Nate draws upon data, creates models and poses theories that take me half a day to figure out. OK, I'm dumb.  

  But I watched the Gallop Daily for several months. It only told me that trends over weeks and months made the results of any few days or even a week plus of days worthless.  A few days or a week before election day might be significant but I'm at a loss to find any single day result 'interesting'.

 Do we even know if Gallop can explain why specific days of the week swing to favor McCain and other days give Barack an edge? Hey, which day is the 'best' day?

  See, I thought the added value from Nate (over single polls) was the freedom he gave us to ignore why one poll showed a tie, one showed Barack up by 9 points and a third shows a 16 point lead.  It would be easier to rely on 1 Poll but I thought I was goin' crazy at Nate's site because "easy was bad".

Michael,

There are several different types of problems or errors that arise with tracking polls, and the commentary seldom distinguishes.  The one you always hear about is statistical error.  But there is also sampling error and problems that arise from the nature of the question asked.

The statistical error (what you always read about, that the poll result is "accurate" plus or minus 3% or 5%) is purely a statistical result of the fact of sampling.  If you have a bowl full of 100,000 black and white marbles, the only way to know exactly how many are black and how many are white.  If you want to know "within a range of error" how many are black and how many white, you can draw a sample and count that.  But it is in the very nature of randomness that if you take a sample, count it, put it back, re-mix the bowl, and take another sample, and then another, and then another, you are not going to get the exact same result each time.  Your sample of samples will eventually form a normal (bell curve) distribution around the actual figure.  How tight the bell curve is to the actual figure is a function of the size of the sample.  The larger the sample, the tighter the bell curve to the single correct figure.  The percentage cited as the error is the width of the bell out to about 3 standard deviations, but even that is not absolute because sometimes you are going to get a result that is very far from the mean or correct value just as a matter of chance -- an "outlier."

This all pre-supposes that the sampling technique is perfect so that the only error is the inevitable uncertainty that comes from the fact of sampling, an error that can be ignored if day after day the poll shows the same thing.  If, however, the sampling technique is itself flawed as a result of correlations between the methods used to select the people to be questioned or self-selection processes that skew the pool based on who is willing to speak to a pollster, then there is a systematic bias in the poll, and the statistical error described above is the uncertainty about the systematically biased outcomes.

Then there is the third problem about just what the questions actually mean in terms of predicting elections and whether the structure of the poll questions is introducing some bias.  For example, the typical question is, "If the election were held tomorrow, who would you vote for, Barack Obama, Democrat, John McCain, Republican, or undecided."  But people don't necessarily answer truthfully, for a variety of reasons (e.g., the "undecided" is typically thought to overstate the number who are actually undecided) and it is not clear that what they answer on a given day is completely predictive of what they would do.  After all, the election is not in fact being held tomorrow and the hypothetical nature of the question tends to muddy the results.  The assumption, and it is an assumption, is that the voters on one side are no more likely to give answers different than their ultimate behavior and that the errors therefore wash out.  But not necessarily.  Sometimes too the other questions asked in a poll can influence the way in which people answer the tracking question.

If you look at the array of polls, it is clear that there are biases that result from some combination of sampling technique and questions asked.  Rasmussen seems almost always to have bigger differences between candidates than other polls.  This was true during the Democratic primaries.  I don't know why, but the fact that it is consistently so evidences that Rasmussen's techniques are different in some way, as each pollster's technique is different in some way.  It even appears to be the case that the poll results for FOX and WSJ, tend to be more favorable toward the Republican or more right-wing candidate than the others.  Why?  Some type of bias due to sampling most likely, but in a consistent direction.

What Nate Silver is taking not of is that the Gallup result that might have been taken to indicate some trend in McCain's direction turns out most likely to have been a case of the first type of error, a random variation from the mean.  That was fairly predictable as the most likely case given that none of the other polls was showing the same relative movement.

Nate, if you are up to it, some explanation of what you think are the consistent biases of the various pollsters and why those biases are there would be very helpful.

  roidubouloi, I'd not dare to articulate my thoughts as you did, But yes, I do suspect why polls fail and the charm of Nate's sausage making is that he makes a good case for why he's not giving a snapshot but plugging into the future.  (All the more reason I was surprised he found a 'daily' to be interesting)

  Regarding "the consistent biases of the various pollsters" it seems they are learning and adjusting and we may not know till November if they had the correct recipe.  Also, consider Barack is daring to go where 'conventional wisdom' would have warned otherwise.  The 'old' wisdom of the traditional polls?  

  And he seems to be making progress if not nailing down what many called lost causes.  

  Another good question, "Is Obama paying much attention to the commercial polls or does he think he has better info from the inside?" Jeeze, they may be laughing at the 'reality' we all soak up but hope their secret formula stays in house....

Obama and McCain certainly have their own polling, and it is a big part of their expenses, tailored to their strategy.  The poll more heavily where they plan to contest and may ignore other places altogether.  They may depend on the public polls for the "overall national percentage" but that would be about it.  They are also polling issues and all sorts of non-issue issues (like race, age) to try and find things that can be spun in a campaign and used to advantage.  The importance of insightful polling is part of why Mark Penn got to have such an important role in Hillary's campaign.

See below for some real polling analysis from Nate, but this Gallup poll , which is being cited by Andrew

As the primaries dragged on--and on and on and on--many pundits worried that it was damaging Barack Obama's