Measuring lost votes by mail

Abstract

The rise of mail balloting has led to concerns that procedural requirements can lead to “lost votes by mail.” We theorize how procedural requirements can affect the incidence and form of lost votes and highlight three measurement issues with equating lost votes and rejected mail ballots. First, coverage: Not all rejected mail ballots are documented. Second, substitution: Some people whose mail ballot is rejected may subsequently successfully vote, in particular if they were notified in time to take action. Third, deterrence: Others may not return their mail ballots if they expect them to be rejected. While rejected mail ballots could over- or underestimate lost votes, a case study of Pennsylvania’s 2022 general election reveals at least 47% more lost votes than rejected mail ballots. These lost votes could prove electorally consequential in Pennsylvania given the number of mail ballots cast and the substantial partisan splits on mail versus in-person ballots.

Procedural requirements on mail balloting reduced counted ballots by more than previously thought in Pennsylvania’s 2022 election.

INTRODUCTION

The growing use of mail balloting poses a unique challenge for election administration. In short, the availability of mail ballots promises to increase voter participation (1, 2) and satisfaction (3). Yet, precisely because mail ballots are sent to individuals outside a polling place, states must impose a series of procedural requirements on mail balloting (4). Further, its very nature makes election administrators and voters alike dependent on the performance of the United States Postal Service (USPS) (5, 6).

In recent years, the mix of procedural requirements used for mail balloting across states has become the focus of substantial public debate. The core dispute centers around the familiar trade-off between voter access and electoral integrity. In general, procedural requirements ensure that mail ballots are cast by their intended recipients in a timely manner. However, Stewart (7) developed the concept of “lost votes by mail” to highlight that these requirements may prevent eligible voters from successfully casting ballots too.

Building on Stewart (8), we theorize about how particular procedural requirements affect the incidence and form of lost votes by mail. We focus on three relevant factors: (i) the extent of rejection of returned mail ballots, (ii) the ease of substitution from a rejected mail ballot to cast a counted ballot, and (iii) the degree of deterrence from returning a mail ballot that may otherwise be rejected. We also consider the circumstances under which these lost votes by mail could prove electorally consequential.

Given our theory, we argue that previous approaches used to measure lost votes by mail are incomplete. In a leading study, Stewart (8) constructs a national estimate of lost votes by mail based in part on the aggregate number of rejected mail ballots reported in the Election Administration and Voting Survey (EAVS). Other scholars have taken an alternative approach, using individual-level administrative data on mail ballot rejections (9–12). However, both approaches necessarily assume that aggregate counts or administrative data have complete coverage of rejected mail ballots. Further, neither approach recognizes that rejected mail ballots may understate or overstate lost votes depending on the relative amount of substitution and deterrence. Rejected mail ballots will overstate lost votes when there is more substitution among voters with rejected mail ballots than deterrence among voters with unreturned mail ballots. Alternatively, rejections will understate lost votes when there is more deterrence than substitution. For both reasons, scholars cannot compare state mail balloting regimes by simply counting rejected mail ballots.

We use Pennsylvania as a case study to correct the measurement of lost votes by mail. We gather separate individual-level administrative data on both the status of requested mail ballots and the turnout of registered voters in the state. As a result, we can measure both substitution, by observing whether voters who submit a rejected mail ballot go on to cast a valid vote, and deterrence, by modeling why registrants who request a mail ballot do not return that ballot. Further, data publicly revealed during recent litigation in Pennsylvania enable us to both evaluate the coverage of documented, rejected mail ballots and compare how rates of substitution vary by whether local election officials notified voters of rejections. To be clear, the magnitude of lost votes in Pennsylvania reflects the specific state and electoral cycle we study. However, we expect that the incomplete coverage of rejected mail ballots, the relationship between notice and substitution, and the relationship between ballot deadlines and deterrence are each more generalizable.

We make three points about lost votes by mail. First, we demonstrate that not all rejected mail ballots are documented as such. While Pennsylvania reported rejecting about 23,000 mail ballots, we identify at least 5000 additional rejected mail ballots using a variety of public sources. Second, we show why some documented rejected mail ballots should not be considered lost votes. Overall, we find that about 15% of voters whose mail ballots were rejected nonetheless cast a valid vote. We also show that voters are more likely to avoid a lost vote when counties notify people about rejected mail ballots with enough time to correct them. Third, we show why some mail ballots that are not returned, and thus not rejected, should nonetheless be considered lost votes. In particular, we estimate that about an additional 11,000 people in Pennsylvania who requested a mail ballot were deterred from voting because their mail ballots would have otherwise been rejected as late. We find more deterrence for people who request their mail ballots close to the election, and in particular in the last 2 days before the request deadline.

Together, clarifying the measurement of lost votes helps refine both the study and assessment of voting by mail. In total, we conclude that Pennsylvania had at least 11,000 more lost votes than the roughly 23,000 rejected mail ballots documented in the EAVS (13). Further, we show how these lost votes could be electorally consequential because many voters in Pennsylvania vote by mail and mail and in-person voters have distinct preferences. In our conclusion, we propose a number of ways to reduce lost votes by mail by focusing on the broader system of mail balloting. This includes reducing the number of deficient mail ballots initially received, improving the identification and notification of remaining deficiencies, and making it easier to substitute if a ballot is rejected.

Lost votes by mail

We first consider the theory of a lost vote by mail before evaluating how lost votes have been measured in the literature.

Theory

The concept of lost votes was first developed in the wake of the contested 2000 presidential election. An estimated 4 to 6 million in-person presidential votes were “lost” because of a combination of registration problems, polling place operations, faulty voting equipment, and confusing ballot design (14).

As vote mode subsequently shifted to include substantial voting by mail, a related concept of lost votes by mail emerged (7). Lost votes by mail are distinct from lost votes in person because voting by mail requires a different series of steps than voting in person. Depending on the state, some registrants must specifically request a mail ballot. States also impose a series of a procedural requirements on mail ballots to maintain electoral integrity outside of the polling place. In general, a voter must mark a received mail ballot and potentially place it in an inner envelope, complete an accompanying affidavit, and return it all by the relevant deadline. Election officials then determine whether the voter satisfied the requirements necessary for the ballot to count.

We build on Stewart (8) to theorize about how particular procedural requirements affect the incidence and form of lost votes by mail. Stewart offered a typology of reasons for lost votes by mail: votes lost by the postal service, because a mail ballot request or mail ballot is sent but not received by a voter or election official; votes lost by election officials, because a mail ballot request is not fulfilled; votes rejected, because of the voter affidavit, ballot envelope, or ballot deadline; and votes lost at tabulation, in the form of residual votes (8). In particular, we identify three relevant factors about procedural requirements that affect the likelihood of a lost vote. As in Stewart’s typology, the first is the extent to which returned mail ballots are rejected for failing to satisfy the procedural requirements. However, not all rejected mail ballots are lost votes because affected voters sometimes substitute by taking subsequent steps to cast a counted ballot. After experiencing a mail ballot rejection, a voter might correct their initial mail ballot, submit a new mail ballot, or shift to voting in person. Further, people may be deterred from returning their mail ballots because they cannot satisfy the procedural requirements. Although the literature on mail balloting has yet to account for deterrence, the concept is central to the literature on voter ID laws. In that context, not all people affected by the ID law cast a rejected provisional ballot (15, 16). Analogously, not all people affected by a ballot receipt deadline cast a mail ballot that is rejected because it is received too late.

