Vorlon, I just read

[agcatter]

that Democratic pollster Mark Mellman doesn't think much of Gallup's likely voter numbers this far out.  If I'm not mistaken, I think this jives more or less with what you have said.  I seem to remember you saying that it's not designed to work well this long before the election - that it's only accurate at election time.

[The Vorlon]

that Democratic pollster Mark Mellman doesn't think much of Gallup's likely voter numbers this far out.  If I'm not mistaken, I think this jives more or less with what you have said.  I seem to remember you saying that it's not designed to work well this long before the election - that it's only accurate at election time.

Point #1

Mellman works for Kerry, any poll showing a Bush lead is fatally flawed.

Bush pollster Matthew Dowd however is of the opinion that all polls showing Bush ahead are methodologically perfect.

But serioulsy, Mellman works for Kerry, what did you expect him to say "Ya, Gallups right, we're getting F$%king killed in Ohio and Missouri, we're thinking of pulling out of those states..."

Point #2

Re Gallup's likely voter screen.

2 months out ANY likely voter screen, by any company, is, well, a crap shoot.

Every reputable pollster, including Gallup, will tell you that 2 months out ANY LV screen is a shot in the dark.

Gallup's model screens for BOTH past voting intentions AND current level of interest/excitement about the race.

This can, and often does, produce larger swings than actually occur in the electorate when you are 2 months out.

Here is an example:

I am doing a poll to determine if the St Louis Rams or SF 49ers have the most have the most fans, but I am limiting my survey to "Likely" football fans - those fans who are in the top 55% of all fans in terms of "excitement" about their team.

It is the 2nd week of the season and the 49ers just beat the Rams 57-0.

Needless to say, 49er fans are more excited, Ram fans are not.  If I did a poll just after the game, I bet the 49ers would do well, the Rams, not so well.

Fast Forward to just before a big 49ers/Rams playoff game.

Here one could reasopnable expect before the big game (ie just before the actual election) that a "likely" voter screen would worh much, much more accurately as both teams fans would have similar levels of excitement not artificailly pumped up or down by a good/bad regular season game.

Gallup has changed the way the have done things versus 2000.

In 2000 they were using the same 55% turnout assumption many months out for the actual election date, and Gallup showed huge week to week swings that just were not there as Gore and Bush alternately had good weeks and supported popped in and out of the "likely" voter pool.

In 2004 Gallup has extended their likely voter screen out to 75% or so which is a reasonable "possible" voter poll level this far out.

Bottom Line....

Gallup is indeed a bit more volitile this far out than other polls due to the way they do their screening.  

I would note however that the interest levels being shown by the electorate right now are very close to where they typically are in say mid October by historical standards.

Gallup is a very very good poll.

I never take any ONE poll as the sole and only truth.  When Gallup says Bush is up 8 in Ohio - that means something, but it doesn't mean the race in Ohio is over.

I think, as I have previously posted, Bush has a "real" lead of 3-4% Nationally, and right now an extra 3-4% of lead that is post convention froth that will likely fade away pretty fast.

These Gallups are likely a bit high in Ohio and Missourri.  

I think Bush is up 4% or so in Ohio, maybe 6% or so in Missouri.  Pennsylvania is indeed close and behaving pretty much as I expected, Washington baring a Chernobyl style meltdown goes to Kerry

[agcatter]

Interesting also, is that I read Gallup will not be doing a daily tracking poll this election as they did in 2000.  They are saying that instead they will be coming out with weekly state polls.

Fine with me.  That daily tracking poll last time out was swinging all over the place.  Plus 10 for Gore and then plus 8 or 9 for Bush two wks later.  Talk about a making your head spin...

[The Duke]

Vorlon,

Is your map based on Bush +8 or Bush +4?  Or based on state numbers solely and not national numbers at all.

[The Vorlon]

Interesting also, is that I read Gallup will not be doing a daily tracking poll this election as they did in 2000.  They are saying that instead they will be coming out with weekly state polls.

Fine with me.  That daily tracking poll last time out was swinging all over the place.  Plus 10 for Gore and then plus 8 or 9 for Bush two wks later.  Talk about a making your head spin...

Gallups methodolgy is not built to do a tracking poll.  Its designed to do a three day snapshot.

Their state polls were very good in 2002 doing senate and gov races.

[agcatter]

Yes, their state polls were indeed good in 02.

Zogby on the other hand....well in my state he had Kirk up over Cornyn 2 days before the election.  Haha.  Kirk only lost by 12.

[agcatter]

Correction.  He had Cornyn up, but only by one point.

[CARLHAYDEN]

Point #1

Mellman works for Kerry, any poll showing a Bush lead is fatally flawed.

Bush pollster Matthew Dowd however is of the opinion that all polls showing Bush ahead are methodologically perfect.

But serioulsy, Mellman works for Kerry, what did you expect him to say "Ya, Gallups right, we're getting F$%king killed in Ohio and Missouri, we're thinking of pulling out of those states..."

