Democracy Lost: A Report on the Fatally Flawed 2016 Democratic Primaries – IV. Documented Types of Voter Suppression and Election Fraud in the 2016 U.S. Presidential Primaries – D. INACCURATE VOTING MACHINE COUNTS

D. INACCURATE VOTING MACHINE COUNTS

1. Exit Polls and Computerized Vote Counts

The computerized vote counts in the 2016 Democratic Party Presidential primaries have, in many states, differed widely from the vote totals predicted by the exit polls conducted by Edison Research. These discrepancies were overwhelmingly to Clinton’s benefit. The vote counts for the Republican Party Presidential primaries; however, with the exception of the primaries in two states, West Virginia and Texas, with very large discrepancies (18% and 10.6% respectively) going against Trump, have closely matched the results of their corresponding exit polls.

This section will examine the results of the exit polls and their disparities with computer vote counts; show that there is only one legitimate explanation why the exit polls for the two parties differed; and, lastly, show that the common attempts to explain away the discrepancies between computer vote counts and exit polls, fail for these primaries.

Why did the computerized vote counts only match the exit polls for the Republican Party Primaries and not for the Democratic Party primaries? More to the point, why does this question even arise?

In the United States computerized election vote counts are essentially unverified. Audits of any election, if conducted at all, are hidden from view— only four states specify that observers can verify the markings on the ballots1 (only one state for primary elections).2 Primary election results are only audited in 13 states.3 Only six of these states require that the results of the audit and the data be made public.4  Only one state, for any election,5 is even experimenting with performing audits according to the best practices of a risk-limiting audit.6

To answer the second question posed above, as our computerized vote counts are non-transparent and largely unverified, the comparison of official vote counts with exit poll results is a standard method used to establish the veracity of election results. As USAID stated in their 2015 booklet “Assessing and Verifying Election Results”, [e]xit polls are powerful analytical tools … [a] discrepancy between the votes reported by voters and official results may suggest that results have been manipulated, but it does not prove this to be the case.”7

Presidential 2016 Primaries. Exit Polls versus Reported Vote Counts

The unverified computer vote counts for these primaries differed widely from the Democratic Party exit polls but closely matched the Republican Party exit polls. Illustrative of the difference between the disparities found in the Democratic Party primaries and the lack of such disparities in the Republican Party primaries are the results from the Massachusetts primary.

The explanatory notes for this table and associated article can be seen at: http://tdmsresearch.com/2016/06/14/the-suspect-massachusetts-2016-primary/

In the March 1, 2016 primary in Massachusetts, the exit polls projected a Sanders win by 6.6% and yet he lost by 1.4%, for a total discrepancy of 8%.  The exit polls for the Republican Party race; however, closely agreed with reported vote totals with the largest discrepancy among the five candidates at less than 1%.8

This difference between the two parties is made remarkable by the fact that as in all primary states (for the exception of the primary in South Carolina where the Democratic Party and the Republican Party primaries occurred on different dates), the exit polls for both parties were conducted in the same precincts and at the same time,9 with the same interviewers, and using the same methodologies.

The number of respondents is the primary determinant for the exit poll’s margin of error (MOE). Noteworthy is that the accuracy of the exit polls for the Republican Party, with 800 respondents, was achieved with about 500 less voters filling out the anonymous exit poll questionnaires than the Democratic Party with 1,297 respondents.

To answer the first question: Why did the computerized vote counts only match the exit polls for the Republican Party Primaries and not of the Democratic Party primaries?

As the exit polls for both parties were conducted at the same places and times variations in the conduct of the exit poll cannot explain the difference between the exit polls of both parties. The usual assertion that exit polls may be wrong because respondents, more enthusiastic for a particular candidate, would more likely to agree to be polled also does not apply to these elections. Similar to the many Democratic Party voters demonstrating enthusiasm for candidate Sanders, Republican Party voters demonstrated a great deal of enthusiasm for candidate Trump (see section Addressing Criticisms of Exit Polls below for a more detailed examination of this claim). Thus, we would expect to see consistent exit polling results that were greater than the computer vote count for Trump if this explanation held true, which in fact was not the case.

The only remaining explanation is that computers were accurately counting the votes in the Republican Party primaries but not in the Democratic Party primaries. No other explanation appears to be viable.

The Massachusetts story is repeated in other 2016 presidential primary states. As the Democratic Party primaries table below demonstrates, in 21 of 25 primary states for which exit polls were conducted, the unverified computer count totals differed in one direction—in Clinton’s favor. In ten of these primary results, the discrepancies exceeded the augmented margin of error (MOE) for their respective exit polls (see discussion below for an explanation of the augmented MOE applied). In sharp contrast, the discrepancies in the Republican Party primaries were as likely to favor Trump as the other candidates (13 and 10 respectively).

The explanatory notes for this table can be seen at: http://tdmsresearch.com/2016/06/20/democratic-party-table-2016-primaries/

With the exception of the primaries in two states, West Virginia and Texas, that saw very large discrepancies (18% and 10.6% respectively) going against Trump, the Republican Party races exhibited what should be expected with impartial vote counts and exit polls; sometimes the discrepancies favored Trump, and sometimes went against him.  The average of the discrepancies favoring Trump was 2.7%.  The average of the discrepancies going against Trump, excluding West Virginia and Texas, was almost identical at 2.8%.  These averages discrepancies, less than half of the average of their margin of error (7% and 8% respectively), demonstrate that indeed, exit polls can be very accurate.

