Polls are not science

[Darby]

The are imperfect methods or approximations, conducted by imperfect (not omniscient) people.

And try as they might, is there really any way to know how a nation of 290 million will vote until the people actually really vote?

[Gman]

1. No, but polling, done correctly

will give you a very good approximation.

[SheilaT]

2. As Gman said

The right kind of polling, done correctly, is incredibly accurate. Some election cycles back all the networks did their own individual polling of "key" precincts, and it enabled them to call each state almost as soon as the polls closed. But more recently they stopped doing their own separate polling, pooled their resources and all such polling was done by Voter News Service (I think that's what it was called) who of course had it right in 2000, but took a huge hit for being correct. The result? No exit polling any more, which makes it much easier for evil people to steal elections.

Fifty years ago Isaac Asimov wrote an eerily prescient short story called "Franchise". It's in his short story collection Earth is Room Enough which is out of print, but can probably be found in a library or gotten on Amazon. Read it. It's about polling taken to its logical extreme.

[sonicx]

3. only a few are polling with correct weights...

fox is weighted ok, zogby is weighted the same as 2000's exit polls, Rasmussem is weighted around what zogby is.

the rest are unweighted and oversample republicans, thus the 11 point leads.

[whistle]

4. First, lets look at the basics....

National Voter Turnout in Federal Elections: 1960–2000
Year VAP RV T/O TO/VAP
2000 205,815,000 156,421,311 105,586,274 51.3%
1998 200,929,000 141,850,558 73,117,022 36.4
1996 196,511,000 146,211,960 96,456,345 49.1
1994 193,650,000 130,292,822 75,105,860 38.8
1992 189,529,000 133,821,178 104,405,155 55.1
1990 185,812,000 121,105,630 67,859,189 36.5
1988 182,778,000 126,379,628 91,594,693 50.1
1986 178,566,000 118,399,984 64,991,128 36.4
1984 174,466,000 124,150,614 92,652,680 53.1
1982 169,938,000 110,671,225 67,615,576 39.8
1980 164,597,000 113,043,734 86,515,221 52.6
1978 158,373,000 103,291,265 58,917,938 37.2
1976 152,309,190 105,037,986 81,555,789 53.6
1974 146,336,000 96,199,0201 55,943,834 38.2
1972 140,776,000 97,328,541 77,718,554 55.2
1970 124,498,000 82,496,7472 58,014,338 46.6
1968 120,328,186 81,658,180 73,211,875 60.8
1966 116,132,000 76,288,2833 56,188,046 48.4
1964 114,090,000 73,715,818 70,644,592 61.9
1962 112,423,000 65,393,7514 53,141,227 47.3
1960 109,159,000 64,833,0965 68,838,204 63.1

VAP - voting age population 18+
RV - registered voters
T/O - voter turnout
TO/VAP - percent turnout to voting age population

These data can be found by state using GOOGLE and search by "voter registrations & turnout - year". The polling science is to determine a sample size and draw a random sample by state of registered voters (available to all polling firms who represent candidates - phone numbers can be assigned with 86%+ accuracy using address data bases). The calls are made through large phone centers so these can be done in a relatively short space of time (3 to 5 days).

The sample size is the big variable. Polls generally use a sample of as little as 600 interviews to about 1,300 interviews. But generally they run around 1,050 or so. These sizes are used to make sure that the completed calls can be made in the quickest time possible before factors come up that change voters minds. But, these sample sizes are stated to have 95% confidence levels which means that if repeated say one hundred times in an identical manner (different samples but the same procedures) you'll wind up with the same range of results 95 times out of 100 and that range will generally be as much as +/-4.1% (8.2 point spread) with the lower number sampled to +/-2.9% (5.8 point spread) with the upper number sampled. You can tighten up on both the range and the confidence levels by increasing the sample size, but you really have to control the procedures to make sure the larger numbers are done indentically and no short-cuts or cheating occurs. For example, a sample of 22,000 completed qualified random calls would inprove the confidence level to 99% (meaning of 100 repeated samples of the same size you get the same range of results 99 times) and the error rate would be in a very tight +/-1.0% (only a 2.0 point spread). This is very good, but the effort and resources expended are 20 times or more to get that level of accuracy.

If a tradition poll costs a candidate or sponsoring media company $10-$12 per completed interview or $10,000 to $12,000, that cost would mushroom to $200,000 to $240,000 or more. We also see how easily the opinions of voters change, by issues, advertising, scandals, so on, so the norm has come down to more frequent polls, smaller sample sizes, and respondent to the issues, then measure the potential voter reactions. The other thing is that rather than have an informed voter get a clear picture of what each candidate stands for, keep the air charged with uncertainty as much as possible right up to the voting date among the undecideds and pray that they'll vote there conscience and pick the right man for the job. So polling can be science, but there are a lot of junk polls going on. I like the internet polls that are controlled because internet users are generally better informed, more intelligent and likely to give honest answers, even when they are DU'ing or freeping a poll, as long as there is a control that only lets them participate once (email checks) and only accepts their participation if invited and uses some type of profile screener questions.
O8)