Point #2

Re Gallup's likely voter screen.

2 months out ANY likely voter screen, by any company, is, well, a crap shoot.

I think we strongly disagree on the efficacy of a voter screen for likely voters.

While (from what I have read of your postings) we agree that "interest" and "following the election closely" techniques constitute a poor screen, I believe that historical voting is a very good screen (i.e. most of those who voted in 2000 and are still alive to vote in 2004 will vote in that election, and most of those who did not vote in 2000 but are legally eligible to vote in 2004 will not vote).

The 'historical' screen is of course NOT perfect, for a number of reasons, mosty prominently, that people LIE about whether they actually voted in the previously cited election.

Adding a demographic screen to the 'historical' screen provides a vast improvement.  

Now, instead of playing football with the matter, lets look at relevant hard data.

First, the highest turnout in a Presidential election in the past 50 years (using VAP) was in 1960 of about 50%.  The lowest turnout about 49% in 1988 (about the same in 1996).  Since the voting age was dropped nationwide to 18 in 1972, the turnout has never exceeded 56% nationwide (of VAP).  So, a screen showing more than 60% of VAP turnout is is a little hard to believe.

Second, the data is very clear that turnout is highly correlated with SES.  This is well established.  The University of Michigan SRC has a large amount of data on this.  So do a number of other sources.

Third, you are correct that polling does become a little more accurate in the last month before the election.  One of the factors is that in most states, voter registration cuts off about a month before the election.  

[The Vorlon]

I think we strongly disagree on the efficacy of a voter screen for likely voters.

While (from what I have read of your postings) we agree that "interest" and "following the election closely" techniques constitute a poor screen, I believe that historical voting is a very good screen (i.e. most of those who voted in 2000 and are still alive to vote in 2004 will vote in that election, and most of those who did not vote in 2000 but are legally eligible to vote in 2004 will not vote).

The 'historical' screen is of course NOT perfect, for a number of reasons, mosty prominently, that people LIE about whether they actually voted in the previously cited election.

Adding a demographic screen to the 'historical' screen provides a vast improvement.  

Now, instead of playing football with the matter, lets look at relevant hard data.

First, the highest turnout in a Presidential election in the past 50 years (using VAP) was in 1960 of about 50%.  The lowest turnout about 49% in 1988 (about the same in 1996).  Since the voting age was dropped nationwide to 18 in 1972, the turnout has never exceeded 56% nationwide (of VAP).  So, a screen showing more than 60% of VAP turnout is is a little hard to believe.

Second, the data is very clear that turnout is highly correlated with SES.  This is well established.  The University of Michigan SRC has a large amount of data on this.  So do a number of other sources.

Third, you are correct that polling does become a little more accurate in the last month before the election.  One of the factors is that in most states, voter registration cuts off about a month before the election.  

I think you re over-stating the degree to which we disagree.

Of course you try to narrow your poll down to people who will actually vote, this falls under the catagory of "bloody %^%&ing obvious"

The problems is in the screen, and while I agree you must have a screen, exactly how you screen is far from clear.

There are basically 4 groups of voters:


1) The hard core partisans who always show up - this gets you to about 35% Turnout

2) Add in the the "Good Citizens" who are not particularly political, but usually vote beacuse they are, well, "Good Citizens" - This gets you to +/- 43% turnout or so.

3) The "intermittent voters" these folks represent about 20-22% of the population and about half of these folks vote.  Adding in the 10-11% actual turnout gets you to about a 53% or so turnout.

This is thre group that is a real bugger to screen.  Extending your turnout model to say 75% or so will get you all of these voters, but you also get a lot of voters who will not in fact actually vote.

The other option is to ask a variety of interest questiosn to sorth the half of this group who will vote from the half who will not.

2 months out, this is, in my view, a crapshoot.  

It's still better than NOT trying to do it, but I think any fair minded person has to acknowledge there are substantial uncertainties associated with it 2 months out.

These uncertainties diminish substantially the closer you get to the actual election.

4)  The final group are the un-registered and un motivated.  These folks almost never vote and you can pretty safely screen them out.

We don't know what turnout is going to be.  55% or so seems a reasonable guess, but hey if it was 58% or 52% I would not be stunned.

We know it's not going to be 40%, and we know it's not going to be 70%.

We do the best we can to sort out the "true" 55% who will vote, and generally speaking the good firms do a pretty decent job.

But 2 months out, it's still a crapshoot

[CARLHAYDEN]

Ok.

Basically we agree.

It seemed to me that in your previous post (responding to another poster) that screening was pretty iffy.

I agree that on the face of it, pollsters showing a turnout of over 60% of VAP probably have a very poor screen.

Essentially my point was a multi-stage screen.

First, is the respondent registered to vote.  If yes, go to screen two (if no, drop from main sample).

Second, does the respondent say they are likely to vote.
If yes, go to screen three (if no drop from main sample).

Three, did the respondent vote in the last election for the office being polled?  If yes, include in main sample (if no, weight the responses using democraphic information before including in main sample).

