Welp this is annoying, but we found an issue with how our model was making state-by-state forecasts. The error is fixed. The topline effect is not very large and mainly helps "no majority" slightly. A few state forecasts have changed more noticeably. 1/
projects.fivethirtyeight.com/2020-primary-f…
projects.fivethirtyeight.com/2020-primary-f…
Specifically, the demographic regressions we use as a compliment to the polls weren't calculating properly since some code changes we introduced on 2/5 (after IA). As a result, the state forecasts were underdispersed (too conservative) especially in states with little polling. 2/
Without demographics the model basically defaulted to a geographic prior, which tends to underrate Biden/Bloomberg in the South and overrate them (especially Biden) outside the South. It also tended to overrate Sanders in Southern states and underrate him in Western states. 3/
So, for example, the model now has Biden doing better than before in South Carolina and Alabama, but Sanders doing better than he was before in Nevada and California. 4/
The district-by-district forecasts were also underdispersed. That hurt candidates who projected to get <15% statewide, but who were potentially strong in some districts and thereby could expect to pick up some district delegates. 5/
The fact that you can pick up district delegates even in states where you don't get 15% statewide makes it slightly harder for candidates to win by runaway delegate margins. So reintroducing the proper amount of district-by-district variance tends to help "no majority". 6/
My sincere apologies for not catching this sooner. Bugs can happen but I usually have a good eye for when something doesn't look right. Thought the changes were just the model reacting to IA; didn't realize I'd also introduced a problem with the demographic regressions. Shoot. /7
