Thread by @lxeagle17

🚨NEW GEORGIA RUNOFF MODEL🚨

In the precinct model made by me and @ADincgor, we've now accounted for the individual-level shift observed by NYT/Siena, thanks to precinct race and party adjustments made by @joe__gantt.

Forecast (Jan. 2): D +4.25

docs.google.com/spreadsheets/d…

Okay, now, let's go into some of the methodology behind the race/precinct adjustment for in-person EV, made after a conversation with @Nate_Cohn. It's not random at all, and it is described in great detail in the spreadsheet, so I encourage you to take a read if you're curious.

Our original precinct projection weighted everything by geographic splits only, which captured a lot of the picture, but did not capture the reality behind individual voter attitude shifts, which did not show up in precinct or county level data. This fixes that.

The upshot is that Democrats benefit from more diverse precincts because of the increased Black vote share. Moreover, they benefit from a significant increase in the fact that the Dem primary voters in the January runoff are more likely to return their ballots than in Nov.

The early vote is now at D +11, which actually lines up pretty well with Cohn's D +13 estimate. We maintain that election day turnout should be proportionate to the November ratios and are projecting the same splits on election day as we saw in November, though this is noisy...

The forecasted margin, modeled using Biden vote share margin, is D +4.25.

The Ossoff/Biden averaged forecast is D +3.31, and the Ossoff forecast is D +2.36.

I'll be reporting the Biden margin in my tweets, because I believe it's a better reflection of the electorate party lean

I hear your arguments about why we should use the Ossoff margins; I just personally disagree. My goal isn't to be conservative or bullish, because although I, as a Democrat, want to steel myself against a loss (and it's *not* safe D), I want to give the most accurate picture.

But if you want to see the Ossoff/Biden averaged forecast or the Ossoff forecast, please take a look at the spreadsheet. It will be updated as regularly as the others. I am personally much more confident in our data here after these adjustments, and will be reporting using this.

The original models, which only projects the November electorate by county/by precinct, are still there in the spreadsheet under different tabs and are clearly labeled. They will be updated regularly as well.

An immense amount of thanks to @ADincgor, the co-creator of these models, to @joe__gantt for the precinct adjustment, and to @Nate_Cohn for talking through our model's shortcomings with us and explaining why such adjustments were necessary to better reflect the electorate.

Some shortcomings with this model include the fact that it doesn't use individual data, so we may still risk missing individual level shifts. If NPAs break disproportionately R, then this would miss that, and that is a possible danger.

And we make a lot of assumptions about eday

But I don't really know how best to account for the former without individual level data that we just do not have, and I definitely can't account for the latter because we just do not know how election day looks, so the safest assumption is to assume proportionate turnout.