Simple Tests For The Extent of Vote Fraud With Absentee and Provisional Ballots in The 2020 US Presidential Election

Simple tests for the extent of vote fraud with absentee and provisional ballots in the 2020 US
presidential election
John R. Lott, Jr., Ph.D.

Crime Prevention Research Center
Forthcoming, Public Choice
Abstract
This study reports three tests measuring vote fraud in the 2020 US presidential election,
although they provide inconsistent evidence. To isolate the impact of a county’s vote-counting
process and potential fraud on candidates’ vote margins, I first compare voting precincts in a
county with alleged fraud to adjacent precincts in neighboring counties with no allegations of
fraud. I compute the differences in President Trump’s vote shares on absentee ballots in those
adjacent precincts, controlling for the differences in his vote shares on ballots cast in person. I
also control for registered voters’ demographics and compare data for the 2016 and 2020
presidential elections. When I examine Georgia and Pennsylvania separately, weak evidence of
vote fraud on absentee ballots is found. However, combining the samples produces significant
results and implies at least 10,000 additional votes for Biden in Pennsylvania’s Allegheny and
Georgia’s Fulton counties. I then apply the same method to provisional ballots in Allegheny
County, where, contrary to state law, voters were allowed to correct alleged defects in
absentee ballots by submitting provisional ballots on Election Day. My analysis finds that such
permission contributed to a statistically significant additional 6,700 votes for Biden. Finally, vote
fraud can show up as artificially larger voter turnouts, higher rates of filling out absentee ballots
for people who hadn’t voted, dead people voting, ineligible people voting, or payments for
votes. The estimates for Arizona, Georgia, Michigan, Nevada, Pennsylvania, and Wisconsin
combined indicate an average of 255,000 excess votes for Biden.
Keywords: Vote fraud, absentee ballots, voter turnout rate, provisional ballots, presidential
election
JEL Classification: D72, D73, K16
Electronic copy available at: https://ssrn.com/abstract=3756988

1. Introduction
Brennan and Buchanan (1984) compared voting to cheering at a sporting event – one shouts in
support of one’s team without any demonstrable impact on the outcome. Perhaps we can push
their comparison even further. Sometimes fans (as well as voters) are convinced that the
referees make the wrong call.
This paper offers a unique method of testing the existence of vote fraud by comparing adjacent
precincts – ones in counties with alleged vote fraud to ones where no fraud is alleged. The
method applies to any state with fraud allegations in at least part of the state and to precinct-
level voting data on absentee and in-person voting. I then apply the test to allegations that
some heavily Democratic counties adopted different rules to process provisional ballots.
(Voters receive provisional ballots when questions arise over their eligibility to vote.) Finally, I
report an additional test by looking at voter turnout rates.
In 2005, the bipartisan Commission on Federal Election Reform, chaired by former President
Jimmy Carter and former Secretary of State James Baker III, concluded that “Absentee ballots
remain the largest source of potential voter fraud.” Intimidation and vote-buying also were key
concerns of the commission: “Citizens who vote at home, at nursing homes, at the workplace,
or in church are more susceptible to pressure, overt and subtle, or intimidation. Vote-buying
schemes are far more difficult to detect when citizens vote by mail” (Carter and Baker 2005, p.
46).
Worries about vote-buying have a long history in the United States. They helped drive the move
to the secret ballot, which US states adopted between 1888 and 1950. Secret ballots made it
harder for vote buyers to monitor the candidates whom voters actually supported. Vote-buying
had been pervasive, with voter turnout falling by about 8% to 12% after states adopted the
secret ballot (Lott and Kenny 1999). Absentee ballots reverse that trend, making it easier to buy
votes. The Carter-Baker report identifies additional instances of absentee ballot fraud in
modern times; other recent examples can be found (Lott 2020a; Crime Prevention Research
Center 2020).
Lyndon Johnson was both a victim and perpetrator of fraud in his 1941 and 1948 Senate races
(Caro 1982, pp. 736-738; Caro 1991, Chapter 13). The fraud allegedly involved vote-buying,
stuffing ballot boxes, and destroying votes. Vote fraud in Chicago until the 1980s is well
documented (Bauer 2018). The city’s Democratic political machine added dead people to voting
rolls and then had people vote in their names. Precinct captains also would vote on behalf of
individuals who didn’t show up at the polls, and they would “help” senior citizens in nursing
homes cast absentee ballots. The nursing homes cooperated with Democratic party activists
because they wanted good relationships with the city.
Much of the work on vote fraud in the United States involves audits that look at double voting
or dead people voting (Goel et al. 2020; Wu et al. 2020; Viebeck 2020). Other studies count the
number of convictions for vote fraud (Heritage Foundation 2022). But both approaches are of

very limited value. Many types of vote fraud are hard to detect. 1 It is difficult to catch someone
who fills out and mails in absentee or mail-in ballots for others. Providing provisional ballots for
individuals who are told that they have already voted often doesn't even avoid the problem of
voting an absentee ballot in someone else’s name. Less than half of people vote in most
elections, so only a small portion of fraudulent votes can be detected, replaced, and corrected.
Finally, vote-buying also is tough to detect because buyers and sellers of votes both gain from
the transaction.
It’s not just Republicans who are concerned about absentee ballots. Indeed, almost all
European countries have adopted stricter anti-fraud voting rules than the United States (Lott
2020a, 2021c). For example, 35 of the 47 countries in Europe ban absentee voting entirely for
citizens living there.2 Another ten countries allow it, but require voters to show up in person
and present photo IDs to pick up their ballots. Six of those ten countries limit the practice to
people in the military or a hospital, and they require third-party verification. Another 16
European countries ban absentee ballots for voters living abroad on Election Day. Similar
requirements are imposed in other nations.
Developed countries, with few exceptions, did not adopt emergency voting measures during
the coronavirus pandemic. Poland allowed mail-in ballots as a one-time option for everyone
during 2020, as did Bavaria for a run-off election and two cities in Russia (Flis and Kaminski
2020; Lott 2020b). France adopted more limited exceptions, temporarily allowing sick or at-risk
individuals to vote absentee.
Many countries have learned the hard way about what happens when mail-in ballots aren’t
secured. They have discovered how hard it is to detect vote-buying.
France banned mail-in voting in 1975 because of massive fraud in Corsica -- voters cast multiple
votes with stolen or purchased postal ballots. Mail-in ballots also were cast for dead people
(Briquet 2016; World Today News 2020; New York Times 1973).
1 Other countries, including the United Kingdom and France, have learned about the difficulty of uncovering vote
fraud firsthand. The United Kingdom's Electoral Commission provides examples of vote fraud for in-person voting
(United Kingdom’s Electoral Commission 2017, 2018). As the Electoral Commission describes it: “Later in the day
the same voter attended again and sought to vote again, this time in his own name. Due to certain physical
characteristics of the voter (he was very tall and wore distinctive clothing) and the vigilance of the presiding officer
he was suspected of having already voted earlier and formally challenged.” In another UK case from 2017, police
caught a person voting multiple times only because he bragged openly about it on Twitter.
2
Eight European countries allow proxy voting as a substitute for absentee ballots, permitting voters to designate
others to vote in their names (Lott 2021c). But proxy voting is very strictly regulated by requiring photo IDs and
signed request forms. In Poland, the proxy voter must have a power of attorney and obtain special permission
from the municipal mayor. In France, voters must visit a municipality’s office in person with proof of identity and a
reason for absence (for example, a certified letter from an employer or hospital) and then nominate a proxy. Proxy
voting avoids the problem of absentee ballots being unsecured. Proxy voting requires that the proxy vote be cast in
person in a voting booth. Switzerland is the only European country adopting a comparatively liberal proxy voting
policy, requiring only a signature match.

The United Kingdom, which allows postal voting, has uncovered notable mail-in ballot fraud
cases. Prior to recent photo ID requirements, six Labor Party councilors in Birmingham won
office after what the judge described as a “massive, systematic and organised" postal voting
fraud campaign (Britten and Jones 2005). The fraud apparently occurred with the full
knowledge and cooperation of local Labour Party officials. "Widespread theft" of postal votes
(as many as 40,000 ballots) was documented in areas with large Muslim populations because
Labour’s organizers were worried that the Iraq war would induce such voters to oppose the
incumbent government.
In 1991, Mexico mandated voter photo IDs with thumbprints and banned absentee ballots. The
then-governing Institutional Revolutionary Party (PRI) had for many years engaged in fraud and
intimidation for mail-in ballots to win elections (Lott 2006). Only in 2006 were absentee ballots
again allowed, and then only for voters living abroad who requested them at least six months in
advance (McKinley 2005).
Unsecured absentee ballots create the potential of either introducing fraudulent ballots or
destroying legitimate ones. Some safeguards, such as matching signatures, can minimize the
problems, but even signature requirements are not as secure as requiring voters to present
government-issued photo IDs. Nor do the safeguards prevent ballots from being destroyed.
One of the 2020 election controversies was that states like Georgia, Nevada, Pennsylvania, and
Wisconsin did not match signatures on the outer mail-in ballot envelopes to voters’ official
registration records (Navarro 2020). In other states, particularly Pennsylvania, accusations were
raised of accepting absentee ballots not enclosed in outer envelopes displaying the voter's
signature and that signatures on those envelopes weren’t checked (Navarro 2020, pp. 13-14).
While such questions received the most attention in those swing states, they weren’t limited
solely to them. In Missoula County, Montana, which held an entirely mail-in ballot election in
November 2020, an audit found that 4,592 out of the 72,491 mail-in ballots lacked envelopes—
representing 6.33% of all votes (Lott 2021b). Absent an officially printed envelope containing
registration information, the voter's signature, and a postmark indicating that a ballot had been
cast on time, election officials cannot verify a vote’s legitimacy. It is a violation of Montana’s
election laws to count such votes, yet such votes were counted.
Measures to prevent vote fraud are important for ensuring accurate election results and
encouraging voter participation in US elections (Lott 2006). For example, after Mexico
instituted strict anti-fraud provisions in 1991, voter turnout increased substantially, rising from
59% in the three presidential elections before the reforms to 68% in the three elections
afterwards. That increase occurred despite massive stuffing of ballot boxes with fraudulent
ballots and registration lists filled with non-existent people (Rossie 1988; Uhlig 1991).3
3It is only in recent years that reports of vote-buying allegations have become more common in Mexico
(Associated Press 2005; Associated Press 2018; Linthicum 2018).

