Unrivaled Politics

Facts and Data First. Politics Second.

Category: Analysis of Voting and Elections

Look-Back Assessment of the Pennsylvania 2024 Voter Turnout Model: Findings and Interpretive Discussion

Apr 7th, 2025
As described in a previous Unrivaled Politics blog post, the Pennsylvania Voter Turnout Model is designed to forecast participation rates in an upcoming election for Pennsylvania registered voters. The model was developed in advance of the November 5, 2024 election and has been applied to generate turnout predictions for that election. This article assesses the performance of the turnout model via a look-back comparison of projected to realized outcomes from that election.

The model segments voters based on their history of participation in prior elections; on active vs. inactive status as assessed by election officials; and on other voter characteristics including party affiliation, county of residence, and age. In the current version of the model each segment’s projected turnout rate is its observed participation rate from the 2020 general election.

The model was developed using data on historical participation rates are obtained from the Pennsylvania Department of State’s Full Voter Export (FVE) file as of August 12, 2024.¹ Specifically, the calculated turnout frequencies are those observed for the subpopulation of the full FVE sample whose presence predates (whose registration dates are prior to) the registration deadline for the 2020 general election. ²

As with any projection model based on extrapolation from historical experience, realized outcomes may diverge from projections. In our context, the realized participation rate within a given voter history segment may diverge from the outcome projected by the model.

For instance, realized outcomes may diverge because of how voters react to the particular candidates or issues specific to each election cycle. Also, the composition of the voter population may change over time in ways that affect the segment participation rate. In the latter case, changes in size or demographic composition of the segment population might be observable, but the effects of these on participation rates may not be foreseeable.

That said, there are good reasons to expect projections to be reasonably accurate, especially in regard to rank ordering of participation likelihoods across voter history segments. Voters who have tended to vote sporadically in the past typically will continue to do so. Those who rarely have missed voting in an election almost surely will vote in the next one. Inactive voters have been classified as such by election officials for a substantive reason (such as failure to respond to an address verification notice), so that inactive status should correlate with relatively low likelihood of voting in the next election.

Performance analysis results by party affiliation

Projected outcomes for our assessment of model performance are obtained from applying the model to the population of registered voters in the FVE file as of October 21 (the registration deadline for the November election.) Realized outcomes–whether and how each registered voter cast a ballot in the election–are obtained from the January 20, 2025 FVE file.³

Table 1 shows the gap between projected and actual turnout in 2024 by party affiliation: turnout rate differentials in panel A and projected versus actual counts of voters in panel B. The Democrats achieved a turnout rate 1.5 percentage points higher than predicted (79.4 versus 77.9 percent), with actual turnout exceeding projected by about 56,000 voters. The Republican turnout rate (85.7 percent) was almost identical to predicted.

Although this might seem to be a small differential, in a closely contested battleground state like Pennsylvania even small turnout differences might swing an election one way or the other. In this case, the end result was that the Republican vote count in 2024 exceeded the Democrat count by about 52,000 instead of the projected 111,000 voters.

Table 1: Prediction error by party affiliation

Turnout percentage

Party
actual projected gap
Democratic 0.794 0.779 0.015
Republican 0.857 0.858 -0.001
Other 0.678 0.694 -0.014

Turnout count

Party
actual projected gap
Democratic 3,004,427 2,948,447 55,980
Republican 3,056,493 3,059,585 (21,211)
Other 925,607 946,908 (3,092)

Source: January 20, 2025 PA FVE file; Pennsylvania Voter Turnout Model

What factors lay behind the higher-than-predicted turnout of Democrat registered voters in 2024? One plausible explanation derives from the fact that between November 2020 and November 2024, the Democrats experienced a net loss of about 230,000 party-affiliated registered voters.

It is conceivable that these exiting voters had been less inclined to vote in 2020 compared to others within the same segment—a conjecture consistent with these voters’ apparent dissatisfaction with the party. With their departure, the party’s overall turnout rate would then rise.

It also is possible that some of the lift to Democratic turnout came from crossover voters—disaffected Democrats who had low turnout in prior elections but opted to vote for Trump in 2024 (while not changing their affiliation.) This conjecture is consistent with reported findings from national polls showing increased support for Trump within certain demographic segments of Democrat voters in 2024, such as younger men.

Also in Table 1, we see that turnout was lower than predicted among unaffiliated voters. The realized turnout rate for this cohort was 63.2 percent, lower than their predicted 65.4 percent, with actual turnout falling short by about 45,000 voters.

Reportedly, many voters from across the political spectrum in 2024 viewed both Harris and Trump as significantly flawed candidates. This view may have been particularly strong among independent voters and those not affiliated with either of the two major parties, providing a potential explanation for the relatively low turnout among these voters.

Drilling down further

While, the accuracy of projected turnout rates by party affiliation in an upcoming election is of foremost interest, having direct implications for the election outcome, accuracy within distinct voter segments is also of interest. Segment-level analysis may suggest ways to improve overall accuracy of the model. It also provides insight into robustness of the model–how accurate its projections would be if the composition of the population were to shift.

