Exit Poll Versus Reported Vote Count
By Theodore de Macedo Soares
The 2020 Texas Democratic Party presidential primary was held on March 3, 2020. Election results from the computerized vote counts differed significantly from the results projected by the exit poll conducted by Edison Research and published by CNN at poll’s closing. According to the exit poll Sanders was tied with Biden but lost in the unobservable computer counts by 4.5%.
In this election candidate Sanders saw the largest discrepancy between the exit poll and computer vote counts. His projected vote proportion fell 4% in the vote counts—an 12% reduction of his exit poll share. The combined discrepancies between the exit poll and the vote count for candidates Sanders and Biden at 4.4% significantly exceeded the 2.9% margin of error for the exit poll difference between the two. The discrepancies between Sanders and Bloomberg at 4.9% was double their respective margin of error. See table below.
There is good reason to believe that the exit poll just prior to publishing showed a Sanders win in Texas.
As explained by Joe Lenski, executive vice president of Edison Media Research in a 2016 interview with The Washington Post, as soon as polls close in a state Edison Research alters the exit poll in accordance with incoming vote counts. They are hired by the networks to predict the winners and losers in an election as soon as possible and to provide the proportion and voting patterns of various demographics and their views on topics of interest. The incoming vote counts are useful to them to better predict the results of the unobservable computer counts. They were not hired to ascertain their accuracy.
Texas as in a few other states such as New Hampshire and Florida have the great majority of the state’s polls closing an hour earlier than the remainder. This is great for Edison Research because they can use that hour’s access to the tabulating votes from much of the state to adjust their exit poll prior to first publishing after all polls close in those states. In Texas, as the final vote count shows, candidate Sanders was losing the state and they likely used these results to downgrade a Sanders win to a tie with Biden in their first published exit poll. Edison Research and or the major networks with access to this unpublished poll would be able to confirm if it indeed showed a Sander’s win and by how much.
The United States remains one of the
few major democracies in the world that continue to allow computerized vote
counting—not observable by the public—to determine the results of its
Countries such as Germany,[ii]
Ireland, Spain, Portugal, Italy, Denmark, Sweden, Finland and many other countries
protect the integrity and trust of their elections with publicly observable
hand-counting of paper ballots.[v]
 Exit poll (EP) downloaded from CNN’s website by TDMS on election night, March 3, 2020 at 9:03 PM ET. Candidates’ exit poll percentage/proportion derived from the gender category. Number of EP respondents: 3,130. As this first published exit poll was subsequently adjusted towards conformity with the final computerized vote count, the currently published exit poll differs from the exit poll used here and available through the link below.
 Candidates’ percentage/proportion of the total computer-generated vote counts derived from reported counts (94% reporting) updated on March 6, 2020 and published by The New York Times. Total number of voters: 3,290,429
 The difference between the exit poll proportion and reported vote proportion for each candidate (subtracting values in column two from the values in column three). A positive value indicates the candidate did better and received a greater share of the total reported count than projected by the exit poll. For example, as candidate Sanders, reported percentage/proportion of the total vote decreased by 4% compared to his exit poll share this value is negative.
 This column shows the percentage increase or decrease from the candidate’s exit poll projection (difference in column four divided by exit poll proportion in column two). Shown, to simplify the table, only for candidates with greater than 4% share in the exit poll.
 This column presents a distinct Margin of Error (95% CI) for the exit poll (EP) differences between candidate Biden and Bloomberg versus each of the other candidate’s EP results. This margin of error (MOE), for example, between Biden and Sanders is 2.9% and the MOE between Bloomberg and Sanders is 2.3%. For simplicity MOE only shown for candidates with greater than 4% share in the EP. As this election involves multiple candidates the common method of ascertaining an MOE of the poll and then doubling it to see if the difference between two candidates is significant is replaced by a more appropriate method that directly calculates a distinct MOE for the difference between any two paired candidates. MOE (95% CI) calculated according to multinomial formula in: Franklin, C. The ‘Margin of Error’ for Differences in Polls. University of Wisconsin, Madison, Wisconsin. October 2002, revised February 2007. Available at: https://abcnews.go.com/images/PollingUnit/MOEFranklin.pdf
 The discrepancies between the exit poll and the reported computer-generated vote counts comparing candidates Biden and Bloomberg with each of the other candidates (subtracting each candidate’s difference between exit poll and computer count from Biden’s difference of 0.3% and in a separate column from Bloomberg’s difference of 1.0%. If the MOE is greater than the discrepancy it the discrepancy is not significant as it can be explained by the MOE. Conversely if the MOE is smaller then it cannot explain the discrepancy and another explanation is required. As shown in the table the combined discrepancies between Warren and Biden and separately between Warren and Blomberg are smaller than their respective MOEs and thus not significant. The combined discrepancies between Sanders and Biden and separately between Sanders and Blomberg are significant and cannot be explained by their respective MOEs.
[i] Fittingly, according to a recent Gallup World Poll, only 40% of Americans say they are confident in the honesty of U.S. elections. Finland and Norway with 89% of their citizens expressing confidence in the honesty of their elections along with the citizens of 25 other countries have greater confidence in their elections than do Americans.
[ii] In 2009 the Federal Constitutional Court of Germany ruled that every important aspect of an election must be observable by the public and thus “meet the constitutional requirements of the principle of the public nature of elections.” The Court explicitly ruled that no amount of voting machine testing, security requirements, and licensing procedures can compensate for this constitutional requirement. With this ruling, Germany abandoned inherently unobservable computerized vote counting and reverted to the hand-counting of every ballot in the precincts in which they were cast and in the plain view of the public.
The court also noted that while vote fraud with hand-counted ballots would be easy to detect, “programming errors in the software or deliberate electoral fraud committed by manipulating the software of electronic voting machines can be recognized only with difficulty.”
[iii] During the 2007 presidential election, eighty-three municipalities (France has 36,569 municipalities) were allowed to use voting machines. Due to security concerns and the inability of voters to determine if their votes are counted correctly a moratorium, that remains today, prevents additional municipalities from introducing voting machines. In the 2012 elections only 64 municipalities continued their use. The French government desires a total ban on their use.
[v] The United States’ long ballots–containing federal, state, and local races–are commonly cited as being unwieldy for hand-counting. The use of Sweden’s method of providing different colored paper ballots for federal, state, and local races that are then sorted prior to hand-counting addresses this objection and allows for at least the hand-counting of federal elections with only three races per ballot.
Edited March 31, 2020. Thanks to Chris Weng for spotting error on table for Bloomberg/Biden MOE.
Texas 2020 Democratic Party Primary Exit Poll. Published by CNN at poll’s closing on election night.
Comments made on this or related pages that may be helpful to the reader (comments made on another page will open a new tab):