Courts frequently have rejected Republican challenges to the 2020 presidential vote, citing the
lack of evidence of enough fraud to alter the outcome in a particular state. Republicans
sometimes argued that since their observers couldn’t watch the vote counts, they couldn’t
provide such evidence without investigations backed by subpoena power. Still, while some
courts agreed that irregularities had occurred in 2020, they weren’t willing to grant discovery
unless Republicans first presented enough evidence of fraud that could overturn the election.
Republicans thus faced a kind of Catch 22.
The following sections provide mixed evidence on the existence of vote fraud in the 2020 US
presidential election. First, precinct-level estimates for Georgia and Pennsylvania are
inconsistent in revealing vote fraud with absentee ballots, although the evidence of broken
rules is more robust for provisional ballots. The results show evidence of vote fraud with
absentee ballots when the data from both states are combined. The precinct-level approach
can be applied to other states. I then look at all swing states at the county level to ask if
counties wherein fraud was alleged experienced higher voter turnout rates.
2. Georgia
The Trump campaign and its supporters focused primarily on what they claimed were voting
irregularities in Fulton County, which includes Atlanta.4 In Georgia’s certified ballot count,
former Vice President Joe Biden led President Trump by 12,670 votes (BBC 2020). Biden won
Fulton County by a margin of 243,904 votes and he out-polled Trump on the country’s absentee
ballots by 86,309 votes.5
Part of the controversy surrounding Fulton County’s absentee ballots arises from a burst water
pipe at a vote counting facility that resulted in the removal of poll watchers. According to the
Chair of the Georgia Republican Party, David J. Shafer, “counting of ballots took place in secret
after Republican Party observers were dismissed because they were advised that the tabulation
center was shutting down for the night” (Letter dated November 10, 2020, from Doug Collins
and David Shafer to Georgia Secretary of State Brad Raffensperger, p. 3).
If election workers processed absentee ballots when Republican observers were not present,
can statistical evidence of bias be found in those absentee ballot counts? While in-person
voting took place at the precinct level, absentee votes were counted at one common facility at
the county level. If the type of fraud that Mr. Shafer worried about actually occurred, it would
have affected only Fulton County’s absentee ballots.
To examine that possibility, I gathered precinct-level data for Fulton County and four
Republican-majority counties that border it wherein no allegations of fraud were raised: Carroll,
4 Donald J. Trump and David J. Shafer v. Brad Raffensperger et al., Fulton County Superior Court, December 4, 2020
(https://cdn.donaldjtrump.com/public-files/press_assets/verified-petition-to-contest-georgia-election.pdf).
5 November 2020 differed from previous elections in that respect. For example, in 2012, while Obama captured
64% of the total vote in Fulton County, he won barely a majority of the absentee vote (50.89%) (data from Clark
Bensen at Polidata).

Cherokee, Coweta, and Forsyth.6 Precincts adjacent to one another on opposite sides of a
county’s border are relatively small areas and should be similar demographically. The idea is a
simple one: compare Trump’s share of absentee ballots in adjacent precincts on opposite sides
of a county’s border. The comparison in Georgia is between precincts in Fulton and the four
other just-named counties as well as between precincts in those four counties that are adjacent
to one another. Comparing a county in which fraud was claimed to ones without it is simpler
than comparing counties where hard-to-specify, varying degrees of fraud may have
materialized.
The model is given as:
((A/TA)i0 – (A/TA)i1) = ß((P/TP)i0 – (P/TP)i1) + 𝛿 D*Xi + εI,
where A = absentee ballots for Trump, TA = total absentee ballots for both candidates, P = in-
person votes for Trump, and TP = total in-person votes. The superscripts 0 and 1 indicate
adjacent precincts in neighboring counties and the subscript i represents the set of precincts
being compared. The binary variable D = 1 if one of the adjacent precincts is in Fulton County
(in that case Fulton County is superscript 0); D = 0 otherwise. The error term is εI.
The null hypothesis is that 𝛿 = 0, meaning that precinct pairs in which one is a Fulton County
precinct are no different from other pairs. The alternative hypothesis is that 𝛿 < 0, implying that
precinct pairs in which one is the Fulton County precinct undercounted Trump’s absentee
ballots.
Matching precincts across county lines was a very time-consuming process. While each county
(or, in the case of Pennsylvania, each township) had its own precinct maps, comparisons across
the political boundaries had to be done by hand. The goal is to compare the precincts of Fulton
County that are most similar to precincts in nearby counties with no fraud allegations. In that
way, we can isolate the impact of Fulton county’s vote-counting process (including potential
fraud).
In 2020, Fulton County contains 384 precincts and Cherokee County had 42 precincts. In one
case, Fulton County precinct ML02A matches up with four different precincts in Cherokee
County (Mountain Road 28, Avery 3, Union Hill 38, and a small portion of Freehome 18). 7
Any difference in Trump's share of the absentee ballots would not have been caused by the
general shift to absentee voting among Democrats, because Democrats encouraged their voters
to vote by absentee ballot at both the national and state level. The absentee voting rate
shouldn't differ substantially between two precincts next to each other and similar in terms of
6 Corrected data were not available for Fayette County but including those observations resulted in no changes in
the statistical significances reported in either Tables 1 or 2.
7 The other counties are matched west to east and south to north. For a related discussion, see Bronars and Lott
1998.

their in-person voting support and demographics. After all, Democrats cared about winning the
state, not a county, or even a precinct.
In precincts with alleged fraud, one would expect Trump's proportion of absentee votes to be
depressed, even when such precincts returned in-person Trump vote shares similar to
surrounding counties. It would be suspicious if the shift happened only among absentee ballots
on opposite sides of a county line.
One challenge when entering a standard intercept term in the model specified above can be
the direction of difference between the precincts. If I estimate a typical regression equation
with an intercept term, either subtracting control precinct A from B or B from A, the order
chosen flips the signs for a given data point. But a way can be found for ensuring the same
result no matter the ordering of the between-precinct comparisons: exclude the
intercept.8 The results are robust and unaffected when I drop the intercept term from the
regressions.
What is the problem with an intercept for the control precincts? It implies that two identical
adjacent precincts have a predictable difference in voting behavior. A linear regression without
an intercept is robust because it is symmetric around the origin and thus unaffected by the data
points surrounding it.
I ran the test on data from both 2016 and 2020 elections. No accusations of fraud involving
absentee ballots were voiced in 2016, so one should expect the absentee ballot percentage for
Trump in Fulton County’s precincts to behave no differently than the adjacent precincts in
Carroll, Cherokee, Coweta, and Forsyth counties. In 2016, the average difference between
Trump’s share of the absentee ballots and his share of the in-person votes in the adjacent
precincts was only 0.75 percentage points.
In Tables 1 and 2, if the estimate of the “Difference in Trump’s percentage of the two-candidate
in-person vote” between the two adjacent precincts equals one, it means that the differences
in the percentage of the in-person vote that Trump received in the adjacent precincts would
perfectly track the difference in the absentee ballots. In the estimate for 2016, the coefficient is
0.88, and the 95% confidence interval ranges from 0.684 to 1.076, so it is not statistically
different from one.9 But for the 2020 data, Trump’s share of in-person votes did not align as
closely with the differences in absentee ballots, which can be seen in the smaller coefficient of
the control variable for Trump’s share of in-person votes. Indeed, the coefficient for 2020
(0.6059) has a 95% confidence interval of 0.3851 to 0.8267, so it is statistically significantly
8 The geographical ordering convention in (Bronars and Lott 1998) was applied, but by excluding the intercept, that
ordering doesn’t matter for the results reported herein.
9 The source for the 2016 precinct borderlines was obtained here: http://rynerohla.com/index.html/election-
maps/2016-south-atlantic-republican-primaries-by-precinct/

