Thread by @lxeagle17

@Nate_Cohn is right -- the early electorate is probably closer to D+13, as his tweet says, than D+6 (our model).

Here's the question: What's our model trying to do?

Are we trying to forecast the electorate? Or are we projecting the November electorate onto the current one?

If we're forecasting, we're going to be way off. We don't have the data for it, and I don't have the time to pay the $250 for the voter file and go through it and build a model by Tuesday.

If we're going to project the November electorate onto the current one, then some questions need to be asked.

As @Nate_Cohn so kindly explained to me, there is an individual level shift that doesn't show up in the precinct data, but let's think of it like this, in his words...

"Take two precincts, one 75-25 D and one 75/25 R in the general. Imagine now that Ds turn out at 50%, Rs 40%. Truth is D +11; your model would find D +6".

So clearly, even projecting the November electorate with precinct data will end up being a touch conservative on Democrats.

One way to account for this is to use Biden margins instead of Ossoff/Biden margins. Why? Because Ossoff lost a lot of votes due to undervotes, per @GarrettHerrin, which is not going to happen this time around. So we need a proxy to understand how Democratic the electorate is.

Biden voters being more motivated thus far than Trump voters shows up in individual-level data that does not show up at a precinct level.

So perhaps it's better to *also* report Biden margins along with our current Biden/Ossoff split, and we will be doing that.

The second question is what happens with margins in the model. My take is that nothing we do can properly account for this fully. We could, of course, add in race details at precinct levels as a proxy and go from there. That *could work*, but I don't know if I have time...

And even this, for example, really misses some of the movement within race and the turnout variation there.

I'll see if I can get it to work, but I don't know if we can account for this fully...

But one thing we'll be doing in the precinct model is adding a tab for what Biden margins would be, because that at least accounts for some of the party turnout enthusiasm gap that we see right now visible in the single-voter files and accounts for the elimination of "undervotes"

Beyond that, I can't do much more on this given the time and constraints we have.

The model will be up, and I'll continue to work on and update it. I'll see if I can get the race data to work, though IDK if it will. This is still work I do in my spare time, of course 😂

But we'll have to accept a few things on it:

1) It will miss out on some individual movement with data, which will end up being conservative on Democrats (fine by me!).
2) It's based off public and free data.
3) It provides a rough picture of November projected onto January.

Our model will likely go into election day predicting D+1.5 or something. This is based on some assumptions from November that may not hold, of course, but whatever -- it's the best we can do and it gives you a rough picture.

As always, use the model as a guide. Nothing more

Thanks for sticking with us through it all, y'all. Four more days...

And, of course, a huge thanks goes to the model’s co-creator, @ADincgor. And to Nate Cohn, @joe__gantt, @dooraven, @ForecasterEnten, and @thunderousprof for talking through the approaches.