The explanatory notes for this table can be seen at: http://tdmsresearch.com/2016/06/20/republican-party-table-2016-primaries/

In the Democratic Party primaries the 21 primary states where the discrepancies between the exit poll and the vote count favored Clinton the average discrepancy was 6.9%.  The top ten averaged a large discrepancy of 10.6%.  The average of the four discrepancies favoring Sanders was 3.2%.

Discrepancies between Computer Counts and Exit Polls Likely Larger

To understand why the exit polls may have substantially underestimated Sanders’ projected vote totals in many of these states, it is important to understand the mission of Edison Research, the firm that conducts exit polls for the major news networks. As widely stated, their purpose is to supply accurate estimates of electoral outcomes, including the demographics and opinions of the electorate to inform the networks and through them, the public. Edison Research assumes the computer vote counts are correct and their mission is successfully accomplished when their estimates match the official computer vote counts.

For this reason, the first published exit polls are progressively altered to conform to the incoming computer vote counts until all discrepancies between the first published exit polls and the final computer vote counts are reduced to approximately zero. Only these exit polls, now matching the final vote counts, are presently available to the public on the various networks websites.

To fulfill their mission Edison Research conducts the exit poll surveys according to the well-established science common to all surveys including exit poll survey methodology scientifically honed through decades of practice. As a necessary part of estimating the electoral results they also include the results of their telephone surveys of absentee/early voters which in some states may be a significant part of the electorate.

The reason Edison Research may have actually underestimated candidate Sanders exit poll results in many of the primary states is that their scientifically conducted exit polls estimates are further modified with the “aggregate of pre-election polls…[and] actual vote returns from sample precincts and county vote returns as provided by the Associated Press and location election officials.”10

The Effect of Pre-election Polls

The reason the inclusion of pre-election polls may have resulted in underestimating Sanders’ exit poll results is that in 19 of the 25 primary states the aggregate of the pre-election polls11 conducted within a week or so of the election of each primary state, projected a lower margin for Sanders than projected by the first exit polls published by the networks. In thirteen of these states the under projection was large with an average of 11%. A large miss is illustrated by the pre-election polls prior to the primary in Michigan on March 8, 2016. The aggregate pre-election polls from March 1-7 had Sanders losing by 22%. The losing margin increased to 29% in the period March 4-7.12 The exit polls; however, projected a Sanders’ win by 6.2% and the final vote count had Sanders winning by 1.6%.

Without access to Edison’s raw exit poll data it is unknown how much weight or percentage points they subtracted from Sanders’ exit poll results in order to more closely conform to the pre-election polls. Assuming, for illustration purposes, that Edison had subtracted half of the average March 1-7 pre-election polls (11%) from Sanders’ exit poll results the actual exit poll totals for Sanders would have shown a winning margin of 17.6%. The resulting discrepancy with the computer count would have been a whopping 16% instead of the 4.6% shown using the adjusted exit polls.

As another example, the March 4-13 pre-election polls for the Ohio primary on March 15, 2016 averaged a Sanders’ loss by roughly 14%.13 Similar calculations as above would have yielded an exit poll showing Sanders winning this state by 3.2% and the resulting discrepancy with the unverified computer vote count showing a Sanders loss by 13.8% would have been another whopping 17% instead of the already large discrepancy of 10% using the adjusted exit polls.

Again, these illustrations depend on how much weight Edison applied to these pre-election polls, which could have been more or less than the examples illustrated above. As the pre-election polls, however, generally underestimated Sanders’ exit polls’ electoral strength the majority of the discrepancies going against Sanders were likely larger than shown using the adjusted exit polls. The discrepancies favoring candidate Sanders may have instead favored Clinton.

The Effect of Vote Counts

As Mr. Lenski of Edison Research explained, when states such as “New Hampshire, Kentucky, Indiana, Florida and Texas (and others) have split poll closing times where part of the state closes at one time and the rest of the state closes at a later time…we may have quite a bit of actual vote returns to use in our estimates that are broadcast…when all of the state polling locations have officially closed.”14 Although Mr. Lenski did not explicitly state that in other states without split closing times they also used vote returns acquired from the precincts they were polling or from other sources, there seems no reason why Edison would not use them if available.

The problem with altering exit polls with vote returns from computer counts that overwhelmingly favored Clinton compared to the states’ exit polls is that such returns would have depressed Sanders’ exit poll totals even further, resulting in an even larger discrepancy between the exit polls and the vote counts. When one adds the effect of pre-election polls the resulting combined discrepancies may have been very large indeed.

Take Texas, for example, a state with split poll closing times. The exit polls had Sanders losing the state by 22.7%. The unverified computer vote count resulted in a 32% Sanders loss. The question becomes how much of the 22.7% of Sanders losing exit poll totals had been bumped up to match the incoming vote totals from this state? If the 22.7% exit poll total had been bumped up by, say, 10% to come closer to the incoming votes it means that the exit poll with the pre-election poll adjustment was actually 11.7%. The RCP pre-election polls averaged a Sanders loss by 30.3%.15If a hypothetical 15% of the pre-election polls had been included to increase his loss margin to come closer to the pre-election polls it is then possible that Sanders unadjusted exit poll may have even projected him winning the state of Texas by 3%. The large existing discrepancy of 9.3% between the exit polls and the computer vote counts would have instead been a mammoth 35%.