The model’s projections proved highly accurate for the segment consisting of voters who turned out for the primary preceding the general election. As shown in Table 2, the gap between actual and projected turnout rate for these voters is no greater than 6-tenths of a percentage point in each party affiliation category.

Table 2: Model performance for the segment of primary voters

Party
actual projected gap
Democrat 0.985 0.979 0.006
Republican 0.985 0.983 0.002
Other 0.949 0.953 -0.004

Also noteworthy, but for a different reason, is the model’s performance within the inactive voter segment (exclusive of primary participants). Here, a substantially lesser degree of accuracy is observed. As shown in Table 3, actual exceed projected turnout rate for these voters by multiple percentage points in each party affiliation category.

Table 3: Model performance for the segment of inactive, primary non-participants

Party
actual projected gap
Democrat 0.432 0.37 0.062
Republican 0.515 0.395 0.12
Other 0.378 0.337 0.041

Because these prediction gaps are positive and large, they have a material influence on model performance within the subpopulation comprising non-participants in the primary. This is seen in Chart 1, which shows the prediction gaps for that population overall and by active versus inactive, by party affiliation. (Percentage point gaps are shown in panel A and count gaps in panel B.)

Chart 1: Actual minus predicted turnout rate by party affiliation and active versus inactive status

Turnout percentage

Vote count

These results suggest that after 2020, the inactive voter classification process became less precise at distinguishing voters who had exited the Pennsylvania voting population (among those who had not participated in the primary.) Accordingly, in 2024 relative to 2020 there would have been more “false positives” (willing and able to vote) within the inactive segment. Conversely, there would have been more “false negatives” (unable or unwilling to vote) among those classified as active.

Let us now drill down yet further within the cohort of active, primary non-participants in 2024, distinguishing between those who voted by mail-in, absentee or provisional ballot in 2020 and others. Within this cohort, those who had voted by mail-in, absentee, or provisional ballot in 2020 had weaker 2024 turnout than those who had voted at the polls in 2020, relative to what the model had projected. This was the case within each party affiliation category, as shown in Chart 2.

Specifically, among Democrats, actual 2024 turnout exceeded projected by nearly 3 percentage points for those who had voted at the polls in 2020, while actual was 6-tenths of a percentage point below projected for the mail, absentee or provisional cohort. Among Republicans, actual 2024 turnout was about equal to projected for those who had voted at the polls in 2020, while actual was a full 6 percentage points below projected for the mail, absentee or provisional cohort. The latter cohort also underperformed relative to the at-polls cohort among voters in the “other” party affiliation category.

Chart 2 similarly shows that active, primary non-participants in 2024 who had been registered but did not vote in 2020 had weaker 2024 turnout than those voting at the polls in 2020, relative to what the model had projected. Again, this was the case within each party affiliation category.

Chart 2: Prediction error by party affiliation and 2020 method of voting

(cohort of active, non-primary participants in 2024)

Turnout percentage

Vote count

Interpretive discussion

The factor or factors underlying the unexpectedly high turnout of inactive voters are not directly observable. As noted previously, their higher turnout suggests that being labelled inactive had become a less precise indicator of exit from the state’s voting population. But why might this be the case?

One plausible scenario is that the state has had less success after 2020 at identifying, and classifying as either ineligible or inactive, individuals who had relocated to other states. Hence, assignment of voters to inactive status became more idiosyncratic (such as due to voters relocating within the state and failing to update their registration data) and thus less reflective of actual exit from the voting population.

Consistent with this scenario is the fact that in early 2023, Florida, Ohio, Virginia, and West Virginia each pulled out of the cross-state data cooperative known as the Electronic Registration Information Center (ERIC).⁴ States that participate in ERIC share their voter registration data and receive notification about voters who appear to have relocated to other states. Each of these states that exited ERIC is a relatively common destination for voters relocating from Pennsylvania. Their departure could have impaired Pennsylvania’s efforts to maintain the accuracy of its voter registration database, including by labelling potentially relocated voters as inactive.

No-excuse mail-in ballots, first introduced in Pennsylvania for the 2020 election, may have further compromised the state’s ability post-2020 to identify voters who had moved out-of-state. No-excuse mail-in voting may encourage or facilitate cross-state voting, a phenomenon examined in a 2024 study from the National Bureau of Economic Research.⁵ Cross-state voting by relocated voters via mail-in ballot could preempt their being assigned to the inactive list or being removed from the registration rolls.

This line of reasoning suggests that the state’s ability to identify inactive or exited voters might have been undermined primarily among the population voting by mail during 2020. Thus, this reasoning is consistent with the finding that in 2024, active primary non-participants who had voted by mail in 2020 had substantially lower turnout rates than those who had voted at the polls in 2020, relative to the model’s projections.⁶

The preceding scenarios are clearly plausible. Confirmation of their explanatory relevance, however, requires gathering more direct evidence.