different from one.10 However, Trump's shares of absentee ballots cast in Fulton and other
counties are not statistically different in 2016 or 2020.
Place Tables 1 and 2 here
Compared to 2016, the in-person vote shares in 2020 explain only about two-thirds of that
year’s variation in absentee ballot voting in adjacent precincts.
I control for demographic variables to help account for any differences that might still exist.
Georgia collects precinct-level information on registered voters’ racial and gender
demographics by precinct. Table 3 accounts for the differences in the adjacent precincts based
on detailed demographic information instead of the change in the in-person difference in
Trump’s share of the votes. It supplies information on the differences between the precincts in
the population percentages of Black males, Black females, Hispanic males, Hispanic females,
Asian males, and Asian females. Table 4 enters those variables as well as the “Difference in
Trump’s percentage of the two-candidate in-person vote.” Thus, the table adopts three ways of
accounting for differences in Trump’s share of the absentee ballot vote: the geographic
closeness of relatively small areas (voting precincts), differences in Trump’s share of the in-
person vote, and differences in the demographics of registered voters.
Place Table 3 here
The results are not consistently significant in answering whether Trump's absentee votes
uniformly were lower in precincts bordering Fulton County than in the precincts just across the
street in neighboring counties. The estimates for the Fulton County effect range from 0.3% to
11.5%, but only one of those estimates is statistically significant. In nearly all of the estimates,
the race and gender demographics variables are not statistically significant, although that is not
surprising given how highly correlated those variables are. That makes it difficult to interpret
individual coefficients on the demographic variables. However, they are statistically significant
as a group in Table 3 (a joint F-test for the demographic variables generates a value of 4.17,
which is statistically significant at the 5% level).
That result indicates that the demographic values are worth including and, moreover, that
Table 4 is the preferred model. But all models agree that Trump’s absentee ballot share was
depressed in Fulton County precincts.
Place Table 4 here
The empirical results imply an unusual drop-off in Trump’s share of the absentee ballots for
Fulton County, which ranges from zero to 11.53 percentage points. Given that 145,267
absentee ballots were cast for Trump and Biden in Fulton County, the largest estimate of 11.53
10The average difference between Trump’s share of the absentee ballots and his share of the in-person votes in
the adjacent precincts was -6.04 percentage points.

percentage points equals approximately 16,749 votes, which is 32% more than Biden’s margin
of victory over Trump.
The average of the various estimates (7.81%) implies an unusual drop in Trump’s share of the
absentee ballots in Fulton County alone. It amounts to 11,350 votes or 90% of Biden’s vote lead
in Georgia.
DeKalb County also drew charges of vote-counting irregularities, but it no adjacent Republican
counties are available to test those claims. With 128,007 absentee ballots cast for the two
major-party candidates in DeKalb, fraud also could have been significant in determining the
election outcome.
While some critics of the research methods employed herein argue that errors correlate across
precincts within a county,11 it isn’t a concern here. The estimates look at the difference
between adjacent precincts across county lines. But, again, there’s no apparent reason why
Democrats would treat absentee ballots differently in two adjacent precincts having similar
political and demographic makeups. Still, I reran the estimates in the first four tables clustering
standard errors by county. That change actually makes some of the estimates more statistically
significant.12
Given that I am measuring how the vote Biden-Trump vote gap between adjacent precincts is
changing over time, the results are equally consistent with vote fraud in Republican counties
and the destruction of Democrat absentee ballots. But no allegations of such fraud have been
made for Republican counties.
If Biden was helped by fraud in Fulton County’s in-person voting, the estimates reported herein
would underestimate the amount of fraud associated with absentee ballots. Not just in Georgia,
but also in Nevada, Pennsylvania, and Wisconsin, critics alleged that large numbers of in-person
voters were not legally registered (Navarro 2020). In Fulton County, Georgia, 2,423 voters were
not listed on the State’s records as registered, and 2,560 felons voted even though they had not
completed their sentences.13,14
3. Pennsylvania
11 Austan Goolsbee in a Tweet regarding the present research project wrote: “you don't seem to be clustering the
standard errors at the county level and instead treating every precinct as though it is independent”
(https://twitter.com/Austan_Goolsbee/status/1344361588535521280).
12 Rerunning the estimates by clustering standard errors by county pairing doesn’t fundamentally alter the results.
The coefficient for county fraud remains statistically insignificant in Table 1, with a p-value of 0.035%. The t-
statistic for Table 2 becomes 0.31 (probability for a two-tailed t-test = 0.79), for Table 3, it is 4.08 (probability for a
two-tailed t-test = 0.055), and for Table 4 it is 0.50 (probability for a two-tailed t-test = 0.664). A one-tailed t-test is
actually the more appropriate statistical test for the various results that we are showing, but reporting those
wouldn’t really change which results we note are statistically significant.
13
The Superior Court of Fulton County, State of Georgia, Trump v. Raffensperger, December 4, 2020.
https://www.democracydocket.com/wp-content/uploads/sites/45/2020/12/Trump-v.-Raffensperger.pdf
14
Survey and other evidence indicate that felons vote almost exclusively for Democrats (Lott 2007b, pp. 182-184).

The Trump campaign focused its allegations of voting irregularities in Allegheny County (which
includes Pittsburgh) and Philadelphia County (which matches the boundaries with Philadelphia),
although it also raised concerns about the counties surrounding Philadelphia. In Pennsylvania’s
initial ballot count, former Vice President Joe Biden led President Trump by 81,361 votes. Biden
won Allegheny and Philadelphia Counties by margins of 146,706 and 471,305 votes,
respectively, and led the absentee vote by margins of 206,505 and 310,553 votes. Unusually
large numbers of provisional votes were cast in those same counties, with Biden leading the
tallies by 1,489 and 9,045, respectively.
Many concerns arose about possible vote fraud in both counties. Republican poll watchers
complained that they were too far away from the ballots to observe the process closely (Li and
Ramey 2020; Payne 2020). Trump’s lawyers said that voters in Pittsburgh and Philadelphia
submitting invalid mail-in/absentee ballots were notified and allowed to correct the defects by
casting provisional ballots on Election Day. In contrast, election officials in Republican-leaning
counties followed Pennsylvania’s election laws more strictly and did not notify voters about
similarly defective mail-in/absentee ballots.15 Complaints arose when voters were allowed to
cast provisional votes because they supposedly had requested a mail-in ballot already, even
though some voters claimed to have not done so.16 That process raises concerns that someone
other than the registered voter may have voted by submitting an absentee ballot in that
person’s name.
While affidavits attested to such potential problems, an open question has been whether the
issues were widespread enough to alter the 2020 election’s outcome.
To explore that question, I adopted the same approach with precinct-level data as I did for
Georgia. I collected voting data from adjacent precincts in Allegheny County and the four
Republican counties that border it: Beaver, Butler, Washington, and Westmoreland. My
comparisons are between Allegheny and those four adjacent counties, and also between the
four counties themselves when they are adjacent to one another. However, unlike Georgia, I
could obtain the breakdowns of absentee and provisional voting only for Allegheny County in
2020. The estimates reported below thus apply to that single year, In addition, while large-scale
fraud allegedly occurred in Philadelphia County, no Republican counties are adjacent to it,
thereby limiting the analysis further.
The precincts in the relevant Pennsylvania counties encompass very small, homogenous areas.
For example, Allegheny County contains 1,323 precincts – on average, a different precinct every
half mile. The more rural, less populous counties also contain large numbers of precincts:
Westmoreland 307, Washington 180, Beaver 128, and Butler 111. 17
15 Rudy Giuliani, Trump Campaign News Conference on Legal Challenges, C-SPAN, November 19, 2020
(https://www.c-span.org/video/?478246-1/trump-campaign-alleges-voter-fraud-states-plans-lawsuits).
16 See the complaint filed in Trump v Boockvar et al. in the United States District Court for the Middle District of
Pennsylvania (p. 48).