Of course, all these calculations are hypothetical. Without access to the raw exit poll data from Edison Research with a detailed accounting of the adjustments made, these calculations will have to remain as speculations. They are presented as illustrations of the possibility that the first exit polls published by the networks soon after the closing of the polls for the state may have significantly understated Sanders’ electoral strength. The actual scientifically conducted exit polls may have shown much greater totals for candidate Sanders and may have even shown that he handily won the 2016 Democratic Party primary contest against candidate Clinton.

Addressing Criticisms of Exit Polls

Determining Total Survey Error

Exit polls like all surveys are susceptible to the following possible sources of error:  Coverage error (not able to sample a selection of the population), nonresponse error (not able to poll all persons in the chosen sample), sampling error (when the survey sample is different than the population being measured), and measurement error (from inaccurate responses).16

The results of surveys of all kinds, such as the pre-election polls referenced in this section, are generally published with an associated margin of error that is primarily associated with the sample size—the larger the sample, the smaller the margin of error. The other possible sources of error, difficult to quantify and formulate mathematically are almost always left unaddressed. The fact that exit polls are singled out for this fault does not diminish the validity of this observation.

The best approach would be to determine the total survey error (TSE) for any given survey. If the actual population values being measured were known then the results of the survey could be compared to determine how close the sample-based survey came to these values. The resulting margin of error for all the possible sources of error could then be determined. This was the approach taken in the margin of error that was applied to both the Democratic and Republican Party primaries.

Exit polling for the Republican Party primaries, as noted above, was fairly accurate with only two states exhibiting suspiciously large discrepancies (Texas at 10.6% and West Virginia at 18%) going against Trump. Excluding these anomalous and suspicious results from the calculations the Republican Party primaries yielded 106 separate exit poll results for the six main candidates (Trump, Cruz, Kasich, Rubio, Carson, and Bush). To determine the total survey error their exit poll survey estimates were compared to their vote totals.17 This comparison determined that by increasing the standard statistical margin of error by 32%, the percentage of exit polls exceeding this augmented MOE was 4.76% (5 of 106) which is just under the expected 5% in an exit poll with a confidence interval of 95%. Accordingly, this increased MOE represents the total survey error of the exit polls for the Republican Party primaries.

As mentioned above, since in all primary states (with the exception of South Carolina) the exit polls for both parties were conducted in the same precincts, at the same time, with the same interviewers, and using the same methodologies, the increased MOE also represents the total survey error for the Democratic Party primaries.

The bottom line: the Republican Party Primaries had two states exhibiting large discrepancies exceeding the total survey error, whereas the Democratic Party primaries had 10 states that exceeded the TSE. Since all the potential sources of error common to surveys including exit polls are accounted for with the application of this augmented margin of error, the current use of these exit poll results to assess the veracity of the unverified computer vote counts should be harder to criticize.

Youthful Voters not the Cause of Discrepancies between Exit Polls and Computer Counts

Nate Cohn, a writer for the New York Times, recently wrote an article theorizing that the reason “why…exit polls so tilted toward Mr. Sanders,” was due to the fact that “young voters are far likelier to complete the exit polls than older voters.” Although he acknowledged that Edison Research corrects for this usual and well known fact, Mr. Cohn asserts that their correction “falls short.”18

Mr. Cohn’s assertion is a variation of the popular criticism of exit polls that centers on nonresponse error as the cause for the discrepancies between the exit polls and the final vote count—the voters of a particular candidate may be overrepresented because they are more enthusiastic for their candidate and thus more willing to fill out the exit poll’s anonymous questionnaires than the voters for another candidate. This criticism is generally put forth without any evidence and has been refuted as the explanation for the discrepancies in previous elections. 19 Additionally, Edison Research works very hard to reduce such errors by taking note of the characteristics of the nonresponders, such as their gender and age range, and adjust their exit polls results accordingly.20

According to Mr. Cohn, the discrepancies between the exit polls and the computer vote counts that in 21 of 25 states favored Clinton were caused by an overrepresentation of younger voters in the exit polls. As younger voters overwhelmingly voted for Sanders, the theory goes, the alleged uncorrected higher proportion of younger voters participating in the exit polls distorted the exit polls in Sanders’ direction.

Actual 2016 Democratic Party primaries data does not support Cohn’s theory:

  1. The proportion of youthful voters in these primaries had no correlation with the discrepancies between the exit polls and the unverified computer vote counts that favored candidate Clinton.
  2. The enthusiasm of youthful voters for candidate Sanders had no correlation with the proportion of youthful voters in the different states.
  3. Higher proportions of youthful exit poll respondents voting for Sanders, instead of correlating with higher discrepancies, actually correlates with lower discrepancies; the opposite of what would hold according to Cohn’s asserted explanation for the discrepancies.

