004. So what's going on? We see that LV's favour McCain and in general Republicans. Have the Obama capmpaign put the screws to pollsters and threatened them like they did Jewish groups in Ny????

What reason would their be for so many RV polls at this point itn the race?

By the way the RCP average was still off by .9% in favour of the democratic although it had Bush winning by 1.5%, he won by 2.4%.

ALL the polls in the final RCP average were LV screened. None were RV's.

People need to be very cautious if pollsters don't start bringing in LV screening especially with tracking polls. They lose legitimacy if they don't at this point in the race.

____________________

s.b.:

Judas, no one counts polls with twice the participants at double those that don't. it just isn't done that way.

____________________

judas_priest:

s.b.:
A number of polls may be waiting until after the debates to shift to the LV screen. I know that the site fivethirtyeight.com had announced quite a time ago that it would use RV poll results rather than LV earlier and then shift after the first debate - that is, from polls that offer both. Also, perhaps some pollsters are aware of the possible changes in turn-out patterns and being cautious.

BTW, a .9% variation is probably within the margin of error even for an average of a number of polls. Also, the Bush results are shaped by the fact that the GOP had a considerably better GOTV operation thandid the Dems in 2004. There are some indications that the shoe may be on the left foot this year rather than the right one.

Again, there are far more polls out there than the daily trackers, and chosing such a sub-set of polls sounds suspiciously like cherry-picking data.

Conclusion: you are trying to extract far more meaning from your selected data than is really there. As one of my graduate school professors used to say, you are "grabbing the data by the neck, shaking it and saying 'speak to me.'"

____________________

judas_priest:

It may not be done that way, but it is in fact mathematically correct. BTW, fivethirtyeight.com, whose math is probably the most sophisticated out there, weights polls according to sample size. Since MOE is the inverse of the square root of the sample size, the fact that Rasmussen has almost four times the sample size suggests strongly that its weighting should be just under two times that of the Battleground poll.

____________________

s.b.:

.9% is very close true Judas. Please know I have had the same thoughts about sample size. However Rasmsussen and Gallup would skew any agregate model of polls if sample size were a factor. Also different weightings, and screens are used so just simply saying well this one has twice as many people facotored into the poll (different than #sampled) so it should count twice as much doesn't work.

To wait until after the first debate to use an LV screen is highly unusual especially since it will be only five and a half weeks until the election before those polls come out.

As i have stated, Gallup said they would switch over after the conventions but hasn't eventhough thte conventions were much later this year.

It is unheard of to have so many RV polls out there this late in the race and I think people need ot use caution drwing any conclusions from them.

For example if obama is ahead by 2% in the averages on the day of the election but half the polls are RV and he loses by 3%, will there be civil unrest and howls of racism, aka riots, when really its just a function of pollsters not screening repondants and a normal general shift to the Republican ticket on game day from the polls. This may be especially true if Party id weightings aren't current and are skewed to Democrat in polls or if undecideds are reassigned.

I have serious concerns about what is being stirred up in the MSM with regards to race. I think this in and of itself produces a silent McCain vote. It's not racism, It's the accusation of racism thrown around so often in the MSM. The sooner pollsters start using a screen the better for the entire nation.

____________________

judas_priest:

Once more, dear friends, into the breach:

The RCP collection of tracking polls deliberately omits the Kos/Research2000 poll, also LV, which in today's release (for the three days immediately prior - the same as the others - shows a five point Obama lead. Its sample size is 1100, larger than the battleground. Research2000 has an excellent track record. I do not know about the pollster for battleground.

Plugging that one in, and weighting by the sqare root of the sample size, we get Obama with a 1% lead.

____________________

s.b.:

Judging by the baseless accusations, vitriole and hatred that comes out of an awful lot of Obama supporters mouth's any time he is questioned, Nov 5th could be a very bad day for America if pollsters don't try to be more accurate, have an LV screen, weight party id appropriately and not skew their results to Democrat.

If RV's and dubious party id weighted polls are counted, there is a distinct possiblity that Obama could be ahead by 2% on lets say RCP or other aggregates and lose by 2%. All hell will break loose. I am getting very concerned.

