In 2016, Donald Trump rode a perfect wave of anti-Obama sentiment and economic disenfranchisement to the greatest upset victory in American Presidential history.

That’s a commonly held narrative about the 2016 election. But does the data back it up? I conducted a number of statistical tests and mapping exercises to find out. Predictably, Trump’s margin of victory is correlated with anti-Obama sentiment; but surprisingly, there was only a weak relationship with economic performance.

My core research question: How did President Obama’s approval rating in a state and that State’s GDP change in the year prior to an election impact voting behavior in the 2016 presidential election?

The dependent variable of this analysis is the state-by-state percent of votes for Trump in the 2016 election (represented by the database-friendly label “gop_vtsh” in the figures below).

My predictors:

1. Change in state GDP in the year before the election (“gdp_change”)
2. President Obama’s approval ratings by state (“obama_app”).

Mapping

First, I mapped my three variables. Trump’s vote share was weakest on the East and West coasts and strongest in the South and plains states (Figures 1-3). These core regions of Trump support also tended to have GDP growth below the median and low Obama approval ratings.

Graphing

Next, I made a scatter plot matrix (Figure 4). It confirmed that GDP change and Obama’s approval rating were both negatively and significantly correlated with Republican vote share at the p<0.01 level.

Where things got interesting: I noticed a distinct cluster in the scatter plot of Obama approval and GDP change. When I brushed these, I realized that they were the same states identified as outliers in the Obama Approval boxplot map: Obama’s home states and states with dynamic urban economies. Regimes regression revealed this cluster had very different behavior from the other states: they were less likely to support Trump than would be expected based on their GDP.

For more detail, I replicated the Obama Approval vs GDP Change graph in a normal scatterplot (Figure 5). I found that the Chow test produced a p-value of 0.0000 and that both the p values on the urban cluster and its complement remained significant. This increased my confidence about differential trends.

Finally, I replicated the scatter plot of GOP voteshare as a function of GDP change, brushing the same states (Figure 6).

Again, the Chow test detected a significant difference, and the regression lines of the two groups remained significant at the p<0.05 level. Interestingly, the clustering did not impact the relationship between Obama’s approval rate and GOP vote share (Figure 7).

Conclusions

My analysis had 3 main findings:

1. Obama’s approval rating and GDP change in the year prior to the 2016 election had substantial and significant negative correlations with Trump’s vote share in that election.
2. There is a cluster of urban states with dynamic economies that have different trends than the rest of the states. If we do not control for this when we use GDP change to predict Republican vote share, we may end up underestimating.
3. The clustering did not impact the relationship between Republican vote share and Obama’s approval rating, perhaps indicating that Obama’s approval was a more robust predictor of vote share than GDP change.