In the last election, Initiative 1433 gave Washington voters the chance to raise the statewide minimum wage to $13.50/hour by 2020. Voters obviously also had the chance to elect the next president. It turns out that (in King County) there was a strong positive correlation (R =.894, R^2=.798)* between voting for Trump and voting against a minimum wage increase at the precinct level. In other words, as you can see in the above chart, areas with more Trump voters were much more likely to vote "no" on I-1433.
Regression models can be interesting, but they can be even more informative when you physically map out where they fail. Below is a map of the residuals, which shows us where our simple regression model either overpredicted or underpredicted the proportion of voters in that precinct that would vote against I-1433. Green colors represent places where people voted in favor of increasing the minimum wage more than we might expect given how they voted in the presidential election, yellow areas voted as expected, and redder areas represent places that were more against the minimum wage increase than their Trump voting predicts.
This shows us that other factors that can influence how people feel about increasing the minimum wage beyond just how they feel about Trump, and these factors are spatially related**. For example, it seems that people from higher-income areas such as Mercer Island, and Seattle's Northwest coastline were more against a minimum wage increase than their Trump voting would predict, and people in the interior and South were more for the minimum wage increase than their Trump voting would predict. If we wanted to improve our regression model, we might add income information to our model and potentially find that I-1433 voting behavior at the precinct level was driven both by average income and by proportion of Trump voters.
Data:
Results
Spatial Data
Note: The final election results will be in on November 30th, 2016 and this will be updated after that to reflect the final results.
*Using a weighted least squares regression to correct for heteroskedasticity observed in an linear least squares regression (Breusch–Pagan p < .05).
**Moran's I, a measure of spatial autocorrelation, indicates that there is significant clustering occuring in the map of the residuals. (I = .197, p <.001). Local indicators of spatial assocaition showed what you'd expect: the clusters in the suburbs to the Northeast and in South Seattle are significant, as is the cluster on Mercer Island.