____________________

judas_priest:

Not using the Sq rt of the ample size and using the r2000 poll also yields a 1% margin for Obama.

But again, using only the tracking polls is ignoring a good deal of data that just happens to show better results for one of the candidates.

To continue on the sample size weighting topic, imagine three polls, one of which has a sample size of 2500 and the other two, 250 each. Are you stating that each should be weighted equally?

____________________

s.b.:

The Daily Kos poll has a weighting for Party id that is improbable to say the least. It shouldn't be counted by anyone. It's propaganda and yes they set the parameters if they are paying Research 2000. It's utter nonsense, much like the CBS NYT poll with an 11% spread for Party id and +20%Dem generic congressional vote. Again NYT paying that bill. Nonsense.

____________________

s.b.:

I was just using the polls released today sans Daily Kos. I think ARG has a 2% Obama lead today. They also had Kerry winning in 2004. needless to say, the race is narrowing again, if it was ever not narrow.

And we need to start seeing LV screens on all polls or not count them.

____________________

s.b.:

Actually they could be counted in their own averages or aggregates but not with LV polls. Some sites actually had two diffferent averages those for RV and those for LV. They should be separated and only LV paid attention to very shortly.

____________________

judas_priest:

If you wanted to adjust the RV samples for supposed bias toward Obama, adjust his figures downward, using the figures for the average difference in those polls that publish both. But ignoring the other data smacks of selecting polls by their results.

BTW, each polling firm uses a different LV screen and we do not know that the screens they use match the reality on the ground for this election. My own belief is that RR, for example, will underestimate the Obama vote by 1-1.5% (assuming that Rasmussen hasn't changed his LV model from last time.) (You may have realized that I was pprofessionally trained in this area, although I no longer do it on a regular basis.)

Also, from what I have learned about Gallup's past LV screens, they are seriously flawed. Again I don't know what they will be using this year, since all reputable pollsters try to improve their models based on additional information.

____________________

s.b.:

Judas, the democrats always say their GOTV efforts will be better and that young voters will make the difference and that they have registered so many new voters.

It's an old song. They thought McGovern would win because of new registrations and the youth vote too.

McCain now has the same money as Obama and there is no reason to believe that voter patterns will shift dramatically in this election.

Blacks voted for Gore by 95%, Kerry by 93% and I believe all non whites voted for Dukakis 90%, so a similar number for Blacks, which weren't counted sperately that year. My point being the voting paterns will be similar to previous years.

In fact the youth vote is declining as a function of demograohics aka the baby boom and baby boom echo.

I don't believe the hype and that fact that pollsters seem to is worrisome.

Again people and the MSM are already spinning racism as the reason if Obama loses, when their is no emperical evidence to suggest that a far Liberal senator from the North east with another senator from the North east as a running mate should ever win as a Democratic president. It's never happened.

Kennedy only won because Johnson carried Texas, Carter won Georgia, Clinton Arkansas. Southern governors win for the Dems period.

But its already being spun as racism. It should be a great concern to pollsters. AS i have said, I believe there is a silent McCain vote, as there always is for Republicans, and I believe things could get ugly Nov 5th in America if people aren't very cautious about drawing erroneous conclusions from skewed polls.

____________________

Allen:

I personally think weighting polls by the number of respondents is reasonable. If the only source of error were sampling error, weighting by sample size would be mathematically correct. While there of course other sources of error, it is reasonable to assume that the polls with larger sample sizes have been more carefully constructed and therefore would have smaller errors. This is especially true for the state polls. For example, the one-day polls have smaller sample sizes but probably higher errors from non-response (no one answered the phone) bias. election-projection.net also weights its state polls by # of respondents. As a practical matter, this seems like a good approach to me.

____________________

s.b.:

Yes Judas all pollsters ahve a different LV screen which is why their results can't be weighted based on final sample size, which is also often weighted.

The data should be left alone as much as possible, but screening non voters is a necessity.

____________________

s.b.:

Judas weighting polls on aggregates by sample size is a weighting on top of a weighting on top of a screen, on top of any possibly sampling error. It's like a rounding error, on top of a rounding error.