Given our theory, we expect the extent of lost votes to differ markedly across states. Beyond the number of mail ballots cast by state, the particular procedural requirements adopted by a state can generate different levels of initial rejection, subsequent substitution, or deterrence. In short, we theorize that the relative mix of rejection and deterrence will depend on whether voters can verify procedural requirements ex ante. We expect relatively more deterrence than rejection when voters can verify requirements ex ante, and the converse when voters cannot.

To illustrate this claim, consider the different processes used by states to verify mail voters’ identities. Having some process is necessary when voting outside of a polling place to maintain electoral integrity. To that end, some states require voters to sign affidavits on mail ballots in the presence of a notary. We expect such a requirement to cause relatively high amounts of deterrence, as we anticipate that many voters who cannot access a notary will stop the process of mail balloting. Other states verify whether the voter’s signature on the affidavit is sufficiently similar to their signature on their registration record. While we expect that few voters are deterred from submitting mail ballots because they must sign an affidavit, signature matching has the potential to generate a substantial number of wrongful rejections, which voters are unlikely to anticipate (17).

Similarly, different approaches to notifying voters of a mail ballot rejection and allowing for the correction of an error could affect the relative level of substitution. For example, some states begin processing mail ballots upon receipt or before Election Day, while others wait until Election Day. As a result, some people whose mail ballot is rejected will have more time to take corrective action (18). For another, states offer voters different types of opportunities to correct any deficiency. Some permit voters to quickly take corrective action by phone or text, others rely on mail or in-person visits, and still others require either submitting a new ballot or simply voting in person (19).

Finally, we also expect the electoral consequences of lost votes to vary by state, even among those with similar procedural requirements. In general, two conditions must both be present for an election rule to affect election outcomes: The preferences of those affected and unaffected by the rule must be different and a sufficient number of people must be affected by the rule (20). Many election rules plainly fail to satisfy at least one of these conditions, meaning that the rule is unlikely to be electorally consequential. In the case of mail balloting, though, table S9 shows increasing vote mode splits, in which people who vote by mail disproportionately support Democratic candidates compared to people who vote in person. The larger this vote mode split in a state, the greater the potential for electoral consequences when uniformly applied procedural requirements generates lost votes.

Measurement

The EAVS is used to study many aspects of American election administration (21–24), including lost votes by mail. However, there are a set of overlapping concerns about the ability of the survey to measure the full extent of lost votes by mail, as we conceive of it.

In short, the EAVS asks election officials to provide a battery of jurisdiction-level statistics. Stewart used the introduction of the EAVS to study lost votes by mail in the 2008 election (7) and then updated the analysis for the 2016 election (8).

The latter analysis incorporated information from the EAVS that about 380,000 mail ballots were reported to not be counted, which represented about 1.1% of the approximately 33.5 million mail ballots cast (25). It also mentioned, although did not incorporate, the 8.2 million mail ballots reported in the EAVS as unreturned. In a similar vein, more recent research used the EAVS to conclude that the national mail ballot rejection rate declined from about 1 to 0.8% between 2016 and 2020 (26).

Figure 1 illustrates some limitations of the current use of the EAVS for the measurement of lost votes. It compares the reported percent of mail ballots transmitted that were rejected (y axis) or went unreturned (x axis) in the 2022 general election. In general, we expect that measurement error causes Fig. 1 to understate the true rejection rate in many states. The EAVS suffers from known issues of incomplete reporting (27, 28). For example, the EAVS reports that Illinois rejected 1.3% of mail ballots. However, Illinois has publicly confirmed that its rejection rate was actually twice that (2.6%) (29); the result is that Illinois looks typical in Fig. 1, yet is actually more like an outlier.

Fig. 1. Voting by mail outcomes by state.

Even if the EAVS offered an accurate accounting of rejected and unreturned mail ballots, the type of data illustrated in Fig. 1 cannot capture the patterns of substitution or deterrence that are relevant for measuring lost votes. For one, the EAVS does not differentiate between people who did and did not take subsequent steps to cast counted ballots after mail ballot rejections. However, substitution could prevent many rejected mail ballots from turning into lost votes (30). To see why this limitation of the EAVS matters for evaluating mail balloting policies, consider the case of Texas. In 2021, Texas adopted a new voting law (SB1), which required mail voters to provide the same identification number on their mail ballot, either a driver’s license number or the last four digits of a Social Security number, as they had on their initial voter registration application (31). SB1 also instituted a ballot curing process (32). Figure 1 shows that Texas voters experienced one of the highest mail ballot rejection rates (2.83%) in 2022, the year the law went into effect. However, we cannot learn how SB1 affected lost votes using EAVS because the reported rejections do not account for any subsequent substitution.

Similarly, the EAVS also does not allow researchers to learn more about why mail ballots are going unreturned. Figure 1 shows that many states experience an order of magnitude more unreturned mail ballots than rejected mail ballots. This is not only true in states that conduct vote-by-mail elections but also often in states in which registrants must request a mail ballot before the election. Thus, if deterrence is causing even a relatively small share of these unreturned ballots, it still could be responsible for a substantial share of the total lost votes by mail. Yet, we have no clear way to differentiate between ballots that went unreturned due to deterrence from those that went unreturned because the person decided to substitute into in-person voting or was no longer interested in voting.

A final limitation of any aggregated data on rejected or unreturned mail ballots, such as the EAVS, is that it can obscure subgroups of voters who disproportionately experience lost votes by mail. Several recent studies of mail balloting use individual-level administrative data to overcome this particular limitation. These studies generally find that people with demographics predictive of lower turnout (e.g., young, non-white, have voted in fewer previous elections) cast a disproportionate share of rejected mail ballots (9–12). However, this work has not taken the next step of linking administrative data on mail ballots with administrative data on turnout. As a result, it does not account for the substitution and deterrence necessary to estimate the total number of lost votes by mail.

Pennsylvania case study

We use Pennsylvania as a case study to correct the measurement of lost votes by mail. We first discuss the relevance of new data sources, then detail the state’s particular mail ballot policies, and finally consider the generalizability of our case study.

Data

Several features of Pennsylvania’s 2022 general election are helpful for measuring lost votes. First, Pennsylvania provides administrative data on mail ballot outcomes that can be linked with turnout outcomes. We can thus measure how many people with rejected or unreturned mail ballots were nonetheless able to successfully vote.

