There were a few tweaks made for this model run. First and foremost, there was a bug! Some of the weighted calculations weren’t applying correctly, meaning lower quality polls were having more influence than they should have. That has been fixed.
There has been a change made to the Markov chains in the election nigh simulation as well. There was a somewhat static boost to a candidate’s probability of winning states they weren’t expected to in the previous version. Let’s say a random state initializes in the hypothetical election night. If Harris won that state, and she wasn’t expected to based upon the state-level probability, then she would get a boost influenced by the state’s PVI times a certain value. That would lead to a posterior probability update for the remaining states, each time sequentially as a new state runs, and all states would be affected in the same way.
The new version is based on geographic clusters and will most heavily affect nearby states with similar attributes. This is a somewhat qualitative and subjective process that could include more empirical foundations in future updates to the model.
Thanks for reading Partisan-Gravity!