You end up with the wrong answer unless you only round the final number.

It's too much messing with the data. it becomes hocus pocus, more art than science.

Polling is a social science often criticized by hard scientists for way too much subjective input. The less the better. That's how I feel about weightings.

____________________

s.b.:

By the way, I distincly remember the Princeton meta analysis being wrong in 2004 and calling it for Kerry.

____________________

Allen:

@s.b.

You misunderstand the point of weighting. Its goal is to combine multiple polls to come up with the best aggregate estimate. Certainly if you don't understand the mathematics, then it all looks like hocus pocus, but I guarantee that some sort of weighting will give you a better estimate than no weighting. The art/science comes in figuring out what weighting is optimal. Like many issues that come up in statistical polling, that question cannot be answered precisely, but that does not mean one answer is just as good as another, and it certainly doesn't mean that no weighting is better than a carefully thought out system of weighting.

____________________

judas_priest:

s.b.:

The battleground poll you think is so superior has a fixed weighting of D+3. In other words, it is eactly the same as Research2000 but with different, more or less equally inaccurate weighting. The solution is not to toss data based on one's subjective estimate of which ones are "acceptable" but to include all which meet certain pre-selected criteria.

I was only talking about measuring the sample MOE using weights based on sample size. The other errors are extra and, in general, not measurable up front. But see fivethirtyeight.com for his method of weighting pollsters by prior accuracy.

BTW, RR is the most Republican in results of the major polling forms, and Battlegrouond is out there too. Not that they are necessarily wrong, but their results are skewed in that direction compared to the "average" poll.

If you go back and look at the Princeton metaanalysis, you will find that the model was dead on, having nailed all states and DC. Itwent off only because Sam Wang decided to use his judgment and weight undecideds in a different way. That changed only one state but it changed the outcome. If he had simply used the figures from his model he would have nailed all 51 electoral entities.

He is currently showing CO and VA as blue and NV and NH as red. NC is toss-up. All the other states are as one would expect. He uses all polls in this site's database, except that he excludes Zogby interactive, which is not based on a properly drawn sample.

____________________

judas_priest:

@s.b.:

The prior years "but X will be different," memes were based on hope (like the Gooldwater "hidden conservative" vote), not data. There is actual turnout data and polling data suggesting the turnout model should be different this year. And for the first time in a long time the Dems will have a good GOTV operation, not just a hopeful wish. Not that McCain won't, but at the least the difference in the ground game will very likely not operate to the GOP's advantage as it has in the past.

____________________

Robert:

Dear S.B. and judas priest:

I note in your exchanges references to RCP. I have been unable to figure out the criteria RCP uses for inclusion in their averages? On several ocasions, RCP has been inconsistent in the dut-off dates that determine which polls they include in determining their averages (i.e., some older polls are included, some newer ones ar not). What is their policy? Do you know? I've looked on their web page and can't find anything.

____________________

s.b.:

Allen thanks for the condescension. You're an Obama supporter right. I do understand the Math. I understand that the more you wave a magic wand over raw data and call it weighting the more experimentor error is introduced into the result. I do believe that the vast majority of pollsters agree with me on not weighting polls in aggregates based on sample size. This site does, RCP does etc.

____________________

judas_priest:

@Robert:

The link below is to a discussion RCP's selection of tracking polls. (It does not apply to RCP'suse of state polls.)

http://www.fivethirtyeight.com/2008/09/rcp-and-r2k-tracking-poll_6506.html

The proprietor is a Obamanik, but one who tries hard to be objective about his use of data, if not unbiased in his preferences. He also reveals completly the details of what goes into his models. That site, along with this one and the princeton metaanalysis site (http://election.princeton.edu/faq/) are the statistically most sophisticated sites that I have seen. There is a detailed discussion of the methodology on fivethirtyeight under theheading FAQs. It helps if you have a math/stat background (I have the equivalent of an undergraduate math major, athough I haven't used the stats much in the past 25 years) - basic stats through regression analysis would be good.