Moreover, during the course of federal litigation over mail balloting rules in place for the election (33), Pennsylvania county boards of election were required to disclose additional data beyond what is captured by the EAVS or administrative data. These disclosures are helpful for at least two reasons. For one, the responses were provided by the counties directly, who retain the rejected ballots, rather than the state. For another, the responses were signed by attorneys who had to certify, subject to sanction, that the disclosure “is complete and correct” to “the best of the person’s knowledge, information, and belief formed after a reasonable inquiry” (34). As a result, we believe the litigation data allow us to observe the true number of rejected ballots, even if those ballots were not documented in administrative data. Further, information on whether and when county election officials notified voters of mail ballot deficiencies allows us to study how such policies affect whether rejected mail ballots become lost votes. We further detail both sets of data in Materials and Methods, including how we obtained the administrative data from public records requests and the litigation data from a federal court’s public docket.

Rules

To vote by mail in Pennsylvania, registrants must first request a mail ballot. Pennsylvania accepts mail ballot applications up until a week before Election Day (35). For the mail ballot to be counted, a voter must sign and date an affidavit, place the ballot in an unmarked secrecy envelope, and ensure that election officials receive the ballot by 8 p.m. on Election Day (36). Under these rules, there are at least six causes of mail ballot rejection: (i) an unsigned affidavit, (ii) an undated affidavit, (iii) a misdated affidavit, (iv) a so-called “naked” ballot not placed in a secrecy envelope, (v) a ballot placed in a secrecy envelope containing markings, and (vi) a late ballot. In addition, a voter who did not provide their driver’s license number, the last four digits of their Social Security number, or a copy of their photo ID with their mail ballot request, must provide a copy of their ID when returning their mail ballot (37).

There is no common protocol in Pennsylvania to notify voters of rejected mail ballots. Instead, county officials take vastly different approaches. Depending on the county, election officials might contact voters with a defective mail ballot, post a list of defective mail ballots, or take no proactive action. Moreover, even counties with a notice protocol may not discover a mail ballot deficiency with enough time for a voter to take action. Pennsylvania did not allow election officials to process mail ballots received during the 2022 general election until Election Day. Thus, a county’s ability to learn about deficiencies on mail ballots depended on the extent to which they engaged in what is generally known as “preprocessing.” Preprocessing refers to actions election officials take before Election Day to identify problems with mail ballots without opening the mail ballot return envelopes. Many Pennsylvania counties do not engage in preprocessing, but some do. During the 2022 general election, the most common form of preprocessing focused on the affidavit on the outside of the mail ballot return envelope. Election officials could identify an unsigned, undated, or misdated affidavit in preprocessing because doing so does not require opening the return envelope itself (38). A few enterprising counties also weighed return envelopes to infer whether it included an inner secrecy envelope (38). We summarize the substantial litigation over Pennsylvania mail ballot rules in the Supplemental Materials.

Generalizability

As with any single case study, there are questions about how much we can generalize what we learn about lost votes by mail in Pennsylvania to other states. For the reasons previously discussed, the coverage issues that we document in the EAVS are likely present in many other states, and in particular on questions in which state election officials are reporting quantities collected by county officials. In general, we also expect that the rates of substitution and deterrence in Pennsylvania will be more generalizable than what we learn about initial rejections or electoral consequences.

We learn about substitution, and how it is affected by notice, by comparing the rates of substitution in Pennsylvania counties with and without notice policies. This within-state variation in notice policies is particularly useful for learning about the effects of notice regimes while holding fixed many of the other policies that are most likely to affect substitution rates. These effects are likely to be informative of how much additional substitution we are likely to observe in states with notice, but without ballot curing, and states without notice.

The mail ballot request and receipt deadlines that drive deterrence in Pennsylvania are also similar to the policies used in other states. In general, the USPS recommends that mail ballot applications be completed at least 15 days before Election Day (39). However, many states do not follow the recommendation (40).

However, the procedural requirements leading to mail ballot rejection in Pennsylvania are somewhat distinct from those in other states. For one, Pennsylvania rejects undated and misdated mail ballots, a policy that best as we can tell is unique to the state. For another, few other states reject mail ballots not returned in a secrecy envelope, at least as of the 2020 election (41). On the other hand, Pennsylvania does not engage in signature matching like many other states (42).

We also suspect that lost votes by mail may be more electorally consequential in Pennsylvania than most other states, a point we discuss in detail below. In short, Pennsylvania has the combination of more mail balloting, more mail ballots turning into lost votes, and a greater partisan split in mail versus in-person ballots than many other states.

RESULTS

Our primary dataset is mail ballot applications linked to voter registrations. The mail ballot data allow us to report descriptive statistics on mail ballot usage overall and by demographics.

Overall, more than 1.4 million Pennsylvania registrants requested a mail ballot in the 2022 general election. Based on Table 1, the vast majority (about 87%) of mail ballots were returned. However, about 11% of mail ballots were not returned and about 1.7% were rejected. Further, among the voters whose mail ballots were rejected, the most common reasons varied from issues with the affidavit (33.7%), the secrecy envelope (34.8%), and receipt timing (21.2%).

Table 1. Mail ballot applications by status and rejection reason.

	n	%
All ballots	1,403,760	–
By mail ballot status
Accepted	1,223,423	87.2%
Not returned	157,170	11.2%
Rejected	23,167	1.7%
By rejection reason
Affidavit	7811	33.7%
Secrecy	8072	34.8%
Late	4900	21.2%
Other	2384	10.3%

Table S6 further details the rates at which mail ballots go unreturned and returned mail ballots get rejected by demographics. Consistent with the literature (9–12), registrants most likely to be minorities, young registrants, and people who previously had not voted by mail experience a disproportionate share of both rejected and unreturned mail ballots. Table S6 also shows the relative reasons for rejection differ by group. For example, both young registrants and first-time mail voters are particularly likely to have a mail ballot rejected because it was received late. In contrast, Black registrants were the most likely to experience a rejection because of an affidavit error or the lack of necessary ID.

These descriptive statistics on unreturned and rejected mail ballots, though informative, should not be confused with lost votes. Instead, we offer them to motivate three questions about (i) the coverage of rejected mail ballots as well as the relationship between both (ii) rejected and (iii) unreturned mail ballots and lost votes. We then conclude by considering the potential electoral consequences of lost votes by mail.

Measurement of lost votes

In this subsection, we first show that the coverage of rejected mail ballots in Pennsylvania’s administrative data is incomplete. We then measure the extent of substitution following a rejected mail ballot, including the extent to which notice to voters of a ballot rejection can avoid a lost vote. Last, we estimate the extent to which mail ballot request and receipt deadlines contributed to deterrence, as voters may not return a mail ballot in the face of a likely rejection. For all three reasons, we should not interpret Table 1 as showing that there were just over 23,000 lost votes by mail. Instead, we will ultimately conclude that there were likely at least an additional 11,000 lost votes.

Not all rejected mail ballots are documented

We take multiple approaches to estimating the extent to which Pennsylvania’s administrative data on mail ballot rejections is incomplete. We initially show that aggregate statistics are similar across the EAVS and state administrative data. However, both sources are missing two types of mail ballot rejections. For all counties, we focus on mail ballots rejected because of an issue with the affidavit, which we shorthand as affidavit rejections. We compare affidavit rejections as reported by the state in the EAVS to affidavit rejections publicly disclosed by local election officials in federal litigation. We then focus on select counties and compare all mail ballot rejections as reported in the EAVS to the same statistics reported during the county canvass, in local newspapers, or in conversations with county election officials. We ultimately conclude that there are at least 5000 rejected mail ballots in Pennsylvania not otherwise documented.

