Wow... A question I
might be able to answer..
A poll has two types of error.
Systematic error:
Systematic error is when there is just something "wrong" with the way the poll is constructed.
Probably the most common type in here is an improperly drawn sample.
To pick on one of the polls I love to hate, CBS news does somethng structually, fundementally wrong in that they use pure simple "random digit dialing" to get their 1000 or so phone numbers.
What is so bad about this you may ask....?
In a "perfect" sample ever VOTER (not phone number) would have an equal chance of being contacted.
Married couples (ie two votes) tend to have one phone number.
Single people living alone also tend to have one phone number.
Married people break strongly for the GOP, while singles break strongly for the Dems.
Because the ONE single voter has ONE phone line, while the TWO married voters have ONE phoneline, the single voter is twice as likely to be contacted as the married voter..
So CBS news structuarally, inherently, as a matter of poll design, has twice the chance of reaching single people (a strongly pro Democratic voting block) as they do of reaching married people (a strongly pro GOP block)
Guess what - CBS polls always wildly favor the Democrats... this is not a fluke... it is a structural flaw with their poll...
Of course this is a huge over simplification, in reality you have to take into account "second" lines for the teenagers, cell phones, etc... but the illistration is valid...
(Enroll in "Advanced Polling 623" - "how to draw a representative phone sample" - for more information)
A second structural error is a bad or biased question wording.
The classic example of this is the pro life / pro choice polling. (A poll that is, IMHO, just about impossible to do accurately)
For example a question asking if you :
"Favored the continued state funded and sactioned slaughter of unborn children"
will likely produce a somewhat different result that asking:
"Should the oppressive authority of the state be used to force women against their will to bear children"
Both of the above are examples of structural or systemic error.
Random Error
The second type is just "random" error due to sampling, it has not bias that favors one side over the other.
If you flip 100 coins in the air you don't get 50 heads and 50 tails, you usually get 53/47 or something like that, the result is no more likely to favor heads than tails.
Lets use an example of a poll that is 53% for A, and 47% for B with a margin of error of +/- 3%
There is a 68% chance that "A" is between 51.5 and 54.5%
There is a 68% chance that "B" is between 45.5% and 48.5%%
In this example, there is a 95% chance A is between 50 and 56%, with a similar 95% chance "B" is between 44% and 50%
As a very rough rule of thumb, if the "gap" between candidate A and B is less than half of the margin of error the two are pretty much tied.
If A beats B by more than 1/2 the MOE A is likely ahead by a bit
If A beats B by more than the MOE you can be pretty sure they are ahead... 19 times out of 20 anyway..
Typically, Systemic errors, in all but the very best polls, are larger than random errors.
Systemic errors are why there are University polls showing Bush +4 in New Jersey.
This concludes the introductory lession of Polling 101 - are they any questions...?
