Rasmussen polls from Tx and California

[agcatter]

Just released.  No surprises.  Cal - Kerry 49 - 41, Tx - Bush 55 - 38.

My question.  All these polls taken over 30 days in May.  What the hell good is a poll taken over a 30 day period for crying out loud?

[I spent the winter writing songs about getting better]

Kerry ahead in CA, Bush ahead in TX. In other news today, the sky is blue.

[Fmr. Gov. NickG]

This is just the subsample from these states of Rasmussen's national daily track.  I'm not really sure what to make of it.  Is each state being independently weighted?

[agcatter]

I am not sure how this is being done.  Maybe Vorlon is around tonight.

[MODU]

I still say that California will be the surprise of the 2004 election by tipping to Bush.    700K move votes from the 2000 election is all he would need.  hehehe

Hey, one can always speculate.  

[I spent the winter writing songs about getting better]

Bush would need to win with at least 55% nationwide to get 700k to switch. It is a lot, even in California.

[The Vorlon]

I am not sure how this is being done.  Maybe Vorlon is around tonight.

BTW - New York is posted 57/34 - no surprise here either

http://www.rasmussenreports.com/New%20York%20Jun%202.htm

How valid a cumulative result for each state from his national sample is a statistically complex question I am actually going to have to think about.

I am just about 100% certain he does NOT do a hard weight by Party ID on these sate polls - I am just about positive the whole sample would really "break" from a randomness point of view.

Assuming he is using these samples but not weighting by party ID, My initial guess is that "strictly" by the book it should not work, but that over 30 days it should be fairly close. (other than, obviously, the timelyness issues)

There are two types of sample types Rasmussen might be using for his polls.  One is called a "cluster sample" which I think would cause a fair number of problems used this way. (not 100% sure on this - I need to think about it some more, and look up a few bits of math my memory is a bit fuzzy on)

The other one is something called a EPSEM sample, which should actually work pretty well the way he is doing things. (Again, other for the timelyness issue - folks interviwed on May 1st obviously could have changed their minds by now)

I think it is more likely he is using a "cluster sample" however, but I don't know for sure.

I have sent Scott an email to ask him.

For anybody interested, here is a decent link to how telephone samples are actually generated.

They kinda gloss over a few of the really nasty math bits (the way they explain it is close, but not exactly correct), but its a good overview.

The first four chapters are just fluff, the guts are in chapter 5, you can just skip there.

http://www.genesys-sampling.com/refernce/genmeth.htm

[agcatter]

Vorlon, in your correspondence with Rasmussen, have you been able to find out what percentage turnout he is basing his numbers on.  I think you once said he overhauled his turnout model since he missed in 2000 - revised it up I presume.

[CARLHAYDEN]

I am not sure how this is being done.  Maybe Vorlon is around tonight.

BTW - New York is posted 57/34 - no surprise here either

http://www.rasmussenreports.com/New%20York%20Jun%202.htm

How valid a cumulative result for each state from his national sample is a statistically complex question I am actually going to have to think about.

I am just about 100% certain he does NOT do a hard weight by Party ID on these sate polls - I am just about positive the whole sample would really "break" from a randomness point of view.

Assuming he is using these samples but not weighting by party ID, My initial guess is that "strictly" by the book it should not work, but that over 30 days it should be fairly close. (other than, obviously, the timelyness issues)

There are two types of sample types Rasmussen might be using for his polls.  One is called a "cluster sample" which I think would cause a fair number of problems used this way. (not 100% sure on this - I need to think about it some more, and look up a few bits of math my memory is a bit fuzzy on)

The other one is something called a EPSEM sample, which should actually work pretty well the way he is doing things. (Again, other for the timelyness issue - folks interviwed on May 1st obviously could have changed their minds by now)

I think it is more likely he is using a "cluster sample" however, but I don't know for sure.

I have sent Scott an email to ask him.

For anybody interested, here is a decent link to how telephone samples are actually generated.

They kinda gloss over a few of the really nasty math bits (the way they explain it is close, but not exactly correct), but its a good overview.

The first four chapters are just fluff, the guts are in chapter 5, you can just skip there.

http://www.genesys-sampling.com/refernce/genmeth.htm

Thanks for the study.

Let me point out a few problems with the data base.

First, until recently, nxxs were issued in blocs of 10,000 (no 1,000), hence a "communicty" for 500 would have as many possible numbers assigned as a "community" twenty times as large.

Second, not all residential land lines are equal.  Some are dedicated lines for internet and/or fax use.  Others are 'teen' lines.

Third, the study does not indicate if 'distinctive ringh' numbers are counted the same as primary numbers (if so, it would distort the results).

Fourth, it is not at all unusual for the nominal telephone customer to be different from the actual customer.  Widows keep the telephone in the name of the deceased husbands to cite but one example.

Firth, a number of telephones in rural areas are 'foreign exchanges' maintained for convenience, and the residents do not actually reside where the nxx would indicate.

There are other problems, but I thought I'd just cite a few.

[© tweed]

I'm banking on no 60% for Bush in Texas.

[The Vorlon]

Thanks for the study.

Let me point out a few problems with the data base.

First, until recently, nxxs were issued in blocs of 10,000 (no 1,000), hence a "communicty" for 500 would have as many possible numbers assigned as a "community" twenty times as large.

Second, not all residential land lines are equal.  Some are dedicated lines for internet and/or fax use.  Others are 'teen' lines.

Third, the study does not indicate if 'distinctive ringh' numbers are counted the same as primary numbers (if so, it would distort the results).

Fourth, it is not at all unusual for the nominal telephone customer to be different from the actual customer.  Widows keep the telephone in the name of the deceased husbands to cite but one example.

Firth, a number of telephones in rural areas are 'foreign exchanges' maintained for convenience, and the residents do not actually reside where the nxx would indicate.

There are other problems, but I thought I'd just cite a few.

The study BTW is 5ish years old, so yes, a lot of the technical points you raise are valid.  It's just an intro, and the only one I know of on the internet I can link to. - It's a low level intro..

But yes the "science" of building a random sample is held together with Bubblegum, duct tape, chicken wire, and string.. (don't tell anyone.. ok..?)