EAVS

Table S2 shows how many total mail ballots were rejected, overall and for specific reasons, in both the EAVS and Pennsylvania administrative data. The table makes it clear that the state responded to the EAVS on behalf of counties based on the same administrative data used to construct Table 1. The state appears to have reported that there was no data available whenever the administrative data contained no cases of that type. The concern is that, for some counties and some types of ballot rejections, there were rejected mail ballots that were never documented, or only partially documented, in the administrative data and thus never reported on the EAVS.

More specifically, the EAVS reported that 65 of Pennsylvania’s 67 counties rejected a total of 23,393 mail ballots. However, there was substantial heterogeneity across counties in which type of ballot rejections were documented and thus represented in the total. For example, only 48 counties reported late mail ballots, only 56 counties reported mail ballot affidavit issues, and only 53 reported secrecy envelope issues. While nonreporting was more likely in small-population counties, it also affected counties with substantial population too. Allegheny County offers an extreme example of this phenomenon. On the basis of the EAVS, the second largest county in the state had no data available on the number of mail ballots rejected for secrecy envelope issues. A related challenge is to identify underreporting. For example, Lehigh County implausibly reported rejecting just one mail ballot because of a problem with the affidavit. Given both underreporting and nonreporting, the EAVS statistics must be an undercount of the number of mail ballots rejected overall and for specific reasons.

Affidavit rejections

We expect more under reporting of misdated and undated ballots than other types of rejected mail ballots. For one, it was uncertain whether these ballots would be rejected until just before Election Day. Relatedly, the state did not provide guidance about how to record these ballots in the state system until after a court determined they would be rejected. Their guidance was to record the misdated and undated mail ballots in the statewide system using the code designed to document mail ballots rejected for having unsigned affidavits. This combination of late guidance and atypical documentation drives our expectation that misdated and undated ballots will be less likely to be recorded in the state system, and thus also be less likely to be included in the EAVS than other types of rejected mail ballots.

Litigation over Pennsylvania’s mail ballot requirements provides a vehicle for observing misdated and undated ballots not documented in administrative data or reported in EAVS. As described above and in Materials and Methods, county election officials disclosed counts of misdated and undated mail ballots during the course of federal litigation. Ultimately, the top of table S3 shows that counties reported about 10,576 misdated or undated mail ballots. As a point of comparison, the EAVS only documents 7894 mail ballots being rejected for any affidavit issue.

The state’s guidance for how to record misdated and undated ballots makes it impossible to use the litigation data to calculate precisely how many of these misdated and undated ballots went undocumented in the state system. Ideally, we would calculate how many more misdated and undated ballots a county reported in their response to the discovery request than they recorded in the state system. However, the aforementioned documentation protocol means that ballots recorded as misdated or undated cannot be isolated from ballots recorded as unsigned. In addition, because the discovery requests did not ask counties to report the number of unsigned ballots, we do not generally do not know the total number of ballots a county rejected because of any specific affidavit issue.

We can nonetheless use the litigation data to calculate the minimum number of undocumented undated and misdated ballots. When the difference between the number of misdated and undated ballots reported by the county in the discovery request and the number of affidavit rejection recorded in the administrative data is positive, it represents a lower bound on the number of mail ballot rejections that are missing in the administrative data. Across all counties, at least 3017 undated or misdated mail ballots were not documented in the administrative data. This lower bound assumes that all of the deficient affidavits documented in the administrative data were specifically rejected for undated or misdated affidavits. In practice, we know that the opposite was more likely to be true; several counties only recorded mail ballots rejected for having unsigned affidavits in the state system. Hence, we assess that the number of undocumented undated or misdated mail ballots is likely substantially larger than this lower bound.

The remainder of the table separately considers the coverage of affidavit rejections in the top five counties by mail ballots returned and all other counties. Overall, the top five counties are responsible for about half of all returned mail ballots. The top five counties contribute a smaller absolute number of undocumented affidavit rejections, but a slightly higher rate of underreporting than all other counties.

Large counties

Table S4 takes a complementary approach to measuring coverage by focusing on the coverage of all mail ballot rejections, for any reason, but only in the five counties with the most mail ballots returned. For each type of mail ballot rejection, it presents both the quantity as reported in the EAVS and the quantity, if available, as reported in a county canvass, local newspaper, or directly from a county election official.

For example, there were at least 1702 mail ballot rejections in Allegheny County not reported in the EAVS. The substantial underreporting was the result of the county’s unique approach to curing defective mail ballots. In short, the county returned defective mail ballots to voters rather than directly rejecting them and marked them as pending. For example, the county did not report a single mail ballot as rejected for a missing secrecy envelope. However, a news article from 5 days before Election Day reported that the county returned 683 mail ballots for that reason (43). While voters may have cured some of these returned mail ballots, it is likely that even more defective mail ballots were received in the final 5 days than cured. Further, as discussed above, Allegheny County reported rejecting 1009 misdated or undated mail ballots in the subsequent litigation. As a result, we conclude that the total number of lost votes by mail in Allegheny County is at least the combined 1886 mail ballots mentioned in the newspaper article and litigation.

Overall, table S4 identifies about 3200 lost votes not documented by rejected mail ballots in the EAVS. In addition to the expected affidavit discrepancies, the EAVS is also missing some mail ballot rejections for missing secrecy envelopes. For example, in Philadelphia, the county canvass revealed that there were 1946 mail ballots with secrecy envelope issues, not the 1820 reported in the EAVS (44). Similarly, the canvass in Bucks County revealed 551 mail ballots with secrecy envelope issues, not 65 (45). Montgomery County election officials believe the EAVS correctly captured defective secrecy envelopes, although they adjusted the number of defective affidavits. Similarly, Chester County supplied individual-level data, which generally matched the EAVS, but for an adjustment for the number of late ballots.

Not all rejected mail ballots are lost votes

Coverage concerns aside, the discussion of lost votes typically focuses on the documented number of rejected mail ballots. However, rejected mail ballots are not equivalent to lost votes because some people whose mail ballot is rejected go on to substitute, either by shifting into in-person voting or correcting their initial mail ballot by submitting a new one.

To see this, Table 2 details the ultimate turnout outcome of people whose mail ballots are rejected. The top row of Table 2 reports that about 15% of people who cast a rejected mail ballot ultimately cast a valid vote.

Table 2. Mail ballot rejections by turnout.

		Turnout
	n	In person	By mail	% Voted
All rejected ballots	23,167	2507	979	15.0%
By rejection reason
Rejected for affidavit	7811	1341	448	22.9%
Rejected for secrecy	8072	840	455	16.0%
Rejected for late	4900	59	40	2.0%

The remaining rows of Table 2 look at how turnout outcomes vary based on the rejection reason. Our expectation is that substitution rates will positively associate with the amount of time people have to take action. On the basis of the particulars of preprocessing discussed above, this leads us to expect more substitution among people whose mail ballot is rejected because of an affidavit than people who return a ballot outside of a secrecy envelope. Consistent with this expectation, we observe that about 23% of people whose mail ballot is rejected because of an affidavit go on to successfully vote compared to about 16% of people whose mail ballot is rejected because of a secrecy envelope. We are also not surprised that about 2% of people whose mail ballot is rejected as late go on to vote. That is because a mail ballot is not marked as late until after Election Day, although a voter may anticipate their mail ballot being late and take further action.

