The Changing Racial Landscape of the Democratic Primary Race and How it Affects Bernie

In the context of lily-white Iowa and New Hampshire granting Bernie Sanders early life and a subsequent stream of southern states producing a near insurmountable delegate lead for Hillary Clinton, race has undoubtedly proved a defining element of the 2016 Democratic nomination fight. While Sanders has consistently gained more support from the white Democratic electorate, Clinton has won by even greater margins among black voters. That advantage may shrink soon, as exit polls have shown that black support for Sanders typically doubles or triples outside of the South, the area to which the race is moving–more Northeastern and Midwestern states remain on the Democratic calendar than Southern ones. At the same time, more Western states and thus more Hispanics will vote, representing a new electoral test for Bernie with a group that leans more heavily towards Clinton.

One of those challenges will arise when Arizona votes on March 22nd, a much more diverse state whose outcome will offer another indication if Sanders has any last-gasp chance for the Democratic nomination (even if he somehow wins, his path stays extremely narrow). Alongside the Grand Canyon State, two more states in Utah and Idaho, both of which favorable to Sanders due to the key racial component of this nomination race, will vote. Given this crucial demographic aspect, I will briefly review how these three impending contests will play out according to the state’s racial composition, and that of past state electorates.

Several racial variables can prove pertinent here: those relating to prior electorates–the share that whites, blacks, and Hispanics individually made up in electorates for states that have already voted (only for states with exit polls)–and those from Census data–the share that each of these three races makes up of the total population in all states.

Table 1. Source: U.S. Election Atlas, CNN exit polls, U.S.Census Bureau.

Correlating all of these variables, which the above table shows, the relationship between Sanders’s share of the vote and black percentage of a state’s total population emerges as the strongest absolute one. This produces a very strong negative correlation of -0.88; the greater a black population lives in a state–not even its portion of the electorate–the lesser vote share it gives to Sanders. This strongest racial variable predictor can also be understood in the below plot.

Graph 1. Source: U.S. Election Atlas, U.S. Census Bureau.

Considering the low levels of black population–as expressed in the below table–in each of the three states imminently voting, this would seem to avail Sanders. Arizona contains the highest percentage at 4.7 of the three states, while the other two hover around one percent. In states with at least as small black populations that have already voted, Sanders has averaged 63.78 of the vote.

Table 2. Source: U.S. Census Bureau.

However, Hispanic/Latinos will now increasingly occupy a greater presence in state electorates, starting with the three Western states voting on Tuesday, and represent another demographic disadvantage for Sanders. Only five of the 26 states that have held Democratic elections contain a larger Hispanic population than these three new states. On average for these five states, the Clinton vote share has been 54.62–barely above her national average at 53.72, and thus indicating perhaps not as much of an advantage for her. At the same time, among the six states with the highest portions of Hispanics making up their state electorates (and with exit poll data available), Clinton has averaged 61.4.

As can be seen in the below plot, although most states that have voted had very low Hispanic portions of their electorates, those that had higher ones tended to vote less in favor of Sanders. At -0.14, the correlation is hardly strong however.

Graph 2. Source: U.S. Election Atlas, CNN exit polls.

The idea of Clinton having a stronghold among Hispanic voters comes from robust support from them at the earliest stages of the primary race, as well as from national level polling and recent election results broken down by race. But on those last two points, ambiguity in past months has plagued the debate over which candidate claims greater support from Hispanics. It started with entrance polls for the Nevada caucus showing Bernie having won Hispanics by an eight-point margin, though that result was likely inaccurate for a variety of reasons. Then, the leadup to the Illinois primary–one of the more Hispanic-heavy states in the country–included wildly divergent levels of Hispanic support in different polls. Even in the past week or so, a polling firm in Morning Consult had Sanders leading among Hispanic voters by six points, and a few days later released a poll in which Clinton had an 11-point edge among the group. Earlier than that, YouGov had Clinton up eight points. In sum, there are clearly significant issues in polling the Hispanic population in the U.S. right now that likely have contributed to this lack of clarity.