Doug Johnson Hatlem, in an article published in CounterPunch, pushed back on this theory by noting that states such as Alabama, Georgia, Ohio, and South Carolina with lower proportions of young voters, had high discrepancies between the exit polls and vote counts favoring Clinton while states such as North Carolina, with higher proportions of young voters, had small discrepancies.21 Here, in the following first graph, the observations of Mr. Hatlem are extended to include the Democratic Party primaries of all states for which exit polls were conducted.

(1) If Mr. Cohn is correct then one should see a clear correlation between the proportion of youthful voters and the discrepancies—the greater the proportion of younger voters in the exit poll, the greater the discrepancy. Actual data, however, shows that there is near zero correlation between the proportion of exit poll responders age 18-29 and the discrepancies found in the 25 primary states for which exit polls were conducted:

Each dot on the graph represents exit poll data for a state in the 2016 Democratic Party primaries. The leftmost dot, for example, represents the results of the primary election in Pennsylvania. The dot marks the proportion of voters age 18-29 at 11% on the horizontal x-axis and the discrepancy between the exit poll and the computer vote count at 2.6% on the vertical y-axis. The first thing to notice is that the linear trend (the dotted line) is almost horizontal indicating no relationship between the increasing proportion of youth and the increasing discrepancies in the various states. The second thing to notice is that the dots are widely scattered around the trend line. If there were a strong positive correlation, the dots would be bunched close to the trend line and it would be pointing steeply upwards; showing that greater proportions of youthful voters resulted greater discrepancies in favor of Clinton.

This chart shows that as the proportion of voters aged 18-29 increases in 25 primary states, there is almost no increase in the discrepancies between the exit polls and the unverified computer vote counts. As the results for every state are spread-out in the graph there is also a near zero (R =0.061) correlation between the proportions of this age group in the various states and the discrepancies. The proportion of the discrepancies that can be explained by the proportion youthful voters is also almost zero at 0.37% (R2=0.0037). Contrary to Mr. Cohn’s theory, the proportion of youthful voters represented in the exit polls is not correlated with the exit polling discrepancies and does not explain why Democratic candidate Sanders’ computer vote counts were so much less than predicted by the exit polls in so many states.

Increasing the age range to 18-44 also does not support Mr. Cohn’s assertions: there is a near zero (R=0.074) correlation between the discrepancies of the exit polls with the unverified computer vote counts. The proportion of the discrepancies that can be explained by the proportion of youthful voters is also almost zero at 0.55% (R2 =0.0055; graph not shown).

(2) If Mr. Cohn is correct the higher level of youthful enthusiasm for Mr. Sanders should result in larger proportion of youthful voters participating in the exit polls in the different states. Actual data, however, shows that there is near zero correlation between the enthusiasm of youthful voters and their proportion in the exit polls for all states.

This chart shows that as the enthusiasm level increased for Mr. Sanders—as his share of the youth vote increased among exit poll respondents—there was no accompanying increase in the proportion of youthful exit poll respondents. In Mississippi (represented by the leftmost dot), for example, Sanders received 49% of the exit poll vote of the respondents age 18-29 while the proportion of this age group participating in the exit poll was estimated at 15% by Edison Research. In contrast, Vermont (represented by the rightmost dot) with the same proportion of youthful voters (15%) had 95% of this age group stating they voted for Sanders. Similar to the chart above, the almost horizontal trend line shows no correlation between the increasing exit poll vote share for Sanders and the increasing proportions of youthful voters.

Increasing the age range to 18-44 also does not support Mr. Cohn’s assertions: there is a near zero (R2 =0.0042) relationship between the level of enthusiasm for this extended age group and their proportions in the exit polls of various states (graph not shown).

Contrary to Mr. Cohn’s theory, the level of enthusiasm for Mr. Sanders was not at all correlated with any increase in the proportion of youthful voters in the exit polls.

(3) Finally, If Mr. Cohn’s and the standard explanation for vote count/exit poll discrepancies is correct, higher levels of youthful enthusiasm for candidate Sanders should result in them being more willing to fill out the exit poll questionnaires and thus increase the discrepancies between the exit poll predictions and the unverified computer vote counts. Actual data, however, shows just the opposite: as Senator Sanders share of the youth vote among exit poll respondents increased, the discrepancies decreased.

This chart shows that as the enthusiasm level increased for Mr. Sanders—as he received a greater share of the youthful exit poll vote—the discrepancies between the exit polls and the computer vote counts decreased. In Alabama (represented by the topmost dot), for example, the state with the highest discrepancy at 14%, Sanders received 55% of the exit poll vote of the respondents age 18-29. Oklahoma (represented by the bottommost dot) with the lowest discrepancy at -6.1% (favoring Sanders) received 81% of the exit poll vote.

In contrast to the previous charts the trend line instead of being horizontal is pointed steeply downward indicating a negative correlation—as the youth support for Sanders increases, the discrepancies between the vote counts and the exit poll results decreases. The dots representing the various primary states are bunched closer to the trend line indicating greater correlation between the variables (R=0.60).

At R2=0.365, about a third (36.5%) of the discrepancies can be explained by Sanders’ age group results. These results are statistically significant (p-value=0.0013) with a very low probability that the relationship occurred by chance. Additional statistical details, related explanations, and data source for this and the other graphs above can be found at: http://tdmsresearch.com/2016/07/23/youthful-voters-exit-polls/.