____________________

s.b.:

Judas, I don't toss out any data, on the contrary. I toss out ludicrous weighting practices placed on top of the raw data. I do have a little bit of respect for the Rasmussen party id weightings as they are based on 45,000 respondants. The new 6 week weightings are based on around 25,000. The recent Dem advantage is about 3%, as the average over 6 weeks is 5.5%. However the Dem advantage did go up this week in response to the market turmoil.

But in all honesty, I would prefer they didn't weight their data at all. Their sample is quite large over three days. i don't see why the data can't stand on it's own.

And yes Judas I've heard a lot about the ground game, but honestly a lot of it seems like unco-ordinated busy work. It's mostly out of state college kids and sometimes even high school kids knocking on doors. It's even Acorn which has been brought up on charges in several states for voter registration fraud.

You know what the Republican response was in Colorado when asked about their lack of ground game? They said our ground game lasts 72 hours. The party workers do it. We'll get our voers out. They seem unconcerned.

In Virginia the Republicans feel that neighbors will be more pursuasive than out of state college kids harrassing voters. I tend to agree with them.

I honestly don't buy the hype. Remember that Axelrod has never run a national campaign. The republicans always do better at GOTV than Dems. McCAin is coordinating with state and local parties, sharing space and resources and man power with them. Obama is not. Huge advantage McCain.

Obama is relying in part in uncoordinated efforts of 527's such as unions to get out the vote. Republicans don't.

I don't see a huge advantage to Obama in on the ground organizing. I see a lot of wasted time and energy and money. But this is a matter for speculation. We will see.

I still maintain even from the primary data, not including caucuses, there is no emperical evidence to suggest a greater percentage of youth or black turnout this general election cycle.

____________________

s.b.:

Judas let me give you a perfect example of the hocus pocus I am talking about and this is for you too Allen because the Math is so tough.

CNN/usa Today Gallup in its last poll before the election in 2004 published its results at 49-49 tie. This however was not the result they actually got. The result they got, with internal weighting was 48 Bush 46 Kerry, a 2% margin for Bush and very close to the actual result.

However they decided that their internal weithing was not enough and that the would redistribut the undecided vote get this, I know the Math is hard here, 90% in favour of Kerry. That's right in a close election they decided to give the undecided vote to Kerry 90%-10%. WOW holly friggin hocus pocus!

Like I said I know the Math is tough but I'm pretty sure arbitrarily alloting the undecided vote to Kerry 90% introduced some experimentor bias into this poll.

Here's the link for you. And this poll had a real big sample size too so it would have weighted heavily in any aggregate that took that into account on top of internal weighting and redistributing undecideds.

here's the link http://www.usatoday.com/news/politicselections/nation/polls/usatodaypolls.htm

It makes me wonder if the undecideds are being similarly redistributed by Gallup this time around. They don't say.

____________________

s.b.:

Robert I too have notinced this about RCP and it ticks me off. I have responded to their survey and let them know this. Incinsistency in their removal of polls by date is the biggest error in their averages and especially their graphs over time. Is it two weeks, two days, one week? It seems arbitrary at times. In the primaries it seemed to favour Obama, less so now, but the polls right after the convention were dropped after a couple of days, after the Rep convention not, making Obama's rise and fall in the graph seem sharper than it actually was. Who knows.

____________________

s.b.:

Sorry final gallup poll without redistributing undecided 49 Bush 47 Kerry 2% spread

Actual result 50.7 Bush 48.3 Kerry 2.4% spread

Less Hocus Pocus = more accurate poll with correct election result.

____________________

judas_priest:

@s.b.

The polls we are reading today are not distributing the undecided. If four years ago a polling firm distributed them 90-10 that is an error in judgment, not an error in polling. The best single guide to where an undecided voter will end up is his/her party ID. If this is independent there are other, more time consuming methods. Absent such data it is probably best to publish numbers showing the undecided. If the customer wants to eliminate the undecideds, absent other information they should be split either 50/50 in in proportion to the state or national vote. This has the advantage of not chnaging things much. But if Party ID is available the best course is the assume a break toward that party but not 100%.

