Party switching – the Case for Racial Resentment

A lot has been written about the relationship of racial resentment and support for President Trump. As Michael Tesler has noted in this blog, views on race have mattered more in electing Trump than Obama, perhaps because they have funneled a feeling of white vulnerability. If, as The Nation writes, the phenomenon Trump has “accelerated a realignment in the electorate around racism”, it offers some explanation for the Republicanization of some minority-devoid stretches of the country. The reverse has received less attention. If racial animus has the power to draw those with a high propensity of racial resentment into the Republican tent, can it then function as a wedge-issue, potentially alienating traditionally Republican voters from Republican candidates? Tests of this hypothesis are difficult, mostly because Big-N-data on resentment that would permit county-by-county analyses are hard to come by, and we have had only few elections since 2016. Over the last months, PredictWise has collected such data designed to answer this and similar questions. [Details on the ongoing PredictWise data collection can be found here.]

Using algorithms we have developed for our rather successful prediction of the 2016 presidential election to adjust for non-response (a legitimate concern in all surveys) coupled with demographic breakdowns of the voting population derived from Big Data on all registered voters in the US, we can monitor racial resentment at very granular levels. Specifically, we collapse elements of racial animus – the standard racial resentment battery plus items capturing standings on police interactions with African Americans as well as standings on African-Americans athletes kneeling through the national anthem at NFL games – into a single score and use demographics, partisanship, and gps-coordinates-based urbanicity measures to project these scores, ranging from 0 to 1, onto geographic units we are interested in. The end-result is a good proxy of “racial resentment spread”.

One of the first test-cases of the reverse of this hypothesis (after the 2016 General Election) presented itself in the Special Senate Election in Alabama between Doug Jones and Roy Moore on Dec 12, 2017. If it is, in fact, true that racial resentment has the power to serve as a wedge issue, we should see a systematic movement toward the Democratic side of counties in which racial resentment is less widely spread, at least compared to the 2012 general election (all results hold for 2016 to 2017 but are less robust since Trump also ran on a similarly racially driven message as Moore). First, we plot our estimates of racial resentment spread. Indeed, racial resentment varies widely from county to county in Alabama.

Figure1
Figure 1: Spread of Racial Resentment by County in Alabama; the state-level median is colored white, counties with high levels of resentment appear red, and counties with low levels of resentment appear blue

Racial resentment is much less widely spread in more populous counties, such as Mobile County, Madison County, or Jefferson County. But, some less populated counties also display lower degrees of racial resentment, such as Macon County in the south of Alabama. In contrast, racial resentment is widespread in the central Northern part of the state, particularly in De Kalb County, Winston County, and Blount County. But, the variation in and off itself is not very meaningful. We want to test if low levels of resentment can predict Democratic gains in the electoral map. If we plot changes in Democratic two-party-vote-share 2017-2012, a suggestive story emerges: These dynamics indeed correlate with our county-level racial resentment scores. In counties less plagued by racial resentment, voters switched to Democratic Senator Dough Jones in droves.

Figure2
Figure 2: Changes in Democratic votes hare 2017-2012 by County in Alabama versus Spread of Racial Resentment by County in Alabama

Of course, these are binary relationship, and are further complicated by concerns of endogeneity that arise because we project our racial resentment scores with some county-level demographics that at the same time may be governing the changes in vote share. But, if we do account for county-level demographics, this relationship prevails. Specifically, our resentment scores predict movement toward the Democratic camp conditional on partisan composition as well as age-, gender- and race distributions in these counties. We predict that 1 percentage point less spread of racial resentment is associated with a 0.4 percentage point gain of Doug Jones vis-à-vis Barack Obama in 2012 in these counties.

The verdict on the national level may still be out. And there were certainly more issues in the Alabama special election than race, with serious charges of sexual misconduct level at Moore. But, with all of the caveats of a single election, as far as Alabama in 2017 is concerned, the flipside of the racial-resentment-as-new-cleavage hypothesis holds true: Not only do voters prone to racial resentment flock to the Republican tent, Republicans who hold more moderate views on race are alienated from the Republican core as well. Whether these dynamics cancel out or give one party a systematic advantage will be answered by future elections. In Alabama, these dynamic certainly played into the hands of Democrats.

Notes on methodology: We surveyed about 5,000 American smartphone users once a month between October 2017 and January 2018. We then modeled racial resentment using a Bayesian hierarchical Item Response Theory model that in essence collapses six items measuring racial animus into one score, including perceptions of racial minorities and work ethic, historical discrimination and equality of opportunity, current discrimination and equality of opportunity,
racial minorities and social welfare net (i.e. traditional racial resentment battery), racial minorities and crime, racial minorities and police brutality and racial minorities and anthem protest. In the same step, we model these scores on the basis of age, gender, education, race, party identification, state, family composition and urbanicity, and project the estimates onto the likely voter population using a full cut of the TargetSmart voter file.