Key lines from Jay Cost at RealClearPolitics with emphasis added:
I think the problem is that we do not have a reliable metric to measure the state of the race. Polls are of limited utility for gauging Iowa Democrats. This is a subject I discussed earlier in the year. There are two problems. The first is devising a sample of voters. Turnout in the Iowa caucuses is difficult to measure because it takes a good degree of devotion to participate. [...] A poll of Iowa Democratic caucus goers does not really mimic the process in which they participate. [...] Democrats begin by standing in an area designated for their first choice candidate. Then, for thirty minutes, they either persuade or are persuaded by others to switch their choices. At the end of the half hour, electioneering is halted and caucus officials count the number of supporters that each candidate has. Candidates who have less than 15% or 25% are deemed not to be viable. And so, another thirty minutes for electioneering is once again granted. The supporters of nonviable candidates must find new candidates to support, team up with supporters of other nonviable candidates to make their candidate viable, or abstain.
In short, the polls tell us that Hillary Clinton, John Edwards, and Barack Obama all have a large enough group of supporters in Iowa that they could win. They really don't tell us anything else, since the polling firms have no real ability to model what's going to take place.