17 https://www.butlercountypa.gov/DocumentCenter/View/1982/Precincts-List?bidId=

The results reported in Table 5 show that in 2020, Trump’s percentage of absentee votes was
smaller in border precincts within Allegheny County than in the precincts just across the street
in neighboring counties. Trump’s share was just 0.25 percentage points lower on the Allegheny
County side, a difference that is not statistically significant.
To the extent that any fraud occurred with in-person voting, the estimates here will
underestimate the prevalence of fraud in absentee ballots.18
Place Table 5 here
Because of the concerns expressed about the provisional ballots relied on to solve problems
with improperly voted absentee ballots in Allegheny and Philadelphia counties, I compared the
previously described adjacent precincts to examine the possibility of provisional ballot fraud.
Table 6 is the same as Table 5, except that it applies to provisional, rather than to absentee
votes. The estimate implies a 3.6 percentage point reduction in support for Trump in precincts
adjacent to Allegheny County than in Allegheny County itself, but the result is not statistically
significant. One obvious reason is the small number of observations: 53 of the 87 precincts have
no provisional ballots for Trump and, since one cannot divide by zero, those vote share
differences are not defined.19
Place Table 6 here
An alternative approach that avoids the loss of those observations is to look at the rates at
which provisional ballots were issued. In that case, a clear difference exists. Approximately 1.5%
of the votes in border precincts on the Allegheny County side were cast by provisional ballots,
which is 3.2 times the 0.48% rate in the surrounding counties’ adjacent precincts. That
difference is statistically significant at below the 0.1% level on a two-tailed t-test.20
Table 7 compares across adjacent precincts the percentages of Biden’s votes that were on
provisional ballots, after accounting for the same difference for Trump. The share of Biden’s
votes from provisional ballots is about 0.65 percentage points higher in Allegheny County than
in the adjacent precincts, which amounts to roughly 2,800 more votes for Biden. If the same
pattern held for Philadelphia, that would mean another 3,925 votes. Adding together the
estimates for the two counties, our test implies approximately 6,700 extra ballots for Biden.
That total represents about 8.3% of Biden’s vote margin in Pennsylvania.
18 Republicans argue that in-person vote fraud is a problem in Pennsylvania. One instance is the 2020 conviction in
of a Philadelphia Judge of Elections charged with election fraud for allegedly stuffing ballot boxes on behalf of
Democratic candidates in three different races (Meyer 2020). The former president’s lawyer, Rudy Giuliani, also
claimed that people from New Jersey voted illegally in Philadelphia (Sobey 2020).
19 I also applied the same method to Georgia data, but with so few places there allowing provisional-ballot voting,
only 12 observations remained, none of which were located in Fulton County.

20 The provisional-ballot rate is slightly higher for the entire county: 1.98%.

Place Table 7 here
As a control, I tried running this comparison for Georgia. No one claims that Fulton County used
provisional ballots to correct absentee ballots, so one would not expect a statistically significant
result for that state. Indeed, those results were statistically insignificant, with a t-statistic of
only 0.27.
Finally, I reproduced the estimates from Tables 5, 6, and 7A with information taken from
Polidata on the racial demographics of voting-age populations in the relevant Pennsylvania
precincts (see Table 8). While information on voters’ gender wasn’t available, data from the
2010 Census was available on the differences between the precincts in the voting-age
population percentages of Blacks, Hispanics, and Asians. The results are similar to those
reported before, although the estimates corresponding to Table 5 become statistically
significant at the 10% level on a one-tailed t-test; the results corresponding to Table 7A remain
significant, implying a swing to Biden of about 8,500 votes.
Place Table 8 here
Overall, Sections 2 and 3 provide evidence of vote fraud in 2020, but it often isn’t statistically
significant. The evidence regarding provisional ballots is the strongest.
4. Combining Data for Georgia and Pennsylvania
It made sense to examine Georgia and Pennsylvania separately because some information was
available only for one state but not the other. Georgia, for example, published more detailed
demographic information at the precinct level than Pennsylvania did. In addition, the
allegations regarding provisional ballots were specific to Pennsylvania. But it is possible to
combine some of the data for the two states in our examination of absentee ballots. One
reason for doing so is to ask whether the failure of some estimates to reach statistical
significance arises from the small numbers of observations for each state.
Interestingly, the results reported in Table 9 show statistically significant evidence of large-scale
problems with absentee ballots. The estimates presented in column 1 correspond to those
reported in Tables 2 and 5, while the estimates in column 2 correspond to the first part of Table
8. I couldn't reproduce Table 3’s specification because racial data by gender is not available for
Pennsylvania. The one difference from the earlier analyses method is that I now enter fixed
effects for the two states.21
Place Table 9 here
21Clustering by county reduces the absolute t-statistic for a two-tailed t-test in column 1 of Table 9 from 2.73 to
1.83 and column 2 from 2.89 to 1.96.

The estimates in Table 9 imply a drop-off of between 3.5 and 3.9 percentage points in Trump’s
share of absentee ballots for the precincts in the counties where fraud was alleged. For
Allegheny and Fulton counties, the models imply that Biden received an extra 10,000 to 11,300
votes. Applying those estimates to Philadelphia County adds another 10,800 to 12,000 more
Biden votes, for a total of 20,800 to 23,300.
5. Voter turnout rates
I have accounted for three differences between voting precincts: geographically contiguous
areas, in-person vote shares for Trump to account for other political differences, and
demographic variables. But it’s possible that other differences across county lines may explain
variations in how absentee ballots were counted in 2020. It isn’t readily apparent what those
differences would be because the push for absentee ballots by Democrats appears to have
been a statewide and national effort. If once compares two adjacent precincts that have the
same demographics and same support for Trump, it would seem that Democrats would
undertake the same efforts in both precincts to get absentee votes for Biden. Still, even if the
absentee-ballot factor is independent of other voting irregularities, another qualitatively
different test might help make alternative explanations less plausible.
Vote fraud can elevate official voter turnout statistics. Fraud can take many forms: higher rates
of filling out absentee ballots for people who hadn’t voted, dead people voting, ineligible
people voting, or even bribes to encourage legally registered people to vote. However, the
official turnout rate also will be reduced to the extent that votes for opposing candidates are
lost, destroyed, or replaced with ballots filled out for the other candidate.
Republican plaintiffs in the United States District Court for the District of Arizona claimed that
up to 94,975 voters returned absentee ballots that were marked as unreturned and, hence, not
counted. Peter Navarro’s (2020, p. 9) election report characterizes those lost or destroyed
ballots as “consistent with allegations of Trump ballot destruction.”
No one claims that Republicans systematically encouraged people to submit affidavits about
vote fraud, specifically in places where Democratic voter turnout had been heaviest.
Other scholars have looked at voter turnout in Russia as a proxy for fraud (Mebane and Kalinin
2009). They examined high turnout rates and strange patterns in them, such as percentages
being reported in round numbers. Research on Bolivia studied unusual reporting of late-
counted votes (Idrobo et al. 2020).
Numerous contributions to the literature ask why voters vote. More competitive (“close”)
political contests generate more voter participation and larger campaign expenditures
(Bursztyn et al. 2018; Matsusaka 1993; Lott 2000). Turnout also varies between the two major
US political parties based on whether they control other offices (Jung et al. 1994). Voting
regulations regarding mail-in ballots, absentee voting, voter IDs, same-day voter registration,
and ballot harvesting all can impact voter turnout rates (Alvarez et al. 2008; Lott 2007a).

To examine voter turnout, we start with county-level election participation in six states wherein
Trump advisor Peter Navarro (2020) claims voting irregularities occurred in 2020: Arizona,
Georgia, Michigan, Nevada, Pennsylvania, and Wisconsin.
For example, a legal challenge in Georgia’s Fulton County Superior Court by State Republican
Chairman David Shafer and President Donald Trump revealed hundreds of thousands of
possible extra votes: 40,279 people who had moved within the state without re-registering;
4,926 voters who registered in another state after they had registered in Georgia; 305,701
people who, according to state records, applied for absentee ballots past the application
deadline; 66,247 ineligible voters under 17 years of age; 2,560 felons; 8,718 who remained on
the voting rolls after they were dead; and 2,423 who were not listed on the state’s voter rolls.22
In Nevada, more than 42,000 people voted more than once. 23 In testimony before the Senate
Hearing on Election Security and Administration, Jesse Banal, lead counsel for the Trump
Campaign, compiled the list by reviewing voter registrations and finding the same name,
address, and birthdate for some registered voters. In some cases, two registrants might have
the same last name, same birthdate, and same address, but one is “William” and the other
“Bill” – they, in fact, are the same person. More than 1,500 dead people allegedly voted.
Another 19,000 votes didn’t live in the state (excluding military personnel or students). In
excess of 1,000 voters listed non-existent addresses.
Similarly, 28,395 people allegedly voted without identification in Madison and Milwaukee,
Wisconsin. Republican lawyers claimed that 200,000 absentee ballots did not submit the proper
signatures.24 Payments to Native Americans to vote allegedly were “orchestrated by the Biden
campaign . . . [with] Visa gift cards, jewelry, and other ‘swag’” (Navarro 2020, p. 8; Bedard
2020).
Another reason for higher turnouts could be a much lesser absentee ballot rejection rate. For
example, Ballotpedia notes that in the 2016 general election, Georgia rejected 6.42% of
absentee ballots, but that rate was only 0.60% in 2020 – a difference of about 76,971 votes. 25
Other swing states also saw drops in rejection rates, although they were much smaller than
Georgia’s. Pennsylvania’s rejection rate went from 0.95% in 2016 to 0.28% in 2020, a difference
22 Donald J. Trump and David J. Shafer v Brad Raffensperger et al., Fulton County Superior Court, December 4, 2020
(https://cdn.donaldjtrump.com/public-files/press_assets/verified-petition-to-contest-georgia-election.pdf).
23 Senate Hearing on Election Security and Administration, December 16, 2020 (https://www.c-
span.org/video/?507292-1/senate-hearing-election-security-administration).
24 Ibid.
25 Election results, 2020: Analysis of rejected ballots, Ballotpedia, December 23, 2020
(https://ballotpedia.org/Election_results,_2020:_Analysis_of_rejected_ballots). The number of absentee ballots

cast (1,322,529) is from the Georgia Secretary of State’s website
(https://sos.ga.gov/index.php/elections/number_of_absentee_ballots_rejected_for_signature_issues_in_the_202
0_election_increased_350_from_2018).