Table 3 further distinguishes between turnout outcomes based on the local notice policies disclosed by county election officials during the federal litigation discussed above. Consistent with our expectations, it shows that people who return a mail ballot with a disqualifying affidavit were about 14 percentage points more likely to successfully vote if they lived in a county with a notice program than if they lived in a county without one. In comparison, the same quantity is only about 4 percentage points for people who had a problem with their secrecy envelope. Last, people who submitted late mail ballots were slightly less likely to successfully vote if they lived in a county with a notice program than if they lived in a county without one, although we suspect this result is an artifact of a data error in Centre County, which did not have a notification program.

Table 3. Mail ballot substitution by notice.

		Turnout
	n	In person	By mail	% Voted
Rejected for affidavit
With notice program	4709	1122	228	28.7%
Without notice program	3102	219	220	14.2%
Rejected for secrecy
With notice program	5350	779	181	17.9%
Without notice program	2722	61	274	12.3%
Rejected for late
With notice program	3573	33	2	1.0%
Without notice program	1327	26	38	4.8%

Table S7 reports how substitution after rejection varies by demographics. Some of the demographic trends of substitution are similar to the demographic trends of rejected mail ballots. For example, young people and first-time voters are least likely to substitute. However, some trends are different. Black voters, for example, are more likely to substitute following a rejection than any other registrants. The difficulty with interpreting these descriptive statistics is that individuals are differently situated with respect to notice depending on the county that they live in. Across the state, counties have different notice programs, different demographic profiles, and attract a different amount of campaign outreach. The racial difference is at least partly attributable to the fact that Black registrants are more likely to live in jurisdictions with notice programs and benefit from Democratic campaign outreach. This also likely explains, at least in part, why Democratic registrants are more likely to substitute than Republican registrants.

Some unreturned mail ballots are lost votes

In contrast to rejected mail ballots, unreturned mail ballots do not typically appear in calculations of lost votes. The difficulty of unreturned mail ballots for understanding lost votes is that unreturned ballots represent a combination of three different phenomena. First, there are registrants who abstain from returning a mail ballot because they are not interested in voting. Second, there are registrants who substitute into voting in-person despite receiving a mail ballot, either because they decide they prefer to vote in person or are concerned they will not be able to successfully satisfy the mail ballot requirements. Third, there are registrants who are deterred from returning a mail ballot because they are similarly concerned about satisfying the mail ballot requirements but are not able to substitute into voting in person. Below, we detail the incidence of both substitution and deterrence among registrants who do not return their mail ballots. We ultimately conclude that the Election Day receipt deadline deterred about 11,000 people from voting.

Substitution

Table 4 reports the turnout records of voters who did not return their requested mail ballots. The first row focuses on all requesters, showing that about a third of requesters who do not return their mail ballots nonetheless vote. By definition, these are not lost votes.

Table 4. Unreturned mail ballots by turnout.

			Among unreturned
	Distributed	Not returned	% No vote	% Vote
All mail ballots	1,403,880	157,170	62.4%	37.6%
Where was mail ballot sent?
To zip of registration	1,313,892	138,055	59.3%	40.7%
To rest of county	18,199	4057	69.5%	30.5%
To another PA county	21,743	5308	85.0%	15.0%
To another state	40,88	8648	90.3%	9.7%
To another country	2676	742	95.3%	4.7%
Unknown	6482	360	76.9%	23.1%

To test whether the accessibility of in-person voting relates to the likelihood of substitution, the remaining rows of Table 4 distinguish between requesters based on where they had their mail ballots sent. The table groups requesters by whether their ballots were sent to their zip code of registration (zip), elsewhere in their county of registration (county), elsewhere in Pennsylvania (PA), to another state (US), to another country, or to an unknown address. The assumption is that, on average, people will find it more difficult to vote in person on Election Day the further they send their ballots from their home. Consistent with this assumption, Table 4 shows that the likelihood of substitution decreases when the requesters sent their mail ballots further afield.

Figure 2 further illustrates how geography interacts with the timing of mail ballot requests in shaping the likelihood of substitution. Figure 2 focuses on mail ballots requested in the last 28 days before the request deadline. The figure is divided into four panels based on the same geographic variable used in Table 4, excluding the small number of mail ballots sent internationally or to unknown addresses. As the panels move from left to right, mail ballots are being sent further from registrants’ address of registration. All panels highlight that an increased share of mail ballots go unreturned when requested closer to the deadline. However, consistent with Table 4, people who send their mail ballots to their zip code of registration engage in more in-person substitution than people who send their mail ballots elsewhere. Correspondingly, the figure shows a positive association between sending a mail ballot further from an address of registration and the share of unreturned mail ballots. Together, this suggests that people who send their mail ballots further afield are most affected by the Election Day receipt deadline because they face the most time pressure and are least able to substitute into in-person voting.

Fig. 2. Share of turnout outcomes by date of mail ballot request.

Deterrence

The previous subsection focused on requesters who did not return their mail ballots but nonetheless voted. In this subsection, we focus on the converse: requesters who did not return their mail ballots and did not ultimately vote. In theory, nonvoting could be explained by either abstention or deterrence. However, while the former should not be considered a lost vote by mail, the latter should be.

Estimating the extent of deterrence among people who do not return their mail ballots requires us to a formalize a model of why ballots go unreturned. Our model focuses specifically on the subset of people who requested a mail ballot in the 4 weeks before the deadline. As previously discussed, Pennsylvania allows people to request a mail ballot up until 7 days before Election Day, even though the USPS recommends a deadline of 15 days in states with an Election Day receipt deadline (39).

Figure 3 highlights that a substantial number of mail ballot requests come after the USPS recommended timeline. The figure counts the number of mail ballots requesters by week of request, differentiating between those requesters who successfully voted (black) from those who did not (gray). The distribution is bimodal because Pennsylvania registrants can either make an annual request to receive a mail ballot for all elections that year or request a mail ballot for a specific election (46). Ultimately, election officials received more than 100,000 mail ballot applications during the week before the deadline, or less than 2 weeks before Election Day. Of these applicants, about 16,000 did not successfully vote.

Fig. 3. Mail ballot requesters by week of request.

Our model estimates the extent of deterrence by relating patterns of in-person substitution, discussed above, and nonvoting to variation in the degree of risk of returning a late mail ballot. More specifically, let $N_t$ be the number of people requesting a mail ballot on day $t$. We observe a set of turnout outcomes, $Y_t$, for all of the people who request mail ballots on day $t$. We define $Y_t=(Y_t^m,Y_t^s,Y_t^a)$, which are the share of these mail ballot requesters at time $t$ who produce a returned mail ballot ($Y_t^m$), an in-person vote ($Y_t^s$), and abstain ($Y_t^a$), respectively.

