My 2019 paper, “The Density Divide,” makes a case for the importance of selective migration as an explanation for the striking relationship between population density and party vote share. A good question to ask when you see a correlation between two variables is whether you’re looking at a “treatment effect” or a “selection effect.” When I began looking into explanations for this bizarrely tidy pattern, I wasn’t sure which was more likely. The first thing I did was try to find evidence for the treatment effect. I wanted to see if there’s reason to believe that living at higher population densities makes us more liberal and Democratic and that living at lower population densities makes us more conservative and Republican.
There’s plenty of evidence that disembedding from one community and re-embedding in another will tend to shift our opinions in the direction of the new community. If you know any humans, it’s pretty obvious that we’re conformist creatures worried about status and prone to adopt opinions in the vein of locally prevailing sentiment. But the harder I looked for evidence that moving to a city makes us more liberal or that moving to the country makes us more conservative, the weaker it seemed. Opting into a new community affects our opinions, for sure. But what explains why we opt out of one community and opt into another? As I dug deeper and deeper, it became clear to me that the treatment effects probably swamped the selection effects.
However, I didn’t arrive at this conclusion because there was clear, straightforward empirical evidence to this effect. It was more a matter of what C.S. Peirce called “abduction,” which amounts to hypothesis generation through well-informed hunches. I wanted to marshal as much evidence as I could for my hypothesis, but I’m not a quantitative methods guy, so I tried to be clear that my aim was to develop a theory and show that it’s worth taking seriously, but that I wasn’t claiming to have shown that it’s correct. Because I couldn’t! The right sort of studies simply weren’t available.
This is what I ended up saying about the treatment vs. selection effect issue in a short section titled “Selection Effects Don’t Explain Everything”:
Separating “selection effects” from “treatment effects,” untangling the web of reciprocal causal influence, and establishing the relative explanatory importance of all the variables at work is a forbidding analytical task, and we’ve only just begun to make serious headway on it. My claim is emphatically not that only sorting and self-selection matter. Rather, my claim is that the extent to which urbanization sorts populations has yet to be fully appreciated, and that many of its implications for political polarization have been overlooked or under-emphasized. My aim, as a theoretically inclined, cross-discipline synthesizer, is to connect enough unconnected dots to lend credibility to these claims and persuade specialists in empirical methods that it would be worth their time to drill down into the relevant data, formulate crisp, testable hypotheses, and see how well the picture I present here holds up.
Well … guess what?
In a new paper, “Urban–Rural Residential Mobility Associated With Political Party Affiliation: The U.S. National Longitudinal Surveys of Youth and Young Adults,” Markus Jokela, a psychologist and epidemiologist at the University of Helsinki, takes a good crack at it.
Selective residential mobility is one possible explanation for the development of urban–rural differences in political attitudes (Motyl, 2016; Wilkinson, 2018). People may actively seek, or passively drift, to neighborhoods that are populated by like-minded others; many observable neighborhood characteristics, such as the types of cars or number of churches, can provide cues about the residents’ political leanings that people notice when selecting neighborhoods (Gebru et al., 2017). Even if people did not specifically seek for neighbors with similar political views, individuals with different political attitudes tend to prefer different kinds of neighborhoods (Motyl et al., 2019), which may steer political segregation (Carlson & Gimpel, 2019). Personality traits have also been shown to cluster geographically (Rentfrow & Jokela, 2016; Rentfrow et al., 2015) and to predict selective residential mobility across urban and rural regions (Butikofer & Peri, 2017; Campbell, 2019; Jokela, 2020). Given that political attitudes are correlated with personality traits (Gerber et al., 2010), selective residential mobility associated with political party affiliation seems plausible.
… But selective mobility cannot be taken as the only explanation for the evolution of urban–rural differences. Living in an urban versus rural community might differently mold people’s political views in response to political views of their neighborhood peers (Johnston et al., 2004; Lazer et al., 2010) or other neighborhood characteristics that correlate with political attitudes. Changing political geographies could also be driven by third variables that influence migration or subsequent development of political attitudes (e.g., lifestyle factors, family structure) or by changes in the policies pursued by political parties that appeal differently to urban versus rural voters.
However, Jokela notes, there have been no studies using individual-level longitudinal data that track both migration patterns and party identification over time. So Jokela fixes that and then fits a model from the data to estimate the size of the effect on urban/rural party vote share difference. Here’s what he finds:
The current study provides direct longitudinal evidence on how political party affiliation of individuals predicts their subsequent residential mobility across urban and rural regions of the United States. Compared to those supporting Democrats, individuals supporting Republicans were less likely to move from rural areas to major cities and were more likely to move away from major cities. Simple simulation models based on the regression models indicated that, between ages 20 and 40, selective residential mobility would have widened the political divide by 6 percentage points in major cities (from 50% Democrats vs. 40% Republicans to 53% Democrats vs. 37% Republicans) and by 6 percentage points in rural areas (from 50% Democrats vs. 40% Republicans to 47% Democrats vs. 43% Republicans).
That’s the ball game in the passage I’ve emphasized in bold. If this weren’t true, the theory of the Density Divide would be falsified. But this amounts to solid, direct evidential support. The fact that Republicans are more likely to move away from major cities is a bonus, but isn’t strictly necessary, since there’s much more rural-to-city migration than city-to-rural migration. That’s what urbanization is. Over the longer-run, the really important thing is that Republican types are less likely to urbanize.
Because very few people move after the age of 40, “most of the correlation between urban–rural residence and political affiliation emerged between ages 20 and 40.” Jokela notes that the simulated correlation between urban-rural residence and party is slightly less than half of the observed correlation. That would suggest that treatment effects matter, too. But that doesn’t mean that selection effects account for less than half of the density divide. He writes:
However, it must be emphasized that the simulation was very simple and did not consider any possible compounding or long-term effects of selective residential mobility, which may have underestimated the contribution of selective mobility on urban–rural differences. For example, spouses resemble each other in political attitudes (Alford et al., 2011), which might increase the overall geographic clustering of party affiliation beyond that estimated for single individuals. Intergenerational transmission of party affiliation from parents to offspring (Kandler et al., 2012) could also compound the effects of parents’ selective mobility over the long term.
So he’s probably understating the effect.
Additionally, official geographic categories are an absolute conceptual disaster that fuck up the usefulness of basically any data set containing them. The Census categories for “urban” and “rural” are exceedingly crude and count places most of us would consider very small, rural towns as “urban.”
Anyway, the participants in the panel Jokela is using (the National Longitudinal Study) are apparently coded as either rural, urban, or major city. Here, “rural” means either Census-rural or not in a Metropolitan Statistical Area, which is basically a commuting zone consisting of one or more counties containing at least 100,000 people that includes a “central city” of at least 50,000. (“Central city” is an old OMB category that has been replaced by “principal city” in recognition of the fact that many metro are no longer organized around a single “core,” but may be polycentric, organized around several anchor municipalities.) “Urban” is either Census-urban (could be a smallish town) or inside an MSA, but not the central city (could be a farm in the sticks.) “Major city” is an MSA’s central city.
But “major city” doesn’t mean what you’re probably imagining, either. In this schema, I live in a major city, since Iowa City is larger than 50,000 and the MSA is larger than 100,000. Because any area inside the MSA is coded as “urban,” any farm or tiny town in Johnson County is going to count.
That’s why Jokela is tracking moves to “major cities” from rural areas and moves out of “major cities” to anywhere else. A rural-to-urban move could be a move from a farm outside an MSA to a farm inside and MSA. And an urban-to-rural move could be a move from a farm in an MSA to a higher-density small town of 2000 people outside an MSA. That’s clearly not what we’re interested in. Rural-to-major city definitely picks up an increase in the population density of residence. But Jokela is also tracking moves away from major cities to other urban areas (there’s not enough major city-to-rural to matter statistically), and this might be misleading.
For example, lots of University of Iowa students are from suburbs of Chicago that are much denser and more populous than Iowa City. Moving from Evanston to Iowa City takes you from a city with an overall population density of over 9,000 people per square mile to a city with an overall density of a little less than 3,000 people per square mile. But Evanston is not the anchor of the Chicago MSA, so it’s not “central city.” There’s no guarantee, then, that moves out of “central cities” to urban areas outside the central city of an MSA aren’t moves from lower to much higher densities.
Anyway, the data Jokela’s working from just isn’t going to pick up a lot of moves relevant to the density divide thesis. The fact that a model fitted to this data nevertheless explains nearly half the urban/rural difference in party vote share is significant.
More detailed indicators of the urban–rural continuum would be useful to characterize the selective mobility patterns with better geographic resolution. In the present analysis, selective residential mobility related to political affiliation was observed mainly between major cities versus other locations (i.e., rural or other urban regions). This is in line with population-level results showing that the association between population density and votes for Democrats is nonlinear, that is, particularly strong for the most densely populated cities (Rodden, 2019; Wilkinson, 2018).
I’m not sure it’s fully clear to Jokela how much the garbage-fire nature of the geographic categories weaken his finding. Unless there’s a similar dataset that contains the participant’s physical addresses, I’m not sure how you’d correct for it.
I find it interesting that this paper appears in the journal Social Psychological and Personality Science. As far as I can tell, there’s not a whit of social or personality psychology in it. So what’s it doing there? In “The Density Divide,” studies by Jokela are my main source for evidence that personality traits (especially Openness to Experience) that predict social liberalism or conservatism also predict the propensity to migrate at all. And I also lean heavily on studies by this paper’s recieving editor, Jason Retfrow (some of them co-authored with Jokela), on the geography of personality and partisanship. In his summary section, Jokela writes:
[T]he longitudinal data demonstrated that selective residential mobility was associated with political party affiliation, but these data cannot determine whether selective residential mobility is caused by political attitudes; there may be some other characteristics that determine people’s mobility decisions and correlate with political attitudes that were not assessed in the present study. However, political attitudes would still get geographically sorted even if they were not the causal factors of migration decisions and merely correlated with the causal factors. Future studies should assess whether factors such as lifestyle preferences, neighborhood perceptions, or personality traits might help to explain the selective migration associated with political party affiliation.
My hope is that the journal this paper is published in means that he’s planning to connect more of the empirical dots between personality, partisanship, and selective migration. I hope! Please, Markus?