Well, my initial reaction to the second debate -- that Obama won, yet Romney did well enough to maintain his support -- seems to be bearing out in the daily trackers, in spite of the media and blogs coalescing around how an idea that Romney was beaten badly, what with the "act of terror" drama and the awkward "binders full of women" phrasing. Three days of post-debate polling are now a part of Gallup's seven day rolling average of likely voters, and there's really been little shift in the last few days. As for Rasmussen's 3 day roller (which, you may remember, I'm always skeptical of for methodological reasons) has Romney over Obama by 1%, 49-48; this is pretty much where it was after the first debate (Romney over Obama, 49-47).
Some differences are so small that a sample of a thousand people is not going to give you a confident read. If the margin of error on a poll of a thousand people is 3%, it takes exhorbitant resources to whittle the margin of error down further. It's not a linear relationship, it works on squares. If you want to cut the margin of error in half, you can't just double your sample, you have to quadruple it. If you want to whittle that 3% MOE down to 1% (cutting it down by two-thirds) you have to interview nine times as many people. Every organization has priorities, and I doubt that a more precise read on the daily preferences outweighs staffing decisions.
In line with this... Not sure if you saw Nate Silver on the Daily Show this week. Stewart asked Silver about the limits of polling, and Silver pointed to the 2008 Senate race recount, and the slim margin by which Al Franken won. And on top of that, there's no way the best, most expensive telephone poll in the world could have anticipated the issue of ballot format on voting (I'm talking about the 2000 Palm Beach County butterfly ballot). All those Buchanan "voters" would have been telling pollsters they would vote for Gore. Because that's what they tried doing. (Yes, you could actually do an in-person interview with an actual ballot, or do it online, but I don't see that happening.)
And yet, we have to remember that respondents in polls are people; Presidential election results are based on people grouped into states for the purposes of the Electoral College. So the national polls remain practically meaningless; and if I were running a national polling organization I would undersample the "safe" states, and oversample the "battle ground" ones, and re-weight the data to put them in proper proportion. (Read the disclosure... I need to add that I don't even know what my company is doing in its polling.) Two hundred surveys in Ohio is more valuable in a projection than two hundred surveys in New York.
