Some dood commentating on 538 had a good explanation of why the national polls might skew Roms while the state ones don't.
I wonder how much of the State-level / National Polls difference may be due to asymmetric distribution of errors among small sub-samples.
1. Suppose a national poll manages to contact 4 Hispanic voters in Texas (it "should" get 15 out of 1000 nationally - about 1.5 percent of US likely voters are Texas Hispanics! - but it's a difficult group to reach). And it "should" find that 3 support Obama and 1 supports Romney.
In reality, the group is so small it would not be much more likely to get this result than any other. And the mistake of getting 1 Obama is not cancelled out by a mistake of getting 5 Obama; 100% is a ceiling which makes the error asymmetric.
So, when a national poll "reads" Texas, it often gets a wrong Texas, with an unrealistically low Hispanic support for Obama, one which is even more pro-Romney than it is in reality. Similarly California's O support is estimated too low, etc.
2. State-level polls of the same states will not commit the same mistake simply because their sub-samples of the typical Demographics to that state would be much bigger.
3. This is a polling reality for Obama more than for Romney. Obama's support is more concentrated in groups that are (a) heavily skewed towards him, (b) are hard to poll. For this reason, they would tend to display asymmetric error that favors Romney in smaller samples.
4. In short, Obama's support is more granular and for this reason he does better in more granular polls. It could be as simple as that.
http://fivethirtyeight.blogs.nytimes.com/2012/10/26/oct-25-the-state-of-the-states/?comments#permid=38
http://fivethirtyeight.blogs.nytimes.com/2012/10/26/oct-25-the-state-of-the-states/#more-36719