2016 White and Non-White Vote by County Project

[Adam Griffin]

Probably, but I'm not sure when I'll begin. I spent more time on this one than I care to admit.

[snip]

Very cool! I think links are able to be posted at either 12 or 15 posts, for the record. You can also just put a space in the URL to get around that.

One question regarding the trend you also saw; white voters being more Democratic in heavily minority counties. Did you observe this trend/limit it to rural counties as well? It all but makes sense for this to occur in urban areas given what we know about white urban voters, but I observed this as well in many Deep South rural counties. I did some very basic regression analysis (I'm not nearly as statistics-savvy as some think) with Georgia's counties and there did appear to be a trend.

What initially caught my eye was Georgia: since the Secretary of State publishes both voter registration and voter turnout by race for each county/precinct, I was able to definitively see a trend in select counties where the white/black combined electorate exceeded 95%. The broader trend also appeared throughout the Black Belt to some degree, but I did not have such explicitly accurate data to work with elsewhere.

There weren't many other opportunities outside of the South (other than the Southwest, and again, the Latino issue was a huge one that consistently messed with my numbers beyond the point of confidence; ditto for Native American counties) to really look at it and see if it was broadly associated with larger non-white populations or just black populations.


As far as converting your data into a map how I did: there are a few tutorials on the web for doing such with Fusion Tables, but the gist of it is:

You'll need to first upload a shapefile (can be in CSV or KML format) containing all of the geography/data points for each county to be mapped. You can do this by going to Google Drive, then clicking "New", going to "More" and then selecting "Google Fusion Table". You should be able to download the exact shapefile I used from here. If that for some reason doesn't work, though, then there are plenty of county map shapefiles available; you may have to convert others to KML or CSV for it to work with Fusion Tables, though.

Then, create a CSV spreadsheet with FIPS county codes (as well as the county names ideally in a separate column for easy reference). The FIPS codes you'll need to use are the five digit ones, with state (first 2 digits) and county (last 3 digits) FIPS codes merged together (01001 = Autauga, AL; 06107 = Tulare, CA; 42133 = York, PA; etc). In each county's row, input your data. Save the spreadsheet and upload the 'table' to Fusion Tables. It's worth noting that Fusion Tables does not like explicit percentages and may not work with them; this is why all of my results are formatted in decimals (i.e. "56% = 0.56"), so you'll want to do the same.

With your CSV spreadsheet open, go to "File - Merge", and select your shapefile map as the file to merge with your CSV sheet. You'll need to match the columns from each sheet via the menu there; in my shapefile map linked above, the correct column is entitled "name". For your CSV sheet, it'll be the name of the column that has your FIPS codes in it (BTW: you need to specify a name for each column in your CSV sheet in its first row).

If all has worked according to plan, then your data should be merged successfully. From there, you'll go to the "geometry" tab at the top, click on "change feature styles" and navigate to the "fill color" option under 'polygons'. You can use a gradient (like what I did in my first map) or a series of buckets (what I used to show the straightforward blue, purple and red in the second map) by configuring the exact numerical cut-offs or ranges for each, and telling it which column (the one with your results) to use.

[Adam Griffin]

^^^ That's very cool! You should really make your own thread here with the map since it's a separate project. That way, it'll get the attention it deserves.

[Adam Griffin]

The margins were pretty close in all of these, and I suspect that Clinton won the White vote in the city proper for all of these cases, but the other suburbs in the county put Trump over the edge.

Per this model, Trump won the White Vote in Marion County, IN by 3.2%, Wayne County, MI by 0.4%, and Fulton County, GA by 2.5%.

I'm curious as to what number you have for Clayton County, GA: looks like it's in the 30s on your map. I had 50% in 2012. I'm thinking we both may be a bit off of what was the case in each of these elections. The problem with counties like Clayton (in general; at least with my map) is that the white share of the vote/population is so small that even small changes in non-white vote share can dramatically affect the white Democratic percentage. The great thing about GA specifically - and I think I mentioned it before - is the Georgia's voter turnout by race and by county stats from the SoS, if you haven't already checked it out and/or used it in your figures.

On the other hand, if there was to be a large swing among whites to Trump in any metro ATL county, then Clayton would probably be the most likely culprit.

[Adam Griffin]

^^^ FYI, the white vote in ATL proper is immensely Democratic: we're talking anywhere from 70-80% Democratic. It's the non-ATL Fulton whites that drag the number down. Only about one-third of whites in Fulton County live in the city limits.

There's a huge gulf in the voting patterns of whites in the city limits versus remainder of the county. You're basically looking at non-ATL Fulton whites as a whole being comparable to the surrounding metro (25-30% D); when combined with ATL proper whites, it gives you an approximately 40-45% D white population.