However, the clearer indication of this group’s political leanings probably come from the states that have already voted. The below table shows the margin of victory for Clinton among Hispanics with large enough Hispanic populations:

Table 3. Source: CNN exit polls.

Despite the suspiciously large positive margin for Sanders in Nevada, Clinton has still beaten him by an average of 17.25 percentage points among Hispanics through the four state contests with exit poll data on the subgroup. There’s plenty of variance in these numbers, however, and it seems the margin will grow smaller than the one in Texas or Florida; perhaps Sanders does better with all minorities outside the South, and not just with blacks as noted before. Nevertheless, it remains fairly clear that Clinton has a stable advantage among Hispanics in this race.

In order to estimate where Clinton and Sanders’s shares of the vote will fall in the three states voting on Tuesday, plotting their vote shares in previous states against the state population demographics of them–rather than their electorate demographics–is the best way to go about this; the already known state-wide racial compositions can then be used to see where the vote shares might fall.

The below plot of the Hispanic portion of a state and the vote share for Sanders reveals a very weak correlation of -0.08. Similar to what can be done and in conjunction with Graph 1 above in having the black percentage of the state population, however, knowing the Hispanic portion can indicate where Sanders’s vote share might more or less land.

Graph 3. Source: U.S. Election Atlas, U.S. Census Bureau.

Applying the demographic characteristics from Table 2 to Graphs 1 and 3, it seems that Bernie’s percentage vote share will fall in the low to mid 40s in Arizona, the high 50s to low 60s in Utah, and possibly in the high 50s but more likely in the 60s and even 70s range in Idaho. At the same time, the polling leading up to these contests points to a smaller lead for Sanders than these estimates, which likely rely too much on the smaller black populations across all three states and the smaller than average nonwhite populations in two of the three.

Advertisement

The Racial and Economic Aspects of Trump Support

Each new point in the GOP primary calendar seems more important than the previous one, but after the latest slew of state contests winnowed the field down to three candidates, as well as eliminated the presumptive establishment favorite, that feeling has never been stronger.

In John Kasich, Ted Cruz, and Donald Trump we have, respectively, a staunchly conservative governor that for some reason moderated his tone during an appeal to the most ideologically extreme of his party (during primary season, as opposed to the general election, when candidates typically change their approach in front of a more moderate audience), one of the most extreme conservatives in the U.S. Senate but the type of candidate the establishment usually successfully screens out from the nomination, and almost assuredly a living anomaly of a presidential candidacy that has defied well-established principles of political science.

A contested convention has never occurred in the post-1972 McGovern-Fraser reforms era that democratized the nomination process, so if Trump does not meet the 1,237 delegate threshold for the nomination, there’s some guidance as to how it would play out–namely that it favors the frontrunner–but there’s also a chance we experience the brute institutional force of one segment of the two-party system. In other words, the outcome is still up in the air, making it very difficult to attach some kind of probabilistic estimate to the GOP wresting the nomination away from essentially the guaranteed delegate frontrunner–Trump.

Regardless, through the 29 states that have held elections, several patterns have already emerged to explain the electoral outcomes we’re seeing–especially with the surprising success of Trump.

Racial animus

Prominence of white racial identity that becomes more pronounced in states with higher non-white populations represents just one quantifiable racial component to Trump’s support. Another one cannot be so easily accessed with conventional methods.

Due to the effect of social desirability bias, survey research often fails to properly tap into a sensitive issue like one pertaining to race relations, unable to reveal indication of racial hostility in a respondent. The advent of Google search data can help solve this problem. Drawing from a paper that measures racial animus with this Google data and uses it to show that racial attitudes cost Barack Obama more of the national vote share than studies using surveys previously concluded, I use the same proxy for racial animus–racially charged language, such as the word “nigger,” in searches (see page 3 for more on how it’s defined)–to see whether it explains Trump’s vote share across different states. I restrict the frequency of the search data from the start of the Obama presidency to the present.

