- Good for Trump
- Trump +3 in the LA Times tracking poll
- Clinton (only) +3 in ME
- Clinton (only) +4 in the NBC News/SM national poll. This is a big national poll that typically causes jumps in the model in one direction or another, and she's been holding at +6-7 in this poll for a while. This one probably moved the needle.
- Good for Clinton
- Clinton +8 in VA
- Trump +12 in KS
- Trump +6 in TX
- Clinton +25 in CA
- None of these states will be decisive in the electoral college
One key way my model differs from 538 is in the choice of poll used. Not a choice based on which pollsters or which states (we both include them all) but on how many candidates the poll asked about. Specifically, when given the choice I opt for head to head polls, and 538 opts for 4 way (Clinton/Trump/Johnson/Stein) or 3 way (Clinton/Trump/Johnson) polls when available. 538 then subtracts totals from third (and fourth) party vote share to reflect how they think votes will move from those 3rd/4th party candidates by election day. I find this needlessly complex.
Voters who say they support 3rd party candidates early in the cycle tend to coalesce around either the D or R candidate as election day approaches (although for Gary Johnson that hasn't happened yet). This is a natural and expected process. Were I so inclined, I might indulge in the luxury of supporting whomever I Truly Like, but then as election day approaches I decide I can't abide Trump, and I vote for Clinton.
Of course, there will still be some people who vote for Johnson/Stein/Misc on Election Day. I've made the modeling decision (actuarial judgement) that those people would be reasonably approximated as the people who respond 'neither' when presented with a head to head poll. For one, people who answer "none of those" when only given two options are going to be the diehard 3rd party voters who likely will persist until election day. And for another, this is a functionally a head to head election and should be thought of as such. I'm sure Johnson and Stein voters are bristling, but the simple reality is that the winner of this election will be Clinton or Trump and it's silly to pretend otherwise.
A good model should reflect a logical construction of what's being modeled. This is a head to head race, we should use head to head polling.