Vorlon said these polls are unreliable for small states, since they were set to get a sample right for the entire nation, and thus it's unlikely they got a good sample for a small state, with the correct numbers of men, women, whites, blacks, ect. (Same applies for blacks/whites, rural/urban, young/old, etc)
I expect that for his National polls Rasmussen is using something called a "cluster sample" bought from a firm that specializes in producing telephone calling lists (very creatively named "Survey Sample International") to have his robots make his calls from.
Cluster samples are designed to give you representative sample for the entire population, but not within small sub sets.
For example, the sample he is buying is balanced to indeed give you 51% Women and 49% men nationally, but it is not guaranteed to give you 51%/49% in say Ohio or New Jersey..(same applies for young/old, black/white, etc)
In the big states (California, texas, Ohio, MIchigan, etc) his subsets might work ok, but after that it gets pretty dicey IMHO.
The dividing line is VERY VERY roughly 12 EVS or so
If I were to take a guess.... (and labled as such) based on what I know of how Cluster samples usually fill out when you actually try to call them. the following is likely...
Bush will do worse than expected in the EST time zone, about right to a bit (?) better than expected in the central time zone, and also better than expected in the mountain and pacific time zones...
The problem is that when you do a national sample, in the small states you sample very few people. Out of a 1500 sample, you would talk to maybe 9 people in Montana, which needless to say is NOT a representative sample of Montana. You add up a bunch of small UNrepresentative samples, and you end up with a BIG unrepresentative sample..
The smaller the state, the bigger the potential distortion.
Bush +2 in Washington State anybody?
If Nevada, which is a small state in the Pacific Time zone, comes in Bush +9 I think we just toss the entire Rasmussen lot out the window from Coast to Coast (I am tempted to do that anyway BTW) ... we will see.
There is a possibility he is using a different type of sample which should give good results, but I think it is unlikely he is using this type of sample for his national polls from which these state "polls" are extracted.
Again this is a semi-educated guess... , but I'd be willing to bet a reasonably fresh bag of donuts....
These errors (if they exist) are over and above the BIG sampling random error from only 400 sample size, which will be in additional to any survey design error...
Remember the +/- 5% applies to both candidates, so the margin of error on the "lead" is actually 10%
Statistically there is a 68% chance the reported "lead" is within 5% - ie one out of 3 (32%) will report a lead that is wrong by 5% or more.
If Rasmussen does 40 of these state polls we can expect 10-15 of them to be wrong by more than 5%, and that is assuming ONLY normal statistical errors, without any additional errors due to the way he is actually doing these polls.These things are pretty darn dicy in my estimation...
Maybe Rasmussen is just doing this to save money?
These "polls" cost him nothing - he has already collected all the data, he just has to add them up and he gets "free" polls to use... ?
Link to Arkansas Polling