Age is the demographic characteristic widely acknowledged as the primary determinant of candidate Sanders’ vote share—the younger the age group, the higher his vote share. This chart shows that this fact is not responsible for increasing the disparities between the exit polls and the computer vote counts. Contrary to the assertions by Mr. Cohn and others, the increasing strength of candidate Sanders youth vote is associated with decreasing the disparities. Why this is the case remains a question to be addressed later.

In summary: the analyses presented do not support Mr. Cohn’s or the standard assertions made to explain away the discrepancies between the vote counts and exit poll results. In the 2016 Democratic Party presidential primaries the proportion of youth participating in the exit polls did not drive the discrepancies. The level of enthusiasm among younger voters did not increase the exit poll participation of younger voters—it had no relationship whatsoever to the proportion of youth participation in the various primary states. And finally, the increasing level of enthusiasm among younger voters instead of increasing the discrepancies as their theory would hold, was found to significantly decrease the discrepancies between the unverified computer vote totals and the exit poll projections.

Nate Cohn, of the New York Times, and various other media pundits, have continually attempted to undermine the use of exit polls as a means to verify the results of computer vote counts that are not verified by any other means. These pundits do not hesitate to label those concerned with election integrity as “conspiracy theorists.” Along with offering little or no evidence in support of their positions, the common weakness in their commentaries is their unspoken blind faith in unverified computer vote counts.  Such unquestioning faith apparently prevents them from even acknowledging the possibility that the unverified computerized vote counts may be questionable and suspect.

It is to the very strong statistical evidence for this notion—that electronic vote counts may have been tampered with in the 2016 Democratic presidential primaries—that we turn in the next section of the report.

2. Clinton’s Vote Share Increases with Precinct Size, Independent of Demographic Factors

Section Summary

Within a given location with a shared history and similar demographics, we would expect a reasonably low level of variation from neighborhood to neighborhood in terms of a particular candidate’s share of the vote by percentage. At the precinct voting level in large, diverse counties, however, Election Justice USA has found a strong and disturbing statistical pattern in seventeen to twenty states. We believe the pattern to be highly indicative of potential election fraud. The pattern occurs even when carefully controlling for demographic factors such as age and race. With some notable exceptions, this pattern appears in states where exit polls by Edison Research or Capitol Weekly (California) showed high levels of discrepancy with the unverified machine count results.

The Pattern: Hillary Clinton’s share of the vote by percentage increased strongly and remarkably smoothly as precinct size by total vote increased.

Our method of analyzing data is based on the Law of Large Numbers. As a sample size grows, its average rapidly approximates the average of the population being measured. Toss a coin repeatedly, and, as the number of tosses increase, the increasing sum of the tosses rapidly approaches the average of the two-sided coin—50% of the tosses average heads or tails.

The same principle applies to the election results between candidate Sanders and candidate Clinton. As the number of votes accumulate for each candidate, their individual vote share should rapidly approximate their average share of the vote from the entire state. This was not the case in states with large exit polling discrepancies favoring Clinton, however. That Clinton’s vote share rises smoothly with precinct size raises the strong possibility of voting machine tampering.

Fritz Scheuren, professor of statistics at George Washington University and the 100th President of the American Statistical Association (ASA), states: “as a statistician, I find the results of the 2016 primary voting unusual. In fact, I found the patterns unexpected [and even] suspicious. There is a greater degree of smoothness in the outcomes than the roughness that is typical in raw/real data.”22

The argument Election Justice USA is advancing suggests that an algorithm may have been applied to electronically counted votes. The proposed algorithm would have increased Clinton’s share of the vote and decreased Sanders’ share of the vote by an increasing percentage as precinct size by total vote increased. Because the final numbers would be algorithmically related to the actual vote total, they would remain random in a way that would avoid detection by election fraud analysis tools. The logic is simple: discrepancies and irregularities are easier to conceal in precincts with more votes, and, in cases where a limited number of precincts can be targeted, the larger precincts yield a greater number of votes to work with.

Methodology

The largest county from each Democratic primary state (except Mariposa County, AZ) were examined at the precinct level. Mariposa County Board of Elections (BoE) does not post precinct-level data on their website. Furthermore, all FOIA attempts made to the Mariposa County BoEabout acquiring detailed election results were not answered. The county election results were each sorted by precinct size–number of Democratic ballots cast. These results were then evenly distributed into seven blocks representing a group of precinct sizes. Based on a wide range of differing state-to-state precinct sizes, these seven blocks could not be standardized across all counties. Instead the grouping sizes were determined based on having ample precinct representation within each block. For example, in Marion County, Indiana, the blocks ranged from less than 100 to 300 or more ballots cast with a separation of 40 ballots. However, in East Baton Rouge Parish, Louisiana, the precinct sizes on average were much smaller than in Marion County. The seven blocks ranged from less than 50 to 175 or more ballots cast with a separation of 25 ballots.