of 17,361 votes.26,27 Nevada’s rejection rate dropped by 0.6 percentage points, a difference of
4,143 votes. Michigan is the only other swing state for which Ballotpedia measures rejected
absentee ballots; its rate essentially was unchanged from 2016 to 2020, falling from 0.49% to
0.46%.
To test whether counties with alleged fraud had higher voter turnout rates, I examine the
change in voter turnouts between the 2016 and 2020 general elections, by county, for the
following swing states: Arizona, Florida, Georgia, Iowa, Michigan, Nevada, North Carolina, Ohio,
Pennsylvania, and Wisconsin.28 I focus on swing states because I expect turnout rates to be
higher there, given that election campaigns tend to focus their resources there and, moreover,
voters are more motivated to turn out. I attempt partially to account for other cross-state
differences by comparing turnout rates in 2016 and 2020. But even within the sampled states,
three were rated the closest "toss-ups" by USA Today: Arizona, North Carolina, and
Pennsylvania. I also control for those three states separately.
The question is whether a larger increase in turnout rates is observed in the counties where
voter fraud was alleged relative to other counties. I already have discussed counties where vote
fraud was alleged in Georgia (Fulton and DeKalb) and Pennsylvania (Allegheny, Centre, Chester,
Delaware, Montgomery, Northampton, and Philadelphia). Other counties were under critics’
spotlights in Arizona (Apache, Coconino, Maricopa, and Navajo) (Davidson 2020; Smith 2020),
Michigan (Wayne), Nevada (Clark and Washoe) (Bedard 2020), and Wisconsin (Dane,
Menominee, and Milwaukee) (Bauer 2020).29
To examine differences in county turnout rates, I control for each county's turnout rate when
Trump ran in 2016. I also look at how heavily Republican or Democrat the counties are, based
on whether they voted for Trump or Biden. I classify those counties that Trump carried as
Republican counties and Biden's as Democratic ones. Since the turnout change may differ for
Democratic and Republican counties, I identify the counties supporting Trump and Biden with
26 I obtained the number of absentee ballots cast in Pennsylvania for Biden and Trump from Pennsylvania’s
Secretary of State
(https://www.electionreturns.pa.gov/General/SummaryResults?ElectionID=83&ElectionType=G&IsActive=1).
27 While it isn’t necessary for the results shown here, a higher turnout rate also could also show up in
manufacturing false ballots. A possible example occurred in Atlanta, where, as noted, election officials ordered
ballot-counting stopped because of a water leak (Chung 2020). The officials reportedly told observers that vote-
counting would start up again the next morning. Then once poll watchers, observers, and the media left, the vote-
counting continued with surveillance video catching large boxes of ballots being pulled from underneath a draped
table. (Trump Campaign lawyers present video 'evidence' of ballot fraud, Senate Judiciary Subcommittee,
December 4, 2020 https://www.youtube.com/watch?v=LJ0xDWhWUxk).
On the other hand, Fulton County Elections Director Richard Barron, a Democrat, claims that observers decided on
their own to leave the building in Atlanta (Associated Press 2020). Similarly, Gabriel Sterling, Georgia’s voting
system implementation manager, says that Georgia Secretary of State’s investigators were present even if political
observers weren’t on the spot (https://twitter.com/GabrielSterling/status/1334825233610633217?s=20).
28 I exclude the two states regarded as longshot swing states (New Hampshire and Texas) (King 2020).
29 See also
https://web.archive.org/web/20201111220325/https:/www.facebook.com/permalink.php?story_fbid=153929728
6270372&id=573103029556474.

two separate variables. When Biden won a county, the values for the Republican variable are
zero. Similarly, when Trump won, the values for the Democratic variable are zero. Elsewhere
those variables equal Trump's share of the vote minus Biden's share. Since I don't have strong
prior beliefs that the change in turnout is linear with respect to how partisan the county was, I
try including the squares of the partisan county measures. Enthusiasm for the two major
candidates might be increasing at either an increasing or a decreasing rate, but entering the
squared terms does not affect the alleged fraud variable (see Table 10).
Place Table 10 here
The regression takes the form: (Turnout Ratei2020 - Turnout Ratei2016) = ß1(Vote Fraud County
Dummyi) + ß2(Vote Share in Republican Counties Trump Won) i + ß3(Vote Share in Democrat
Counties Biden Won)I + ß3(Other Factors)I +  + εI, where  is the constant and ε is the error
term. The null hypothesis is that ß1 = 0.
I rely on data from the US Census Bureau's 2019 American Community Survey for median
household income, the percentage of the population that is female, different racial groups
(Black, Hispanic, Asian, and two or more races), the highest level of education attained (high
school graduate, some college, bachelor’s degree, and graduate or professional degree), and by
ten-year interval age groups (18 to 19, 20 to 29, 30 to 39, 40 to 49, 50 to 59, 60 to 69, 70 to 79,
80 and up). As voter turnout rates vary by gender, race, and age, all of the estimates account
for those simple demographics.
Finally, the last regression in Table 11 includes state fixed effects. The state dummy variables
may serve as a proxy for how states changed their election procedures between 2016 and
2020. However, any procedural changes implemented, such as mail-in ballots, would have
affected all counties in a given state, making it more difficult to detect increases in fraud in the
counties where fraud was alleged.
Place Table 11 here
The estimates displayed in Table 11 start from the simplest specification to one with more
controls. They imply that the counties with alleged vote fraud had between 142,000 and
368,000 excess votes – in other words, a 1.1 to 3.1 percentage point increase in turnout. The
county fraud variable's coefficient is statistically significant in every case, at least at the 8% level
on a two-tailed t-test. All estimates cluster standard errors by state (see Moulton 1986, but a
more skeptical approach is provided by Abadie 2017).
The first specification indicates that the more heavily Republican counties experienced higher
voter turnout rates in 2020 than in 2016, although the effect is not statistically significant. More
heavily Democratic counties show a slight drop in turnout, except for the counties wherein
fraud was alleged, but the change is not statistically significant in the Democratic counties
where fraud wasn’t alleged. The F-test implies that Democratic and Republican counties did not
behave differently in terms of voter turnout rates. The joint F-test for the gender, race, and age

variables is statistically significant at the 0.35% level. The second specification adds the other
demographic information on income and education, but those independent variables have little
impact on the results. The size and statistical significance of the fraud coefficient are reduced
slightly. But the primary effect is to lower the statistical significance of the joint F-test for the
demographic variables, so that it is now significant at the 8.9% level.
The next four estimates look at the percentages of Democratic or Republican voters in the
counties and their squared values. The voter turnout rate increased at a decreasing rate in the
most Republican counties and fell at an increasing rate in the most Democratic ones, but
neither effect is statistically significant. The joint F-tests also indicate that the results are not
statistically significant.
The fourth specification adds the changes in turnout between the 2012 and 2016 presidential
elections to explain the change between 2016 and 2020, but that is not statistically significant
and doesn’t affect the other estimates. The fifth specification includes a dummy for the states
rated "toss-ups" by USA Today on the morning of the November 3rd, 2020, election. That
variable implies a three-percentage point larger increase in turnout in the swing states (Arizona,
North Carolina, and Pennsylvania) and, moreover, the marginal effect is statistically significant
at better than the 0.01% level. Including that variable cuts the size of the fraud effect in half,
although it still significant at the 4% level or better on a two-tailed t-test.30 Including state fixed
effects cuts the size of the fraud variable further, but it remains statistically significant at the 8%
level on a two-tailed t-test and implies an excess of 144,000 votes.
As one test of the sensitivity of sample selection, I re-estimated the models in Table 11 on just
the two states that I examined earlier – Georgia and Pennsylvania – as well as the controlled-for
swing states (Florida, North Carolina, and Ohio) (see Table 12). In all but one of the
specifications, the results imply even larger increases in turnout in counties allegedly engaging
in vote fraud; the marginal effects consistently are statistically significant at least at the 8% level
on a two-tailed t-test. In the first five specifications, the estimates on the county-level fraud
variable imply excess votes for Biden of between 2.21 and 4.34 percentage points, or about
70,000 to 79,000 votes. The total combined win margin for Biden in Georgia and Pennsylvania
was 92,334.
Place Table 12 here
Again, the results reported above underestimate the extent of fraud if votes for opposing
candidates are either lost, destroyed, or replaced with ballots filled out for the other candidate.
It would also underestimate the amount of fraud to the extent that illegal votes are being cast
in the comparison counties.
30I also ran a specification analogous to the fourth specification with census data but entering a dummy for the
three states that were rated immediately before the election as the most likely toss-ups. The results remained
stable, although the coefficient on the county fraud dummy is reduced from 0.015 (significant at the 3.8% level) to
0.012 (with a t-statistic of 2.53) – a value that is midway between those shown in specifications 4 and 6, implying
about 143,000 excess Biden votes in the counties where fraud was alleged.

