done!

should be able to endorse in applied stats if you want

Prop 50 goes into effect regardless of this TX decision, if that's what you mean

But to be clear for those who don't click through, the model is:

logit(y_it) = alpha_i + delta_t + eps_it

alpha_i is ~PVI, delta_t and eps_it are drawn from distributions calibrated to past House elections

Goal is to be just a little bit "smoother" than basic uniform swing, but still be very transparent + give people a way to edit in their own assumptions/handicaps/etc

Just an approximate probabilistic model for win prob. District baselines are averaged presidential margins, but parameters (size of district vs national swings) come from a model fit to past house elections.

Details in one of the tabs: docs.google.com/spreadsheets...

Prob. the baseline being used. 'Baseline' column is Cook PVI as of July, and adjustments are based on DRA's calculations on something approximating that PVI. So the size of the shift in CD1 is smaller than that.

If I add R+4 to CD1, so R+10 overall, and make CD3 R+10, net chg is still <R+0.1 seat

Net change in avg. Dem seats by state:

Yeah. I mean you need to use good ML methods in the DML but in many ways a few highly predictive confounders is an easier setting

DML should avoid shrinking in a way that biases the causal estimates - that's in many ways it's raison d'être.

Agree data-driven covariate selection is (very) risky, but that's a separate q from DML's goal of safely and efficiently controlling for (the right set of) many covariates