Reid Wilson analyzes the polling split:
Pollsters on both sides try to persuade public surveyors that their voter-turnout models are more accurate reflections of what's going to happen on Election Day. This year, GOP pollsters and strategists believe those nonpartisan pollsters are adopting Democratic turnout models en masse.
Regardless of the cause, strategists on both sides acknowledge the difference in their internal polling. Republicans believe Democrats are counting far too much on low-propensity voters and a booming minority turnout that isn't going to materialize on Election Day. Democrats believe Republicans are hopelessly reliant on an electorate that looks far more like their party than the nation as a whole. The day after Election Day, somebody's pollsters are going to be proven seriously wrong.
Douthat believes a Romney win is possible:
Some partisan groups aside, pollsters have every incentive to get things right, and the fact that there are more polls in the field predicting the kind of turnout Democrats want is good evidence that such turnout might well be forthcoming. But it isn’t dispositive evidence. As Dan McLaughlin has pointed out, both Gallup and the similarly Romney-friendly Rasmussen have a pretty good track records when it comes to projecting the likely final breakdown of a presidential-year electorate, and the other national polls (Fox, Pew, Washington Post/ABC) showing the race tied are presumably using demographic projections closer to Gallup/Rasmussen than to the state polls showing Obama with a slight lead. Factor in the evidence, from surveys and state-level data, that the president is underperfoming relative to 2008 in the early voting, and there’s a coherent, plausible, data-driven case — not just a ranting, emotive, “it feels close” pundit’s case — that the state-level polls are projecting a liberal-tilting electorate that won’t actually turn out.
The Dish 