MARION COUNTY, IN

No. of Precincts

68

99

115

101

72

51

91

No. of Ballots Cast

< 100

100-139

140-179

180-219

220-259

260-299

300+

EAST BATON ROUGE PARISH, LA

No. of Precincts

34

73

53

39

43

19

54

No. of Ballots Cast

< 50

50-74

75-99

100-124

125-149

150-174

175+

In Ohio, for instance, Hillary Clinton outperformed Edison’s initial full exit poll by 10%. In Cuyahoga County (Cleveland), the gap between Hillary Clinton’s election day share of the vote and Bernie Sanders’ election day share of the vote moved by 11.4% in Clinton’s favor from the smallest grouping of precincts by total votes to the largest grouping of precincts.

In numerical form, Bernie Sanders’ share of the vote dropped from 43.8% of the vote in the smallest seventy-seven precincts (0-75 total votes) to just 38.1% of the vote in the largest eighty precincts (200 or more votes). Clinton’s share of the vote increased over the same precincts, smallest to largest by total vote, from 56.2% to 61.9%.

The argument Election Justice USA is advancing suggests that an algorithm may have been applied to electronically counted votes. The proposed algorithm would have increased Clinton’s share of the vote and decreased Sanders’ share of the vote by an increasing percentage as precinct size by total vote increased. Because the final numbers would be algorithmically related to the actual vote total, they would remain random in a way that would avoid detection by election fraud analysis tools such as a Benford’s Law Test.23 The argument is not based upon which candidate won or lost a particular state. Because the Democratic primaries awarded delegates proportionally, winning or losing a state by a greater or lesser percentage had a tremendous impact on the number of delegates each candidate accrued for the overall pledged delegate race.

North Carolina, where Clinton won by 14%, shows what a large county looks like where the precinct share increase pattern does not appear:

The actual difference in the vote share gap between the smallest grouping of precincts for Mecklenburg County (Charlotte), North Carolina is less than 1% – 38.4% in a grouping of the smallest twenty-three precincts for Sanders versus 38.0% for Sanders in a grouping of the largest thirty-six precincts.

A different way of graphing the same data also shows a small amount of expected variation with relatively horizontal, parallel lines.

While Mecklenburg County, North Carolina is known to be highly racially polarized, with the majority of its public schools including 80% or more students from a single racialized group,24 variation between precinct size throughout the county results in very little variation in terms of vote share by precinct size groupings. Edison Research’s initial full exit poll for North Carolina was correct within less than two percentage points.

Even when controlling for factors such as age and racialization, Election Justice USA’s analysis of large counties in the following primary states shows that Clinton’s share of the vote by percentage increased substantially, grouping by grouping, as precinct size by total vote increased: Alabama, California, Georgia, Illinois, Indiana, Kentucky, Louisiana,Maryland, Mississippi, Missouri, New Hampshire, New Jersey, New Mexico, New York, Ohio, South Carolina, Tennessee. In these seventeen states, precinct level analysis of the largest counties by population show that the gap between Clinton and Sanders’ share of the vote by percentage changed in Clinton’s favor by more than 10% from the smallest grouping of precincts by total votes to the largest grouping of precincts by total votes. In extreme cases, the gap increased in Clinton’s favor by more than 60% from the smallest to largest precinct groupings.

In an additional four states (Massachusetts, Pennsylvania, Maryland, and West Virginia), the pattern appears in weaker form. Clinton’s vote share gap increased by between 5% and 10% in Pennsylvania and West Virginia’s largest counties from smallest to largest precinct size. In the case of Massachusetts, the pattern appears only in the more specific location of college or university towns. Fourteen primary states showed no meaningful indication of this pattern.

Discussion

Previous analyses claiming fraud in prior elections or in the 2016 Democratic primary have noted substantial differences in vote share by percentage for particular candidates between counties or smaller localities using hand-cast, hand-counted ballots and those using electronic voting machines. One example includes the 2008 Democratic primary in New Hampshire where the analysis has been dismissed as “simply an artifact of the demographics in New Hampshire of the towns that had voting machines versus those that voted on paper.”25 To test and avoid similar criticism, Election Justice USA has analyzed precinct level voting results of fifty counties in states that held Democratic Presidential primaries in 2016. Our analysis has included the largest county in thirty-five of the thirty-six primary states. Election Justice USA could not obtain precinct level results from Maricopa County, the largest in Arizona.26

Because an enduring narrative theme of the 2016 Democratic Primary has been Hillary Clinton’s strong advantage with non-white voters, particularly with older black voters, Election Justice USA has controlled for demographic factors, particularly race, in a wide variety of ways. These controls began with our decision to focus on the largest counties by state. Almost without fail, these counties are the most demographically diverse within a given state. In several important instances, these largest counties by state include a majority population of non-white residents. Our analysis includes all ten of the most racially segregated cities in the United States.27

Election Justice USA has further controlled for race in New York City where the New York Times website includes data for all five thousand precincts in terms of racialized majority. In Louisiana, Election Justice USA obtained racial registration data by party for every precinct in the state and has included it in our graphical analysis as well as in further linear regression analysis. In Cook County, Illinois and Wayne County, Michigan, controlling for race was made further possible by the fact that separate data is kept in each county for Chicago and Detroit and for their far whiter suburbs.

Specific analysis for each of the twenty states in question is included in Section III of this report. The remainder of this section references specific states in order to support the generalized case and to describe controls for wealth, age, and race. Before conclusions are drawn in the final section, insights from the discussion that follows are brought into conversation with conclusions from the previous sections (on exit polling) and with a nod toward the next section (hackable voting machines) of this report.