Stability of the Rank Ordering

Although, there are aspects of divergence between actual and modelled turnout rates, as described above, the model reliably differentiates turnout likelihood across voter turnout segments. This is evidenced by stability of the rank ordering of turnout rates across modelled segments by party affiliation, as seen in Chart 3.

The five pairs of bars in Chart 3 depict projected and actual turnout rates for five turnout likelihood segments of voters. The first segment from the left comprises primary participants. The next three segments classify non-participants in the primary who are active into three ranges of projected turnout likelihood: high (90 percent or greater turnout likelihood); high-mid (at least 80 but less than 90 percent); and lower (below 80 percent). The last segment comprises inactive voters.

Both projected and actual turnout rates decline across segments moving from left to right in the chart. In most cases, the changes in actual and projected turnout rates closely align across adjacent segments, demonstrating stability and reliability of the model’s rank ordering of turnout likelihoods.

For each party affiliation, the largest misalignment occurs between the lower propensity range of active voters and the inactive segment. Here, the model’s projected decline in turnout rate overstates the actual decline, consistent with the earlier discussion on model accuracy.

For Democrats and Republicans, the next largest divergence occurs between the high and high-mid propensity ranges of active voters. Here, the model’s projected decline modestly overstates the actual decline for Democrats and understates it for Republicans. For the Other affiliation category, the second largest divergence occurs between the high-mid and lower propensity ranges, where the model understates the actual decline in turnout rate.

Chart 3- Stability of rank ordering by model turnout propensity

Concluding Summary

Projected voter turnout rates by party affiliation from the Pennsylvania Voter Turnout Model were close to actual. For Democrats, the actual turnout rate exceeded projected by 1.5 percentage points, while for Republicans, actual and projected were virtually equal. The model’s understatement of Democrat turnout conceivably might be tied to an exodus of less engaged voters from the Democratic party after 2020.

The model also reliably differentiates turnout likelihood across turnout propensity segments. While the model exhibits a high degree of accuracy, one should bear in mind that even small deviations in turnout can swing an election one way or the other in a closely contested state like Pennsylvania.

The model appears to be least accurate at projecting turnout rates of inactive, primary non-participants. This outcome suggests that Pennsylvania has had less success since 2020 at identifying and classifying as inactive, individuals who had relocated to other states.

One factor that may have impeded the state’s ability to identify these voters is the departure from ERIC in early 2023 of Florida and a few other states that Pennsylvanians commonly relocate to. No-excuse mail-in ballots, first introduced in Pennsylvania for the 2020 election, may have further hindered the state’s ability to identify voters who had moved out-of-state.
1. The FVE data can be accessed here . ↩︎
2. Thus, the estimation sample comprises only voters who were present in 2020 and have survived through 2024. We are unable to estimate a model using the full population that was present in 2020 because we lack access to archived FVE data from 2020. ↩︎
3. The FVE data distinguishes between four methods of voting: at polls (in -person machine voting), provisional ballot, mail-ballot (no excuse mail-in), and absentee ballot (including military and other traditional absentee ballots). A missing value indicates non-participation. ↩︎
4. See Miles Park, “Republican States Swore Off a Voting Tool. Now They’re Scrambling to Recreate It.” NPR.org, October 20, 2023. (Link) ↩︎
5. See Gordon B. Dahl, Joseph Engelberg, Runjing Lu, and William Mullins, “Cross-State Strategic Voting.” National Bureau of Economic Research, Working Paper 30972, September 2024. (Link) ↩︎
6. Similarly, the departures from ERIC may have curtailed the state’s ability to identify inactive or exited voters among the cohort who were registered but did not vote in 2020. As noted, this cohort had weaker 2024 turnout than those voting at the polls in 2020. Alternatively, this cohort may simply have become even less engaged as of 2024 than they were in 2020. ↩︎
Swing or Trajectory? The Shifting Political Landscape of Pennsylvania

Mar 16th, 2025
In the 2020 Presidential Election in Pennsylvania, Joe Biden prevailed by about 80,000 votes over Donald Trump. In 2024, against Kamala Harris, Trump prevailed by about 120,000 votes, thus gaining about 200,000 votes on net.

Might the 2024 outcome signify a fundamental shift in Pennsylvania’s political landscape? This note delves into this question from a data-driven perspective.

Inferences from registration and participation data

Two key factors indicate that at least a part of the Republican gains in 2024 are more than just fleeting and will be challenging for the Democrats to turn back.

First, between November 2020 and November 2024, the number of voters registered as Democrats declined substantially, while the counts of registered Republicans and of those unaffiliated with either major party increased, as seen in chart 1. Moreover, those increases were a continuation of gains accrued between 2016 and 2020 by those groups.