6. Subsequent research
Eggers et al. (2021) respond to two of the three tests presented here. They object to a
December 29th, 2020, version of the current paper where I entered an intercept term in
analyzing the precinct-level data. As discussed above, those initial specifications were
influenced by the ordering of precinct comparisons (whether one subtracts control precinct A
from B or subtracts B from A). But the January 6th, 2021, revised version of the paper corrected
that problem by removing the intercept term. Unfortunately, despite being immediately sent a
copy of the revised study and not submitting their paper until the end of February 2021, Eggers
et al. completely ignored the revised discussion for mail-in/absentee ballots. They have offered
no response to the statistically significant results regarding provisional ballots.
Finally, my regressions in the January 6th version of my paper used the change in voter turnout
rates between the 2016 and 2020 elections. As shown previously, even with many other
determinants held constant, alleged fraud has a statistically significant and large effect on voter
turnout. The authors ignored that information, despite having been made aware of it when I
engaged in a public debate with one of the authors in competing mid-February 2021 op-eds in
the Washington Times before they submitted their paper for review (Grimmer and Hall 2021;
Lott 2021a).
7. Conclusion
The precinct-level estimates for Georgia and Pennsylvania provide some evidence of vote fraud
in the 2020 US presidential election, but the sizes and statistical significances of the effects are
not consistent. However, combining the Georgia and Pennsylvania samples implies additional
mail-in/absentee votes for Biden. Allegheny and Fulton Counties alone account for at least
10,000 votes The evidence is much stronger for concluding that Democrats treated so-called
provisional ballots differently than Republicans did.
The results reported herein might underestimate the extent of vote fraud because they assume
that no fraud occurred with in-person voting. Yet, given that both results measure how the
Biden-Trump vote gap between adjacent precincts is changing over time, they are equally
consistent with more vote fraud in Republican counties than in Democratic counties.
Nevertheless, no such misbehavior in Republican counties has been alleged either in terms of
fraud or treating provisional ballots differently in 2020 than in previous elections.
The voter turnout rate data provide stronger evidence of significant excess Biden votes in
Arizona, Georgia, Michigan, Nevada, Pennsylvania, and Wisconsin. The estimates imply that the
counties wherein vote fraud was alleged returned between 142,000 and 368,000 excess Biden
votes. While the findings reported here are dramatic, they may be underestimates because the
voter turnout estimates do not account for ballots cast for the opposing candidate that are lost,
destroyed, or replaced with ballots filled out for the other candidate. While it would involve a

lot more work, it would be possible to apply the precinct level tests to compare turnout rates in
adjacent precincts across all the counties where fraud was alleged.

Acknowledgements
I want to thank Luis Locay, Sherwin Lott, Maxim Lott, Roger Lott, and Carl Moody for their very
constructive discussions. I originally did this research while working for the Office of Legal Policy
at the US Department of Justice, but it purely reflects my personal views. It did not represent
work done by or for the US Department of Justice, and the DOJ did not approve it.

References
Abadie, A., Athey, S., Imbens, G. W., & Wooldridge, J. (2017). When should you adjust standard
errors for clustering? MIT Working Paper (October).
Alvarez, R. M., Bailey, D., & Katz, J. N. (2008). The effect of voter identification laws on turnout.
Social Science Research Network.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1084598
Associated Press. (2005). Mexican senate approves mail-in absentee ballots for Mexicans living
abroad. The Arizona Republic, April 28.
http://www.azcentral.com/specials/special03/articles/0428mexicovote-ON.html
Associated Press. (2018). Observers say vote buying strong ahead of Mexico election.
Associated Press, June 26. https://apnews.com/article/cd1f3c2f00584bdb896f54701dcb1555
Associated Press. (2020). Surveillance tape of vote counting breeding false fraud claims in
Georgia. HuffPost, December 4. https://www.huffpost.com/entry/video-georgia-election-false-
fraud_n_5fcac976c5b619bc4c330575?guccounter=1&guce_referrer=aHR0cHM6Ly9kdWNrZHVj
a2dvLmNvbS8&guce
Bauer, K. (2018). Chicago and rigged elections? The history is even crazier than you’ve heard.
Block Club Chicago, October 24.
Bauer, S. (2020). Wisconsin issues recount order in 2 counties as Trump wanted. Associated
Press, November 19. https://apnews.com/article/wisconsin-recount-2-counties-
f408a7b43deb96e2ac7ff0b24a2f968a
BBC. (2020). US election 2020: Biden certified Georgia winner after hand recount. BBC,
November 20. https://www.bbc.com/news/election-us-2020-55006188
Bedard, P. (2020). Pro-Biden effort offered Native Americans $25-$500 Visa gift cards and
jewelry to vote. Washington Examiner, December 3.
https://www.washingtonexaminer.com/washington-secrets/pro-biden-effort-offered-native-
americans-25-500-visa-gift-cards-jewelry-to-vote
Brennan, G., & Buchanan, J. (1984). Voter choice: Evaluating political alternatives. American
Behavioral Scientist, 28(2), 185-201.
Briquet, J. L. (2016). Expatriate Corsicans and the Vote au Village: Mechanisms of control and
expressions of sentiment (Nineteenth-Twentieth Centuries), Revue française de science
politique (English Edition), 66(5), 43-63.

Britten, N., & Jones, G. (2005). Judge lambasts postal ballot rules as Labour 6 convicted of poll
fraud. The Telegraph (United Kingdom), April 5.
https://www.telegraph.co.uk/news/uknews/1487144/Judge-lambasts-postal-ballot-rules-as-
Labour-6-convicted-of-poll-fraud.html
Bronars, S. G., & Lott, J. R. (1998). Criminal deterrence, geographic spillovers, and the right to
carry concealed handguns. The American Economic Review, 88(2), 475-479.
Bursztyn, L., Cantoni, D., Funk, P., & Yuchtman, N. (2018). Polls, the press, and political
participation: The effects of anticipated election closeness on voter turnout. University of
Chicago Working Paper.
Caro, R. A. (1982). The Path to Power: The years of Lyndon Johnson. New York: Alfred a Knopf
Incorporated.
Caro, R. A. (1991). Means of Ascent: The years of Lyndon Johnson. New York: Alfred a Knopf
Incorporated.
Carter, J., & Baker, J. (2005). Building Confidence in U.S. Elections: Report of the Commission on
Federal Election Reform. Center for Democracy and Election Management, American University.
https://www.legislationline.org/download/id/1472/file/3b50795b2d0374cbef5c29766256.pdf
Chung, F. (2020). ‘Slow leak’: Text messages cast doubt on Georgia officials’ ‘burst pipe’ excuse
for pause in counting. News.com.au, November 12. https://www.news.com.au/world/north-
america/us-politics/slow-leak-text-messages-cast-doubt-on-georgia-officials-burst-pipe-excuse-
for-pause-in-counting/news-story/19176f5113512210517c82debe684392
Crime Prevention Research Center. (2020). UPDATED: Examples of Vote Fraud for Mail-in
Ballots. Crime Prevention Research Center, May 27.
https://crimeresearch.org/2020/05/examples-of-vote-fraud-for-mail-in-ballots/
Davidson, J. (2020). In Nevada, a corrupt cash-for-votes scheme is hiding in plain sight. The
Federalist, November 18. https://thefederalist.com/2020/11/18/in-nevada-a-corrupt-cash-for-
votes-scheme-is-hiding-in-plain-sight/
Eggers, A. C., Garro, H., and Grimmer, J. (2021). No evidence for systematic voter fraud: A guide
to statistical claims about the 2020 election. Proceedings of the National Academy of Sciences
of the United States of America, 118(45).
Electoral Commission. (2017). Analysis of cases of alleged electoral fraud in the UK in 2016:
Summary of data recorded by police forces. UK Parliament (March).
https://www.electoralcommission.org.uk/sites/default/files/pdf_file/Fraud-allegations-data-
report-2016.pdf

Electoral Commission. (2018). Analysis of cases of alleged electoral fraud in the UK in 2017:
Summary of data recorded by police forces. UK Parliament.
https://www.electoralcommission.org.uk/sites/default/files/pdf_file/Fraud-allegations-data-
report-2017.pdf
Flis, J. & Kaminski, M. (2020). Multi-player electoral engineering and COVID-19 in the Polish
presidential elections in 2020. R&R Mind and Society.
Goel, S., Meredith, M., Morse, M., Rothschild, D., & Shirani-Mehr, H. (2020). One person, one
vote: Estimating the prevalence of double voting in US presidential elections. American Political
Science Review, 114(2) 1-14.
Grimmer, J., & Hall, A. B. (2021). Dismantling Trump’s election fraud claims. Washington Times,
February 8. https://www.washingtontimes.com/news/2021/feb/8/dismantling-trumps-
election-fraud-claims/
Heritage Foundation. (2022). A sampling of election fraud cases from across the country. Online
database at Heritage.org/votefraud. Heritage Foundation. Last checked January 3, 2022.
Idrobo, N., Kronick, D., & Rodriguez, F. (2020). Do shifts in late-counted votes signal fraud?
Evidence from Bolivia. Social Science Research Network.
Jung, G. R., Kenny, L. W., & Lott, J. R. (1994). An explanation for why senators from the same
state vote differently so frequently. Journal of Public Economics, 54(May), 65-96.
King, L. (2020). Battleground states: These are the 12 states that will determine the 2020
election. USA Today, November 3.
https://www.usatoday.com/story/news/politics/elections/2020/11/03/key-states-watch-
include-south-carolina-georgia-arizona-and-iowa-democrats-must-win-four-seats-
take/6085491002/
Li, S. and Ramey, C. (2020). What are election observers? Role at crux of Trump lawsuits in
Pennsylvania. Wall Street Journal, November 10. https://www.wsj.com/articles/what-are-
election-observers-the-role-at-the-crux-of-trump-lawsuits-in-pennsylvania-11605053759
Linthicum, K. (2018). In Mexico, voters are wooed with gift cards, washing machines and cheap
tortillas. Los Angeles Times, February 28. https://www.latimes.com/world/la-fg-mexico-tortilla-
war-20180228-story.html
Lott, J. R. (2000). A simple explanation for why campaign expenditures are increasing: The
government is getting bigger. The Journal of Law and Economics, 43(2), 359-393.

