1 Introduction 🗳️

The inspiration for this project came during the 2018 midterm election when I came across a Twitter account called @PizzaToThePolls. On Election Day, Pizza to the Polls was tweeting about delivering pizzas to polling locations across the United States. These tweets, like the one shown below, contained street addresses of polling locations where pizzas were being delivered to motivate voters to stay in line to cast their votes.

There are good reasons to be curious of polling locations. Lengthy lines clearly suggest that there is not enough voting booths or locations to accomodate the size of the voting population in the district. One could either interpret this as an unfortunate result of local government’s lack of resources or, less benignly, a predetermined outcome of an attempt to suppress voter turnout. Since there is neither a public databasenor has there been a study attempting to collect polling location addresses at national level, voters do not have a full picture of voting access.1

I have collected, parsed, and geocoded 169 unique street addresses contained in the tweets posted by Pizza to the Polls. Unfortunately, the sample size is too small for the data to be used to draw generalizable inference on voter access. Therefore, I have narrowed down the scope of this project to answer the question, “What does the data tell me about Pizza to the Polls’s level of outreach?” I find that districts that contacted Pizza to the Polls for pizzas were holding key Senate races and where women candidates were running for office.

The rest of the report proceeds as follows:

  1. briefly introduce Pizza to the Polls and their activities;
  2. explain the data collection method;
  3. create an interactive map of the polling addresses;
  4. conduct exploratory analysis (bag-of-words and regression);
  5. present recommenations for how Pizza to the Polls can improve its outreach for the upcoming election;
  6. discuss the limitations and avenues for future research.

2 Pizza To the Polls

logo

Pizza to the Polls (abbreviated PTTP) is a 501(c)(4) nonprofit organization that receives reports about long polling lines on various social media platforms and requests pizza deliveries to polling locations.2 Although PTTP is not officially affiliated with any political party in the United States, there is an obvious partisan flavor to their Twitter activity (more detais below). To be clear, unlike 501(c)(3) organizations, 501(c)(4) organizations are allowed to engage in lobbying activity so long as the purpose is to promote their social welfare objectives. They do not lose their tax-exemption status from engaging in electoral campaign activities (i.e. activities influencing electoral results or advocating for a candidate) so long as those activities remain secondary to their primary missions.3

3 Data

The data used in the analysis comes from four different sources: tweets posted by Pizza to the Polls, election results from Wikipedia, Cook’s Political Report rating for election competitiveness, and women candidate profiles from Center for Women in American Politics.

3.1 Searching for tweets

First I made a query asking how many tweets were posted per hour by PTTP from 00:00 of November 6th, 2018 to 02:00 of November 7th, 2018.4 Looking at the hourly counts, I decided to restrict the timeframe from 11:00 and 23:59 of November 6th, 2018. In total, I retrieved 28,329 raw tweets. The figure below shows the hourly distribution.

Next, I processed the texts to remove any special characters, url’s, and @’s, and parsed the text to find street addresses. This reduced the number of relevant tweets to 1,097, which is about 5% of the original sample of tweets. The figure below shows the hourly distribution of tweets containing street addresses.

3.2 Geocoding addresses

I used Google’s Geocoding API to parse the addresses and locate the geocoordinates of the addresses. From the API response, I scraped the street number, route, state, county, zipcode, longitude, and latitutde.

According to my sample, on 2018 Election day PTTP delivered pizzas to 169 unique addresses across 37 different states and 101 different cities.5

3.3 2018 midterm election data

I scraped the 2018 midterm congressional and gubernatorial election results from their Wikipedia pages. For all House, Senate, and gubernatorial races, I scraped the winning candidate’s and the second runner’s names (candidate1 and candidate2), parties (party1 and party2), and final voteshares percentages (share1 and share2). There are four congressional districts that had special elections prior to November, namely OH-12, DC-98, PA-18, TX-27, AS-98, AZ-08, GU-98, VI-98, MP-98, PR-98. Twenty-three states did not have any Senate elections, and twenty states did not have gubernatorial elections.