Chart 1: Pennsylvania registered voter counts by party affiliation

Source: PA Department of State 2016, 2020, and 2024 Voter Registration Statistics

Going into the November 2020 election, the Democrats had enjoyed an advantage of 686,000 more registered voters compared to the Republicans (rounding to the nearest 1,000). Because of each party’s gains and losses over the interim period, as of November 2024, the Democrats’ registered voter advantage had fallen to 286,000.

Second, based on extrapolation from 2020 turnout rates (using statistically modeled turnout likelihoods), the Republicans had a substantial turnout rate advantage over the Democrats going into the November 2024 election.¹ Democrat registered voters had an expected turnout rate of 77.9 percent, compared to 85.8 percent for the Republicans and 69.4 percent for unaffiliated and other voters.

Consequently, the Democrats’ shrinking margin in raw number of registered voters enabled the Republicans in 2024 to surpass the Democrats in participation-weighted count—the number of voters expected to participate in the November 2024 election. As shown in chart 2, the expected count of participating Republican voters, based on modeled turnout likelihoods, exceeded that of the Democrats by about 110,000, presaging a Republican victory.

Chart 2: Predicted (turnout-weighted) 2024 counts by party affiliation

Source: Unrivaled Politics data turnout model, from PA voter registry

Democrat turnout in 2024 ended up higher than expected

Model-based predictions, of course, are subject to error. Chart 3 shows the gap between predicted and actual turnout in 2024 by party affiliation: the turnout rate differentials in panel A and predicted versus actual counts of voters in panel B.

The Democrats achieved a turnout rate of 79.4 percent, higher than the predicted 77.9 percent, with actual turnout exceeding predicted by about 56,000 voters. The Republican turnout rate (85.7 percent) was almost identical to predicted. The end result was that the Republican vote count exceeded the Democrat count by about 50,000 (instead of the predicted 110,000).

What factors lay behind the higher-than-expected turnout of Democrat registered voters in 2024? One plausible explanation is a “selection effect” from voters who had left the Democratic Party (by switching to unaffiliated or Republican) between 2020 and 2024. It is conceivable that these exiting voters had had been less inclined to vote in 2020 compared to others within the same segment—a conjecture consistent with these voters’ apparent dissatisfaction with the party. With their departure the party’s overall turnout rate would rise.

It is also conceivable that the lift to turnout came from disaffected Democrats who had low turnout in prior elections but opted to cross-party lines and vote for Trump in 2024, without changing their affiliation. This conjecture is consistent with reported findings from national polls showing increased support for Trump within certain demographic segments of Democrat voters in 2024, such as younger men.

Chart 3: Prediction error by party affiliation

Chart A: Turnout Percentage Chart B: Vote Count

Source: Unrivaled turnout model (from PA voter registry)

Independent and cross-over voters necessarily account for the remaining margin

The analysis thus far has established that if voting had been strictly along party lines, with Trump and Harris tied among independent voters or those not affiliated with either major party, then Trump would have won by 50,000 votes. It then follows purely mathematically that the residual 70,000 vote portion of Trump’s winning margin must have come from getting more cross-over votes from Democrats than Harris received from Republicans, or from dominance among unaffiliated voters, or from some combination thereof.

In this sense, most of Trump’s margin appears to have come from swing voters, who tend to be less wedded to a partisan point of view and comparatively responsive to individual candidates and the candidates’ stands on key issues. Underscoring this point is the fact that in the Pennsylvania senatorial race, the Republican candidate Dave McCormick garnered only about 15,000 more votes than the incumbent, Democratic candidate Bob Casey Jr., mostly due to Trump voters who had opted not to support McCormick.²

On the role of mail-in voting in 2024

Reportedly, many voters from across the political spectrum in 2024 viewed both Harris and Trump as significantly flawed candidates. As was seen in Chart 3, turnout was lower than expected among unaffiliated voters, perhaps for that reason. However, Republicans did not have noticeably lower turnout, and Democrats had higher turnout relative to model predictions.

Potential explanations for Democrats’ higher-than-expected turnout have already been offered. But what might explain the sustained high turnout of Republican voters?

One possibility is that Republican voters who may have disliked Trump still voted for him because of concerns about what a Harris presidency might bring (in terms of a doubling down on Biden’s most divisive policies). Another is that the unrelenting, brutal lawfare and perceived weaponization of justice against Trump motivated Republican voters to circle the wagons, so to speak.

Another important factor to consider is stronger embrace of mail-in voting by the Republican party in 2024 relative to 2020. Mail-in voting can bolster participation by making the process more convenient for some voters and less susceptible to Election Day attrition. Thus, greater receptivity toward mail-in voting may have helped boost Republican turnout and offset lack of enthusiasm for Trump among some elements of the party.

