Well if you’re going to get scooped on a story, at least let it be by somebody you read and respect, with better data.

Jonathan Rothwell, formerly at Brookings and now a senior economist at Gallup, has published a working paper, Explaining Nationalist Political Views: The Case of Donald Trump, which sorts through many of the questions I raised in my first piece about the 2016 Election. The Washington Post published a summary of Rothwell’s findings, but I recommend reading the paper itself, as qualifiers usually don’t come through clearly enough in mainstream media coverage of academic research.

Rothwell used Gallup Daily Tracking survey microdata from approximately 87,000 interviews conducted between July 2015 and July 2016, in which American adults were asked how favorably they viewed Trump, as well as a series of identifiers, such as political views and party affiliation, race/ethnicity, educational attainment, occupation, and more. Rothwell then cleverly linked the responses to sub-county level geographies to compare with Raj Chetty’s economic mobility data.

Rothwell’s use of Gallup survey data gets around many of the primary vote data limitations I discussed in my piece. For example, Rothwell was able to directly examine the relationship between a respondent’s view of Trump and his or her socioeconomic and demographic characteristics, rather than having to infer connections, as I did, based on what a county, as a whole, looks like and how residents of that county voted, in the aggregate. Rothwell’s analysis provides a degree of precision that is not possible using county-level vote totals and Census data. Further, Gallup data made it possible for Rothwell to apply weights to the sample to make it nationally representative.

There are many interesting aspects of Rothwell’s analysis, but the key passage is on p. 11 of the paper:

“These results do not present a clear picture between social and economic hardship and support for Trump. The standard economic measures of income and employment status show that, if anything, more affluent Americans favor Trump, even among white non-Hispanics. Surprisingly, there appears to be no link whatsoever between exposure to trade competition and support for nationalist policies in America, as embodied by the Trump campaign.”

The reference to trade competition is in response to a popular argument that Trump’s support is fueled by workers–specifically, white, male workers–in economically distressed communities tied to long-term stagnation or decline in manufacturing and other “blue-collar” employment, a perceived impact of trade liberalization. Rothwell’s analysis found no such evidence, when controlling for demographic characteristics, party affiliation, etc.

And later (p. 12):

“. . . this analysis provides clear evidence that those who view Trump favorably are disproportionately living in racially and culturally isolated zip codes and commuting zones. Holding other factors constant, support for Trump is highly elevated in areas with few college graduates, far from the Mexican border, and in neighborhoods that standout within the commuting zone for being white, segregated enclaves, with little exposure to Blacks, Asians, and Hispanics.”

There’s plenty to debate in the paper–check out twitter for scholarly disagreement about model specification, robustness, and multicollinearity–but I don’t find much that’s debatable in Rothwell’s conclusions, based on my analysis of the county-level presidential primary returns, flawed as they may be as a data set. In fact:

  • White Alone, Not Hispanic or Latino share of total population and educational attainment among White Alone, Not Hispanic or Latino males age 25 or older are statistically significant predictors of Trump’s share of the total primary vote at the county level, consistent with Rothwell’s findings.

Here is the urban-rural split of Trump’s share of the total primary vote compared to Clinton and Sanders for counties included in my data set where White Alone, Not Hispanic or Latino residents make up 50% or more of total population and 50% or more of White Alone, Not Hispanic or Latino males age 25+ have no college:


And, to Rothwell’s point about contact theory (pgs. 8-9, 12), here is the same table as above, but this time showing counties where White Alone, Not Hispanic or Latino residents make up 90% or more of total population and 65% (mean + 1 SD among majority white counties) or more of White Alone, Not Hispanic or Latino males age 25+ have no college:


Only 200 counties, but go back and compare these tables to the one in my last piece and you’ll find that my analysis is generally consistent with Rothwell’s findings.

  • Using data from EMSI, which provides estimates of total employment by industry for small counties where QCEW data is suppressed, I can find no statistically significant relationship between Trump’s share of the total primary vote and any measure of manufacturing employment I could think of testing, including the industry’s current and past shares of total employment in the county, change in industry employment over various time periods, or change in number of male workers in the labor force relative to manufacturing jobs available.

Again, generally consistent with Rothwell’s findings.

There are some interesting exceptions and regional differences, especially when comparing Trump to the other two main candidates. I’ll get into that next time.

