NATIONAL CHANGE-OF-ADDRESS PROCESSING AND AUTOMATIC REREGISTRATION

The NVRA, while mainly about facilitating registration for eligible nonregistrants, also requires states to maintain an accurate, up-to-date voter database. For accurate list maintenance, states can use the permanent change-of-address (COA) requests submitted to the U.S. Postal Service (USPS). Individuals submit such requests so that the USPS can forward their mail to the new address while they make necessary adjustments. The USPS maintains the last 48 months’ requests, called NCOA data, which approximate 160 million COA records with accurate old and new residences and when the individual moved/made the request. These data (1) determine the ARR treatment and (2) are the primary source for this paper.

California’s law (CA Elec Code § 2225) requires the Secretary of State to match the statewide voter file to NCOA data (called NCOA processing). If existing registrants have changed addresses within the state, the Secretary will then notify relevant counties. For voters’ privacy, COA request dates are coarsened to the month of the move instead of the exact date. The Secretary’s office disseminates the data monthly, and each county makes necessary updates to its voter file.

If a county finds that a voter moved within the state, the voter file is automatically updated and a confirmation postcard is sent to the new address. If this update was a mistake, the voter could inform the Registrar using prepaid postage or phone.Footnote ³ The NVRA requires that no list maintenance is performed 90 days before Election Day, setting an exogenous cutoff.

This is an automatic reregistration for registrants who move. It clearly conveys to the affected electorate that they have been subject to ARR, ridding movers of the costs of having to reregister to vote. Figure 1 shows the mailing that is sent out to the movers in its original form.

Figure 1.

NCOA Mailing of Orange County, California, Front and Back

Most importantly, if the voter does not actively refute having moved, the voter file will reflect the new address. This is true even if voters do not return an acknowledgment of residential changes—without a counteraction, the USPS information is treated as a true move. If the voter is a permanent absentee, mail ballots will be sent to the new address. If the voter has crossed precinct boundaries, the new polling place’s roster will have her name printed and not the old one.

Automatic reregistration is an extraordinarily unique and proactive measure to help movers maintain their registration, but its efficacy has never been measured. Note that Huber et al. (Reference Huber, Meredith, Morse and Steele2021) have shown that in Wisconsin, falsely flagging voters as movers depressed turnout, especially for racial minorities. However, the situation is dramatically different in California, where (1) same-day registration and voting are allowed, (2) in-county registrants who were flagged as movers were not removed but kept as active voters, and (3) because these were in-county movers, portable registration still applied (McDonald Reference McDonald2008), ensuring provisional voting opportunities.

DATA AND METHOD

The NCOA-based list maintenance also allows the detection of movers within the state.Footnote ⁴ The data are built on 156 daily “snapshots” of more than 1.5 million unique voters in Orange County, California, from April 26 to December 31, 2018 (Kim, Schneider, and Alvarez Reference Kim, Schneider and Alvarez2020). Carefully documenting the changing data enables the detection of movers because registration records that are matched to USPS data will be automatically updated. Matching the voter file to USPS records allows for complete verification of a move that is not a correction of wrongly entered data. In total, there were more than 100,000 verified in-county movers between the 2016 and 2018 general elections.Footnote ⁵ The dependent variable is the 2018 general turnout.

In Orange County, an NCOA matching and ARR were performed on July 26, 2018, up to the movers who moved before June 15, 2018, as the Secretary of State’s office disseminates NCOA data in the middle of the month. This discontinuity, stemming from the 90-day requirement, creates an interesting quasi-experimental opportunity, as those who have moved in the latter half of June and beyond were not subject to ARR and those who moved in the early half of June were subject to ARR and were notified as such.Footnote ⁶ This is a natural experiment, with ARR as the treatment, and I estimate a regression model specified as

(1) $$ {\mathrm{Turnout}}_i\hskip0.35em =\hskip0.35em \alpha +\beta {\mathrm{ARR}}_i+\gamma \mathbf{X}+{\varepsilon}_i, $$

where $ i $ indexes individual registrants, $ \mathbf{X} $ contains covariates for conditional ignorability, $ \mathrm{ARR} $ is the binary treatment variable, and the average treatment effect is $ \beta $ .

