Preliminary data analysis on the voter registration and suspended registration files during the 2018 election cycle in Georgia indicate differences between the profiles of those who were affected by the state’s strict approval code and those who cleared the application. In this article, we discover that the most distinguished difference between the Georgians who were registered between July 1, 2017 and October 10, 2018 and those who failed the state’s exact match procedure was their self-identified race. Even without determining causality, the total number of those who were placed under the pending list were likely did not produce enough effect to change the outcome 2018 Midterm Election. However, culturally, these findings call upon a decades-old, institutionalized distrust between civil and voting rights among the United States electorate.


On April 2, 2019, Georgia governor Brian Kemp signed HB 316. This bill would remove most of an existing state code about voter registration applications which failed to be completely erradicated before the 2018 midterm elections despite legal and political push-back on the policy. Under this legislation, through the 2018 Midterm Election, the board of registers could place registration applicants on a suspended registration list following failure of an exact match between the applicant’s input data and the applicant’s internal data from the Department of Driver Services (DDS) or the federal Social Security Administration. This policy, found under Article 21, Section 6 of the GA Code, was previously modified on July 1, 2017, under HB 268. Under the 2016 and 2018 election cycles, two versions of the code were in effect.

GA Code § 21-2-220 (2016)

“If an applicant fails to provide all of the required information on the application for voter registration with the exception of current and valid identification, the board of registrars shall notify the registrant in writing of the missing information… In the event the elector does not respond to the request for the missing information within 30 days, the application shall be rejected.”

GA Code § 21-2-220.1 (2018)

“If a completed voter registration application… cannot be verified… the applicant shall be notified that the number cannot be verified… If the applicant has not provided such sufficient evidence or such number has not otherwise been verified on or before the date of a primary or election, the applicant presenting himself or herself to vote shall be provided a provisional ballot… counted only if such number is verified by the end of the time period set forth in subsection (c) of Code Section 21-2-419”

From July 1, 2017, until October 10, 2018 (the last date to be registered in order to vote in the 2018 general election in the state of Georgia), just over 15,000 applicants were added to the list of pending registrants in the state of Georgia for failure to match the state’s information on file. The total number on the list would hover around 55,000 total pending applicants as election day approached. Although these applicants could still vote, they would need to follow strict compliance standards in order to validate thier identity and practice their right. One of these methods includes receiving a provisional ballot and requiring proof of identity within an arbitrary certain date to be determined by the registration office. So far, there is minimal evidence to support that this code was a vetted process meant to facilitate the extended voting process pending registrants must face.


We shall only consider pending applicants and registered applicants who registered to vote for the first time between July 1, 2017 and October 10, 2018. This does not include those who were “registered” between 2017 and 2018 but, based on the voter file, had a date_added value before the July 1, 2017 cut-off. Considering that updating your voter registration is a different process than registering, and under the assumption that the state of Georgia only subjected those who had not already registered to the exact-match policy, we assume that the data we’re examining comes from first-time applicants who are joining the electorate for the first time.

What results is 555,691 records of people who registered on or between July 1, 2017 and October 10, 2018 and 14,926 records of people who registered on or between July 1, 2017 and October 10, 2018 but were put on the “pending” list. From here on, I mostly refer to those people in the former group as just “Registered” and the people in the later group as “Pending.” However, all analysis done is only on this subset of 570,617 combined records.

Who are these applicants?

Surprisingly, there is more similarity among the registered and pending records than there is not. As referenced in the abstract, the striking difference among these groups mostly deal with race. Not just any racial difference, though–a significant imbalance of white versus black people.

They have similar ages

Georgia is a participant of Automatic Voter Registration (AVR), which under the National Voter Registration Act of 1993 (“Motor Voter Act”) requires “states to provide individuals with the opportunity to register to vote at the same time that they apply for a driver’s license or seek to renew a driver’s license.” Therefore, it’s typical to see many young applicants within the span of a year who have elected to register through their driver’s license. However, while this explains the age distribution for the registered applicants, this can’t easily explain why those who are pending are also so young. The exact-match policy that placed them onto the pending list checks to see if your records match with the Department of Driver Services (DDS). Anyone who is registering through their driver’s license already has an application that matches their driver’s license. While registering to vote can be potrayed as a “coming of age” act many young people participate in, the drastic amount of 18 year olds registering compared to every other age being placed on the pending list is interesting.