We expect that the composition of observed turnout outcomes on day $t$ to be a function of the share of people who perceive heightened risk that their mail ballots will not be received by the Election Day receipt deadline. Let $\mathrm{\pi }=({\mathrm{\pi }}_m,{\mathrm{\pi }}_s,{\mathrm{\pi }}_a)$ be a vector of probabilities representing the likelihood that someone who perceives normal risk returns a mail ballot (${\mathrm{\pi }}_m$), substitutes into in-person voting (${\mathrm{\pi }}_s$), and abstains (${\mathrm{\pi }}_a$). Let $p_t$ be the probability that someone who requests a mail ballot on day $t$ faces a heightened risk of a late mail ballot. When a person faces a heightened risk of returning a late mail ballot, let $\mathrm{\delta }=({\mathrm{\delta }}_m,{\mathrm{\delta }}_s,{\mathrm{\delta }}_a)$ be the change in the probability of returning a mail ballot (${\mathrm{\delta }}_m$), substituting into in-person voting (${\mathrm{\delta }}_s$), and abstaining (${\mathrm{\delta }}_a$) relative to when the person perceives normal risk. Under this model, the share of people requesting mail ballots on day $t$ that produce each turnout outcome is the share under baseline risk of a late mail ballot plus the difference in the share under heightened risk multiplied by the probability of heightened risk

$$Y_t^m={\mathrm{\pi }}_m+p_t{\mathrm{\delta }}_m$$ (1)

$$Y_t^s={\mathrm{\pi }}_s+p_t{\mathrm{\delta }}_s$$ (2)

$$Y_t^a={\mathrm{\pi }}_a+p_t{\mathrm{\delta }}_a$$ (3)

Our goal is to estimate the total number of mail ballots that were deterred because they would have been rejected as late, or ${\sum }_t{N}_t{p}_t{\mathrm{\delta }}_a$. We estimate this quantity using the following steps

1) Identify a $\overline{t}$ such that for $t\le \overline{t}$ that $p_t=0$.

2) Calculate ${\widehat{\mathrm{\pi }}}_s$ using mail ballot requests made at $t\le \overline{t}$.

3) Construct $\widehat{p_t{\mathrm{\delta }}_s}=Y_t^s-{\widehat{\mathrm{\pi }}}_s$.

4) Regress $Y_t^a$ on $Y_t^s$ to estimate $\mathrm{\beta }=\frac{\text{cov}({\mathrm{\pi }}_s+p_t{\mathrm{\delta }}_s,{\mathrm{\pi }}_a+p_t{\mathrm{\delta }}_a)}{\text{var}({\mathrm{\pi }}_s+p_t{\mathrm{\delta }}_s)}=\frac{{\mathrm{\delta }}_s{\mathrm{\delta }}_a\text{var}(p_t)}{{{\mathrm{\delta }}_s}^2\text{var}(p_t)}=$$\frac{{\mathrm{\delta }}_a}{{\mathrm{\delta }}_s}$. $\mathrm{\beta }$ captures how many abstentions occur for every one substitution into in-person voting, among mail ballot requestees who face heightened risk.

5) We estimate ${\sum }_t{p}_t{N}_t{\mathrm{\delta }}_a$ using ${\sum }_{t>\overline{t}}{N}_t\widehat{p_t{\mathrm{\delta }}_s}s\widehat{\mathrm{\beta }}$.

Figure 4 illustrates the first three steps of our model. The figure focuses specifically on people who expressed interest in voting in close proximity to the election. It shows the share of mail ballot requestees who substitute into in-person voting by the date of their mail ballot request. To get a baseline rate of substitution into in-person voting when people do not face a heightened risk of their mail ballots being received late, we set $\overline{t}$ equal to 11 October and calculate that about 4.9% of people who requested mail ballots in the week of 5 October through 11 October ultimately cast an in-person ballot. We observe a similar rate of substitution into in-person voting among people who requested mail ballots in the week of 12 October through 18 October, meaning that $\widehat{p_t{\mathrm{\delta }}_s}$ is small for $t$ in that range. However, we start observing more substantial $\widehat{p_t{\mathrm{\delta }}_s}$ for larger $t$. More than 10% of people who requested mail ballots on the final 2 days before the deadline substituted into in-person voting.

Fig. 4. Share of in-person substitution by date of mail ballot request.

The left panel of Fig. 5 visualizes step four of our model. In this graph, the x axis represents the share of people who requested a mail ballot on day $t$ who substituted into voting in person, $Y_t^s$. On the y axis is the share of people who requested a mail ballot on day $t$ who did not vote, $Y_t^a$. Each dot represents the combinations of $Y_t^s$ and $Y_t^a$ on every day between 5 October and 1 November, with a clear positive association between these two variables. The regression line shows that we estimate a $\widehat{\mathrm{\beta }}$ of approximately two, meaning that for every one percentage point increase in the share of mail ballot requestees who substitute into in-person voting we observe about a corresponding two percentage point increase in the share of mail ballot requestees who do not cast a counted ballot.

Fig. 5. In-person substitution and unreturned or late mail ballots.

The right panel of Fig. 5 provides a test of the assumption underlying our model. Namely, that increased substitution occurs because mail ballot requestees face heightened risk that their mail ballots will otherwise be rejected as late. While the x axis remains the same across the two panels, the y axis is now the share of people who requested a mail ballot on day $t$ who cast a mail ballot that was rejected as late. Consistent with our expectation, there is a strong positive association between the number of people substituting into in-person voting and the share of people returning late mail ballots. Only about 0.3% of people who requested mail ballots during the period from 5 October to 11 October returned a late mail ballot, which is the period during which we assume no one faces a heightened risk of returning a late mail ballot. This rate of late ballot return increases about 10-fold to about 2.3% among people who requested mail ballots during the final week before the deadline, which also was the period when we observe the most in-person substitution. That said, the observed 0.38 percentage point increase in late ballots from a 1 percentage point increase in the share of mail ballot requestees who substitute into in-person voting is only a fraction of the increase observed in people who do not vote.

Table S8 shows how we apply what is observed in Figs. 4 and 5 to produce an estimate of the total number of people who were deterred from voting by the Election Day receipt deadline. Ultimately, we estimate that about 11,000 people were deterred, with about 4500 of these deterred ballots being held by people who requested their mail ballots on the last 2 days before the mail ballot application deadline.

Electoral consequences of lost votes

While lost votes by mail are concerning regardless of whether they affect election outcomes (47), this section assesses whether mail balloting procedural requirements could have electoral consequences. We first consider the consequences in Pennsylvania before again noting that many states lack the same conditions necessary for lost votes to change the outcome of an election.

Pennsylvania consequences

Our analysis focuses on registrants who requested a mail ballot but did not ultimately vote. For ease, we refer to these as registrants with an unrealized mail vote. We construct a measure of the difference in the number of Democratic registrants and Republican registrants with an unrealized mail vote. We prefer unrealized mail ballots to rejected mail ballots because of issues with both incomplete coverage and deterrence. However, our measure likely overstates the net partisanship of lost votes by mail because it also includes registrants who abstain from returning their mail ballots.

