Registering Re-Entering Citizens to Vote

We will contact previously-convicted people prior to fall municipal elections, offering them information about eligibility to vote and a registration form and stamped envelope to return that form to the board of elections.

We will scale up the previous pilot study for a larger trial prior to the 2020 general election. We use a between-subjects design with five groups: 1) control (uncontacted); 2) contacted with a letter from our partner organization with felony-targeted introductory language, information about eligibility, a registration form and stamped envelope; 3) treatment 2 minus the felony-targeted introductory language; 4) treatment 2 minus the registration form and envelope; 5) treatment 2 plus a stronger encouragement to register that informs recipients of the importance of voting to elect representatives who will shape civil rights and criminal justice policies.

We are also conducting a second trial focusing on individuals who live in the same zip codes as those in our main study, but who administrative records do not indicate have previous felony convictions. We can use the results of this study to compare responses in our population of interest (those with previous convictions) to those in the broader population. This trial will have two arms: 1) control (uncontacted); 2) a letter from a partner or us with information about eligibility, plus a form + stamped envelope. (treatment arm 3 from the main study).

2019-08-28

2020-11-03

Voter registration (appearing on the state's public voter file) and voting in municipal elections where relevant, as well as longer-term voter participation in the next statewide election.

For the scaled up trial, we will collect the following outcome variables after each state’s voter registration deadline and after the 2020 General Election: voter registration and whether an individual casts a ballot in the election.


For the larger scaled up project we will also collect data for voting in the General Election.

We will randomly divide people in a sample of previously-convicted residents of North Carolina into two groups; the "control" group will not be contacted, while the "treatment" group will be sent a letter telling them about the voter registration process and encouraging them to register if eligible.

For the scaled-up study we will use a between-subjects design with five arms, one of which will be an (uncontacted) control group. The study will be fielded in North Carolina and Texas. We will randomize at the individual level with equal probability across the five treatment groups, blocking on state, using the "randomize" function from the "experiment" library in R. The comparison (no-record) group is randomized (individually) with equal probability into treatment and control.

In North Carolina, people who have been convicted of felonies are eligible to vote after they have completed their sentences (including any probation or parole). North Carolina’s Department of Public Safety (DPS) provides publicly available data containing personal information for everyone who has been in NC DPS custody since 1972. We will use these data to identify people who were convicted of a felony and have completed the terms of their sentence (248,472 records as of early 2019). We will then use the publicly-available North Carolina voter file to identify and remove those who are already registered to vote from the data. The resulting dataset will include individuals who have completed the terms of their sentences and are, therefore, eligible to vote, but who are not registered to vote. These will be the people we seek to contact in our pilot studies.

In our first pilot study, described here, we will draw a random sample of 10,000 people from this dataset of individuals who have completed sentences and are not registered to vote. We will then contact a data vendor to attempt to find current mailing addresses based on their names and date of birth; we will retain everyone matched to an address for assignment to treatment/control. Then, before completing treatment assignment, we will remove individuals residing in the 3rd and 9th Congressional districts. We do this because special elections are being held early in September in these districts, and we worry that our mailer could reach people in these districts after the registration deadline for these elections has passed, thus confusing them about their ability to vote in the special election. We will then randomly assign the remaining individuals in the sample (with valid addresses) to treatment or control conditions with equal probability.

Our treatment will consist of a letter describing the eligibility criteria for registering and voting and encouraging people to register and vote, including a blank voter registration form and a postage-paid envelope for returning it to the local elections office. We will send one letter to everyone in the treatment group; we will not contact the control group. After a pre-specified period of time, we will collect our outcome measures. We will observe whether people have registered to vote by collecting another snapshot of the North Carolina voter file and searching for them. We will also plan to collect voter turnout in the next statewide election. For both outcomes, our first analysis will be a simple difference-in-means comparison between the treatment and control groups.

For the scaled up 2020 trial:
We focus on North Carolina and Texas due to the states’ felony disenfranchisement policies, accessible administrative data, and the availability of potential research partners. In North Carolina and Texas, people who have been convicted of felonies are eligible to vote after they have completed their sentences (including any probation or parole).

North Carolina’s Department of Public Safety (DPS) provides publicly available data containing personal information for everyone who has been in NC DPS custody since 1972. We use these records to identify people who were convicted of a felony and have completed the terms of their sentence.

We then work with a commercial data firm and use the publicly-available North Carolina voter file to identify and remove those who are already registered to vote from the data.The resulting dataset includes individuals who have completed the terms of their sentences and are, therefore, eligible to vote, but who are not registered to vote. These will be the population--of non-registered previously-sentenced people--for whom we will attempt to find addresses and then seek to contact in our study.