The Economic Geography of the 2016 Election

We like to take the occasional detour into politics, especially when there’s an economic development story to be told. In the past we’ve looked at Mitt Romney’s infamous makers and takers argument, Rick Perry’s Texas Miracle, and creative class support for Obama. Successful politicians, our storytellers-in-chief, are particularly adept at turning immensely complicated issues into palatable soundbites because complexity and uncertainty have a way of stoking demand for easy answers–the convenient truths–that shape the narrative, as they say.

Sloganizing complex forces shaping voter perceptions of the economy and their place in it is hardly a new political tactic, but for those of us who pay attention to rural economic development and labor market issues, 2016 seems to be a marked departure from the usual talking points. The 2004 election made “offshoring” of jobs somewhat of a national issue, but I don’t recall much of an urban vs. rural flavor to the debate. The red state/blue state narrative, articulated beautifully by local journalist Bill Bishop in The Big Sort, certainly included economics, but was really more of a statement about social and cultural factors–an epilogue, of sorts, to the culture wars of the 1980s and 1990s. Indeed, we’d still be talking about how Tim Russert broke the Internet on election night if Twitter were around in 2000.

Politicians and pundits love to characterize every presidential election as a critical inflection point, the proverbial crossroads that represents some deliberate attempt on the part of the electorate to choose a distinct path forward. Most elections probably don’t live up to that billing with the benefit of hindsight. We’ll see what the historians say about 2016. But the convenient truths of the 2016 presidential election are not hard to spot, and our ability to sort out what’s real from what’s being used as political cannon fodder may help determine economic and labor market policies affecting Rural America in the next administration.

I’ve been combing through results of the presidential primaries, and, through that process, developed a new appreciation for people who do that for a living. Speaking of, I need to thank Joshua Darr, assistant professor of political communication at LSU, for the pointer to Ben Hamner’s election data warehouse on Kaggle. Most of the major news outlets host interactive maps of election returns, but for obvious reasons don’t make it easy to assemble your own data sets to work with.

There are several caveats to keep in mind when working with primary data. First, primaries are not reliable predictors of turnout for the general election. Turnout for the primaries this year was high by historical standards, approaching the record participation in 2008, but turnout for the general in November could throw off conclusions drawn from primary results in any number of ways.

In addition, the way in which primary elections are held and results are reported in several states make it challenging to assemble a complete and accurate national data set. The majority of states report primary results at the county level, making it easy to match returns to other data sources, such as the American Community Survey. However, some states report results using different sub-state geographies that do not line up nicely with counties, such as congressional districts or townships, or, in a few cases, not at all (e.g., Republicans in Colorado).

So, with apologies to AK, CO, CT, KS, ME, MA, MN, ND, RI, VT, WY, as well as the District of Columbia, I’m going to use a data base of primary results in 2,720 counties, matched to various data from the 2014 ACS 5-Year Estimates, to make a few observations about the 2016 presidential election over the course of the next few weeks. Given the states omitted from the analysis, combined with the inherent shortcomings of primary data, the purists among you will no doubt be left wholly unsatisfied. But we work with what we have. Finally, two pieces of background reading I want to mention because they motivated me to take this on:

Nate Cohn in the NYT: Donald Trump’s Red-State Problem

Alan Greenblatt in GoverningCan Counties Fix Rural America’s Endless Recession?

The overarching narratives I want to explore deal with race/ethnicity and education characteristics of voters supporting the two nominees, Hillary Clinton and Donald Trump, although Bernie Sanders will make a few appearances, as well. In particular, I’m interested in this idea that non-college, working-class whites, reacting, in part, to declining economic prospects, are breaking for Trump in a significant way, and whether or not there are differences between urban and rural counties.

But before we can get to any of that, we need a basic understanding of how people voted in urban and rural counties. The USDA Economic Research Service classifies counties on an urban-rural continuum. The first three rows in the table below are counties located in metropolitan statistical areas (MSAs). Rows four through nine are non-MSA counties, with rows eight and nine representing counties USDA ERS considers “completely rural.” Here’s a breakdown of the popular vote for Clinton, Trump, and Sanders, on that urban-rural continuum:


Approximately 60 million votes were cast in the primaries, which means we have about 87% represented in our data set. Trump received 45% of the Republican primary vote and 23% of the total popular vote. Clinton got 55% of the Democratic primary vote and 28% of the total popular vote. So the county data set we’re using here with 11 states and DC missing is tilted by about one percentage point in favor of Trump, and it short changes Sanders by about two percentage points (he won 22% of the total popular vote). Keep that in mind as we continue this thread, but it shouldn’t have too much bearing on the themes we’ll discuss.

As always, feel free to fire away with questions or comments and I’ll try to work them in. Thanks for reading. More soon.