Here’s a link showing how the hostility indicator maps out across the U.S. Following the hyperlink, you’ll find that the search frequency concentrates in the southern part of the U.S., and reaches its peak in West Virginia. Interestingly, the percentage of black population in a state and the state’s racial animus measure correlates very highly at 0.70 (i.e. black resentment exists in larger quantities where more blacks live).

Across all states, the correlation between racial animus and Trump vote share is only moderately strong at 0.31, though much stronger than the same relationships for the other two candidates in the race: Cruz (-0.05) and Kasich (0.06). Examining this same association by different regions in the country–using the ones defined by the U.S. Census Bureau–brings more interesting results.

Source: Google Trends, New York Times.

As seen in the above plot, the correlation between racial animus and Trump vote does not become much stronger when looking only at states in the South, resulting in a 0.34 correlation. For the West and Northwest, the relationship turns negative, but that’s far from the same case for the Midwest:

Source: Google Trends, New York Times.

Bearing in mind it’s only a sample of seven states, the animus indicator correlates very strongly with Trump support in the Midwest, producing an unexpectedly, very high correlation at 0.91.

States with low racial animus as expressed by Google search data and low support for Trump–Minnesota, Kansas, and Iowa–and those with higher animus indicators and concurrently high vote shares for Trump–Illinois, Ohio, Michigan, and Missouri–fuel this trend. At the same time, the racial animus scores remain far and away highest in the South, in comparing either states that have already voted or across all of them (see below table). The Midwest follows the South in this animus score, though it’s closer to the region with the third highest animus measure–the Northeast–than the first one. On average, the South and Northeast tie for the regions voting for Trump at the highest rates.

Source: Google Trends, New York Times.

As an aside, the bottom row on this table may suggest that Trump might not get as much support from the remaining states in the primary calendar. Given the moderate correlation between his vote share and racial animus in a state, and the higher animus indicator in the states that have already voted relative to the national average, the remaining slate of states have a lesser racial animus score on average (58 in fact) and thus could prove less favorable to him. At the same, the dynamics in the race–namely the number of candidates running–have changed considerably, which could obscure a correlative relationship moving forward.

The economic aspect

Economic difficulties among Americans mark another key way to explain Trump’s strength in the GOP primaries. Using panel data with observations from several months before Trump’s candidacy and several months into it, some have discovered a very strong positive relationship between dissatisfaction with life/economic status in January 2015 and Trump support in December 2015. Similar relationships using vote share and economic indicators at the state level appear as well.

Source: 2016 CNN exit polls, New York Times.

Using exit poll data for percentages that the five main income groups (seen above) made up in each state that had this data (18 of 29 did), the table above shows the correlation between each income level and Trump’s share of the vote in that state. While the relationships are not very strong across the board, a class dynamic in Trump’s support becomes very evident. The higher representation of the lowest income bracket you get in a state, the more that state will vote for Trump. This association progressively weakens as one move up in income levels and then gradually becomes more negatively correlated, to the point where when a state has a higher percentage of top earners (>200k), it becomes least the favorable to Trump relative to the effect that the other groups’ presence has.

Looking beyond characteristics of only the Republican electorate and instead at state-wide economic conditions, the idea of economic struggles and poor economic condition potentially driving Trump support gets thrown into sharper relief. Three variables across states relating to income paint this picture: 1) median household income figures, 2) trajectory of median household income over the last 15 years (the percent change from 1999 to 2014), and 3) the the net change of a state’s national rank in median household income over this same time span. (All data for these three measures from here.)

1) In the most straightforward and frequently used of these measurements, median household income in 2014 ranged from 35,521 to 76,165 among different states, with an average of 54, 963. Out of the 29 states that have held elections this year, Trump vote correlates only weakly with income at the state level with a -0.23 coefficient, though in the correct direction given the hypothesis of economic struggles fueling Trump support: states with lower median income levels tend to vote for Trump more, but not by much. The plot below displays this relationship.

Source: Advisor Perspectives, New York Times.