We used the following steps to obtain our sample:
We downloaded datasets from NC DPS on July 19, 2020 (https://webapps.doc.state.nc.us/opi/downloads.do?method=view). According to the NC DPS website, these files “contain all public information on all NC Department of Public Safety offenders convicted since 1972.” We began with the dataset of all offenders (specifically, “Offender profile”; OFNT3AA1.dat; 1,205,971 rows) and then joined the inmate and supervision profile tables.
We then dropped records of individuals that met the following criteria:
Still incarcerated
On parole
Died in the system
Not convicted of a felony
Still under supervision (parole, probation)
Possible Absconders/ escapees / out of state
From that set we then removed obvious duplicate records using full name and date of birth (many of these were “duplicates” because they lacked first or last names or both), leaving us with a set of 328,171 records.
We then dropped anyone born in 1949 or before bringing us to 299,100 rows (i.e. we retain people who are at present less than age 70). We did this to lower the chances of sending mail to people who are, in fact, deceased.
We then dropped:
281 people still in the dataset with no names (now at 298,819 rows)
9,047 known non-US citizens (now at 289,772 rows).
The final set of 289,772 records represents what we believe to be the pool of individuals who have felony convictions, have completed all terms of their sentence, and who are eligible but have not registered to vote in North Carolina.

From this set we removed all observations that were used in previous pilots of the study. This left us with 136,268 observations, which we sent to our data firm to find mailing addresses. The firm sent back 35,249 observations that had mailing addresses.

We followed a similar procedure in Texas. We requested and received the Conviction Database from the Texas Department of Public Safety in July 2020.

The database consists of a number of data tables, each covering a specific area of the data, including: individual identifying information for all those in the system; information on those who have been in prison; information on those who have been on probation; and information on those who have been subject to neither. As this project is focused on registering those who have previously been convicted of a felony, we drop all misdemeanor charges from the data.

We remove duplicate observations, including only each individual’s most recent experience with the system, and then focus on those observations for which we can identify an end date to their prison, parole, and/or probation and for which those dates are not at some point in the future.

We then merge these with the datasets that include other personal characteristics, yielding 1,937,305 observations.

Finally, we remove any individuals who are under 18 years of age or over 100 years of age, look for and drop any remaining duplicates or observations with supervision end dates in the future, and remove any observations for which the last interaction with the Department of Public Safety was a deferral. The final data set of individuals we believe are eligible to register to vote and therefore eligible to be in this trial includes 1,429,264 observations. We sent a random set of 1,000,000 of these observations to a voter-data firm.

This data firm matched the observations we sent them to the Texas voter file. The dataset they sent back showed that 776,575 from our list were not on the state’s voter file, and, therefore, not registered to vote: these unregistered people were retained for the study.We then dropped from the data any individuals over the age of 70 in order to match our data from North Carolina, which left 625,412 observations.

Finally, we pulled a random sample of 250,000 observations to send to our data firm (along with the North Carolina data) to find mailing addresses. The firm was able to match 163,160 of these observations to addresses and sent us a random sample of 89,751 of those matched addresses.

The observations with matched addresses from North Carolina (35,249) combined with the observations with matched addresses from Texas gives us a total sample size of 125,000 for the main study.

Our data firm also supplied us with a list of 37,000 unregistered North Carolina residents for our second study. In this study we will test the effectiveness of treatment three on individuals who live in similar neighborhoods to the people in our main study, but do not appear to have been convicted of or incarcerated for a felony (given the information we have available to us). These records were generated by drawing all non-registered voters known to our data firm from a set of six zip codes in North Carolina: these were the six most common residential zip codes among our main-study sample. Then, the data firm suppressed any records that appeared on our main-study list, to avoid overlap, and provided us with a random sample of 37,000 of the resulting addresses. We then checked the resulting list against our records from previous pilot studies to ensure that no one in the comparison-group sample was known to have been in DPS custody for a felony, and discarded several hundred observations based on name and address similarities to people that had been in our pilot study samples. This process resulted in a comparison-group dataset of 35,708 people that would be assigned to treatment or control.

Randomization done in office by a computer.

individual

For the scaled-up trial: individual, blocking on state

No

8,621 people for whom we have addresses in the initial sample.

Across the two trials all conditions and control groups, we expect a total sample size of about 160,703 individuals in the 2020 trial.

same as clusters (individual randomization)

people will be randomized with equal probability into treatment/control conditions (so n/2).

People will be randomized with equal probability into treatment/control conditions, blocking on state, for the 2020 expansion.

With respect to power, we have performed power calculations aimed at testing the sample size needed to identify differences between treatment arms. We assumed a baseline control-group registration rate of 2%, which was the average registration rate across all three of our pilot studies. We believe that in our main study, the sample size of 125,000 people, with 25,000 in control, should allow us to detect a minimum treatment effect of .28 percentage points. Comparing the control to any one of the treatment arms (each of which included 25,000 people) will allow us to detect a minimum treatment effect of .35 percentage points. These minimal detectable effects are substantially smaller than the effects we saw in our pilot studies. In our second study, with a total of 35,708 individuals equally split between treatment and control, we should be able to detect a minimum treatment effect of .41 percentage points.