Elections Correlation

In this next section, I’d like to leverage elections data to look at the relationship between the number of polling places in a state and which party won elections in that state. The elections data have been pulled from FEC.gov. In the data, I look at the breakdown, by party, of the House, Senate, and Presidential races and the total number of votes for each in every general election between 2012 and 2018, using that data to match onto the polling places data.

I’ll do so by determining a new variable defined by the overall races won by a political party in a given state. For example, if in New Jersey 2018, Democrats won the House and Senate races, Democrats controlled that election year in NJ. If in Wisconsin, a Republican (Trump) won the Presidential election in 2016, but Democrats overall won the House races, then it’s a “Split Ticket” state.

Let’s start by getting a sense of this new variable. Which party won each state in the past 4 elections?

Now that we better understand how this variable works, let’s correlate this against the number of polling sites available to voters in that state. Now that we have our dataset, we’ll build 4 scatter plots (one for each year) of polling places by number of total votes.

Let’s take another look at our top 10 counts from earlier to see if there’s a correlation with which party won a given state.

Interestingly enough, in the most notable election covered by our dataset – the 2016 election – states that ended up going full Republican had the most polling places per capita. The state with the most polling places per capita in 2016 – Wisconsin – was notable in that it swung for Trump, much to the Hillary campaign’s surprise.

To be honest, this finding surprises me a bit, especially given a lot of the rhetoric that Republican states go out of their way to make it more difficult for individuals to vote. Obviously this doesn’t tell the full story. Just because there are more polling places does not mean that they are accessible, or that there are a lot of people eligible to vote (because of difficulties in the voter registration process). Still it is a metric that helps understand the fuller picture.

Let’s take a look at our “worst” 10 states to see if this trend reverses.

Looking at the bottom polling places makes the argument a bit more muddled. It’s overall a healthier mix of states that went full-Democrat and full-Republican (along with a fair share that had mixed results).

Let’s take a closer look at our House races. Specifically, let’s tie together all three measures we’ve previously been discussing:
1. Number of Polling Sites
2. Number of Votes Cast
3. Total Population

In terms of trend, I can’t see much distinction between states where Republicans won the House versus Democrats. Let’s test this formally using regression. In particular, let’s fit a regression model for House elections, using party, number of votes cast, and population as predictors. We expect, based on the plot above, to see highly correlative relationships between number of votes cast, population, and number of polling places. The real question lies in party.

Here are the results from running a basic regression:

Looking at the model summary, there are a few interesting things to point out about trends in the number of polling sites in a given state.

1. Population | The most important thing to note is that there is only one factor that adequately predicts the number of polling sites: population. The model states that for every additional million people in a state, there will be 577 additional polling sites.
   
2. Total Votes Cast | Now, if you look back at the bubble plot above, you’d probably (like me) assume that the number of total votes cast is also a strong predictor for number of polling sites. As one increases, so does the other. Why then is the p-value (denoted in the model output as PR(>|t|) ) .6867 – much higher than the usual .05 threshold used to determine statistical significance? Well, it’s also true that the number of total votes cast is highly correlated with the population of a state. States with larger populations naturally have more people turn out to vote. Thus, using both features as predictors is redundant.
   
3. Party | The question we’ve all been waiting for. Is it true that the party that wins the House elections in a given state is determinate of the number of polling sites in that state? Turns out, not so much. The p-value for party winner (denoted party_winnerRepublican, representing the base case in which a state goes Republican) is .3694, well above the usual threshold of .05. Thus, we cannot reject our initial null hypothesis that party winner predicts number of polling sites.

Future Analysis

This article really is just a small look at one facet of elections: polling places. Where else should this article go? Well, if I could get my hands on elections data smaller at the county-level or even more granular, I could run a lot of the previous analyses on a much more robust dataset. Unfortunately, that data is expensive.

Otherwise, a good route for this analysis would be to compare the availability of polling places with demographic data on a zip code or county level. The hypothesis here would be that white Americans have better (i.e. more) access to polling places.

Where else should I go with this? Leave your thoughts in the comments.

Blog Post Available on Medium:
https://jschulberg.medium.com/u-s-polling-places-389ca5ed7960