If there are 100 undecided voters in a given poll, and they are 45 Dem, 40 Rep and 15 Ind You should split the Party ID types toward their ID, say at 2:1 and split the Ind's 50-50. These are strictly rules of thumb, not based on any hard numbers. (There may be better methods now, developed since I stopped reading Public Opinion Quarterly.)

If, however, you have more extensive data (issue positions, past voting, degrees of favorable/unfavorable (the University of Michigan used to use (and may still do so) a scale of 0-100 asking respondents to rate their affect toward the candidate that way) these should be factored in. It is rare for media polls to do this, since it costs extra and does not add a great deal of accuracy.

____________________

Allen:

@s.b.

Gallup didn't "redistribute" votes. They projected that the undecided voters would go 9 to 1 for Kerry. They felt that was the most accurate estimate of what would happen based on past elections results. It was however no guarantee that would happen; it was simply their best estimate. To which I say: (a) Every poll and even more so every prediction will be wrong about something. It is inherent in sampling and prediction. Even the "right" answer is only right in part because they got lucky. (b) Who cares about national polls anyway. Their only use is for tracking a trend, and (c) if you want to track a trend, you should either average all polls or handful of carefully selected polls, with some kind of weighting.

FYI, there are lots of different kinds of weighting, that serve different purposes. What Gallup did was not weighting. Weighting is however many times REQUIRED in order to get accurate results, for example, to create a close as possible to a random sample given real-world limitations.

Finally, for what it is worth, IMO, Gallup's LV model is defective. I pay no attention to Gallup's LV polls.

____________________

judas_priest:

Robert,

From what I have heard, calling Gallup's LV model defective is like calling an NBA center tall. They may, of course, change it for this election cycle.

I'm afraid I must return to my real life. Being mostly retired gives me more freedom that I had before, but still the world impinges.

I may be back later tonight (Pacific time)

____________________

Mike_in_CA:

Just to interject here for a moment...

Before I get to my point, I will admit, the discussion on these more sophisticated blog posts pleases me as it far more substantial than the bologna that goes on over in the polling release threads.

I wanted to respond to s.b.'s criticism of R2K's LV model with Democrats outnumbering Republican by 9%, 35 to 26. I agree, this seems unrealistic, but only because it suggests that almost 40% of voters are independent or undeclared or "other" which is just absurd. Only during the period after Nixon's resignation was the number of independents this high. While it is arguable that the Bush partisanship of the last 8 years has likely turned off a great many Republicans AND Democrats (due to the sheer partisan nature of political discourse in this country), its unlikely that 40% of the public is undeclared.

That said, R2K might not be THAT far off. Gallup released numbers today showing Democrats at 49% and Republicans at 39%, a 10% gap, not all that off with the R2K 9% gap.

Now, let's play with numbers,

Say Obama gets 90% of Dems and McCain gets 90% of Republicans, and they split indies evenly 50-50 (this is purely for the purpose of idealized academics, not actual results).

Using Gallup's data, that puts the final result as: Obama 54 - McCain 46.

Now, using R2K's data, with the absurd number of "independents", and the same distribution of support described above, the result is: Obama 53.6 - McCain 46.4.

Almost the same result, despite the huge discrepancy in undeclared voters.

Now, the functional disagreement arises in whether or not you believe that Democrats maintain such an advantage. Gallup and R2K believe, YES. Ramussen and Battleground believe NO. (And Hotline is somewhere in the middle at about +6-7%).

Fundamentally, this is an exercise in opinion, really. Battleground, going off past elections, is assuming that Likely Voters are going to be Republicans. Rasmussen is finding the gap to be about 5%, which could have to do with either THEIR methodology or Gallup's.

Either way you slice it, party ID numbers in LV models are, at best, educated guesses. Let's leave it at that eh?

____________________

s.b.:

Allen it's Gallup's own wording that they redistributed the undecided vote. Read the link

And saying that "projecting" the undecided voters would vote for Kerry 90% is somehow different than "redistributing" is absurd semantics.

Projecting and redistributing and extrapolating and weighting are all forms of the same thing, also know as guessing. Pollsters taking it upon themselves to decide how many voters will be Rep or Dem or Women or young or from San Francisco, or will break for Kerry, unless the data actually tells them that is just plain bad science that wouldn't get published in any scientific journal or peer review.

