It's not impossible to poll the Iowa Democratic caucuses accurately, but it is very hard, and results have to be taken provisionally. The caucus rules skew the math so weirdly that pre-Caucus polls often over-capture momentum surges and understate core strength.

It's clear, from a variety of recent polls, that Barack Obama has marginally improved his standing among likely caucus voters since September, and that Hillary Clinton's standing has marginally dropped. (John Edwards is standing still, roughly).

But a poll of likely caucus goers -- roughly 5% of adults, requiring an initial sample of 3,500 or so -- does not really capture and cannot really project delegate allocation ratios.

One reason is that a surge or decline in turnout can completely muck up the model. And it only takes about 20,000 troops to implement this surge.

More fundamentally, the caucuses themselves can vary wildly. You might show up supporting candidate Green but be forced to support candidate Blue because of the alignments of that particular caucus. Some caucuses may be entirely for one candidate or another; others may be split. .....

And the allocation of the delegates itself is not proportional to population. Bonus delegates are awarded to Democratically performing areas. Certain smaller counties pack the same punch as bigger counties -- a caucus of 50 in one part of the state can yield the same number of delegates as a caucus of 100 in Des Moines.

As a totally unscientific rule of thumb, some analysts tend to subtract three points from Barack Obama's percentage in a good poll -- he does better in urban and suburban areas than he does in rural precincts ... and tend to add a few points to John Edwards's tally. He has many second-tier counties locked up.

Do not put much stalk in Democratic polls.