[Adam Griffin]

I still haven't figured out how to make this into an embedded map (on site or externally), so I'll just link to the Fusion Table in the interim. I'll change the color scheme to Atlas colors once I figure out how how to embed it somewhere.

https://fusiontables.google.com/DataSource?docid=1UVNwaKv9xnAwuLv6Z7tFWxV5pu19qB6CIVGtIUsH#map:id=3

It won't embed here (because this forum is a thousand years old), but:

[Adam Griffin]

I still haven't figured out how to make this into an embedded map (on site or externally), so I'll just link to the Fusion Table in the interim. I'll change the color scheme to Atlas colors once I figure out how how to embed it somewhere.

https://fusiontables.google.com/DataSource?docid=1UVNwaKv9xnAwuLv6Z7tFWxV5pu19qB6CIVGtIUsH#map:id=3

It won't embed here (because this forum is a thousand years old), but:

I'm aware of that feature, but I haven't managed to successfully embed it on any of the other discussion forums I'm on. I've also tried looking for a free website host, but nearly all of those make HTML embedding something you have to pay to enable. How did you acquire the page to host your 2012 map, if I may ask?

FT is basically a proto-GIS solution that to my knowledge Google isn't expanding or servicing anymore. It had a lot of potential but like with many Google services that start out, they simply gave up. It lacks a lot of features, including any reasonable embed options for most forum BB code.

I own the domain/hosting for that website, so I just uploaded a simple HTML page with the embedded code. I can do the same for you if you'd like.

[Adam Griffin]

Thanks! That's a really gracious thing for you to do. Would you like me to invert the colors on the map before it's uploaded to a new page? Also, I have the statewide results calculated, so I could probably make some images of maps to show that as well.

No problem - it's live now. I have your forum profile attributed at the bottom but if you'd like anything changed or added, just let me know.

Either/or on the color scheme. I clarified on your page that it's the traditional color scheme for now, but if you change it, I'll be happy to update descriptions. I'm probably going to cross link our two but presumably in either case, people will be able to read descriptions and not get too confused.

A couple of recommendations, especially when it's in a format like this. Whenever you are "editing" the FT map (i.e. when you click on the layer's arrow and go to "Change Map"), the last positioning of the map before you click "Done" determines the default positioning of the map in the embed. If you notice, the white/non-white maps are at different zoom levels and positions; this is why. On mine, I did my best to line up each map in the same position - for aesthetic reasons, mainly, and to also position the default view where I thought most people would be interested in seeing it. Sadly, I can't remember whether the embed code needs to be re-copied each time when you reposition the map or not - I'm leaning toward no, but if you were to change the positions/zooms of each map and it doesn't reflect there, just let me know.

Another thing you might have noticed is the legend stretching. FT will make the legend width based on the title of the legend, and currently, it's taking up a pretty big percentage of the embed's width. You might want to shorten it to something like "Trump Margin (%)" or "Trump Margin PctPt" or something like that.

[Adam Griffin]

Since I know there was a similar map made for the 2012 election, is there any chance we could get a swing map between the two? I imagine it would track pretty much with affluent and educated communities generally swinging to Clinton and poorer and less-educated communities swinging to Trump, but I'd be curious to know if there are any exceptions.

It's doable in a sense, but as reagente pointed out, we used different ways of generating our maps. Because 2012 was simple enough in terms of a two-way model, I just focused on calculating Obama's share of the white vote directly (due to the fact that in most places, Romney's share would be ([Obama's share] - 1pt). 2016 was more complex and the way reagente did his makes more sense. We also presumably used different data-sets/combinations of data to make our maps.

Nevertheless, I could pretty easily generate a map that shows the difference between Obama and Clinton's shares of the white vote. It should be pretty accurate in the vast majority of cases, but as I've already noticed, we were in disagreement in some select counties based on the margin difference in them.

[Adam Griffin]

^^^ Everything should be fixed/added now.

[Adam Griffin]

Any chance you can make the original sheet(s) you used available for access, where each county is available by row? I tried to download the set from the maps themselves but it formats the data into something that is not usable. I'm not sure if it's a sharing setting that has it hidden via the link or something else. Sometimes you can just link to it directly by using the link available when editing it (as opposed to the sharing link for embeds, etc). Usually the sheets in my projects have something like "Row 1", "Card 1", and (by default) "map of geometry". That first one is the one to which I'm referring. 

[Adam Griffin]

This is really great (though also depressing)! Thanks Reagente and Adam.

I know you've discussed it a bit already, but would you mind explaining your respective methodologies and how they differ to a beginner?

