I think there's some chance I overstated Clinton's White support in a few places. Having spent a lot more time looking at precinct level results, I think the Latino Decisions / AALDEF estimates of Trump's share of the non-White vote are closer to reality than the exit polls (It's probably also the case that past Republicans saw their non-White share of the vote inflated by exit polls).
Accordingly, if I probably over-estimated Trump's non-White support, I by definition, probably under-estimated his White support (and thus overstated Clinton's).
Granted, I don't think this will move the needle too much in the South, since I don't think my African-American support numbers need any revising based on what I've seen. Using the more accurate turnout numbers recently provided by the Census could impact things somewhat, but the direction likely varies by state.
In any case, I'm working on a version 2.0 of the model, and I hope to have it finished before I head back to university at the end of August.
I'd be very interested in seeing what you would generate for 2012 at the county level with your model.
Another interesting thing I just noticed from the discussion above: I had Oktibbeha whites as 23% Obama in 2012; your 2016 model shows 27% Clinton - a deviation from the general rule of thumb of difference between our two models. I remember that when I attempted to generate a "swing" map between our two models, there were a lot of counties in MS that swung to Clinton compared to my Obama 2012 model, which I thought was interesting. I remember having some questions about my initial model in MS when I was working on the project and believe I posted about it somewhere, but I honestly can't recall what exactly.
As far as population standards used for my model: since it's been two years and fifty projects ago, I don't exactly recall all of the parameters of my model. It was heavily adjusted as I progressed and also adjusted for individual states/areas, to the point that it was anything but one-size-fits-all.
My model initially began with VAP as the baseline but I made a series of custom alterations for many states. Especially in the SW, VAP is useless. In states and counties where non-black, non-white populations were non-negligible, I know that I switched to CVAP for the baseline and worked from there. As far as general demographic composition of the electorates, I'm quite confident in my model in the vast majority of counties in the nation - the only areas where I feel results outside the margin of error may exist in any sizable number of circumstances are places such as the Rio Grande Valley/TX and CA.