Wealth

Certain arguments have been advanced against early versions of our argument about Clinton’s percentage share of the vote increasing as precinct size increases. The New York Times’ Nate Cohn and Twitter user @NaphiSoc have made this argument in slightly different ways. After Cohn wrote an article attempting to debunk exit poll “conspiracy theorists,” he went back and forth for several days on Twitter with two co-authors of this report to argue that “racial polarization” was responsible for the phenomenon in question.28 @NaphiSoc put the argument succinctly: larger precincts are in more densely populated areas that generally have more people of color and therefore more support for Hillary Clinton. Smaller precincts are more rural and white.29

Virtually all of our data analysis is intentionally from the largest counties by state, generally anchored by very large, high density cities. The formulation by @NaphiSoc, as such, somewhat misses the point. Very few of the precincts analyzed in what follows qualify as rural, and we are often able to limit our analysis either solely to large cities or two large cities in relation to their less colorful suburbs. The argument by @NaphiSoc overall, however, is a cogent one, even if it is one we anticipated from the outset. Generally, less wealthy people of color can reasonably be assumed to be bunched up in higher density housing areas such as apartments and high-rises. It is logical that precinct size in these areas may be greater, leading to greater support for Clinton than in whiter, less dense population areas.

One problem with this argument already at the theoretical level is that wealthier Democrats were also noted to have favored Clinton.30 On a theoretical level, then, one might expect to see Clinton doing better in both the smallest and the largest precinct groupings, with perhaps a bit of a dip in mid-sized precincts. The line graph chart above for Mecklenburg County shows just such a pattern. And, this is precisely the pattern we see in Wayne County, Michigan. Wayne County includes Detroit (83% Black) as well as its wealthier and whiter suburbs like Grosse Pointe. Grosse Pointe is on the shore of Lake Michigan, it is 92% white, and the median household income is $100,688.31 Grosse Ile Township, highly similar demographically, is situated on several islands just offshore from Grosse Pointe.

While there are plenty of Detroit precincts in both the smallest and largest precinct sized grouping in the following graph, Grosse Pointe or Gross Ile precincts in the smallest precinct grouping outnumber Grosse Pointe or Gross Ile precincts in the largest grouping four to one. As expected, Clinton does a slightly better in both the smallest and largest precinct grouping, while the overall trend is basically horizontal with a slight edge overall asprecinct increases for Bernie Sanders.

Election Justice USA analyzed Wayne County as a whole, Detroit alone, Detroit suburbs alone, and early ballots for Detroit alone. We found no indication in any of the four analyses that Hillary Clinton’s share of the vote increased with precinct size. The initial full exit poll for Michigan as a whole showed Clinton outperforming her expected vote total by 4.6% and for Wayne County by 10%. These numbers put Michigan in the middle range, somewhat high, but not outside the margin of error as figured by sample size alone.

Age

Nate Cohn of the New York Times has attempted to explain exit polling misses in large part by reference to early or absentee voting where older, more party loyal voters tend to dominate. Exit polling, as Cohn somewhat persuasively argues, tends not to include a large enough early voting sample.32 While this argument with respect to exit polling is addressed in more depth in the previous section of this report, it is of interest to Election Justice USA’s study of a precinct size increase pattern favoring Hillary Clinton because of its impact on the question of age. We have attempted to control for factors that might help explain the pattern and where it shows up; age is one potentially explanatory factor.

Election Justice USA has separately analyzed early, absentee, or vote by mail balloting for at least eight large counties or cities in six different states. Capitol Weekly’s exit poll of vote by mail voters in California was particularly important in this regard as it used data directly from California counties to weight its exit poll by age, race, party, and geography (specifically, by congressional district).33 Data from the counties showed that nearly 70% of all early voters in California were fifty-five-years-old or older. The fifty-five and older age demographic generally made up around 35-45% of Edison exit polling for each primary state. Election Justice USA was then able to obtain and analyze early vote by mail returns by precinct, as reported on election night, for multiple California counties with more than one million residents.

Given age similarity on top of geographic, political, and racial or ethnic demographics, we would expect an even greater likelihood that voting patterns would be similar across precinct size within a given county. In Detroit (without its suburbs), Harris County (Houston), and even Cuyahoga (Cleveland), fairly straight, horizontal data trend lines fit the expected pattern. Where Cuyahoga in our initial analysis in this section showed a 11.4% increase for Clinton on election day, absentee voting in Cuyahoga looks untouched, with a very slight trend toward Sanders:

In Fulton County (Atlanta), Georgia and Los Angeles, Orange, and San Diego counties in California, however, there is a dramatic trend in early voting, increasing in favor of Clinton as precinct size by total votes increases.

Edison exit polling missed by a massive 12.2% in Georgia, and while election day voting swung a fair amount in Clinton’s favor (11.4%) according to this analysis, Clinton’s percentage gap over Sanders in the same precincts increased by nearly 23.0% from smallest to largest precinct groupings in early voting totals.