I also consulted the Cook Political Report’s most recent ratings prior to November 6th, 2018 to label whether the race was considered a key seat for the Democratic Party’s takeover of the legislative branch (key).6 If a race was ranked as leaning to either party or a toss-up, I labeled the race as a key race.7 Lists of key House, Senate, and governors seats are presented below.

## Key House races: 
##  AZ-02, CA-10, CA-25, CA-39, CA-45, CA-48, CA-49, CO-06, FL-15, FL-26, FL-27,
## IA-01, IA-03, IL-06, IL-14, KS-02, KS-03, KY-06, ME-02, MI-08, MI-11, MN-01,
## MN-02, MN-03, NC-09, NC-13, NJ-03, NJ-07, NJ-11, NM-02, NV-03, NV-04, NY-19,
## NY-22, PA-01, PA-07, TX-07, UT-04, VA-02, VA-07, VA-10, WA-08
## Key Senate races: 
##  AZ FL IN MN MO MS MT ND NJ NV TN TX WS WV
## Key governor races: 
##  AK CT FL GA IA KA ME MI NM NV OH OR SD WI

Because 2018 was noted for the number of female candidates on the ballot, I also record the number of women candidates who proceeded to the general election for each race. In total, 667 women were on the ballot for House races, 54 women for Senate races, and 52 women for gubernatorial races.

4 Interactive RShiny map

I created an interactive map with tmap package (v2.3.1) in R and published it as a RShiny app. There are markers placed on polling locations where pizzas were delivered; the number on the markers counts the number of deliveries made to that location. The map has three layers. The layers color the districts or states that held key races for House, Senate, and governor seats.

2018 midterm pizza deliveries

5 Analysis

One goal of this project is to learn how PTTP—as a nonprofit organization working towards increasing voter turnout—can improve its level of outreach, which I measure by counting the number of congressional districts that requested pizza deliveries.

Since there is most likely some selection mechanism behind the kind of voters who would know about and contact Pizza to the Polls on Election Day, some of these characteristics should be reflected at the district level such as

The main outcome of interest is whether a district requested and received pizzas from Pizza to the Polls at any of its polling location. First, I will show some pairwise plots for preliminary analysis to see which varaibles are most likely to have strong associations with the outcome and whether there is any strong collinearity among the covariates.

5.1 Pairwise plots

The first panel shows pairwise plots for numerical variables, namely the outcome variable, logged total population (log total pop), logged median household income (log median house inc.), logged Black population share (log Black pop.), logged population share of age group 18 to 24 (log age 18-24). None of the pairs have strong correlation to worry about multicollinearity. log Black pop. and log ages 18-24 are positively and negatively correlated with the outcome respectively.

The histogram below counts how many poling locations were identified via Pizza to the Poll tweets per district. It shows that a vast majority of congressional districts (approximately 80 percent) did not have any of its polling locations identified thorugh our data collection method. On the other hand, the distribution is skewed, suggesting some campaigns had volunteers who actively used the nonprofit service to keep voters stay in the line with pizzas.

5.2 Propensity score outliers

Even though I have access to the number of deliveries, I chose my outcome to simply indicate whether the district had any requests for pizzas to Pizza to the Polls. Most districts did not engage with Pizza to the Polls. And for those that did, I imagine that once someone in the line requested for pizzas, there was a high chance that another pizza delivery was made shortly by a word of the mouth. In light of this, the more interesting outcome to study is the binary outcome. I create another outcome variable, zero for each district \(i\) where \[ \texttt{zero}_i = \begin{cases} \texttt{TRUE} & \; \text{if }\texttt{count}_i = 0 \\ \texttt{FALSE} & \; \text{if }\texttt{count}_i > 0 \end{cases}.\] I will use a logistic regression to model this new outcome variable against election-related predictors, namely key_rep, key_sen, key_gov, women_rep, women_sen, and women_gov.

I am also interested in identifying districts where no pizza deliveries were requested even though their demographics and elections were similar to districts that did have pizza deliveries. These may be districts where more voter turnout could have influenced the result, and that is an outcome that Pizza to the Polls would be interested in investgating. Therefore, I am going to plot the distribution of propensity scores of observations with zero == TRUE against the propensity scores of observations with zero == FALSE.