Lott, J. R. (2006). Evidence of voter fraud and the impact that regulations to reduce fraud have
on voter participation rates. Social Science Research Network (August).
Lott, J. R. (2007a). Evidence of voter fraud and the impact that regulations to reduce fraud have
on voter participation rates. Social Science Research Network.
Lott, J. R. (2007b). Freedomnomics. Washington, D.C.: Regnery Publishing.
Lott, J. R. (2020a, April 10). Heed Jimmy Carter on the Danger of Mail-in Voting. Wall Street
Journal.
Lott, J. R. (2020b). Why do most countries ban mail-in ballots?: They have seen massive vote
fraud problems. Social Science Research Network (October).
Lott, J. R. (2021a). Taking another look at election fraud. Washington Times, February 15.
https://www.washingtontimes.com/news/2021/feb/15/taking-another-look-at-election-fraud/
Lott, J. R. (2021b). A river of doubt runs through mail voting in Montana. Real Clear
Investigations, March 24.
https://www.realclearinvestigations.com/articles/2021/03/24/a_river_of_doubt_runs_through
_mail_voting_in_big_sky_country_769321.html
Lott, J. R. (2021c). Why do most countries require photo voter I.D.s?: They have seen massive
vote fraud problems. Social Science Research Network (May).
Lott, J. R., & Kenny, L. W. (1999). How dramatically did women's suffrage change the size and
scope of government? Journal of Political Economy, 107(6, pt.1), 1163-1198.
Matsusaka, J. G. (1993). Election closeness and voter turnout: Evidence from California ballot
propositions. Public Choice, 76, 313-334.
McKinley, J. C. (2005, June 29). Lawmakers in Mexico approve absentee voting for migrants.
New York Times. https://www.nytimes.com/2005/06/29/world/americas/lawmakers-in-
mexico-approve-absentee-voting-for-migrants.html
Mebane, W. R., & Kalinin, K. (2009). Comparative election fraud detection. Social Science
Research Network. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1450078

Meyer, K. (2020). Philly judge of elections pleads guilty to election fraud, accepting bribes.
WHYY NPR, May 21. https://whyy.org/articles/philly-judge-of-elections-pleads-guilty-to-
election-fraud-accepting-bribes/
Moulton, B. R. (1986). Random group effects and the precision of regression estimates. Journal
of Econometrics, 32(3), 385-397.
Navarro, P. (2020). The immaculate deception: Six key dimensions of election irregularities. The
Navarro Report, December 15.
New York Times. (1973). Corsicans of France are feeling the sting of publicity given to criminals.
New York Times, January 7. https://www.nytimes.com/1973/01/07/archives/corsicans-of-
france-are-feeling-the-sting-of-publicity-given-to.html
Payne, D. (2020). Pennsylvania poll watcher: 'We literally had no input and no ability to watch
anything'. Just the News, November 9. https://justthenews.com/politics-
policy/elections/pennsylvania-poll-watcher-we-literally-had-no-input-and-no-ability-watch
Rossie, C. (1988). ‘Tacos’ stuffed with fraudulent ballots with PM-Mexico-Elections, Associated
Press, July 7. https://apnews.com/article/4be7794766c41f503e5a29f7d6e72034
Smith, A. V. (2020). How indigenous voters swung the 2020 election. High Country News,
November 6. https://www.hcn.org/articles/indigenous-affairs-how-indigenous-voters-swung-
the-2020-election
Sobey, R. (2020). Rudy Giuliani claims Trump campaign has found nationwide Democrat voter
fraud conspiracy plot. Boston Herald, November 19.
https://www.bostonherald.com/2020/11/19/rudy-giuliani-trump-campaign-has-found-
nationwide-voter-fraud-conspiracy-plot/
World Today News. (2020). In Corsica, the tormented history of the vote by correspondence.
World Today News, June 15. https://www.world-today-news.com/in-corsica-the-tormented-
history-of-the-vote-by-correspondence/
Uhlig, M. (1991). Mexico overhauls electoral rolls. New York Times, January 23.
https://www.nytimes.com/1991/01/23/world/mexico-overhauls-electoral-rolls.html
Viebeck, E. (2020). Minuscule number of potentially fraudulent ballots in states with universal
mail voting undercuts Trump claims about election risks. Washington Post, June 8.
https://www.washingtonpost.com/politics/minuscule-number-of-potentially-fraudulent-
ballots-in-states-with-universal-mail-voting-undercuts-trump-claims-about-election-
risks/2020/06/08/1e78aa26-a5c5-11ea-bb20-ebf0921f3bbd_story.html

Wu, J., Yorgason, C., Flosz, H., Handan-Nader, C., Myers, A., Nowacki, T., Thompson, D. M.,
Yoder, J., and Hall, A. B. (2020). Are dead people voting by mail? Evidence from Washington
state administrative records. Stanford University Democracy & Polarization Lab Working Paper.

Table 1: 2016 Difference in Trump’s share of the absentee ballot vote
between adjacent precincts at the border of Fulton, Carroll, Cherokee,
Coweta, and Forsyth counties
Control variables Coefficient Level of statistical
significance for a two-
tailed t-test
Difference in Trump’s 0.8800 0.000
percentage of the
two-candidate in-
person vote between
two precincts
Fulton County Effect -0.0215 0.466
Number of F-statistic = 60.36 R-Squared = 0.7253

Observations 45 Level of significance =
0.0000

between adjacent precincts at the border of Fulton, Carroll, Cherokee,
Coweta, and Forsyth counties
tailed t-test
percentage of the
two-candidate in-
person vote between
two precincts

0.0000

after adjusting for racial and gender demographics of registered voters

tailed t-test
Difference in the 1.6396 0.528

percent of voters who
are Black males
Difference in the -1.8755 0.300
are Black females

are Hispanic males

are Hispanic females

are Asian males
are Asian females

0.0006

Table 4: 2020 Difference in Trump’s share of the absentee ballot vote after
adjusting for racial and gender demographics of registered voters and the
difference in the in-person vote

tailed t-test
percentage of the two-
candidate in-person vote
between two precincts
Difference in the percent -0.5052 0.768

of voters who are Black
males
Difference in the percent 0.8265 0.519
of voters who are Black
females

of voters who are Hispanic
males

of voters who are Hispanic
females

of voters who are Asian
males
of voters who are Asian
females
Number of Obs = 22 F-statistic = 18.98 R-Squared = 0.9156

Level of significance
= 0.0000

between adjacent precincts at the border of Allegheny, Beaver, Butler,
Washington, and Westmoreland counties
significance for a two-tailed
t-test
percentage of the
two-candidate in-
person vote in the
adjacent precincts
Allegheny County -0.0025 0.770
Effect
0.0000

Table 6: 2020 Difference in Trump’s share of the provisional ballots
between adjacent precincts at the border of Allegheny, Beaver, Butler,
Washington, and Westmoreland counties
significance for a two-tailed
t-test
percentage of the
two-candidate in-
person vote in the
adjacent precincts
Allegheny County -0.0362 0.417
Effect
0.0571

Table 7: 2020 Difference in the share of Biden’s votes from provisional
ballots in adjacent precincts
A) Examining Allegheny, Beaver, Butler, Washington, and Westmoreland

counties
tailed t-test
share of Trump’s
votes from
provisional ballots in
the adjacent precincts
Allegheny County 0.0065 0.008

Effect
0.0000
B) Examining Fulton, Carroll, Cherokee, Coweta, Fayette, and Forsyth
counties
share of Trump’s
votes from
provisional ballots in
the adjacent precincts
Fulton County Effect 0.0312 0.795

0.0000

Table 8: Re-estimating Tables 5, 6, and 7A by including Census 2010
precinct demographic data on difference in the percent of the voting-
age population who are Black, Hispanic, and Asian
Regression Coefficient Level of
Estimate on the statistical
Allegheny significance
County for a two-
Effect tailed t-test
Table 5 -0.0125 0.166 Number of Obs = 87
F-statistic = 6.61
Level of significance F-test =
0.0000
R-square = 0.2871
Table 6 -0.04196 0.376 Number of Obs = 34
F-statistic = 2.70
0.0400
R-square = 0.3180
Table 7A 0.0057 0.036 Number of Obs = 87
F-statistic = 15.13
0.0000
R-square = 0.4798

