Yair just published a long post explaining (a) how he and his colleagues use Mister P and the voter file to get fine-grained geographic and demographic estimates of voter turnout and vote preference, and (b) why this makes a difference.
The relevant research paper is here.
As Yair says in his above-linked post, he and others are now set up to report adjusted pre-election poll data on election night or shortly after, as a replacement for exit polls, which are so flawed.
I thought all the effort that goes into predicting election outcomes very shortly before they happen was sort of a waste, but analyzing breakdowns of election results as soon as possible after they’ve happened, well there’s a thing you can do
Z:
Here’s an example of what we did after 2016. There are lots of results there, but we didn’t put in the effort to structure them as we did with Red State Blue State so it’s not really clear what to do with it all.
I would be curious to see those much higher numbers than I typically see for non-college white voters in 2016 would compare to those numbers for 2012.
I have long wondered how many of the supposed Obama —> Trump voters as aggregated in geographic locations are actually explained by a combination of previous Obama voters not voting in 2016, and people who didn’t previously vote much coming out for Trump (in those areas). From a bird’s eye view (without specific data about actual votes cast by specific people in both elections, data which I’ve read are considered rather inreliable due to recall bias, as hard as that might seem to believe), it seems to me the two phenomena would look the same.
Hi Joshua. The blog post links to a spreadsheet showing our numbers, compared to these other sources, going back to 2008:
https://docs.google.com/spreadsheets/d/1Lr1z1UDy9-sCn1rPMmV8H5KTjcFHTjEujoBw1oMPyDo
About white non-college (WNC) voters: according to all the sources, they slightly shrunk as a percent of the electorate from 2012 to 2016, but they voted more for Trump than they did for Romney. Our data says the shift was about 8 points in margin, some of the other sources are slightly higher.
Decomposing how much of this is from turnout versus vote shift is extremely difficult, because the people who chose to vote within the WNC group was different, and those differences move together with their candidate preferences. That’s true within smaller subcells too, however you want to cut it. You can try to model it out but it’s hard and I’ve never been very satisfied with the results, except to say “it’s usually always both.” Panel methods, i.e., looking at the same survey respondents over time, kind of help, but people who answer surveys multiple times aren’t representative.
Thanks!