A polling thread for POLLS (and other numberical oditties)

[Dave]

I've been trying to devise a description of the oddity called polling that would remain true to the mathematics.

Be patient for a paragraph or two.

Scientific statistics
As a scientist, I would devise a premise and a theory so that I could collect data at several different times, vary temperature, vary pressure, and other things called "independent" variables. I wold then match those results to the theory. The theory is a description of how the "stuff" behaves in a reactor at time, temperature, etc... The theory explains the reaction being measured. There is a target value that is "predicted versus actual" and the agreement can be measured. The Higgs Boson is this way.

Polling Statistics
Now polling public opinion is different. It has no model or theory or premise. Nothing explains the way the public votes. Even past behavior is no predictor of future behavior with any great degree of certainty.
This lack of a "theory" is why STATISTICS for polling is different than STATISTICS for science experiments.
There is no predicted value. There is only one result, the election.

What this means is that political polling isn't "obeying" and underlying theory, it is merely a measure of opinion on that day at that time.

Take for example the "Daily Tracking Polls" ...
Gallup and Rasmussen do daily tracking polls and then average the results of 7 and 4 days respectively.
A daily poll allows a smaller sample because the "margins of error" are calculated on the 4 day total.
Averaging a daily poll over 4 or 7 days is supposed to smooth out the variations in the data (we call that scatter) ..
For example:
The current Gallup Daily Tracking poll just swung 4 points for Obama. That's not a good thing for this poll to do.
There is no clear reason for Obama to be "up" in a poll today, no explanatory public event. There is no clear reason for ROmney to have lost 4 points in one day.
That means that the Gallup Daily Tracker is still way to subject to sample variations to be trusted beyond what I said in the other thread --- trust only the aggregate number, the long average that you can see scattered on a graph.
This is also happening in Rasmussen's Daily Tracker.

This isn't fitting measured results to a predicted curve. This is taking samples and then look back for a trend.
Now we've reach that point I made in the other post --- the past is not a predictor of the future
Nothing predicts public opinion ahead of time.
Look at a graph of all the polls against the date they were taken. You cannot draw a line beyond the last polling date. HOWEVER, you can see if a candidate is rising or falling over time.

That brings us back to the 47/47 split that I spoke about before.

There's more involved in the polling than just these few points but this is the start of understanding the special nature of polls and why they don't behave the way we think they might. I will add that Stock Market Statistics are similar to Polling in that they do not have an underlying theory. HOWEVER, there are some very good stock predicting programs out there that do make money most of the time. The thing is, like an election, the stock markets are subject to that variable of public opinion and the public behaves the way it wants -- sometimes illogically, irrationally, and unpredictably.
Hey, we're human, aren't we?
Real science is statistically never irrational, illogical or unpredictable.

[Dave]

Tomorrow, I will answer a question that I'm surprised no one has asked...

Why do the two tracking polls differ? (Gallup and Rasmussen)
Shouldn't they see the same numbers?

[Riverwind]

Simple, Rasmussen is a right winged based poll so the questions tend to be right wing. A poll is only as good as the question asked, given that your poll has the diversity to cover all people. I don't think Rasmussen does that but I do watch it, when I see it even or has a republican up by only a couple points, I know its good for the democrats. So I adjust my numbers based on who is given the poll.

River

[fhunter]

Tomorrow, I will answer a question that I'm surprised no one has asked...

Why do the two tracking polls differ? (Gallup and Rasmussen)
Shouldn't they see the same numbers?

I agree with Riverwind here - it is either different questions asked OR different audience (you can not ask everybody, etc.).

[A-1]

Rasmussen, Rasputin, no difference. Fuck ALL of the Republicans!!!

[Dave]

I don't think that Rasmussen set out to be biased. They do, at least, release enough of their raw numbers for people to understand why they are not identical with Gallup.

There are Republican and Democratic leaning pollsters. So what. AS long as we can look into their methodology and see why, it's not a problem.

No one polling firm can do all the polling we see out there. AND, you don't want that to be the case.

ALSO, there are variations that are truly startling in these two tracking polls that a lean to one political party doesn't explain.

I'll try to explain that tomorrow, most likely late in the day.

[Dave]

One of the reasons that the Gallup Tracking Poll and Rasmussen Tracking Poll differ goes back to the statements I made in an earlier post about the oddity of public opinion poling -- there is no underlying theory that predicts the result.

The math argument:
If I am seeking to find experimental evidence of Black Oil Turning to White Gasoline following the theory that
Gasoline equals Oil times temperature and concentration,
That is a line on a graph that I can compare measure results against.

There is no such theory or equation in public opinion polling.
The "theoretical result" is the one number that happens on election day -- candidate A = 51%, Candidate B - 49%.

That number is composed of ACTUAL VOTERS.
Before that, a poll can measure only the following:
-- ALL citizens including non-voters (not the Actual Voters)
-- Registered Voters (again, not the Actual Voters)
-- Likely Voters (once again, not the Actual Voters)

People can decide to vote or not vote up until the polls close and that is one cause of the uncertainty and of the difference between Gallup and Rasmussen's Tracking Polls.

Each polling firm gets to pick a "model election" for voter turnout. Since no one can predict exactly who will vote ahead of time, the polling firm picks a year -- say 2000, 2004, 2008, 2010, as a model for voter turn out. Yes, congressional elections are valid to pick as a frame for a model of turnout.

In 2012, Rasmussen picked the turnout in the year 2010
In 2012, Gallup picked turnout from the year 2004
PPP has picked the turnout from the year 2008.
(I think I'm right. I might be off here, but the numbers are close)

Each pick is valid and each pick alters the prediction when a poll is taken.

Within that choice of year is an explicit statement of the following:
a) the turnout of White versus Non-white voters
b) the turnout of men versus women
c) the turnout by party affiliation, republican versus democrat (often called enthusiasm)

Again, picking those numbers ahead of the vote will change the result of the poll. This is essential in the mathematics. You cannot take a simple "yes/no" poll of 100 random people. Those results must be fit into a model of the electorate and that electorate is not simple. It is very complex in its makeup.
The way around that is to wait until election day because all of those numbers will be known after election day. however, that means no polling which is stupid. We want to see the polling. So predicting that number is critical to the success of the poll.

There is a central point -- no polling company wants to be known as "the company that missed the result"... every company wants to be the one who hits the percentage right.

[Riverwind]

Dave that is why when it comes time to actually vote even though the numbers will be different both will claim victory because they all put in the dishonest disclaimer, +or- 3% which will cover their errors.

River

[Dave]

Dave that is why when it comes time to actually vote even though the numbers will be different both will claim victory because they all put in the dishonest disclaimer, +or- 3% which will cover their errors.

River

That's true to an extent. It's better than a guess.
There are ways to measure the accuracy of polling in an election but only after the fact when all the votes are in.

However. the error margins are real. I've dealt with error margins from my first year in chemical engineering. Let me assure you that those + or - limits are true. And the single number prediction is true as the most likely number. I said Most Likely Number, that's all it is.

I still have to figure out a way to explain why both Tracking Polls swing so damn much. They shouldn't be doing that.