Gallup Tracking Poll Thread [Obama vs McCain]

[muon2]

I am treating them as independent events.  I'm not presupposing what the true answer is.

You are presupposing the true answer. "...the idea that two polls showing different results would only have 1/400 chance of both being off and having the actual result be in the middle..." You are assuming where the answer is!

You are not treating them as independent events either. "Again, let's take my two poll, 8% off example." They're not independent if you specify where they are with respect to each other, as that statement does.

Statistically they are not independent events. They are independent measurements of the same event - the presidential preference. A comparison of MOE is inadequate, since these polls only use the statistical error to quote an MOE.

A better comparison would have the polls also quote a systematic error. The systematic error measures the expected measurement fluctuations based on the choice of modeling. Since no pollster directly uses a raw random sample, there is a systematic error associated with the poll. My guess is that the systematic errors are comparable to the statistical errors based on the dispersion of polling values seen on any given day.

[muon2]

If Gallup and the other polls show some closing in the margin that may not be to surprising. I've cited the 1976 race previously where Ford closed a similar mid Oct gap to get within 2% by election day. Gore likewise was the candidate of the incumbent party and closed a lesser gap to take the popular vote.

[muon2]

You must be logged in to read this quote.

Could you be bothered to refresh me on what regression to the mean means?  I seem to remember some of this from General Ed statistics at UCSC twenty years ago...

Regression toward the mean is the effect of looking at outliers of a sample and finding them to be closer in the next measurement. This effect is strongest when the samples and the method to collect them are identical and dominated by statistical error. In that case one expects that an outlier is really just a statistical fluke and the next measurement will have high probability of landing closer to the mean.

This does not apply if there are systematic differences in the samples. It also should be used with caution in polls, since it really applies to individual measurements (like batting averages or exam scores), and is less applicable for a measurement which is itself a statistical sample of measurements.

[muon2]

As much as the tracking poll is interesting, at some point we will want to have this entire board semi-retired as we did with 2004. Tracking official job performance is probably more appropriate on the US General board.