During the 2020 election, the ongoing coronavirus pandemic spurred popularity of mail-in voting, due to perceived risks of social contact at polling places. However, the mail ballot option proved far more popular among Democrats in 2020 compared to Republicans. The latter held far less confidence in the security and reliability of mass mail-in voting, which had been newly introduced into Pennsylvania.³

This attitudinal gap between Democrats and Republicans in 2020 was evidenced by the fact that about 58 percent of the votes for Biden were by mail-in or absentee ballot, compared to just 18 percent of Trump votes. In 2024, however, this attitudinal gap had moderated, likely due in part to organized efforts, such as an initiative known as the “Pennsylvania Chase”, that promoted mail-in voting among Republicans.⁴

This change in attitude is demonstrated in Chart 4, which shows the ratio of Democrat to Republican mail-in ballot share (frequency of voting by mail-in ballot) across four segments of the 2024 participating voter population.⁵ The first bar from the left shows the ratio of mail-in ballot shares in 2020 for the subpopulation comprised of voters who participated in both the 2020 and 2024 elections. The second bar from the left shows the ratio of mail-in ballot shares in 2024 for the same subpopulation.

The third bar shows the ratio of Democrat to Republican mail-in ballot shares in 2024 for participants in 2024 who were present but did not vote in 2020. The final bar shows the ratio of Democrat to Republican mail-in ballot shares in 2024 for participants in 2024 who were not present in 2020.

Thus, in 2020 ,the frequency of mail-in voting by Democrats was 2.8 times that of Republicans among voters who participated in both elections. The ratio of Democrat to Republican mail-in voting frequency within this cohort declined to 1.7 in 2024. The other two cohorts had ratios in 2024 of 1.4 and 1.7, respectively.⁶

This analysis suggests that organized efforts to normalize mail-in voting among the Republican electorate had succeeded, whereby they became more receptive to the mail-in option than they had been in 2020. As noted, such increased openness toward mail-in voting could have bolstered turnout.

Chart 4: Ratio of Democrat to Republican mail-in ballot share by voter history segment

Source: Unrivaled voter turnout calculations using the PA Voter Export data file

Concluding summary

Since 2016, there has been a steady increase in the number and percentage of registered voters in Pennsylvania choosing to affiliate with the Republican party. The number and percentage of unaffiliated voters has also grown. There are roughly 220,000 fewer registered Democrats in the state today compared with 2020 (or with 2016).

The Democrats still enjoy a raw numerical advantage in Pennsylvania—the number of registered Democrats exceeds the count of registered Republicans by over 300,000 voters. However, the Republican share of the voter population has clearly been on a upward trajectory.

Moreover, the Republican Party in 2024 benefitted from a substantially higher turnout rate compared to Democrats. Their turnout rate advantage provided them with a slim, turnout-weighted margin, whereby they outvoted Democrats by about 50,000 votes cast. Efforts to make Republican voters more receptive to vote-by-mail plausibly helped bolster the party’s turnout in 2024.

In 2024, Trump prevailed over Harris by 120,000 votes, from which we can deduce that he netted some 70,000 votes from cross-over Democrats and independent voters. However, these should be considered swing voters who might readily reverse course in the next election.

The current, slim, turnout-weighted difference in number of Republican versus Democrat registered voters signifies that Pennsylvania remains a swing state—the decisions of cross-over and independent voters can still swing an election one way or other. Yet there is no question that the Republican Party today is much better positioned for winning in Pennsylvania compared to the recent past and appears to be on a trajectory to strengthen its position.
1. An ex-ante probability of turnout is assigned to each registered voter in 2024 based on a statistical model fit to data from previous elections. These modeled turnout probabilities are then applied to obtain estimates of the number of participants in the November 2024 election, by party affiliation. ↩︎
2. Their lack of support was reflected in their opting not to cast a vote or voting for a minor party candidate. ↩︎
3. Before 2020, vote-by-mail was limited to traditional absentee ballots, which were available only to military and oversees civilian voters or those professing to have a valid reason for being away from their home precinct on election day. Beginning in 2020, no excuse was required to vote by mail, although voters could still opt for the traditional absentee ballot. The vast majority of voting by mail has relied on the new, no-excuse ballot format. ↩︎
4. The PA Chase is an initiative of the of the Citizens Alliance of Pennsylvania. ↩︎
5. Here, and through the discussion below, the term mail-in ballot includes both the traditional absentee and the newer, no-excuse vote-by-mail formats. ↩︎
6. Mail-in voting might be most effective at increasing turnout rates among voters with lower turnout propensities. If voters with high modeled likelihood of participation (90 percent or higher) are excluded from this analysis, then these ratios become 2.1, 1.5, 1.3, and 1.6, respectively. ↩︎
Data from the 2020 Pennsylvania Election Confirm: Elections Can Be Won or Lost on the “Supply Side” of Mail-in Voting

Aug 4th, 2024
Voting at a polling place or filling in and returning an absentee ballot can be regarded as the “demand side” of the voting process. Until 2020, elections in Pennsylvania, and in other states without a no-excuse mail-in voting option, primarily played out on the demand side.