2) Over the last decade and a half or so since 2014, all states experienced a percentage decline in median household income. Importantly, however, these declines vary considerably, from no decline to -24.7% over the time span considered. Once again, a weak but present negative correlation emerges between decline in income and Trump share of the vote in a state at -0.30. But one observation has an outsize effect here on attenuating the correlative strength: the state of Wyoming, which only gave Trump 6.7 percent of its vote–his lowest total thus far–and one possibly affected by a caucus system. Excluding this outlier case, the correlation increases greatly (in absolute value sense) to -0.55, a moderate to strong relationship between the two variables.

Source: Advisor Perspectives, New York Times.

The greater decline in median household income a state has experienced over the last 15 years, the greater share of votes it affords to Trump, thus the negative relationship. Such an association becomes much more pronounced when excluding one of the 29 initial state observations, as seen in the above plot (excluding Wyoming) and with the correlation coefficient.

3) Finally, another way of assessing the effect of rising economic struggles on votes going to Trump focuses on the change in national median income rank from 2000 to 2014 by state. For example, North Dakota rose from having the 42nd highest income level to 13th over this time frame, and so it gets a value of 29 for this measure, while Nevada dropped from 15th to 35th, thus receiving a value of -20 (these two state values represent the range for the variable). A larger (positive) rank change indicates more of an economic/income improvement, while a smaller (more negative) rank change signifies a greater decline in this area.

Source: Advisor Perspectives, New York Times.

The above graph places the rank change and Trump vote in each state side by side (note: the units of measurement for the two variables are obviously different, but the graph should still be interpretable). For reference, the average vote share for Trump is 34.97, and the average income rank change is -1.86 among states that have voted. Even a preliminary look at how these two variables relate reveals that the states with the lowest rank values–those seeing their income levels relative to the nation decline the most as of late–vote for Trump in higher numbers (e.g. Nevada, Georgia, Illinois, Florida). On the other hand, those with better economic fortunes–positive rank changes–vote for Trump relatively less (e.g. Iowa, Oklahoma, Texas).

Source: Advisor Perspectives, New York Times.

Accordingly, there’s a moderately strong -0.46 correlation between Trump share of the vote and the net change in national median income rank over 15 years (as the above plot also shows). In other words, as economic struggles–in terms of household income–increase by state, so too does Trump support to some degree. (Remove the somewhat outlier cases from the Northeast in Vermont and Massachusetts and an even stronger correlation–at -0.55– results). In sum, deteriorating economic conditions, even when measured at a fairly broad level such as the state one, certainly can explain some of the support for Trump during this election cycle.

(Another interesting note/aside: relative to the national average for this rank measure–at -0.42–the states that have already voted contain a lower rank measure–at -1.86–and therefore slightly poorer economic conditions than states on average and worse than those remaining to vote. That would mean the states that have already voted disproportionately favored Trump by this measure to at least some degree. This presents a similar scenario to the distribution of racial animus by states discussed above. Though the same limitations apply here, it seems as though Trump may soon compete in states not as favorable to him as the first 29 were.)

Thoughts on Super Tuesday Part 2

1) For all the constant talk about voters crossing partisan lines and switching allegiances during this election season, it’s usually best to ignore such unevidenced presumptions that constitute media fodder more than anything else. But on the Republican side for the Ohio primary on Tuesday, this dynamic was truly in play. As can be seen below, the non-Republican portions of the electorate for the Ohio Republican primary grew in 2016 compared to prior elections according to exit poll data.

Source: 2008, 2012, 2016 exit polls.

This shift very likely propelled John Kasich more than the other candidates, offering some credence to the stories about Democrats voting for Kasich to stop Donald Trump. 55 percent of Democrats voted for their state’s governor, the highest percentage of the three partisan groups to go with him. Moderates made up a quarter of Ohio’s electorate as well, and Kasich had his best total in that group–along the moderate, somewhat conservative, very conservative continuum–taking 58 percent of it.

In addition, a recent Monmouth poll revealed a much greater inclination for Democrats to vote in the Republican primary than their own: 17 percent of Democrats–compared to 10 percent of Republicans–said they would seriously consider crossover voting. Accordingly, while self-identifying Democrats made up eight percent of the Republican primary electorate, Republicans comprised only two percent of the Democratic electorate.

