Actually, the model here has a slope of 0.87 +/- 0.02, or 7/8. Which means for every 8% fractional increase in Clinton's vote count (say, from 100 votes to 108) one would expect a 7% increase in Obama's vote count (say, from 100 votes to 107). And you start at the lower end with 1 vote for Clinton and 2 for Obama. So, for instance, at 100 Clinton votes you predict 112 Obama votes, but at 1000 Clinton Clinton votes you predict 843 Obama votes.

If you use Obama's vote percentage as a predictor of Clinton's vote percentage you get a regression line that's much closer to 1:1 - this is because it is different to minimize the variation in Clinton's vote given Obama's than Obama's given Clinton. The correlation is 93% (explained to be that high because precinct size correlates with both vote counts) and that should be the geometric mean of the two regression slopes.

In fact, linear regression is not the proper tool here as we're not really trying to use one of them as predictor for the other but rather find a relationship between the two that treats them on an equal footing. Principal component analysis would be much better.

In any case, there seems to be a very slight slope here, favouring Clinton in large precincts.

I think I'm going to replace that chart with one in which Machine vs. Hand counting is represented by different colours.

We have met the enemy, and he is us — Pogo