Most voters had only one option, which was to go to the polls and pull the lever for their preferred candidate. Absentee voting was limited to active military, and to civilians specifying a pertinent reason they would be unable to go to their polling places on election day. Therefore, the process was controlled almost entirely by citizens’ decisions about whether to vote and for whom to vote.

However, the introduction of no-excuse mail-in voting in 2020 in Pennsylvania and other states greatly expanded the “supply side” of the election process. The added supply-side dimensions encompass the mechanics and procedures by which voters apply for mail-in ballots; how these applications are processed; and how the ballots are then distributed, returned, and tallied.

While in some states mail-in ballots are automatically distributed to all registered voters, in many states, including in Pennsylvania, voters must apply, either online or paper, to receive a mail-in ballot. Where this is the case, efforts to encourage and facilitate of mail-in ballot applications from voters, so-called “application harvesting”, can become a critical, “supply-side” activity affecting the outcome of the election.

This article analyzes data from the 2020 Presidential election in Pennsylvania, finding that spikes in paper applications for mail-in ballots from registered Democrats that occurred in the latter half of August into the first week of September may have produced enough votes to swing the election to Joe Biden. The analysis demonstrates the importance of the supply side to the outcome of the election, as the most plausible source of the spikes was application harvesting activity by partisan actors.

Forms of application harvesting

Nonprofit, tax-exempt, and technically non-partisan “get-out-the-vote” organizations as well as the political parties themselves historically have focused on promoting voter registration. With the expansion of mail-in voting, they have allocated increased attention and resources to application harvesting activities.

One important tactic the nonprofit organizations have employed is the mass distribution of paper applications for mail-in ballots. For example, it is well known that two nonprofit “sister” organizations, the Center for Voter Information and the Voter Participation Center, had together engaged in a massive effort to promote mail-in voting for the 2020 general election, by sending out millions of paper applications for mail-in ballots via postal mail to voters in multiple states.

But get-out-the-vote organizations are not the only possible conduit for application harvesting. Other organizations that are ostensibly nonpartisan such as employee associations, professional organizations, and the like, might distribute paper applications for mail-in ballots to their members.

While permissible in Pennsylvania and many other states, such organizations generally are required to conduct their activities in a non-partisan manner. Sixteen states currently restrict or prohibit third-party individuals or groups from distributing absentee and mail-in ballot applications, according to the National Conference of State Legislatures.

The political parties themselves may engage directly or through surrogates in application harvesting activities, including through mailbox stuffing, door-knocking campaigns, or personal, interactive outreach campaigns by party loyalists. Conceivably, personal outreach might be more effective than mass mailbox-stuffing.

For instance, party activists might be supplied with paper applications to personally distribute to family, friends, or neighbors who share their political leanings. The activists not only would seek to persuade these relatives, friends, and neighbors to apply for a mail-in ballot but could supply them with their own set of applications and encourage them to reach out to others in the same way. Such a campaign, if sufficiently scaled up, could be quite effective at generating additional votes.

Is application harvesting legit?

Application harvesting is a legitimate election-related activity provided it is conducted in accordance with established rules and norms. For instance, when engaged in by non-partisan not-for-profit organizations (where permissible by state law), it can lose legitimacy if conducted in a manner that excludes voters from a disfavored party or if coordinated with a favored party.

The legitimacy of application harvesting efforts by political parties or their surrogates could be open to question to the extent they might incorporate overly aggressive or intrusive tactics or violations of privacy. Indeed, it is conceivable that application harvesting activity in some cases might cross the line into questionable tactics.

For instance, an application might be filed by one individual on behalf of another, either directly without informing the prospective voter, or by asking him or her to sign a pre-filled form without adequate informed consent. A nursing home employee might instruct a resident to sign a ballot application that the employee had filled in, a clear violation of trust or fiduciary responsibility, for example.

A fundamental concern around mail-in voting in general is that the voter’s privacy could be compromised—persons voting by mail may be subject to intrusion into their decision-making, from the initial choice whether to vote by mail right up to the point that they seal their ballot. Such a concern encompasses application harvesting. For example, an intrusive family member might apply for a mail-in ballot on another’s behalf, a form of voter impersonation, as part of a scheme to pressure the voter.

Introducing the data

The source of the data for the analysis is a dataset from the Pennsylvania Department of State with detailed information on each approved application for an absentee or mail-in ballot for the 2020 General Election. A link to the dataset, which is housed within the Department’s online Open Data Portal, was shared with the author in response to a “Right-to-Know Law” (RTKL) request.

Individual data items include the date the application was received by the county in which the voter is registered. The data also indicate the means by which an application was submitted — either online, or as a paper application delivered by postal mail or in person at an election office.

In addition, the data distinguish between traditional absentee ballots available to voters and the newly available, no-excuse mail-in ballots (henceforth to be referred to simply as absentee and mail-in, respectively) and further differentiate among types of traditional absentee ballots, such as military, overseas civilian, regular civilian, and so on. The data also provides the voter’s date of birth and party affiliation.