It's stuff like that that make hard scientists cringe at the term Social Science.

BTW: On what parallel universe have undecideds ever and I mean EVER broken for the Democratic presidential candidate 90% to 10% in the last two days of the campaign and why would Gallup feel it was necessary or accurate to do this crystal ball gazing. It has no credibility, no matter what you call it.

I know semantics is tough for you and you might not quite grasp it Allen and I can understand if you don't but redistributing and projecting are the same thing.

____________________

s.b.:

Juda I don't know if any polls today are redistributing undecideds, Gallup is probably as their undecided # is always very low and has been the same all year; but some definatley have been this cycle. My point is just that when you start add errors on top of errors the result becomes muddied. The original point was that I don't think poll results should be further weighted in aggregates by sample size.

These polls are all apples and oranges really. Weighting them by sample size just makes it worse.

If the raw data collected by pollsters could be put into a common aggregate that would be interesting. Then we could have different weighting models applied to all the data to see how the results are effected. But pollsters don't release their raw data.

____________________

Allen:

@s.b.

The basis of Gallup's allocation of undecideds was not Democrat or Republican, it was Challenger vs. Incumbent, and it had a sound basis from results of prior election. In my opinion, the model was not accurate in 2004 because of the "war time" president issue was different than past election. Regardless, if I were making a prediction in 2004, I would also have allocated somewhat more votes to the challenger than the incumbent, but closer to 60:40. 90:10 was clearly extreme. I would also have not used a Gallup LV poll as a starting point for reasons already stated.

That notwithstanding, your criticisms of that particular extreme example of allocating undecided is not a valid basis for your rejection of all things that you call "weighting". Weighting based on demographics is a valid and necessary way to arrive at an accurate "random" sample given the real world limitation of sampling voters. I also do not agree with "a priori" party id weighting, because party id can change. I do think weighting based on a dynamically changing party id, which is what Rasmussen does, is valid and will give more accurate results by reducing the sampling error. The tradeoff is that it can mask rapid changes in opinion, i.e., it results in a lagging indicator.

Ultimately, you have to separate the prediction parts of the polling process from the measurement parts. Anything involving prediction is in part a judgment call or educated guess and is inherently less reliable than the measurement parts. This includes the prediction of LV's and the prediction of how undecided voters will break.

These again should not be confused with weighting for the purpose aggregating polls or making more accurate measurements. This type of weighting both is reliable and mathematically sound.

____________________

Gary Kilbride:

Intrade is unimpressed with North Carolina supposedly a tossup. It's still trading north of 70/30 in favor of McCain. I'd love to steal a cheap price there. My sister lives in Hendersonville and every time I visit her and talk to locals, or merely listen to the radio for 15 minutes, I can't imagine a Democrat carrying that state.

I don't know how anyone can assert Republicans have always been better than Democrats at GOTV. In '98 and '00 it was consensus that Democrats dominated on the ground. In fact, that's what led to the 72 hour strategy in Georgia with Ralph Reed in '02, so successful that it was expanded throughout the South and elsewhere in '04 and '06.

You're always going to be better off with neighbor contact and fellow church members, etc., than strangers awkwardly knocking at your door and checking your name off a list. I totally agree with that. The GOP is probably one cycle ahead in that regard. But the natural enthusiasm this cycle is with the out party with the younger nominee. Everything tends to drift back to the beginning and all year it's been well known that Democrats are more eager to vote than Republicans. The GOP can pretend they've patched it up with Palin, but that's all it will be, a flimsy patch. It's like a football team with a major weakness. You can get away with it briefly, like New England with Cassel against the Jets, but it's silly to pretend it's no longer a decisive factor that won't show up in critical stages down the road.

____________________

Allen:

@Kilbride

NC a tossup based on two polls? election-projection.net also has NC at 70/30, based on weighted poll averages. Interestingly, that gives the same number as Intrade.

____________________

judas_priest:

fivethirtyeight.com has NC 73/27 for McCain

____________________



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