2) One very interesting question that emerged out of Tuesday was why the Democratic unfolded so differently in Illinois and Missouri–where Hillary Clinton only won by two and one percentage points, respectively–than in Ohio–where Clinton finished with a 14-point advantage. For one, the crossover vote that likely occurred in Ohio may have disadvantaged Bernie Sanders in particular. While no crosstabs were made public for the question on seriously considering voting in a primary other than your own, the director of the pollster that conducted the aforementioned poll stated that Democratic voters willing to take a Republican ballot were more likely to be Sanders supporters than Clinton ones.

Looking at exit poll information among this trio of states, a few distinctions between the first two states and Ohio also appear. Firstly, as seen below, the ideological makeup in Ohio was not as liberal as that in the other Midwestern states.

Source: 2016 exit polls.

Charting the percentages that each ideology made up of the electorate, Illinois and Missouri stand out as about five/six percentage points more “very liberal” than Ohio. A divergence materializes on the other end of the ideological spectrum as well, as the Ohioan electorate described itself as more moderate and conservative; while the non-liberal portion of the Democratic electorate was 35 percent in Illinois and 33 percent in Missouri, it was 42 percent in Ohio.

Apart from this, although younger age categories held fairly even across the three states in their distribution of support, the older age brackets broke more towards Clinton in Ohio.

Source: 2016 exit polls.

Voters aged 50 to 64 voted for Clinton by about 19 percentage points more in Ohio than in the other two states, and Ohioans aged 65 or older favored Clinton 15.5 percentage points more than voters of the same age in Missouri and Illinois. It’s worth noting that two oldest age groups comprised 50 percent of the electorate in Missouri, 53 percent in Illinois, and 54 percent in Ohio–their disproportionate share of the electorate being more important than change by state here. Age continues to represent one of the defining divides during this Democratic Primary race, as it still correlates strongly and positively with Clinton vote share (p.98 in the link)–where that increasing Clinton support becomes further pronounced could very well tilt an election more her way.

3) With the suspension of Marco Rubio’s campaign after losing the primary in his home state of Florida, it seems an apt time to question the prevailing theory in political science about how presidential nominations are won in the post-1972 reform era: that “the party decides,” and that a nomination is a product of the coordination between party elites and insiders, so as to put forth a nominee acceptable to all factions part of the broad party coalition. All of this can be reasonably proxied by the amount of public endorsements–especially from party officeholders–a candidate gets. Since 1980, these endorsements have proven more predictive of the eventual nominee than other factors such as media attention, fundraising, and polling numbers.

The primary race on the Republican side appears to cast serious doubt on this theorem. Up until his exit, Rubio had seemingly become the candidate around which the GOP coalesced, garnering several key endorsements after the Iowa caucus and leading in endorsement totals for much of the time thereafter.

Yet this overstates the involvement on the party of the GOP establishment. As indicated by FiveThirtyEight’s endorsement tracker, there has never been a slower pace of granting endorsements on the GOP side–in the sample of primaries since 1980 considered–than in 2016, and thereby never as little effort from the party to influence the primary race. This slow pace materialized during the invisible primary, defined as the year or so before the first elections take place and when party elites traditionally coalesce around an acceptable candidate. The party decides theory specifically points to this time period, and endorsements that occur within it, as what is most consequential for shaping the nomination fight. It was here when the slowest pace of endorsements were given, and when Jeb Bush–rather than Rubio–emerged as the establishment favorite, but not by any substantial margin.

An active approach to influencing the nominating process has almost always reaped the desired rewards for party elites, yet during this cycle they abstained like never before. Rubio’s surge in endorsements did not take place during this all-important invisible primary phase of the campaign. Along with the lack of party action beforehand, it should come as no surprise whatsoever that party insiders have lost control of the nomination, and that the purported establishment pick in Rubio failed, as in fact his path clearly falls outside of the model of past primaries that indicate elite coordination and backing translating to a candidate’s success.