Counts of applications according to the type of application filed indicate that more than 1.2 million paper and about 1.4 million online applications for the no-excuse mail-in ballots were submitted, far outnumbering requests for traditional absentee ballots. Online applications for domestic, civilian absentee ballots were the third largest category, totaling about 326,000.

Chart 1 shows the distribution of applications for mail-in ballots by type of application separately for three age ranges: under 35 (younger voters); over 70 (older voters); and 35 to 70 (mid-age voters). Younger voters favored online to paper applications by roughly a 2 to 1 margin, with the reverse being the case for older voters. Mid-age voters also favored applying online, but by a smaller, 35 percent margin.

Chart 2 shows counts of applications for no-excuse mail-in ballots by application type (online or paper) across five time periods defined as follows:
- Applications received prior to June 1 would mostly (perhaps entirely) have been from voters who had applied to vote by mail in the early June Primary and had opted (by checking a box on the Primary application) to automatically receive a mail-in ballot for the General Election. Thus, it is natural to distinguish this period, which we refer to as the Primary period.
- Ballot printing in Pennsylvania did not move forward until the state’s Supreme Court resolved a dispute over the Green Party candidates would appear on the ballot. That decision was announced on September 17, which we adopt as the end date of the second application period and shall refer to that as the “summer period”.
- The remainder of September constitutes the third period.
- The state’s deadline for applying for a mail-in ballot was October 27, motivating a division of October into two roughly equal-sized periods: the 1st through the 14th, and the remaining days up to the application deadline.
As seen in this chart, the number of online applications associated with the Primary period was about double the number of paper applications associated with that period. The number of online applications dropped off sharply between the Primary and summer periods, but there was just a small decline in the number of paper applications between these periods. Note that Primary period applicants were restricted to voters registered as Democrat or Republican, because in Pennsylvania voters must be affiliated with the party to vote in its primary, while later applicants additionally included those not affiliated with either of the two major parties.

A surprising divergence

Chart 3 disaggregates Chart 2 based on party affiliation: the number of applications by period and type are shown in Panel A for Democrats and in Panel B for Republicans.

For the Primary period, online outnumbered paper applications by more than two-to-one among Democrats, whereas applications were about evenly divided between online and paper among Republicans. These differing shares of online versus paper might reflect socioeconomic and demographic differences by party affiliation among voters who participated in the Primary.

It also is evident from Chart 3 that online applications from voters of both parties as well as paper applications from Republicans dropped sharply between the Primary and summer periods. However, paper applications from Democrats increased between the two periods.

Thus, the path of paper applications from Democrats exhibits a remarkable divergence. It is now evident that the contrasting dynamics seen in Chart 2 for online versus paper applications reflect the divergent path of paper applications from Democrats.

Specifically, paper applications from Democrats rose by nearly 20 percent, whereas those from Republicans plummeted by 70 percent and online applications from both parties fell by about 60 percent. If paper applications from Democrats instead had dropped only half as much as online applications; that is, by 30 percent, Democrats would have had about 100,000 fewer mail-in ballot applications.

That would translate into 90,000 fewer votes (assuming a 10 percent non-return rate of mail-in ballots, which is the non-return rate observed in the data for summer period, paper applications from Democrats.) That number of votes would have been more than enough to swing the state’s Presidential election. In a sense, then, the 2020 election outcome was a done deal by summer’s end, determined by the summer period’s heavy volume of paper applications for mail-in ballots from Democrat voters.

It seems a safe bet that some form of organized campaign had put paper applications in the hands of registered Democrat voters during this time and then encouraged them to fill out and submit the applications. In other words, the divergent path of paper applications likely was a supply-side phenomenon, not a spontaneous wave of voter enthusiasm for Joe Biden. Otherwise, one would expect equally strong momentum for online applications from Democrats.

Drilling down further—weekly application dynamics

The Department of State data contain sufficient detail to support a more granular examination of the timing of mail-in ballot applications, including further exploration of the summer period surge in paper applications from Democrats. The data also provide the ability to drill down to the county level, enabling an analysis of weekly application rates by county. It is advisable to disaggregate this way because a state-level analysis might mask significant differences across individual counties.

The approach we pursue here is to track weekly application rates (measured as each week’s average daily count) by application type, county, and party affiliation, from the week ending June 7 through the week ending October 27. Because overall application volume can differ greatly across counties and between Democrats and Republicans in any given county, the approach involves normalizing the weekly application rate for each party in each county relative to a benchmark week, in order to facilitate visual comparisons of the timing of activity. The week ending September 20 is selected for this normalization.

The dynamics of paper applications

It turns out that counties can be segmented into groups based on visual comparison of weekly application rates for paper applications from Democrat voters, such that counties within the same group have similar dynamics that are distinct from those in other groups. All but one of these groups exhibit comparatively rapid acceleration in applications from Democrats during mid-to-late August, with a peak application rate occurring no later than the week before Labor Day. The primary distinction among the several groups exhibiting this characteristic pattern regards the timing of the peak—the week in which the paper application rate from Democrats reached its apex.