Meanwhile, Capitol Weekly’s massive poll of early voters in California34 found “Clinton getting 55%-to-58% of the early vote in the Bay Area congressional districts and narrower margins in the LA-area.” With more than 10,000,000 residents, just 26.6% of them non-Hispanic whites, Los Angeles County is the largest election district in the United States (New York City is divided up into five counties by borough). Latinos make up a near majority at 48.4% as of July 2015, according to U.S. Census figures. Throughout California, the Capitol Weekly exit polling missed by 15-25%, in line with where the vast majority of larger precincts differed from the smallest precinct sizes.

Orange and San Diego Counties show the same wide discrepancies for early vote by mail voters – an 15.4% Clinton improvement on the gap for Orange County and a 18.2% improvement for the same in San Diego County. Where we would expect election day to be less stable with more varying demographics, all three Southern California coastal counties show further unusual patterns.

Los Angeles County looks as if it was likely tampered with at a 29% improvement for Clinton on the gap from smallest to largest precinct groupings on election day (still a marked amount lower than the early voting improvement). San Diego and Orange Counties, however, show substantial but weak data trend line improvements in Bernie Sanders’ favor for election day. Sanders improved by 6.8% from smallest to largest precinct groupings in San Diego on election day and by 9.4% for the same in Orange County. Since precincts for election day and early vote by mail remained the same for Orange, Los Angeles, and San Diego counties, there appears to be almost no possible demographic argument for Clinton consistently outperforming early voting exit polling in those counties while also improving rapidly and consistently upward from smallest to largest precincts by total votes.

Chicago early voting totals show a discernible but weak improvement in Clinton’s favor by 9.2% from smallest to largest precinct groupings. Since early balloting in Chicago has come into question based on a citizen groups’ observations of the post-election audit, this material is discussed in more depth in §I.D.

Race

Nicholas Bauer, a member of the research and writing team for this report, was the first to notice and analyze the effect of this pattern even when controlling for race down to the precinct level within a given county. He had previously reviewed cumulative vote share – cumulative vote total analyses for earlier primary states, but after the twelve-point exit polling miss in New York, he decided to test the theory at smaller levels within highly diverse New York City. It would not be possible, if the phenomenon showed up in New York City, to dismiss it as an “artifact” of white rural voters preferring Bernie Sanders with non-white urban voters preferring Hillary Clinton.

It was discovered that Clinton’s share of the vote by percentage increased discernibly and substantially as election district (precinct) size increased in four out of the five boroughs: The Bronx, Brooklyn, Queens, and Staten Island but not Manhattan. Bauer then looked at specific, concentrated areas within those boroughs or counties – Harlem (Manhattan), for a high concentration of black voters, and in the heart of The Bronx, for a concentration of Latino voters. The Bronx is New York’s only majority Latino county and is the most diverse county in the United States according to USA Today’s Diversity Index.35

In Harlem, gentrification is turning the area less black and more white and brown, but black residents remain the majority in Central Harlem.36In Central Harlem election districts it was discovered that, as in Manhattan as a whole, precinct size was not a predictor of Clinton’s share of the vote. In fact, we see the same pattern in Manhattan that we see in Mecklenburg and Detroit. Clinton does a bit better in both the smallest and the largest election districts amidst an overall, relatively horizontal data trend line:

Here’s what the pattern looked like, however, in Bronx County on Bauer’s initial analysis:

BRONX NYC

Clinton improved by 18.2% from the smallest precincts (less than 90 total votes) to the largest in The Bronx, a number similar to Clinton’s improvement by precinct size for Brooklyn, Staten Island, and Queens as well.

 

NEW YORK CITY

a

b

c

d

e

f

g

h

i

No. of Precinct
Ballots Cast

< 90

90-119

120-149

150-179

180-209

210-239

240+

DELTA
(b – h)

Bronx – SANDERS

38.0%

35.5%

31.6%

30.0%

29.7%

28.8%

28.9%

9.1%

Bronx – CLINTON

62.0%

64.5%

68.4%

70.0%

70.3%

71.2%

71.1%

Brooklyn – SANDERS

45.7%

43.7%

39.5%

39.6%

38.3%

38.4%

40.1%

5.6%

Brooklyn – CLINTON

54.3%

56.3%

60.5%

60.4%

61.7%

61.6%

59.9%

Manhattan – SANDERS

35.3%

35.3%

33.3%

34.6%

33.0%

33.3%

34.2%

1.2%

Manhattan – CLINTON

64.7%

64.7%

66.7%

65.4%

67.0%

66.7%

65.8%

Queens – SANDERS

41.8%

44.0%

41.9%

40.0%

38.9%

38.0%

34.7%

7.2%

Queens – CLINTON

58.2%

56.0%

58.1%

60.0%

61.1%

62.0%

65.3%

Staten Island – SANDERS

53.2%

51.9%

47.4%

42.5%

40.7%

40.6%

40.2%

13.0%

Staten Island – CLINTON

46.8%

48.1%

52.6%

57.5%

59.3%

59.4%

59.8%

CITYWIDE – SANDERS

44.4%

42.7%

38.6%

36.8%

35.2%

34.8%

35.9%

8.5%

CITYWIDE – CLINTON

55.6%

57.3%

61.4%

63.2%

64.8%

65.2%

64.1%

At that point, Bauer contacted Election Justice USA. One of the first follow-on tests we ran included isolating all Latina and Latino majority election districts within a particular congressional district – CD 15. CD 15 lies wholly within The Bronx. The pattern still showed: