Are Americans substantially more divided based on where they live and their social identities? Or are stories of voters sorted into neat social and geographic enclaves overstated? Seo-young Silvia Kim finds that it is not so easy to predict how Americans will vote based on their demographic groups—and it hasn’t gotten any easier over time. David Darmofal finds that demographics are a bit more predictive of geographic voting patterns, but spatial polarization has not increased markedly over time. They both take the long view, finding that we are not as divided by social groups and geographies as we seem.
Matt Grossmann: Demographic and geographic polarization is overstated this week on the Science of Politics. For the Niskanen Center, I’m Matt Grossmann.
Are Americans substantially more divided based on where they live and their social identities? It certainly seems that way, with our urban/rural divides and our increasing divisions on race and education. It seems like it should now be easy to predict how an individual or a geographic area voted based on a handful of variables, but taking a longer view makes the story more complicated, with the trends much less pronounced.
This week, I talked to Seo-young Silvia Kim of American University about her new working paper with [Yon Zelinsky 00:00:45], The Divided but Not More Predictable Electorate. She finds that it’s not so easy to predict how Americans will vote based on their demographic groups, and it hasn’t gotten any easier over time. Instead, voters are increasingly divided by partisanship.
I also talked to David Darmofal of the University of South Carolina about his [Springer 00:01:04] book with Ryan Strickler, Demography Politics and Partisan Polarization in the United States. He finds that demographics are a bit more predictive of geographic voting patterns, but spatial polarization has not increased markedly over time. They both find the conventional stories of voters sorted into neat social and geographic enclaves to be overstated. Kim says demographics aren’t destiny when it comes to Americans voting and have not become more important over time.
Seo-young Silvia Kim: Try as we might, demographic labels do not give much information about vote choice throughout the last 70 years. We quantify how a well-performing machine learning algorithm does with just five variables: age, gender, race, education, and income, the big five that the people think is demographics. And we find that, on average, you can only predict about 63.5% of the two party vote choices correctly on average, throughout these years.
It also does not increase over the period of 1952 to 2020. So I think this goes against a lot of people’s intuition that the demographic group identities do really determine political behavior such as vote choice. There’s a lot of punditry built around such notions. And also academics, we believe that demographics is a strong and important predictor that we must pay attention to. And given that we believe that demographic sorting has taken place, and that party line voting has increased, it must have been a natural conclusion to say that, based on demographics, we can predict vote choices better. But it’s not really that case.
Matt Grossmann: Darmofal finds that current geographic polarization does not stand out. And even the latest selections may reflect old patterns.
David Darmofal: There’s really been considerable discussion in the popular press about how we’re becoming a more polarized country. And much of this discussion is focused, of course, on geographic polarization, this notion of red states and blue states, and how we’re divided as a country between red states and blue states. What we wanted to do in this book was to place this current geographic polarization debate and discussion in historical context, because we can’t fully understand the present without understanding what came before it and how this contemporary era fits into broader historical patterns.
For example, is this an anomalous period, or is it reflective of past patterns? And one of our core findings is that we find that, contrary to many popular accounts, we’re not living in a particularly geographically polarized time. Americans are no more likely to live in landslide counties won by a presidential candidate by 20 points or more in recent decades than in previous eras.
We also apply methods of spatial analysis to identify the political geography of county level of voting in each presidential election, from 1828 through 2016. So this is basically the entire period of mass voter participation in the United States since the advent of Jacksonian democracy.
And one of the things we find then is that the spatial structure of partisan voting, the locations of Democratic and Republican regimes across the country, has changed gradually over time rather than in a haphazard, election-specific manner.
Another interesting finding is that you can better predict how counties voted in the 2016 presidential election between Donald Trump and Hillary Clinton from how they voted in the 1828 election between Andrew Jackson and John Quincy Adams, then from the 1976 contest between Jimmy Carter and Gerald Ford. Strong Jackson counties in 1828 tended to be strong Trump counties in 2016 and strong Adams counties, strong Clinton counties. In contrast, strong Carter counties in 1976 don’t tend to differ from other counties in their preferences for Trump or Clinton. And that may well reflect that Trump tapped into what Walter Russell Mead and others define as a populist Jacksonian tradition in the United States.
Finally, the focus on blue states and red states really distorts our understanding of the political geography of voting in the US. They’re significant within state variation and partisan voting. And ironically, we’re actually near a high point in the contribution of within state variation in partisan voting, in contrast to between state variation, precisely as our attention as a country has increasingly turned to the state level debate of blue states versus red states.
In short, there’s a lot of truth in former President Obama’s 2004 keynote speech. There’s a lot of Republicans in our so-called blue states and a lot of Democrats in our so-called red states. And if we really want to understand the political geography of elections, we have to look at the sub-state level. We do that in our book by focusing on counties, the lowest level of aggregation for which we have complete data on the contiguous United States for the full period of mass voter participation in the US, and that’s the focus of our book.
Matt Grossmann: Kim and Zelinsky were tracking individual voters since the 1950s, whereas Darmofal and Strickler were tracking county level patterns since 1828. Kim and Zelinsky expected to find demographic sorority, but didn’t find it.
Seo-young Silvia Kim: We were wondering aloud about the diploma divide that was being repeatedly brought up by the media. And we were thinking that yeah, yeah, the marginal distributions and the cross tabs in need seem to suggest a demographic sorting. Now, what happens if we plug that into a prediction algorithm? Does that also show up as we would predict? And we initially thought that this was going to be a really quick exercise, I think. We expected to see results in line with an increasing demographic sorting. We expected this to end quickly. Turns out that we were really puzzled by the initial results.
Matt Grossmann: They found that using demographics, a prediction algorithm, can’t get very far. It needs to know voter’s party.
Seo-young Silvia Kim: What we do is we take surveys in presidential election years, and then take out the voters who say they have voted, and take out those who voted for either of the two major parties. And then we train the machine learning algorithm to try to predict really, really well, given a training set about which party this voter voted for.
And we take that really well-trained machine and then plug a data set that the machine has never seen before to the same machine. And we find that… I like to do this like a score. And then the machine only gets 63.5. That’s a D on my syllabus. Random guess gives you about 50% because most presidential elections are pretty close. So a low 60s accuracy is saying that the five variables of demographics is not doing a whole lot. A party gets you up to late 80s, and then on the score book, that means that the machine learning algorithm on the score book, that means that the machine learning algorithm has jumped from a score of D to a B.
Matt Grossmann: Voter gaps by demographics don’t necessarily mean easier prediction.
Seo-young Silvia Kim: We are looking at a joint distribution of votes as opposed to the marginals, and I think there is no doubt that at the marginal level it does seem to be the case that there is a stronger association for each two party votes. But jointly considered together, you’re taking a voter and saying that instead of just looking at this person as a Latino we’re now going to consider that a Latino person in his 40s earnings such and such income and educated with a college degree. And then overall, the predictive power really balances out and does not give an overpowering predictability to vote choice. I think that’s a major thing. Cross tabs are really about the marginals, but jointly, it doesn’t do very well.
Matt Grossmann: And additional demographics don’t help.
Seo-young Silvia Kim: We did try adding other variables such as religion. Because our data set is limited in the sense that we have to make the variables consistent from the 1950s to 2020, there is some limitation. But what we added was whether the voter was Christian, Catholic, Jewish, or otherwise. We also tried adding preliminary geographics, which was whether the voter lived in the Southern States or not. Neither did really much of anything. The accuracy is still in its 60s. It does not increase over time.
Matt Grossmann: [inaudible 00:10:58] at spatial polarization, also finding no increased predictability or an obvious shift to new partisan regimes.
David Darmofal: We find little evidence that recent decades have seen an increase in landslide counties. Instead, the pattern is consistent with earlier eras. Now, there was an uptick in this in 2016. We, of course, don’t include analyses for 2020 in our book. Our book covers from 1828 through 2016. And it’ll be interesting to see whether that continues to increase. But when placed in a historical perspective, recent decades don’t appear to be a period of increased spatial polarization. And obviously another aspect of this question of spatial polarization is this notion of Democratic and Republican regimes of partisan voting. Looking at these Democratic and Republican regimes over time they change gradually over time and in understandable way. And I just want to briefly discuss the methods here that we used to identify the spatial structure of voting behavior from 1828 through 2016.
What we’re doing is using a set of diagnostics, the Global Moran’s I and the local Moran’s I, to identify whether counties as a whole exhibited similar partisan voting as their neighboring counties, and we’re using contiguous County as our definition of neighbors here. And then if we identify that at the global level or we don’t identify at the global level, although we do in each election, we then look to the local level to identify which counties specifically are auto correlated with their neighbors. And so we can then use these local Moran’s I’s with a Moran scatter plot to identify which counties exhibit similar voting patterns as their neighboring contiguous counties at rates that are more Democratic than the national average, which are correlated with their neighbors at rates less Democratic than the national average, which have higher support for the Democrats than their neighboring counties, which have lower support for the Democratic candidate than the neighboring counties, and which counties are uncorrelated with their neighbors.
And so what we do then use, we use these spatial tests to identify where contiguous counties with similar levels of Democratic or Republican support are located in the United States. And basically, what we’re looking at here is areas of strong Democratic support and areas of strong Republican support. And what we find is that these regimes really change gradually over time in understandable ways. There tend not to be strong election to election fluctuations. That differs for some outlier Presidential elections such as 1964 and 1972, but generally more Democratic parts of the country tend to be more Democratic for multiple elections and the same for more Republican areas in the country. And, interestingly, the relative areas of Democratic and Republican strength remain such whether the country that had a close election or a landslide election.
So, for example, the 1984 and 2000 maps that show spatially auto correlated areas of Democratic and Republican strength are quite similar despite 1984 being a landslide election and 2000 being a quite close election. The mean levels of Democratic and Republican support shifted between those elections, but the relative areas of Democratic and Republican strength remain quite similar.
Matt Grossmann: Density does matter but the urban rural divide started in the 1920s.
David Darmofal: One of the things we do in our book is we do look at density as a predictor of partisan voting in each presidential election from 1828 through the present. And we do find that population density starts predicting partisan voting at the county level beginning in the 1920s or so. So there’s very little consistent pattern in that prior to the 1920s or so. But since then, it’s been a pretty consistent phenomenon where more densely populated counties tend to have higher levels of support for the Democratic candidate than do more less densely populated counties.
Matt Grossmann: But he does find evidence consistent with a shift toward nationalized elections.
David Darmofal: One interesting finding that may potentially speak to the nationalization of the electorate is this. We employ a set of spacial models for each Presidential election from 1828 through 2016, and so the political geography that we identify, the spatial structure of partisan voting that we identify could be produced broadly speaking by any of either of two types of processes. On the one hand, neighboring counties could exhibit similar levels of Democratic or Republican support, because the citizens of these local electorates directly interact with each other and shape each other’s partisan voting behavior. So that’s a behavioral diffusion story and that’s consistent with a spatial lag model where there are spatial dependents pertaining to the dependent variable.
Alternatively, these counties may not, the people may not interact with each other across these county lines. They may be more atomistic units, but they may be responding to national stimuli. And so their correlation would be then produced not by direct behavioral diffusion across these local electorates, but instead due to admitted co-variates from our models. And so any residual spatial auto correlation you find then is consistent with a spatial error model. So we run diagnostics, Lagrange Multiplier Test, to determine whether a spatial lag model or a spatial error model is more appropriate. And what we find is that the spatial lag model, the behavioral diffusion at the local level, seems to be driving the political geography of voting behavior that we identify conditional on our co-variates for most of American history until recent decades. Now, what we’ve done is we’ve seen that in recent elections, the spatial error model is much more applicable, and that could be consistent with the idea consistent with Robert Putnam’s argument that people are interacting less.
… consistent with Robert Putnam’s argument that people are interacting less, they’re responding instead to national stimuli. Obviously Dan Hopkins’s recent excellent book on the nationalization of political behavior in the United States speaks to this as well. It could be that these local electorates are behaving similarly now because they’re all responding to national stimuli.
Matt Grossmann: One thing that could be driving that is polarization based on party identification. [inaudible 00:18:29] Zalensky tried to figure out which factors matter beyond demographics, settling on part.
Seo-young Silvia Kim: Once that we establish the fact about predictability of demographics, we were curious about what variables add strength to prediction, right? And so what we commonly consider are three separate things, explicit parties and affiliation, partisanship labels, whether you identify as a strong Republican, weak Republican, leaning Republican and such. You also have symbolic ideology, whether you think yourself as a liberal or a conservative and the spectrum on that. You also have issue positions, whether you support or oppose abortion and such. So we tried three models in which we added those three separately to demographics. And we find that all of them increase predictability, but partisanship just outperforms the other two.
Matt Grossmann: Party is becoming more predictive and it encapsulates identities and ideology.
Seo-young Silvia Kim: You add a party ID to demographics, and then the scoreboard suddenly jumps to an accuracy of 87.3 ranges. And over time, on average, every year the accuracy increases by .18 percentage points. So in a decade, that means almost two percentage point increase. But now given that it was already pretty high to begin with, I think that’s a remarkable thing. I think the results reflect a partisan, an ideological sorting that we’ve been discussing in the discipline.
That once partisan ship is included, the other variables don’t seem to do well, I think that speaks to the idea of hardness in ship as a super identity, right? When we consider separately operational or symbolic ideology instead of partisanship, it did improve predictability, but it didn’t do as well as simply including the partisanship label. I think that’s a remarkable evidence saying that partisanship is really a super identity that encompasses everything and sends a much, much stronger signal than the ideology variables that we usually consider.
Matt Grossmann: Partisanship’s predictive power could indicate that party is an effective presidential choice, but that’s less likely.
Seo-young Silvia Kim: Well, it’s entirely possible that it’s not that vote choice is determined by partisanship, but you like a certain presidential candidate and then you change your partisan affiliation. In data such as voter files, you hardly ever see people changing their artists and affiliation, especially not to another party. So I think the chances of that are rare.
Matt Grossmann: Voters could be sorting on identifications and views rather than group membership.
Seo-young Silvia Kim: First of all, parties and ideological sorting and stuff [inaudible 00:21:49] taking place. It’s very, very strong. There is no doubt to that. As to social sorting, I think our results can go hand in hand with the fact that there is social sorting because we are using labels instead of identity. So we are saying that you belong to this group, we automatically assume that you carry this group identity. But the low level of strength that we find in projectability might indicate that it’s not really about the objective labels that are put on you. It’s really about how strongly you identify with these groups. So I think our results could go hand in hand with that and say that these objective labels that we really build our discussions around, they don’t seem to be as strong as we believe them to be. Instead, it’s really about how strongly you believe this group [inaudible 00:22:51] and a group voting should be or how you identify strongly with those [inaudible 00:23:00].
Matt Grossmann: Kim says partisan alignment is real, but not realignment based on demographics.
Seo-young Silvia Kim: The two stories of realignment all together. So there is a partisan realignment, the strength of parties. There was an era of independence and then we’re suddenly identifying more and more party, we’re more involved in politics. And that’s the alignment story that I see in one end. And I think that’s definitely true there. But the second realignment would be about social groups. And I think while social groups can have many nuances and layers, because these are the easiest labels we can muster, we tend to focus on those groups such as age and race and gender and so on. And I think in that sense, we are saying that those labels, they might not be doing much. The strength of realignment might not be so pronounced as to be shown in the data.
Matt Grossmann: And she says that geographic areas could still be more predictable based on demographics.
Seo-young Silvia Kim: If you take that to a neighborhood level, I think that’s a completely different story. Let’s say that you’re a Brooklyn resident that has strong opinions about housing prices, prices in affordable housing. And then between that, I think within those small neighborhoods, I think demographics can do a lot to determine vote choices one way or the other. So I think our results can definitely go hand in hand with the results on spatial polarization, because that has a lot of information about where the voters live and we don’t.
Matt Grossmann: [inaudible 00:24:56] agrees that demographics might predict aggregate voting without predicting individual level voting.
David Darmofal: Clearly you could find that demographics would not necessarily increasingly predict voting behavior at the individual level. But you might see if there’s migration or differential generational replacement that depends on demographics, we could see demographics increasingly predicting aggregate voting in a way that it doesn’t, again, increasingly predict individual level voting. In other words, if counties or other units are becoming more demographically homogeneous, you’d potentially see stronger democratic or Republican support in these units, even if it’s not rising in its effects at the individual level. So for example, let’s say there’s Latino or Latina voting. Let’s say that that doesn’t predict voting any more.. Maybe it was rising in its strength for several years, but that’s leveled off now. But if you had increasing in migration of Latino populations, which are still strongly Democratic, you would see potentially a strong relationship between the local demographic makeup of counties and their partisan voting.
Matt Grossmann: But aggregate level relationships can mislead. Black population was associated with Democratic voting when blacks were prevented from voting in the South based on support from Southern Democratic segregationists. After the Voting Rights Act, it was eventually associated based on enfranchised black voters. [inaudible 00:26:28] and Strickler look at demographic predictability over time, finding strong and growing relationships, including black and immigrant populations.
David Darmofal: Other covariates that we do examine as well are, for example, proportion African-American, the African-American proportion of the population in the county and the immigrant populations in counties as well. And so basically what we find is that the proportion African-American in a county, the size of the local county level black population, is a pretty strong and consistent-
Well, a black population is a pretty strong and consistent predictor of partisan voting at the county level, for quite a long time now. Going back, for example, to 1900, we’re finding this consistent pattern. And basically what we’re finding is that counties that have larger African American populations, tend to be more likely to vote Democratic.
And we actually see growing and stronger effect of that in the Sixth American Party system, in the current American party system, since the late 1960s. Interestingly, again, the size of the local African American population is positively associated with support for the Democratic Party at the county level in all of these elections, but it was not, in particularly anomalous elections. So, the 1964 presidential election is the one election that we find, going back quite a ways, in which larger black populations were associated with lower support for the Democratic candidate.
And of course, 1964 is prior to the Voting Rights Act of 1965. And you could well see this as potentially an example of white voters in counties that had large black populations voting for Goldwater. Of course, we don’t have the individual level data, so we don’t know what’s producing that, but it sticks out quite clearly as the one election in which larger black populations were associated with stronger support for the Republican candidate.
And then the other thing we do examine is immigrant populations. The size of the proportion of foreign born in a population, in a county, rather. And there we find pretty consistent results since the Clinton years, basically, after many elections where there wasn’t an effect, whereby counties that have large immigrant populations have stronger levels of Democratic support. And this could speak potentially to Democratic gains as the country has diversified, demographically. And these locations with large foreign born populations are having a particularly strong, pro-Democratic effect of having local immigrant populations.
Matt Grossmann: Kim also says demographics could still be critical to turn out, which might also show up in these geographic divisions.
Seo-young Silvia Kim: Because you’ve already conditioned the turnout, you might find a little relationship with demographics and vote choice, but if demographics has a larger effect on determining turnout, we already know that we see a large turnout differential by demographics. And I think that means that it’s not that demographics is not useful whatsoever, it’s that once these people have decided to turn out, it’s less to be informative about vote choice.
Matt Grossmann: But she says we’ve gone too far in emphasizing demographic divides.
Seo-young Silvia Kim: Journalists, practitioners, they must really consider rethinking the practices of sometimes reading too much into marginal changes in each polling, each post-election cross tabs.
I think sometimes the stability of voting behavior really outweighs what we see to be very marginal changes in each election. And I think this really is compounded by the horse race coverage in primaries and general elections. In my opinion, I think there is also a potential harm by such reporting, that by emphasizing the divides by demographic groups, you create an affective feelings of distance between demographic groups. And I don’t think there has been an adequate assessment of potential negative effects of such media and [inaudible 00:31:24].
Matt Grossmann: She’s now looking at whether our overemphasis has led to polarization.
Seo-young Silvia Kim: I would be interested to see whether the elite communications have really influenced public perceptions of the demographic divide, out of proportion. If that is the case, I think that could be another source of affective polarization and emotional distances between groups. And I think that’s really bad for the democracy. And I want to see if there are ways in which to reduce, and bash those myths if we can.
Matt Grossmann: There’s a lot more to learn. The Science of Politics is available biweekly from the Niskanen Center and part of the Democracy Group Network. I’m your host, Matt Grossman. If you liked this discussion, you may want to check out the podcast, Our Body Politic in our network, including episodes on speaking with Hispanic Republicans and women of color in the Republican Party. Or our prior episodes on explaining the urban rural divide, or interpreting the early results of the 2020 election. Thanks to Seo-young Silvia Kim, and David Darmofal for joining me. Please check out The Divided (But Not More Predictable) Electorate, and Demography, Politics, and Partisan Polarization in the United States, 1828–2016, and then listen in next time.