While lost mail votes were not electorally consequential statewide in 2022, they could be in future statewide elections. Statewide there were about 51,000 more unrealized mail votes among Democratic registrants than Republican registrants. Given that the Democratic candidates for senator and governor won in 2022 by roughly 250,000 and 750,000 votes, respectively, neither race could be affected by unrealized mail ballots. However, Pennsylvania has experienced several tight statewide elections in the recent past. For example, the 2021 state supreme court race was decided by about 25,000 votes, while the 2016 presidential race came down to about 45,000. Thus, it is not far-fetched that Pennsylvania could experience a close enough statewide race for lost votes by mail to affect its outcome.

More generally, electoral consequences are about more than statewide general elections. Figure 6 expands our focus to state legislative contests, comparing the vote margin in state legislative races (x axis) to our measure of differential partisanship among unrealized mail votes in the district (y axis). On the basis of the figure, the people at risk of having experienced lost mail votes are more likely to support Democratic candidates than members of the broader Pennsylvania electorate in 196 of the 203 state districts. Crucially, the Democratic bias in unrealized mail votes is present in the state legislative contests decided by narrow margins. In the five races in which the margin of victory was less than 1000 votes, we observe an average of 220 more Democrats than Republicans with unrealized mail votes. In one race, the Democratic candidate lost by only 76 votes in a contest with 177 more unrealized mail votes from registered Democrats than registered Republicans.

Fig. 6. Potential electoral consequences of unrealized mail votes by state district.

Recent postelection challenges to mail ballot rules support our conclusion that lost votes by mail could be electorally consequential in particularly close Pennsylvania elections. Consistent with our state legislative analysis, there have been at least two recent postelection challenges in which the outcome of an election appears to hinge on mail ballot rules (48, 49).

Consequences in context

The electoral consequences of mail balloting procedural requirements will necessarily vary by state. The likelihood that mail balloting procedural requirements are electorally consequential is increasing in the share of voters who cast mail ballots, the magnitude of the vote mode splits, the likelihood that the requirement causes rejection or deterrence, and the share of voters who are not notified when their mail ballots are rejected.

On the basis of the available evidence, we believe that lost votes by mail are more likely to be electorally consequential in Pennsylvania than most other states, although more state-by-state work is necessary to properly measure lost votes. To start, Pennsylvania has relatively more mail votes. According to the EAVS, mail votes accounted for about 23% of all counted votes in Pennsylvania’s 2022 general election (13). In contrast, the median state had just 12% mail votes. Table S9 shows that the partisan split in mail versus in-person ballots is greater in Pennsylvania than in many other states. Further, on the basis of Fig. 1, Pennsylvania rejects relatively more mail ballots. As explained in more detail above, Pennsylvania does not have a statewide notice policy, which should lead to relatively less substitution, and has a receipt rather than postmark deadline, which should lead to relatively more deterrence. Finally, Pennsylvania, unlike many other states, often experiences competitive statewide elections. Nonetheless, all states typically have some competitive legislative elections that could be affected when all of these other conditions are present.

DISCUSSION

We use Pennsylvania as a case study to correct the measurement of lost votes by mail. To summarize, we first focus on the coverage of rejected mail ballots and show that there were at least 5000 rejected mail ballots in Pennsylvania’s 2022 general election not documented as such. We then consider that some rejected mail ballots are saved from becoming lost votes and some lost votes are not reflected in rejected mail ballots. With respect to the former, we find that about 3500, or about 15% of, people whose mail ballot was rejected nonetheless successfully voted. Further, substitution was twice as likely in counties with notification rather than without. With respect to the later, we estimate that about 11,000 people did not vote because they anticipated it would be received too late to count. Assuming that up to three-quarters of undocumented rejections become lost votes, we ultimately estimate that there were at least about 11,000 more lost votes than documented, rejected mail ballots.

Our specific findings in Pennsylvania help assess the normative implications of the state’s mail balloting regime. In general, there are at least two potential criticisms of our description of votes as “lost.” First, any fraudulent mail ballots prevented by procedural requirements are by definition not lost votes. However, mail ballots rejected in Pennsylvania are unlikely to have been cast by ineligible voters. For example, the date affixed next to a voter’s signature on the mail ballot affidavit is not useful for assessing registrants’ eligibility to vote. Further, secrecy envelopes are not about eligibility to vote either and, moreover, do little for ballot privacy now that Pennsylvania processes mail ballots centrally using machines rather than by poll workers at voters’ polling places. Second, many of the ballots rejected or deterred by the Election Day receipt deadline might be considered the “fault” of the registrants because they waited until so close to Election Day to request the ballots. However, in the same way that that the CalTech-MIT Voting Technology Project considered a registration mix-up a lost vote (14), even though the voter could have updated their address of registration, we consider deterred votes to be lost votes because there are alternative ways of structuring the mail balloting systems that would allow most of these ballots to be counted. While some aspects of our case study reflect Pennsylvania’s unique position in the political landscape, our results are both illustrative and informative of broader dynamics in mail balloting.

One implication of our findings is that scholars should use more caution when working with the EAVS. For scholars, it is helpful to recognize that the EAVS is not so much a simple collection of state data as an attempt to wrangle state data into uniform, national categories, a process that is never perfect. Nonetheless, the EAVS may consider supplementing its instructions to reduce incomplete reporting. In Pennsylvania’s case, the underreporting reflects, in part, the state’s particular treatment of misdated and undated ballots. However, the EAVS appears to undercount mail ballot rejections in many other states too. Table S5 compares the number of rejected mail ballots reported in the EAVS in both the 2016 and 2020 general elections to the number publicly reported by state election officials (50). The EAVS exactly matches state reports in 6 of the 21 states surveyed in 2016 and only 2 of the 23 states surveyed in 2020. The median difference between the EAVS and state reports is 2.3% in 2016 and 2.8% in 2020, showing that the EAVS is more often an underreport than overreport.

Another implication of our findings is that people should be careful when evaluating mail ballot systems based on the share of rejected mail ballots. For example, the Elections Performance Index (51, 52) ranks states in part based on the number of rejected mail ballots, in addition to the total number unreturned. Yet various features of mail balloting regimes can affect the share of rejected mail ballots that become lost votes. Given differences across states, we interpret comparisons of rejected mail ballots over time and jurisdiction (26) with some caution, because it does not consider, or correct for, any differences in coverage, substitution, or deterrence.

A third implication is that lost votes can be reduced by notifying voters of mail ballot rejections with sufficient time to take corrective action. In general, we observe higher rates of substitution when a deficient ballot could be identified during preprocessing than when it could not. However, the highest rate of observed substitution is 28% among voters with an affidavit error in a county with a notification program. In contrast, 82% of North Carolina voters who submitted a deficient, but correctable, mail ballot in the 2020 general election ultimately took such action (30). While there are several potential explanations for why so many more North Carolinians took corrective actions than Pennsylvanians, one structural difference between the two states is that North Carolina had a cure process through which voters could take corrective action without needing to submit a new ballot. As Meredith and Kronenberg (30) note, there are several reasons why a curing process makes it easier to take corrective action. At some point in the electoral calendar, there is not enough time to mail new ballots. Further, some people vote by mail because they find it costly to vote in person. In addition, a curing process can be used to count ballots with errors discovered on Election Day itself.