I could be wrong, but beyond questions of where we got our data for the electorate composition, I think the primary difference is how we tabulate the Non-White Vote (which has resultant implications for the White Vote).

I built in a elasticity to the non-White vote into my model, operating under the assumption that non-Whites in more Republican areas will be more inclined to support the GOP. In counties with large minority populations, this has the potential to impact the White vote quite a bit.

On the whole, I think my model for the most part would lead to reduced Trump margins among White Voters in very Republican areas, and correspondingly increased Trump margins among White voters in very Democratic areas when compared to Fmr. Pres. Griffin's model

Yes, this may be the primary difference. My initial effort used a pretty uniform formula for non-white voters (not completely uniform) in my first pass for each state. I think reagente's formula may shift from county to county, depending; mine was initially customized at the state level as far as turnout rates and persuasion go. For instance, black voters could vary between 85-95% D depending on the state and what exit polling showed in 2008/2012. The same adjustments were also made with regard to other non-white groups, but it was not customized initially on a county-by-county basis.

I then proceeded on several revision binges in a handful of states where I could tell that the numbers (usually among Latinos) from such a uniform approach were off by quite a bit in various areas. This was where I began tweaking with various counties, but even then, it was only in a few states. Texas was a nightmare, and even when I finally "finished", I was not very happy with the output in both North Texas and the Rio Grande Valley; I see reagente's model also shows unusually high Democratic white support in the Valley as well, which I concluded - while likely a real phenomenon based on a broader occurrence of whites being more Democratic in very heavily minority areas (rural or urban) - could still be skewed.

So the baseline was built off of Census data, voter registration statistics by race (where available), exit polling data and something else that now slips my mind in order to determine both turnout and persuasion. Certain elements (such as the percentage of a county's Latino population that likely voted, affected by citizenship and age) were customized based on county or region. Outliers were adjusted based on what made sense with regard to the path of least resistance concerning all possible adjustments in turnout/persuasion across groups.

[Adam Griffin]

I built in a elasticity to the non-White vote into my model, operating under the assumption that non-Whites in more Republican areas will be more inclined to support the GOP.

Do we have any empirical evidence for this? It seems reasonable to make this assumption for Asians and Hispanics, but my guess would be that Blacks vote pretty uniformly everywhere in the country.

This is where I'm having some questions in regard to the model.

First, I'm not entirely convinced that (non-Cuban) Latino voting behavior is phenomenally different from place to place based on the two-party vote of a given area. With the exception of parts of Texas, New Mexico and areas where extensive histories of Mexican settlement alongside Anglos have literally melded the communities together and diffused identities, I'm not of the belief that Latinos adopt native voting behavior in short order.

One element that makes this so difficult is that many heavily-GOP areas with large Latino populations tend to be relatively new phenomena and have very high rates of undocumented individuals; when combined with widely-observed Latino voter apathy among even citizen adults and a disproportionate share of citizens being under 40, it means that you need a very substantial Latino population to even constitute a double-digit share of the electorate (a 30% Latino area that has roots in the area for around a generation may only constitute 10% of voters). Outside of Texas, most of your heavily-GOP, heavily-Latino areas are relatively newfound occurrences. Small shares of the population = more liable to larger discrepancies/margins of error in projections.

Based on a variety of data-points available regarding the 2012 election that allowed me to compare my work in various areas to others' findings and actual voter demographics, I came to the conclusion that my model tended to under-represent Obama's white support by a couple of points more often than it over-represented it. This is a reality when building any model that tries to fit 3,100+ entities into a single set of criteria, even with customizations; unless you have the time or energy to grind away and calculate specific formulas for each individual county, you're going to have some outliers. Usually, these outliers are the result of one or more broader assumptions we make that either has more variance than we thought or that under/over-estimates what we consider to be reasonable/accurate.

I do believe reagente's model is being affected by the same dynamic, albeit in a different area. It's grueling work to eliminate all of these discrepancies and arguably impossible. While not necessarily true, I would assume that he constructed his model beginning with one of two basic/broader assumptions (as in, either he made more assumptions about the white vote in order to fill in the blanks on how non-whites voted, or vice-versa). Knowing which (if either) broader assumption made would further assist in seeing if what I'm thinking might be happening is correct.

In short, I think his model is more likely to over-represent non-white Trump support. I think this becomes more evident in Appalachia in particular. In fact, I believe this occurrence in his model is the same dynamic I faced in the Rio Grande Valley: you have one race that is damn near 100% of the population in many areas, so even slight discrepancies or shifts in both turnout and persuasion among that group (which for said group is minuscule and almost meaningless statistically) can cause very large and wild shifts among the minority group's numbers there. In some cases, the actual minority voting bloc could be viewed as a rounding error in and of itself. If there is a consistency to the assumption that results in a over- or under-representation, then it becomes very visible across a larger area. It reminds me of how I could shift Latino support in some of those Rio Grande counties by a few points...and generate a 50 point swing for Obama's white support. It drove me crazy trying to work it all out.