This should not detract from the heavy blow the party decides theorem will take in relation to the GOP primary during this election cycle, as actual results on the nomination path have contradicted this line of thought. At the same time, a distinction lies in defining this theory–is it that an establishment candidate overwhelmingly wins the nomination, or that this candidate only succeeds upon concerted effort from the party coalition and elites in offering support? If the latter case is true, then I think the theory remains far from disproved–and especially so when on the Democratic side, it’s being continuously borne out: though encountering some obstacles, Hillary Clinton has amassed endorsements at the fastest rate among all past Democratic candidates, and consequently stands in a near-impregnable position to capture the nomination.

4) In the aftermath of polling debacle for the Michigan Democratic primary, many questioned the faith placed in polling and whether polls were accurately portraying future primaries. Never mind this use of an anomaly as a meaningful precedent, the results of this second Super Tuesday of sorts renewed support for the accurate ability of pollsters–which frankly never should have been seriously questioned (especially considering polls have done a fairly good job during this election season).

Source: HuffPost Pollster, New York Times.

The above graph shows the final polling margins in favor of Clinton before Tuesday–measured by HuffPost Pollster’s average, or if they did not have one, the average of the most recent polls–and the actual voting margin in favor of Clinton. The pre-election polls proved very accurate, deviating by an average of 4.1 percentage points and correctly showing Clinton ahead in all five states. Though only five observations, the final polling margins and voting results had a very strong 0.94 correlation (a 1.00 correlation would mean the polls perfectly predicted the actual result).

Where White Identity Activation Benefits Trump Most

While white identity has been established as a strong predictor of support of Donald Trump in this election, its prominence has begun to appear in distinctive parts of the country. In particular, this New York Times piece pointed to an initially counterintuitive development of Trump garnering less support in states that are more white in their population makeup. The author concludes that “an appeal to white identity tends to work better in areas where that identity is felt to be under threat.”

With this interesting idea in mind about where white identity may become central to vote choice, I try to test it. Below is a plot of the percentage of the white population in a state and the percent of the vote Trump won in the state. While not especially so, a somewhat strong relationship results in the negative direction: a correlation coefficient of -0.48. As the white share of the state population increases, Trump’s share of the vote decreases. In other words, a larger non-white presence in the state makes for higher support for Trump, which squares with the theory proposed above.

Source: U.S. Census Bureau, New York Times.

This also excludes an outlier case of Hawaii in terms of where it appeared on the plot (and how much disproportionately more weight would be given to such as small state), but the exclusion changed the coefficient very little.

Several embodiments of this trend come from the South. States such as Mississippi, Louisiana, and Alabama have offered some of the highest levels of voting support to Trump, but have some of the smallest white populations of the states that have already voted. Among the 10 southern states that have held elections thus far, a stronger correlation between percent white and Trump vote–of -0.59–emerges.

At the same time, he’s performed relatively worse in states with some of the largest numbers of whites, such as Iowa, Minnesota, Vermont, Maine, Kansas, Idaho.

The other factor mentioned in the NYT piece that further accentuates white identity among white voters is specifically the size of the black population part of the non-white one. After all, this finding of white identity influence relates heavily to antipathy and hostility towards blacks.

Above I plot the percentage of the black population in states that have held elections and the share of the vote Trump received in them. Once again, a fairly strong correlation between the two variables arises, nearly as strong as the previous one at 0.45 (in terms of its absolute value).

Though certainly not a very powerful indicator (as much as education level would likely be in the Republican race), where white identity gets activated does account for and explain some of Trump’s support during this primary process. Even just visually comparing the below maps that geographically represent where the white population concentrates, where the black population does, and where Trump has fared best, respectively, would return the same result: the more that a non-white population resides in a state–especially African-Americans–the more likely it will vote in greater numbers for Trump.

Source: U.S. Census Bureau; white alone 2014 estimates.

Source: U.S. Census Bureau; black/African-American alone 2014 estimates.

Source: New York Times; margin of victory by county (Trump in red).