Beyond instituting ballot curing, several changes could be made to states’ electoral systems to further reduce the number of lost votes by mail. In broad terms, the number of lost votes by mail is a function of (i) the number of deficient mail ballots, (ii) the identification of deficiencies, (iii) notification of deficiencies, and (iv) the required corrective action. In this context, ballot curing reduces the cost of taking (iv) the required corrective action. However, there are policies and procedures that can reduce lost votes earlier in the process. For example, following the 2022 general election, Pennsylvania election officials redesigned mail ballots with an eye on (i) reducing the number of deficiencies (53). Lyons (53) notes that, among other things, secrecy envelopes are now yellow and watermarked to discourage stray marking, return envelopes highlight where voters should sign and date affidavits, and affidavits partially prefill the current date to remind voters to not write their birthdate. Pennsylvania officials have also made it easier to (ii) identify deficiencies. For example, the state now allows counties to hole punch the return envelope so election workers can identify when the now-yellow secrecy envelope is missing without canvassing the ballot. A preliminary evaluation of mail ballots rejections in the 2024 primary election suggests that these changes modestly reduced the number of rejected mail ballots (54). Beyond these reforms, we also think it would be helpful to improve (iii) the notification of deficiencies by encouraging people requesting a mail ballot to provide election officials with contact information, which could then be used to inform them of any deficiency.

Finally, our findings also highlight the need for more research on the consequences of deadlines for requesting mail ballots. One reason that voters have little time to correct defective mail ballots is that many wait until just before the deadline to request a mail ballot. Moreover, we observed that thousands of people either returned a mail ballot that was received too late to count or who were deterred from returning their mail ballot because it was unlikely to be received in time to count. Ideally, these registrants would have requested these mail ballots earlier to increase the amount of time they have to complete the process. For those people who are motivated to take action by deadlines, moving the deadline would likely reduce the number of lost votes. On the other hand, some people are not motivated by deadlines as much as they realize late in the electoral calendar that they would like to vote by mail. Ultimately, more evaluation is needed on how to set mail ballot deadlines to balance these competing considerations. Such evaluations may find inspiration from existing work that has considered how voter registration deadlines affect voter turnout (55).

MATERIALS AND METHODS

Our analysis relies on two different individual-level administrative data sets provided by the Pennsylvania Secretary of State’s office through public records requests, as well as aggregate data about mail ballot rejections and notice procedures made publicly available during recent litigation.

Individual-level administrative data

Our primary dataset is individual-level mail ballot applications recorded in the Statewide Uniform Registry of Electors (SURE). We limit the data to approved applications associated with a valid voter registration number, which did not result in a voided mail ballot. In the limited circumstances in which a given person was sent multiple nonvoided mail ballots, we select one mail ballot based on the following rule: We sort the data by mail ballot status, with any rejected mail ballots first, any accepted mail ballots second, and any unreturned mail ballots last, and take the first such mail ballot. For example, if a person was sent two mail ballots, one of which was returned and accepted and one of which went unreturned, we would classify the individual as having an accepted mail ballot. We then link the mail data to registration records. The linkage is straightforward because each data source contains registrants’ unique voter registration number.

Table S1 shows that nearly all records in the SURE data are linked to a registration record (all but about 31,000 of the roughly 1,435,000 records in the SURE data). The SURE records that do not link to a registration record likely represent mail ballot applications made by registrants who had their registrations canceled between when they applied for the mail ballot and 9 January 2023, when we received the voter registration data. Given that our primary analysis focuses on mail ballot applications successfully linked to voter registrations, table S1 shows that we will slightly underrepresent cases in which mail ballots went unreturned or were cancelled.

We use several other variables contained in the SURE data and voter registration records. The SURE data contain information about when election officials received a mail ballot application, what address the ballot was sent to, and the status of the mail ballot. While these data also include information about when ballots were mailed to registrants, we avoid using that information in our analysis because the lag between the receipt of an application and the mailing of a ballot varied by county and over time. Separately, the voter registration data contains registrants’ recorded voting history, including vote mode, registration address, party of registration, and date of birth. To proxy for whether someone had the mail ballot sent to their address of registration or some other address, we compare the zip code where the mail ballot was sent with the zip code of the registrant’s registration address (30). If a mail ballot was sent to zip code other than the zip code of the registration address, we note whether that zip code is within the registrants’ home county, whether it went to a zip code within Pennsylvania outside of their home county, or whether it went to a zip code somewhere else in the US or an international location.

We supplement our primary dataset by constructing estimates of registrants’ likely race and ethnicity. To construct these estimates, we geocode registrants’ addresses of registration in order to obtain their Census block or Census tract of residence. We combine information about registrants’ surnames and the racial demographics of their Census blocks to generate probabilities that each registrant is white, Black, Hispanic, Asian, and any other race or ethnicity, using the method developed by Imai and Khanna (56). We also create a sixth category for when the geocode is not of sufficient quality to impute race (57).

County-level litigation data

We supplement the individual-level dataset of mail ballot applications with county-level data on mail ballot rejections. Much of these data were made publicly available during litigation over Pennsylvania’s mail ballot requirements. During the course of litigation, the plaintiffs served on the defendant county election boards a series of interrogatories about the implementation of mail ballot requirements in the 2022 general election. The plaintiffs then attached the county responses to their statement of material facts as part of their motion for summary judgment. We were not involved in the litigation but instead downloaded the responses from the public docket (33).

On the basis of these interrogatories, we identified the number of undated or misdated mail ballots rejected in each county. We only counted ballots that were rejected for being undated or misdated if they would have otherwise counted but for the deficient affidavit. Thus, when a county noted that a ballot was rejected because the affidavit was undated and it was submitted outside of a secrecy envelope or was subsequently cured, we did not count it.

We also use these interrogatories to measure whether a county had a program in place to notify voters if they returned a mail ballot with a disqualifying deficiency. In theory, Pennsylvania’s statewide SURE system would attempt to notify voters in all counties when a county entered the status of the mail ballot as rejected in the system. However, that notification system was of limited utility. It only worked for voters with an email address associated with their voter registration and then only if election officials logged the rejected mail ballot before Election Day. In addition, because Pennsylvania law forbid canvassing mail ballots before Election Day, this could only happen if a county engaged in preprocessing. As a result, we did not consider the 43 counties using only the SURE system to be a specific program of notifying voters of ballot issues in advance of the election. Instead, we only considered a county to have a specific notification program when their response to the interrogatory indicated that they contacted the voter directly (13 counties), posted a list of voters with ballot issues online (3) or in a public place (1), or shared such a list with the political parties (4).

We also rely on additional sources that documented the number of rejected mail ballots by rejection reason in several counties. These sources include Board of Election meeting minutes or transcripts, media reports, and direct communications from counties’ Board of Elections.

Supplementary Materials

This PDF file includes:

Supplementary Text

Tables S1 to S9

References