In my previous post introducing the national precinct map for the 2016 Presidential Election, I asserted that the precinct-level percentage margin swing between the 2012 and 2016 General Elections could be attributed to education level more than any other single variable. Using 2015 American Community Survey census tract-level demographic statistics, I will attempt to provide some evidence to support this claim.
First, a simple visual inspection of the swing map shows the remarkable correlation with education level. The following maps show the precinct swing map overlaid by a census tract shapefile. Setting a cutoff just higher than the mean precinct level of college degree holders at 38%, they display first the precincts in tracts above this cutoff then below it. The relationship isn’t perfect and varies across regions, but it single-handedly explains much of the swing.
There are correlates with college education which also capture much of this divide but have certain drawbacks. For instance, median household income performs well in most areas, but severely underestimates the swing toward Clinton in college towns which have a large number of young students with very low incomes.
To delve into this further and more formally, I’ll turn to some basic regression analysis. Compare the R-squared values, i.e. the percentage of variation in the dependent variable explained by variations in the independent variables, for the following swing models which incorporate only independent variables of the type specified and state-level fixed effects:
|Proportion with College Degree||0.3023|
|Proportion Employed in Scientific or Technological Industries||0.2551|
|Area Classified by Census Bureau as Urban||0.2352|
|Median Household Income||0.2311|
|Proportion Employed in Finance Industry||0.2102|
|Proportion on Medicaid||0.1821|
|Proportion Employed in Construction Industry||0.1776|
|Households on Welfare (TANF, WIC, etc.)||0.1749|
|Labor Force Participation Rate||0.1698|
|Race and Ethnicity (White, Black, Hispanic, Asian, Native American)||0.1690|
|Proportion Employed in Manufacturing Industry||0.1602|
|Gini Coefficient (Inequality)||0.1446|
As shown here, the closest variable to education in terms of explanatory power is being employed in a professional, scientific, or technological industry (as defined by the Census Bureau), which is itself highly correlated with college education–a correlation coefficient of 0.78, specifically–as most positions in these fields require college degrees just to get in the door.
But just how much of a divide was there, exactly? By computing means for the areas above and below certain demographic thresholds, we can get a clearer picture of how totalizing and cross-cutting the education divide was:
|Positive number indicates pro-Clinton swing||Less than a High School Diploma||High School Diploma||Bachelors Degree||Graduate Degree|
|Above National Precinct Average||-6.98%||-10.77%||4.57%||5.46%|
|Below National Precinct Average||0.19%||4.69%||-9.29%||-8.25%|
Areas with high concentrations of various races/ethnicities:
|Positive number indicates pro-Clinton swing||White||Black||Hispanic||Asian||Native American|
|All Majority Precincts||-4.43%||-3.33%||0.98%||2.40%||-11.45%|
|Majority Precincts with Above Avg College Degrees||4.80%||-1.75%||4.94%||4.63%||-1.23%|
|Majority Precincts with Below Avg College Degrees||-12.99%||-3.59%||0.56%||-0.59%||-11.73%|
For every race and ethnicity, a higher proportion of college degree holders correlated with a higher swing toward Clinton, and vice versa for Trump. As I’ll delve more deeply into in another post, there exists a high degree of heterogeneity among every racial and ethnic group when it comes to their swing behavior in the 2016 General Election, and attempts to paint groups with a wide brush belie a tremendous diversity of opinion and belief. The weight of evidence implies economics, education, and class divided people far more than their skin color in terms of their reactions to Clinton and Trump, and they did so in a systematic and symmetric manner; in this way, we’re all more alike than different.