OH & GA Vote by Mail Modeled Projections

Today, Citizen Data is releasing the results of new vote by mail models in Ohio and Georgia.

In both states, we project that vote by mail turnout will be higher than 2016, and, in both states, Democrats are expected to comprise more than half of all mail-in ballots.

However, each state has idiosyncrasies that benefit from state specific modeling, and the types of people who will vote by mail and where they will turn out varies state to state. For example, while urban areas are likely to see high vote by mail turnout in Georgia, rural areas are more likely to see higher levels of vote by mail turnout in Ohio.

Here are some state-specific highlights:

  • Ohio is projected to have as much as 3x the VBM turnout of 2016 (18%), with up to 47% of likely voters voting by mail.
  • Democrats in OH will drive more than 50% of mail-in voting; in contrast, Republicans will drive almost half of in-person voting.
  • VBM rates among likely black voters will be notably low.
  • Rural areas are likely to have higher VBM turnout levels than urban areas — an inversion from the dynamic in other battleground states.

  • We expect a large increase for VBM in Georgia, at 44.1% among likely voters — nearly 10 times the rate of 2016 (which was 4.9%).
  • Democrats will comprise the majority of mail-in ballots, but in-person voters will be more evenly split by party.
  • Likely white voters in the state will have the lowest rates of VBM; Hispanic voters are projected to have the highest rates.
  • Atlanta is expected to have the highest VBM turnout; rural areas, in contrast, will see lower VBM turnout.
To produce these insights, Citizen modeled the likelihood that each registered voter in Ohio and Georgia would vote by mail or in person. To do this, we first created a custom turnout model tailored to each state. For each state, we used a flash-forward model with 2016 turnout and 2018 turnout as separate dependent variables. For each flash-forward model, we trained an ensemble of machine learning algorithms and combined the results of each to generate a final score. Using our live feed of absentee ballot requests, we trained an additional set of machine learning algorithms to predict the likelihood that those individuals who have not yet requested a ballot would do so.
For those who have already requested a ballot, we assumed that they would vote by mail at a rate equal to the mean of that group’s turnout likelihood; e.g., if 1,000,000 people had requested a ballot in Ohio and those million voters had a mean turnout likelihood of 80% in the model, we would project that 800,000 voters would vote by mail. For the remainder of registered active voters who had not yet requested a ballot, we assumed that individuals would request a ballot if their ballot request likelihood as predicted by the model was above 50%. Then, we applied the mean of the turnout likelihood among that group of voters in the same manner as described above; e.g., if another 1,000,000 people were modeled as likely to request a ballot in Ohio, and that group’s mean turnout likelihood was 60%, we would project that an additional 600,000 voters would vote by mail. In this scenario, we would have projected 1.4 million voters would vote by mail.

The insights from these models are part of an ongoing Citizen Data effort to support election administrators and non-partisan non-profits with the credible, unbiased data they need for strategic resource allocation in the months and weeks leading up to the election. On an ongoing basis, Citizen will release updates to these state models as well as new projections in additional states and other key insights and analyses.

For this project, Citizen is partnering with a number of organizations — including the National Vote at Home Institute and the Stanford-MIT Project on a Healthy Election.