Here, the propensity score is the probability of having zero == TRUE. Naturally the mass of bottom histogram is shifted towards the left and the mass of top histogram towards the right. However, there are also a few observations with relatively low propensity scores in the top histogram. I define missed opportunities as districts with zero counts that have propensity score below 0.65. The table below lists these missed opportunities. To make the values of numeric variables more comprehensible, I have binned these variables into deciles (medianHouseInc_q, totalPop_q, blackPopShare_q, age18to24_q).

None of the districts in the table had key House races, continuing the trend we’ve seen in the main result, but they still had woman candidate running from the Democratic party including LA-03, LA-06, NC-07, and NC-12, where the (male) GOP incumbents won at comfortable margin. This suggests to me that these are districts would have welcomed the help of nonprofits like Pizza to the Polls.

6 Lessons for the 2020 election

If I were working at Pizza to the Polls and was tasked with improving the organization’s outreach for the upcoming 2020 election, I would make two recommendations to their operation team. First is to pay attention to key Senate and gubernatorial elections. Although the probability that the Democrats would take the Senate was low, voters waiting to vote for the key Senate elections were more likely to know and contact Pizza to the Polls than those voting in key House districts. On the other hand, Democrats were vying to win key governorships such as Floria. Georgia where Stacey Abrams was running for governor was also the state that had much controversy over deliberate voter suppression before the election. It is unsurprising that the largest number of pizza deliveries to a single polling location was observed in Georgia. Since senators and governors are elected by the whole state, reaching out to candidates running for these offices is a more efficient way of spreading information about Pizza to the Polls.

My second recommendation is to reach out to women candidates. All the districts named as missed opportunities were districts where a woman was running for the House, Senate, or governor. Past studies have shown that women are less likely than men to run for office. Even though 2018 saw a remarkable number of women challengers on the ballot across the country, they were still more likely than their male counterpart to have had experience as elected officials. Women tend to set higher bars on themselves when making a career choice in politics, and they still face explicit and implicit sexism while on the campaign trail and in office. Thus, although many of the missed opportunities were not one of the key races of 2018 Midterm, supporting women breaking into higher offices would naturally align with PTTP’s democratic agenda as evidenced by my bag-of-words analysis. I would recommend Pizza to the Polls to actively engage with campaigns of women candidates and spread awarenss of the organization.

  1. The only paper I know of that involves polling location data is Brady and Mcnulty (2011) “Turning Out to Vote: The Costs of Finding and Getting to the Polling Place” American Political Science Review, but the scope of the data is restricted to Los Angeles County.↩

  2. This article from zapier explains how the organization delivers pizzas.↩

  3. From the IRS website, https://www.irs.gov/charities-non-profits/other-non-profits/social-welfare-organizations.↩

  4. It seems that PTTP does not have geolocation enabled on its Twitter account, so it is not possible to tell the time zone. Even if I assume that the account activity is generated in Pacific Standard Time, the timeframe that I selected would adequately cover the last polling location that was open in the mainland United States.↩

  5. On their website, PTTP says that in 2018 they delivered pizzas to “611 polling places across 41 states.” The discrepancy between my numbers and theirs may be due to their counting pizzas delivered to special elections held at other times of the year.↩

  6. “2018 House Race Ratings,” The Cook Political Report, October 30, 2018, link; “2018 Governor Race Ratings, The Cook Political Report, October 26, 2018, link;”2018 Senate Race Ratings," The Cook Political Report, October 26, 2018, link.↩

  7. I compiled a list of key House, Senate, and gubernatorial races mentioned by various media outlets and were rated as either toss-up, leaning Republican, or leaning Democratic by the Cook Political Report. “The Washington Post’s Senate Race Ratings,” Washington Post, November 2nd, 2018, link; “The Top 10 House Races of 2018,” Washington Post, March 18, 2018, link; “The top races for control of governors’ mansions,” Vox, November 19, 2018, link; “The battlefield to control the House of Representatives is huge,” Vox, Novebmer 19, 2018, link; “The most contested Senate races,” Vox, November 19, 2018, link.↩

  8. The socioeconomic data comes from U.S. Census 2017 American Community Survey retrieved by tidycensus 0.9.2 R package.↩