I have found the technique to be pretty reliable.

[The Vorlon]

It seemed to me that in your previous post (responding to another poster) that screening was pretty iffy.

It ***IS*** pretty iffy, it's just better than the alternative

[CARLHAYDEN]

I think the problem is that many firms only use screens one and two.

Adding on screens three and four (in my experience) vastly improves the results.

Let me give you a hypothetical example.

A. You get 'completes' on a survey from 1000 respondents (You've already established that they are all according to their statements, of voting age).

B. Say 876 of the completes claim to be registered voters, but in fact, only 700 are registered to vote (some people honestly think they're registered to vote when they are not, and some people just lie about it because they don't want the interviewer to think they're not good citizens).

C. Of the 876 repondents who went through the first screen,  700 tell you they are likely to vote (when in fact only about 543 will vote).

D. Of the 700 who passed through the second screen,  about 456 will tell you they voted in the previous election for the office in question (I'm using Presidential election figues).  In fact, about 432 of them voted for President in 2004.  The beauty of this is that not too many people will tell a whole series of lies in a short interview.  Besides, the previous questions have allowed them to establish that they are 'good citizens,' and therefor can explain away their non-vote in 2000.

E. Now, if we assume a turnout of 54.3% of the VAP (i.e. 543 respondents), then obviously we have a shortfall of 87 respondents if we solely relied upon the previous screen.  This means that just under 36 per cent of the people who passed all of the screens except the historical voting one are in fact likely to vote.

However, they are not equally likely to vote.  So, you use historical voting demographic data to make an improved projection of the likelihood of voting (SES).

Finally, you fold the 87 likely respondents (as established with this method) in with the 456 who passed the earlier screens.

While this process isn't perfect (polling is as much an art as a science), it does produce far more accurate results than the simplistic methods used by too many pollsters.

[Gustaf]

I talked to a prominent Swedish pollster this summer and he told me that pollsters nowadays can tell how many people will vote with great accuracy, but not yet WHO will vote, on an individual basis and that this is the big challenge for pollsters in coming years. Don't know how valid that is though.

[CARLHAYDEN]

Well, that really constitutes an interesting cross national question.

In the United States, only about 54 per cent of the Voting Age Population actually will vote in the Presidential election.

In many european countries, where voting is required by law, turnout runs around 85% of those eligible.

So, projecting who is more likely to vote in the United States is more difficult than projecting who is likely to vote in most european countries.

However, people in the United States and northern europe tend to honestly answer their preferences as to how they actually will cast their ballot in the election, whereas in many other countries the respondents are somewhat less likely to tell you their true voting intent.

[CARLHAYDEN]

I think the problem is that many firms only use screens one and two.

Adding on screens three and four (in my experience) vastly improves the results.

Let me give you a hypothetical example.

A. You get 'completes' on a survey from 1000 respondents (You've already established that they are all according to their statements, of voting age).

B. Say 876 of the completes claim to be registered voters, but in fact, only 700 are registered to vote (some people honestly think they're registered to vote when they are not, and some people just lie about it because they don't want the interviewer to think they're not good citizens).

C. Of the 876 repondents who went through the first screen,  700 tell you they are likely to vote (when in fact only about 543 will vote).

D. Of the 700 who passed through the second screen,  about 456 will tell you they voted in the previous election for the office in question (I'm using Presidential election figues).  In fact, about 432 of them voted for President in 2004.  The beauty of this is that not too many people will tell a whole series of lies in a short interview.  Besides, the previous questions have allowed them to establish that they are 'good citizens,' and therefor can explain away their non-vote in 2000.

E. Now, if we assume a turnout of 54.3% of the VAP (i.e. 543 respondents), then obviously we have a shortfall of 87 respondents if we solely relied upon the previous screen.  This means that just under 36 per cent of the people who passed all of the screens except the historical voting one are in fact likely to vote.

However, they are not equally likely to vote.  So, you use historical voting demographic data to make an improved projection of the likelihood of voting (SES).

Finally, you fold the 87 likely respondents (as established with this method) in with the 456 who passed the earlier screens.

While this process isn't perfect (polling is as much an art as a science), it does produce far more accurate results than the simplistic methods used by too many pollsters.

Vorlon, would appreciate your analysis.

[CARLHAYDEN]

Vorlon,

I repeat my earlier request.

Also, with respect to your site, how much was the contribution qualifier, and where should it be sent?

[struct310]

Vorlon,
This would be the year for a Cherynobl style meltdown for Kerry in Washington.

http://seattletimes.nwsource.com/html/regionalpolitics/2001936331_bush23m.html

Democrats have nothing going on in the state like this and I dont have to live there to know that.

Plus these 119000 Republican King county voters that stayed home probably did so because they thought Al Gore had won Florida

[MODU]

Vorlon,

I just heard a guy on the radio saying there were some new polls released today in the "battle ground" states.  Have you seen any pop up on your sites yet?

[ATFFL]

It is about time for the next Strategic Vision release.