I might write an explanation with regard to my county specifically (which is in Appalachia and has a large Latino population) and how the numbers don't necessarily comport with the shifts in voter turnout by race and the margins between 2012-2016 I've observed in the county's voter file and data, but I'll do that later since this is already so long.

[Adam Griffin]

Any chance you can make the original sheet(s) you used available for access, where each county is available by row? I tried to download the set from the maps themselves but it formats the data into something that is not usable. I'm not sure if it's a sharing setting that has it hidden via the link or something else. Sometimes you can just link to it directly by using the link available when editing it (as opposed to the sharing link for embeds, etc). Usually the sheets in my projects have something like "Row 1", "Card 1", and (by default) "map of geometry". That first one is the one to which I'm referring. 

Is this what you are looking for?

Yes, I believe so - thank you!

[Adam Griffin]

Data from the Congressional Cooperative Election Study makes me somewhat more confident in my model's assessment of the non-White Vote. For example, there was a clear rural-urban divide among the Hispanic vote:

The data coming out from this report has very good insight into a lot of things.

By any chance, do they specify how they determine the four categories of urban/rural, and do they provide crosstabs or percentages for the percentage of Latinos that live in each?

I do still want to come back and talk about my county specifically in a bit relative to this model when I get some time this evening, but one thing I'd point out about it now. I'm pretty sure that by any definition, it would fall into either the "Less Rural" or "More Rural" categories, and here's the kicker: 70% of Latino voters in my county were 18-44 years of age.  

I am betting that in a lot of counties where you have less than marginal Latino populations in heavily GOP areas - even in the more rural parts - the 45+ voting patterns aren't going to be nearly as relevant as one might think, meaning your Latino support would skew much more toward the younger preferences than the older ones. Such is the reality in communities where large majorities of your 45+ populations are non-citizens and nearly all of your under 25s are citizens; it's the predominant reason why Latino voter participation lags even Latino CVAP. They're not necessarily less engaged in a fundamental sense, but rather, a much greater share of their eligible voting bloc is under 40 and therefore behaves like any under sub-40 voting bloc with respect to turnout.

[Adam Griffin]

Also (quite a bit off-topic), I'm wondering if you used the same shapefile I sent you?

The reason I ask is because some time ago - seemingly randomly - my embed maps at the default zoom level suddenly began to have "holes" in it where counties should be, especially in and around Appalachia. If you zoom in, the counties are clearly visible but as you zoom out, there are missing chunks. I noticed yours is working perfectly fine, however. It may just be a glitch on my computer/browser's end that's causing the issues for me on mine, but otherwise, I'm wondering if you did/used something different.

[Adam Griffin]

Since I know there was a similar map made for the 2012 election, is there any chance we could get a swing map between the two? I imagine it would track pretty much with affluent and educated communities generally swinging to Clinton and poorer and less-educated communities swinging to Trump, but I'd be curious to know if there are any exceptions.

It's doable in a sense, but as reagente pointed out, we used different ways of generating our maps. Because 2012 was simple enough in terms of a two-way model, I just focused on calculating Obama's share of the white vote directly (due to the fact that in most places, Romney's share would be ([Obama's share] - 1pt). 2016 was more complex and the way reagente did his makes more sense. We also presumably used different data-sets/combinations of data to make our maps.

Nevertheless, I could pretty easily generate a map that shows the difference between Obama and Clinton's shares of the white vote. It should be pretty accurate in the vast majority of cases, but as I've already noticed, we were in disagreement in some select counties based on the margin difference in them.

So this is the closest thing we can get to that: a simple comparison of the white vote share for Obama and Clinton between the two models. Here's the interactive map. Several of the areas that seem out-of-place I know are the result of the models' differences (for instance, my model is probably the one that is off in Clayton, GA & Imperial, CA).

Red counties/negative numbers indicate Clinton improved over Obama among whites; blue counties/positive numbers indicate Clinton did more poorly than Obama among whites. Numbers are shown as percentage point differences (i.e. Obama got 46% of whites and Clinton got 32% = "14").

[Adam Griffin]

I'm archiving all of my FT data as we speak and will download it once it's ready. Admittedly, it may take some time to sort through given I had hundreds of FT data sets and I wasn't very good at organizing them.

With regard to the 2016 white vote numbers specifically, I believe reagente uploaded his data to his own FT account (however, I hosted the page on which it was displayed), so he will need to double-check to make sure this data is secured. FT data will remain downloadable until March 3.