Thus, in these counties, which comprise most of the state, paper applications from Democrats are front-loaded relative to paper applications from Republicans. The largest of these groups, consisting of 19 counties, is characterized by a spike in the former that peaks in the week ending August 30, as shown in panel A of Chart 4.

For the sake of brevity, we shall limit our discussion of application rate dynamics to the latter cohort (as broadly representative of all groups characterized by front-loading of paper applications from Democrats) and the single cohort not exhibiting this pattern; that is, no front-loading. The dynamics of paper applications for the 18 counties comprising the latter cohort are shown in Chart 4, panel B.

Map 1 distinguishes the latter cohort based on distinct, lighter green shading, and likewise differentiates the other counties by degrees of darker shading, according to the week in which the rate of paper applications from the county’s Democrat voters was at its apex. Due to data irregularities, three counties remain unclassified.

Panel A of chart 4 shows a surprisingly steep and voluminous surge in paper applications from Democrats within the last two weeks of August and into the first week of September in the group of counties represented. This compares to a more gradual and modest rise for Republicans. For voters unaffiliated with either of the two major parties (self-declared Independents and others), paper application rates were situated between those of Democrats and Republicans during the weeks of the spike. Thus, we find that the Democrats reaped a bountiful, late summer harvest of paper applications for mail-in ballots in these (and in most other) counties within this three-week period.

In stark contrast, as seen in Panel B, the no-frontloading group exhibits nearly identical dynamics between Democrats and Republicans for paper applications through the week ending September 20. In both cohorts, application rates from Republicans continued to rise after mid-September; those from Democrats did not. As a result, the paper applications from Republicans were comparatively back-loaded.

The dynamics of online applications

No such surge occurred in the case of online applications. In contrast to paper applications, the dynamics of online applications through the week ending September 20 were very similar between Democrats and Republicans, within both county groups, as seen in the two panels of Chart 5.

Therefore, whatever caused the late summer surge in paper applications from Democrats in most counties was not present in the case of online applications or in the case of the “hold-out” group represented in Panel B of chart 4.

Circling back

These weekly dynamics can be mapped back to the pattern observed earlier in the article mail-in ballot for application volumes across the five application periods. Charts 6 and 7 revisit that analysis, now differentiating between the two county groups. Panel A of each chart tracks the volume of applications from Democrat voters, separately for paper and online applications, while panel B does the same for applications from Republican voters.

As seen in the two panels of Chart 6, the county group characterized by the August 30 peak is characterized by a steep rise in the volume of paper applications from Democrats between the Primary and summer periods, and the opposite (a large decline) for Republicans. Thus, the comparatively rapid acceleration in paper applications from Democrats during mid to late August in these counties yielded a correspondingly large increase in application volume relative to the Primary.

In contrast, as seen in the two panels of Chart 7, the “no front-loading” counties where this rapid acceleration did not occur show a substantial decline in paper applications from Democrats between the Primary and summer periods, comparable to that observed for Republicans.

Thus, it is clear that the statewide aggregate jump in paper applications from Democrats between the Primary and summer periods, described earlier in the chapter, is tied to the counties that experienced a surge in these applications during mid-to-late August. In the case of counties that had no such surge, application volumes across the five periods followed similar paths (only the scaling is different) for Democrats versus Republicans, whether for paper or for online applications.

Summing up

Spikes in the rate of paper applications from Democrats during the last two weeks of August and first week of September, distinct from a more gradual ramping up of paper applications from Republicans, occurred in most Pennsylvania counties. In contrast, the dynamics of online applications were similar between Democrats and Republicans.

It seems implausible that these observed spikes in paper applications from Democrats were voter-initiated. In other words, it is difficult to imagine that among Democrat voters, enthusiasm for Biden, the impulse to get an early start, and the inclination to use a paper application simultaneously surged to generate these unique dynamics. If that were the case, why would the zeal have been confined to filing paper applications, while those preferring the online channel lagged behind? Moreover, how would one then explain the presence of a “holdout” group of 18 counties exhibiting no such wave of enthusiasm?

Therefore, we can be fairly certain that the impetus came from the supply side. In other words, some form of application harvesting in the counties exhibiting the spikes almost surely drove these dynamics. However, the particular activities responsible for the spikes, and the number of additional votes that resulted, have not been determined.

And as for the impact on the election outcome? On the one hand, undoubtedly some of the paper applications “harvested” during the summer period would have been from individuals who would have voted anyway, by applying for a mail-in ballot later on or voting in person. On the other hand, almost surely some would have been from individuals who might not have voted at all if not for the harvesting activity, implying some effect on the election outcome. Based on the data presented here, it is reasonable to conjecture that this effect was material.