Predicting Voter Turnout Based on Election Infrastructure and Policy

Voter Turnout and Voter Suppression

Voter turnout refers to the percentage of the voting population that actually voted in an election. A low rate of voter turnout may yield election results that are not representative of the will of the actual populace of the region, thereby diminishing the legitimacy of the democratic process. It is important therefore to vote for policies and infrastructure changes that make the voting process easier.

Voter suppression is an election strategy that aims at reducing voter turnout from specific voter groups to influence election results. The tactics include making the registration and/or voting process more difficult either by policy or infrastructure changes that affect certain demographics more than others. Without the right data regarding the impact of policy changes on voter turnout, it can be difficult to detect and prevent voter suppression.

Our project uses both supervised and unsupervised learning methods to (a) predict voter turnout based on election policy and infrastructure, and (b) determine the most important features that impact voter turnout.

Voter Suppression Case Study: Fair Fight Georgia

Last fall, 2018, the State of Georgia experienced a divisive gubernatorial election between Stacey Abrams and Brian Kemp fraught with explicit and systematic voter suppression [1]. The tactics employed by Brian Kemp and republican operatives included exact-match purging of voters from rolls and premature closure of polling locations. These tactics saw 85,000 voters removed from the rolls in the three months leading up to the election and 667,000 removed during 2017 [2]. All told, more than 10% of eligible, primarily black, voters were removed from the rolls [2]. The systematic and widespread suppression netted Brian Kemp a victory while inspiring the Fair Fight movement. In the spirit of Fair Fight, this project endeavors to predict voter turnout using machine learning techniques based on district-level infrastructure data.

Dataset

We built a per-state election policy dataset using policy information obtained from NCSL. In particular, we included the following features:

1. No excuse early voting
2. All mail voting
3. Poll Station Closing Time
4. Automatic restoration of voting rights upon release from prison
5. Same day registration
6. Civics Course Requirement
7. Online Registration
8. Vote centers
9. No excuse absentee voting
10. Automatic Voter Registration
11. Pre-registration for young voters
12. Voter ID required

Labels: State Voter Turnout Percentage 2018

We further built a per-county election infrastructure dataset which included the following features:

1. No Voters per Polling Location
2. No Religious Congregations
3. Area per Polling Location
4. Percent Broadband Coverage
5. Purge Rate

Labels: County Voter Turnout Percentage 2018

Unsupervised Learning

We performed clustering on the state-wise voter turnout dataset across two dimensions: first we merged together similar features using FeatureAgglomeration for dimensionality reduction; next, we clustered together similar states based on election policy using k-means.

K-means

We employed the elbow method to determine the ideal number of clusters and then performed k-means on the state voter-turnout dataset to group together states with similar election policies:

Ideal number of clusters using elbow method = 3.

Silhouette Score for K-Means (higher is better) = 0.538.

Feature Extraction

Further, to determine which policies have the most impact on voter turnout, we performed Feature Extraction using the SelectKBest method which outputs the K features with the highest scores.

Supervised Learning

We used the per-county election infrastructure dataset to perform linear and lasso regression to predict voter turnout. Lasso regression is used to reduce model complexity and prevent over-fitting which may result from simple linear regression.

Lasso regression puts constraints on the coefficients (w). A penalty term (lambda) regularizes the coefficients such that if the coefficients take large values, the optimization function is penalized. So, lasso regression shrinks the coefficients and helps reduce the model complexity and multi-collinearity. In addition, Lasso uses regularization (L1) that can lead to zero coefficients i.e. some of the features are completely eliminated. So Lasso regression not only helps in reducing over-fitting but it can help us in feature selection.

LINEAR REGRESSION

Accuracy of the linear regression model:

Predicted vs Actual Voter Turnout:

LASSO REGRESSION

Accuracy of the lasso regression model using varying lambda values:

Best Lambda = 0.00001

Predicted vs Actual Voter Turnout:

Feature Extraction

We further used SelectKBest to determine the impact of each of the infrastructure features on voter turnout rate:

Related Work and Conclusion

Keeter et al. predicted voter turnout based on interviews, voter history, and demographics using random forest and logistic regression [3]. Challenor predicted voter turnout using labor force demographics in [4]. Unlike past work that focuses on immutable population characteristics such as demographic information, we predict voter turnout based on state election policy as well as local election infrastructure, including but not limited to distance from polling stations, as examined by [5].

Studying the impact of election policy and infrastructure on voter turnout may yield information that will help policy makers make better policy and budgeting decisions to improve voter turnout and help identify and prevent policy and infrastructure changes that result in voter suppression.

References

[1] M. Astor, “Georgia governor brian kemp faces investigation by house panel,” 2019.

[2] K. Shah, “Textbook voter suppression: Georgia’s bitter election a battle years in the making,”TheGuardian, 2018.

[3] S. Keeter, R. Igielnik, and R. Weisel, “Can likely voter models be improved?”Pew Research Center,2016.

[4] T. Challenor, “Predicting votes from census data,” 2017.

[5] M. Haspel and H. G. Knotts, “Location, location, location: Precinct placement and the costs ofvoting,”The Journal of Politics, 2005.