Name characteristics play a small role

One major reason why someone could be placed on the pending list is due to a unique name or name spelling. If applications are most often input manually onto a computer, characterisitcs like an unusual name, hyphenation, or other forms of punctuation can easily be mistyped. This can disproportionately affect those who are more likely to have these characteristics in their names, for example, apostrophes in African American names and hyphens in the last names of those who are Hispanic. A bootstrapped loop gathering the number of unique names, number of last names with hyphenation, and the number of first names with an apostrophe for random samples of 14,000 records with replacement from each group shows that the average of these group samples have significant differences between them. Even though the number of people on average with apostrophes in their first name in the pending group is double the average of the registered group, that is still only a proportion of about 1%. So, while these differences are there, they are not dramatic. The potential for a name with an uncommon character to be mispelled, although a serious factor, does not give the most likely reason why people ended up on the pending list. To examine this effect more thoroughly, language processing should be done on the composition of the names themselves. A “unique” name might not be mistaken if the people who are manually inputting data are familiar with the origin of the “unique” name.

They are not similar in race

When looking at the marginal proportions of the pending list and the registered list, one stark difference stands out: the proportions of white and black people on the pending list versus the the registered list. Of all people who were successfully registered, 39% were white and 27% were black. However, in the pending list, only 11% were white and nearly 68% are black. If the exact-match policy is no worse than a random mistake on erroneous applications (under something akin to a naive null hypothesis for this problem), then what would cause this difference? Could the small variation in frequency among name characteristics examined above and their correlation with race really cause such a difference? And if it all did come down to the relationship between race and name composition, why are most with unique name characteristics matched, yet we still see this result?

How was turnout affected?

It is impossible to ask every person on the pending list if their status affected their voting practice, as it is to determine causality with the cross-sectional nature of the data. However, looking upon other attempts to examine the impact of pending status on voting behavior (Biggers, Smith), I follow by propensity-matching our two groups and analyzing the differences that still are present even when accounting for imbalanced covariates. When not accounting for these differences (and assuming under a naive null hypothesis that the pending list resembles a simple random sample from the registered group), there is a worse effect on turnout.

Using the pymatch library, I begin by finding a subset of the registered group that most closely resembles the pending group statistically.

No clear relationship between location and turnout

To preface the analysis on the graph below, the data was first filtered by selecting only those counties of which there were more than 20 people on the pending list. Many counties, especially those in rural areas, may only have a handful of people on their pending list. In that case, a turnout rate of 0% or 100% is not just common, but likely. I felt this added unrepresentative noise to the graph, so I excluded these counties.

Each line represents a county. Starting from the left of the graph, each county is placed on a scale of its approximate distance from Fulton County using “distance as the crow flies,” or Euclidean distance, on the centroids of the latititude and longitude of each county. The lighter the hue, the closer that line (representing a county) is to Fulton County. Tracking the lines along the graph, we see there’s no clear relationship between county location and the turnout of the propensity-score matched registered list. However, moving along to the actual turnout of the pending list, we see that there is a small negative relationship between propensity-score matched turnout and actual turnout.

Outliers contain most significant differences

Unlike the graph above, the graph below contains every county’s data. The y-axis represents the expected number of people to vote in that county. This expectation is calculated by taking the turnout rates by county of the propensity-matched registered group and multiplying it by the actual number of those on the pending list for that county. The x-axis represents how many people actually voted on the pending list. The gray line represents the point at which the expected number of voters is equal to the number of voters on the pending list. Any county above this line had lower turnout than expected. The graph on the right is identical to the one on the left; the only difference is that it’s zoomed in to avoid outliers. One can see that as the expected turnout decreases, the actual turnout tends to fit closer. Some counties even appear right on the line. However, outlier counties like DeKalb, Fulton, Gwinnett, and Clayton who had much lower turnouts than expected produce the largest effect. Of the net 3415 less votes than expected in the entire state, 2578 of them came from those four counties. These happen to be some of the most populous counties in Georgia, so this makes sense. However, it’s common knowledge that these counties house some of the most disadvantaged persons and largest minority groups of the state, leading to disproportionate consequences.


In this article, we have observed that the most distinguished difference between the Georgians who were registered between July 1, 2017 and October 10, 2018 and those who failed the state’s exact match procedure was their self-identified race. Even without determining causality, the total number of those who were placed under the pending list were likely did not produce enough effect to change the outcome 2018 Midterm Election. However, counties with the highest populations (like Fulton, DeKalb, Gwinnett, etc.) were disproportionately affected compared to other counties. With propensity-score matching on the registered list, we found negative differences between the actual turnouts and expected turnouts. More thourough analysis with the propensity-matched group will provide insight on how the effect pending status had on the voters of the 2018 election in Georgia compared to those with pending status in the 2016 election.


Biggers, D. & Smith, D.(2018). Does threatening their franchise make registered votersmore likely to participate? Evidence from an aborted voter purge. British Journal of Political Science (2018), 0: 0, 1–22doi:10.1017/S0007123418000157.

Ga. Code § 21-2-220.1 (2018)

Ga. Code § 21-2-220 (2016)

National Voter Registration Act (NVRA) of 1993