I suspect a lot of the state-level stuff is artifacts of the algorithm. The article says that only 2000 people were polled, and there are about 3000 counties,

I suspect a lot of the state-level stuff is artifacts of the algorithm. The article says that only 2000 people were polled, and there are about 3000 counties, so there's a lot of modeling that goes on to get from the poll results to the map. It's a two-step process: (1) build a model which predicts partisanship from individual demographics. (2) apply this to individual-level demographic data and aggregate to get partisanship by county. But! The demographics come from voter files. And voter files are maintained by the states. So the people overall in some county in South Carolina might be quite similar to those in a bordering county in North Carolina, but the ones in the voter file are not. I'd take within-state differences seriously, but I wouldn't take between-state differences seriously. (However! Utah is dominated by Mormons and you really do see differences in a lot of maps between Utah and adjacent states.)