What's going on with weird tail correlations in ?
statmodeling.stat.columbia.edu/2020/10/24/rev
One theory:
If one of your parameters is "expected national vote" and you dial up one state, that parameter pushes down on other states to try to fit to the expected national vote.
Conversation
Negative correlation is weird, it doesn't make sense that winning one state would make a candidate more likely to lose another one.
1
mostly agree, *unless*:
if your main datapoint is "national polls are 50/50 and really good but I don't have a breakdown"
and then you learn that Biden is up 25% in Alaska
you might still believe the 50/50 poll, and so think Alaska must be counterbalanced somewhere