The final data consist of 5,585 movers who were registered at a different address by 2016 general but filed a change of address with the USPS in June 2018. The treatment group (3,294 voters) is movers from June 1st to 14th who have not voluntarily updated their registration records until late July. If the voter has voluntarily notified the Registrar so that the voter roll’s address is already up to date, the mailing is not sent. Therefore all others, including those who moved in early June and possibly disclosed it to the Registrar before July, are put to the control group (2,291 voters).Footnote ⁷ If anything, this will estimate a lower bound of the treatment, as those who voluntarily inform the Registrar are more likely to vote.

The following covariates are controlled: number of times moved in 24 months (once, twice, or three times), distance from home to polls (Dyck and Gimpel Reference Dyck and Gimpel2005; McNulty, Dowling, and Ariotti Reference McNulty, Dowling and Ariotti2009), permanent absentee voting status (Southwell and Burchett Reference Southwell and Burchett2000), age, inferred gender (Blevins and Mullen Reference Blevins and Mullen2015), inferred race (Imai and Khanna Reference Imai and Khanna2016), partisan affiliation, 2016 general turnout history, census block-level median household income of both old and new residences, whether the voter was born abroad (e.g., a naturalized citizen), the congressional district of the new home, and increasing degree of informational cost depending on how voters’ political districts changed.Footnote ⁸ Congressional districts were controlled in the model to account for some hotly contested House races, potentially driving higher turnout. The replication data is available at Kim Reference Kim2022.

I present linear probability model results with a binary treatment specification and robust standard errors for clarity of the coefficient interpretations. The logistic regression results’ average marginal effects are identical to the LPM estimates. Almost all Kolmogorov–Smirnov tests of the pretreatment variables pass, showing a well-balanced dataset—see Table 1.Footnote ⁹

Table 1.

Covariate Summary Statistics and Kolmogorov–Smirnoff Tests

AUTOMATIC REREGISTRATION INCREASES TURNOUT

Figure 2 shows the regression result in the first column.Footnote ¹⁰ Automatic reregistration increases turnout for movers by 5.8 percentage points (95% CI: $ \Big[0.0329 $ , $ 0.0823\Big] $ ). To put this in comparison with nonmovers, 70% of registrants who did not move in the last 24 months voted, whereas 64.0% of the treatment group voted, up 5.9% from 58.1% of the control group.

Figure 2.

The Effect of Automatic Reregistration and Placebo Tests

Note: Average turnout: treated group (64.0%), control group (58.1%), and stayers (70.0%).

The estimated treatment effect is very large. To put this in context, take the landmark results of get-out-the-vote (GOTV) mailing analysis in Gerber, Green, and Larimer (Reference Gerber, Green and Larimer2008). Showing households that voting is public information and their own records resulted in a 4.9 percentage-point increase in turnout. Threatening to expose voting records to neighbors resulted in an 8.1 percentage-point increase. Many GOTV devices are insignificant or low in effect sizes.

For robustness, I perform placebo tests for cases that should theoretically yield null results. An intuitive test for sample self-selection is using turnout of previous elections. Because residential changes occurred after the 2016 general election, ARR should not affect the turnout of previous elections. The placebo tests all pass for 2016.Footnote ¹¹

So why is the effect so large? First, ARR was an official, prepaid postage mailing from election administrators and not a GOTV mailing from third-party civic organizations. Mann and Bryant (Reference Mann and Bryant2020) show that even a simple postcard from official election administrators can encourage voter registration and turnout (2 percentage points), without any legal or administrative process changes—an official “nudge.” Malhotra, Michelson, and Valenzuela (Reference Malhotra, Michelson and Valenzuela2012) also show that although third-party emails made no difference in turnout, emails from official sources increased turnout.

In addition, Orange County in 2018 was highly contentious. One district that had a Republican representative for thirty-five years now elected a Democrat. Voters may have been more responsive to stimuli related to the general election. However, note that the congressional districts of new residences are already controlled for in the model.