Table 9: Combining data for Georgia and Pennsylvania to examine the
difference in the share of Biden’s votes from provisional ballots in adjacent
precincts, specifications analogous to those in Tables 2, 5, and 8
(the level of significance for a two-tailed t-test are in parentheses)
Endogenous variable: Difference in Trump’s share of the absentee ballot vote
between adjacent precincts
Control variables (1) (2)
County where Fraud -.034725 -.03866
alleged (0.007) (0.005)
share of Trump’s (0.000) (0.000)
votes from provisional
ballots in the adjacent
precincts
Difference in Percent 0.14833
of the Voting Age (0.228)
Population who are
Black
Difference in Percent -.50884
Population who are
Hispanic
Difference in Percent -.58047
Population who are
Asian
Georgia state dummy 0.04197 0.05289
(0.001) (0.000)
State Fixed Effects Yes Yes
Number of R2 = 0.5337 R2 = 0.5706
Observations = 109 F-test level of significance = 0.0000 F-test level of significance =
0.0000

Table 10: Comparing voter turnout rates in 2020 swing states
(Arizona, Florida, Georgia, Michigan, Nevada, North Carolina, Ohio,
Pennsylvania, and Wisconsin) Observations = 668
Variable Mean Standard Deviation
Percent Voter Turnout 0.7502149 0.0704998
in 2020 Election
Percent Voter Turnout 0.6979785 0.0757554
in 2016 Election
Republican Counties 0.18628 0.21074
(Trump’s minus Biden’s
share of votes)
Republican Counties 0.0790 0.1228
share of votes squared)
Democrat Counties -0.1369 0.200619
share of votes)
Democrats Counties 0.05894 0.10930
share of votes squared)
County where Fraud 0.02844 0.1664
alleged

Table 11: Did counties accused of fraud have an unusual increase in voter turnout? (Arizona, Florida,
Georgia, Michigan, Nevada, North Carolina, Ohio, Pennsylvania, Wisconsin)
(Clustering by state, the level of significance for a two-tailed t-test are in parentheses, results for
percent in different age groups are not shown)
Endogenous variable: The change in voter turnout rate between 2020 and 2016
Control variables (1) (2) (3) (4) (5) (6)
County where Fraud 0.0308095 0.02964 0.02957 0.03008 0.01487 0.01136
alleged (0.013) (0.020) (0.018) (0.019) (0.038) (0.078)
Republican Counties 0.0135 0.01539 0.03472 0.03455 0.02408 -0.0006199
(Trump’s minus (0.498) (0.342) (0.266) (0.286) (0.487) (0.980)
Biden’s share of votes)
Republican Counties -0.02377 -0.02123 -0.003800 -0.005518
(Trump’s minus (0.637) (0.713) (0.940) (0.926)
Biden’s share of votes
squared)
Probability for Joint F- 0.2745 0.3066 0.1453 0.9804
test for Republican
Counties
Democrat Counties 0.008405 0.00755 -0.02767 -0.02604 -0.005774 -0.02089
(Trump’s minus (0.593) (0.625) (0.520) (0.519) (0.841) (0.344)
Biden’s share of votes)
Democrat Counties -0.05635 -0.05457 -0.01698 -0.02209
(Trump’s minus (0.333) (0.321) (0.649) (0.421)
Biden’s share of votes
squared)
Probability for F-test 0.4537 0.6182
for how turnout rates
vary differently
between heavily
Democratic and
Republican counties
Joint F-test for 0.3527 0.3507 0.5930 0.5947
Democrat Counties
Percent Voter Turnout 0.01525
change between 2012 (0.808)
and 2016 Elections
Dummy for states 0.029997
rated as “Toss-ups” by (0.000)
USA Today the
morning of Nov. 3
2020
Percent Female 0.2022785 0.17352 0.16992 0.17069 0.052754 0.0011253
(0.145) (0.179) (0.135) (0.136) (0.458) (0.989)

Percent Black 0.011938 0.0016099 0.0013234 0.0044581 0.0069364 -0.037937
(0.443) (0.949) (0.957) (0.827) (0.820) (0.073)
Percent Hispanic or -0.032705 -0.04526 -0.04666 -0.04713 -0.06818 -0.02579
Latino (0.149) (0.116) (0.118) (0.117) (0.009) (0.357)
Percent Asian -0.23748 -0.2908 -0.2889 -0.2829 -0.1904 -0.1721
(0.326) (0.143) (0.132) (0.104) (0.162) (0.060)
Percent two or more -3.81e-07 -3.71e-07 -3.62e-07 -3.66e-07 -2.40e-07 4.59e-08
races (0.168) (0.197) (0.217) (0.219) (0.314) (0.687)
Median household -2.27e-07 -2.23e-07 -2.12e-07 -1.75e-07 5.57e-08
income (0.449) (0.479) (0.466) (0.493) (0.745)
Percent High School -0.07682 -0.08142 -0.0819 -0.04632 0.02977
Graduate (0.554) (0.519) (0.514) (0.277) (0.535)
Percent Some College -0.07325 -0.07852 -0.07579 -0.03529 0.03434
or Associate (0.390) (0.319) (0.280) (0.438) (0.605)
Percent Bachelor’s 0.03066 0.02924 0.02505 -0.01028 -0.02025
Degree (0.699) (0.707) (0.722) (0.857) (0.672)
Percent Graduate or -0.06451 -0.07101 -0.06893 0.02340 0.05956
Professional (0.634) (0.587) (0.589) (0.745) (0.459)
Probability for Joint F- 0.0035 0.0890 0.1033 0.1752 0.0017 0.0009
test for Census Age
Groups
State Fixed Effects No No No No No Yes
Constant 0.1139 -0.005957 0.00313 0.00426 0.1353 0.07914
(0.520) (0.966) (0.981) (0.973) (0.054) (0.407)
Number of R2 = R2 = R2 = R2 = R2 = R2 = 0.4214
Observations = 668 0.1074 0.1183 0.1210 0.1216 0.3040

Table 12: Focusing on voter turnout in Georgia and Pennsylvania. Using the specifications shown in
Table 11, though not all results are reported. (Florida, Georgia, North Carolina, Ohio, Pennsylvania)
(Clustering by state, the level of significance for a two-tailed t-test are in parentheses)
Endogenous variable: The change in voter turnout rate between 2020 and 2016
Control variables (1) (2) (3) (4) (5) (6)
County where Fraud 0.0312286 0.040204 0.043622 0.04341 0.022067 0.0006146
alleged (0.041) (0.038) (0.036) (0.039) (0.081) (0.871)

Simple Tests For The Extent of Vote Fraud With Absentee and Provisional Ballots in The 2020 US Presidential Election

Uploaded by

Copyright:

Available Formats

Simple Tests For The Extent of Vote Fraud With Absentee and Provisional Ballots in The 2020 US Presidential Election

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Simple Tests For The Extent of Vote Fraud With Absentee and Provisional Ballots in The 2020 US Presidential Election

Uploaded by

Copyright:

Available Formats

Simple tests for the extent of vote fraud with absentee and provisional ballots in the 2020 US

John R. Lott, Jr., Ph.D.

Forthcoming, Public Choice

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

The model is given as:

((A/TA)i0 – (A/TA)i1) = ß((P/TP)i0 – (P/TP)i1) + 𝛿 D*Xi + εI,

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Place Tables 1 and 2 here

Place Table 3 here

Place Table 4 here

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Pennsylvania (p. 48).

Electronic copy available at: https://ssrn.com/abstract=3756988

Place Table 5 here

Place Table 6 here

only 12 observations remained, none of which were located in Fulton County.

Electronic copy available at: https://ssrn.com/abstract=3756988

Place Table 8 here

4. Combining Data for Georgia and Pennsylvania

Place Table 9 here

Electronic copy available at: https://ssrn.com/abstract=3756988

5. Voter turnout rates

Electronic copy available at: https://ssrn.com/abstract=3756988

(https://ballotpedia.org/Election_results,_2020:_Analysis_of_rejected_ballots). The number of absentee ballots

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Place Table 10 here

Place Table 11 here

Electronic copy available at: https://ssrn.com/abstract=3756988

Place Table 12 here

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Lott, J. R. (2007b). Freedomnomics. Washington, D.C.: Regnery Publishing.

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Electronic copy available at: https://ssrn.com/abstract=3756988

Number of F-statistic = 60.36 R-Squared = 0.7253

Electronic copy available at: https://ssrn.com/abstract=3756988

Number of F-statistic = 58.50 R-Squared = 0.8540

Electronic copy available at: https://ssrn.com/abstract=3756988

Control variables Coefficient Level of statistical

Difference in the 1.6396 0.528

Difference in the -4.4266 0.196

Difference in the 2.7631 0.394

Difference in the 1.1089 0.534

Number of F-statistic = 7.48 R-Squared = 0.7774

Electronic copy available at: https://ssrn.com/abstract=3756988