One could say that movers are, in general, a more peripheral set of voters, albeit temporarily. Many papers such as Highton (Reference Highton2000) and Hansen (Reference Hansen2016) have demonstrated that disruptions of social environments can depress turnout for movers, and Highton and Wolfinger (Reference Highton and Wolfinger2001) shows that residential stability slowly increases turnout. But also note that Highton (Reference Highton2000) argues that the major reason that movers have lower turnout is because of the reregistration burden. Therefore, from a policy perspective, this is an ideal setting to test the effect of resolving registration burdens. Considering that the sample is limited to in-county movers, likely less affected by moving than out-of-county or out-of-state movers, the estimated effect is extraordinary.

Finally, one could also note that because the information costs of learning new political contexts are lower for in-county movers, they may show a differentially large treatment effect. If further subset by disruptions to political environments, although the effect is null for those with minimal disruptions (i.e., within-precinct movers), it is also null for those with significant disruptions such as changes in congressional districts (see Online Appendix G). This rules out the possibility that ARR is entirely driven by nudging, but the effect may be smaller for out-of-county or out-of-state movers.

It is true that, unlike motor-voter type AVR, the ARR is not on nonregistrants but rather existing registrants who have been disrupted. For nonregistrants, automatic registration after interacting with the USPS may not boost turnout as much—these are a more peripheral population who have so far not been convinced by other registration drives such as interactions with the Department of Motor Vehicles. But as aforementioned, the effect is underestimated by including voluntary disclosers in the control group. Thus, the true effect size might be even larger so that even for nonregistrants we may find some effects.

SUBGROUP ANALYSES BY PARTY

Here I discuss subgroup analyses by party affiliation (Figure 3). Although the treatment is not blocked by subgroups, as convenience voting measures are often quoted as benefiting registered Democrats, it is still worth performing analyses by party registration.

Figure 3.

ARR Effects by Party

Note: Average turnout: treated Democrats (68.8%), control-group Democrats (67.4%), stayer Democrats (72.4%), treated Republicans (66.8%), control-group Republicans (58.6%), stayer Republicans (77.4%), treated other voters (56.8%), control-group other voters (49.1%), stayer other voters (59.3%).

In particular, national AVR reform is a major part of For the People Act of 2021. In this sweeping bill, automatic registration is defined as a “system that registers an individual to vote in elections for Federal office in a State, if eligible, by electronically transferring the information necessary for registration from government agencies to election officials of the State so that, unless the individual affirmatively declines to be registered, the individual will be registered to vote in such elections.” Although not explicitly specified, ARR is also applicable by this definition because the USPS is a government agency.

The results are in Figure 3.Footnote ¹² For Democrats, ARR shows no effect statistically different from zero (95% CI: [ $ -0.0226 $ , $ 0.0615 $ ]). The effects are strong for the rest. The point estimate is 8.1 percentage points for Republicans (95% CI: [ $ 0.0376 $ , $ 0.1250 $ ]) and 7.4 percentage points for registrants with third-party or no partisan affiliations (95% CI: [ $ 0.0315 $ , $ 0.1175 $ ]). The placebo tests all pass.

Although it remains to be seen whether we will see similar effects in counties other than this traditional conservative bastion, this is one piece of evidence that ARR, a close cousin of AVR, is not uniformly beneficial to the Democratic party. Indeed, given that the distribution of parties was almost uniform in the dataset—32.4% Democrats and 32.8% Republicans—why the treatment had no effect for Democrats is puzzling.

One possibility is that conditional on being a Democrat, voters choose more to be permanent absentees. In the test (full) dataset, 69.7% (67.0%) of Democrats and 64.2% (65.0%) of Republicans were permanent absentee. So, ARR does indeed have a greater effect on those who are not permanent absentees (8.9 percentage points) compared with those who are (4.1 percentage points), again passing all placebo tests (see Online Appendix E). But although the differential adoption rate of absentee voting is likely a factor, it is still not enough to fully explain the partisan gap.

Another possibility is that base mobilization for Democrats was much stronger—that they voted regardless of the treatment status, stimulated by the Trump presidency. This possibility is more difficult to test but seems unlikely because, overall, Republicans turned out more (76.7%, conditional on registration) than Democrats (72.1%).

A final possibility is that many voters who registered as “No Partisan Preference” options are more Democratic leaning. If only voters with no partisan preferences are considered, the effect estimate is 8.1% (95% CI: [ $ 0.0376 $ , $ 0.1250 $ ]). So although including them could shift the treatment effect upward, it still seems that Republicans benefited more from ARR in Orange County.