In the 2016 presidential election, were outcomes in Wisconsin and Michigan affected by vulnerable vote tabulation technologies? Russian hacking affected the campaign, U.S. intelligence services concluded — but did it affect the vote count? Candidate Donald Trump unexpectedly won both states. His margins over Hillary Clinton were very small: 0.76 percent (22,748 votes) in Wisconsin and 0.22 percent (10,702 votes) in Michigan.
This week, the Intercept leaked an NSA report about Russian “cyber espionage operations against a named U.S. company in August 2016, evidently to obtain information on elections-related software and hardware solutions.”
So were vote tabulations in Michigan or Wisconsin hacked?
[interstitial_link url=”https://www.washingtonpost.com/news/monkey-cage/wp/2016/11/28/new-evidence-finds-anomalies-in-wisconsin-vote-but-no-conclusive-evidence-of-fraud/?utm_term=.189c40a22720″]New evidence finds anomalies in Wisconsin vote, but no conclusive evidence of fraud[/interstitial_link]
Intelligence officials say there is “no evidence” that vote tabulations were hacked. But they haven’t examined the machines. Absence of evidence is not evidence of absence.
To get a better answer, we looked at data from recounts prompted by Jill Stein’s Green Party campaign. We find evidence that voting technologies in Michigan and Wisconsin treated candidates Trump and Clinton the same, at least in places where votes were recounted. Whether votes cast for Trump or Clinton were counted does not depend on which candidate a vote was for.
Presumably if there had been a hack to benefit or harm one candidate, the voting machines would have systematically under- or over-counted one candidate’s ballots more than the other. That didn’t happen.
How we did our research
Our analysis assumes that the recount, when done by hand, accurately counted the votes that had originally been marked on paper ballots. So what proportion of the recounted votes were originally marked on paper? That includes all the votes in Michigan, and many in Wisconsin. Some votes in Wisconsin were cast on a machine, by touching a screen. In Michigan, only some precincts’ votes were recounted; in all those cases workers recounted the ballots by hand. In Wisconsin, the whole state’s votes were recounted, but sometimes those recounts were done by hand and sometimes by machine. We found roughly similar results whether Wisconsin’s votes were recounted by hand or by machine, however.
[interstitial_link url=”https://www.washingtonpost.com/news/monkey-cage/wp/2017/06/05/its-time-to-bust-the-myth-most-trump-voters-were-not-working-class/?utm_term=.fc6044e39471″]It’s time to bust the myth: Most Trump voters were not working class[/interstitial_link]
For each Wisconsin ward or Michigan precinct, we analyzed whether the recount gave each candidate fewer votes, the same number of votes, or more votes than they’d gotten in the original count. For example, if one candidate lost votes while the other gained votes, we label that situation “lost-gained.”
Our analysis examines whether it matters which candidate had which outcome. For example, is the relationship between technologies and the frequency of “lost-gained” outcomes the same as the relationship between technologies and the frequency of which candidate lost or gained — i.e., “Trump-lost-Clinton-gained” or “Clinton-lost-Trump-gained”? In other words, did the recount by the machines systematically give either Trump or Clinton more or fewer votes than the original count did? If not, then the technologies treated the candidates similarly, and we can conclude technology did not affect the election outcome. If it did, then perhaps hacks helped one candidate more than the other.
We also checked for a variety of possible hacks. For example, was voting equipment from one vendor hacked, or only voting machines in one municipality? The election was so close that even a slight nudge might have changed the result.
The analysis gets technical because we need to account for features of localities where the technologies are located. As the figures show, different technologies are clustered in different parts of the states. In Wisconsin accessibility technologies are similarly dispersed. If you’re interested in technical details about the statistical tests, please do check our working paper.
Here’s what we found
In brief, we find no evidence that the voting technology favored one candidate more than the other.
Nor could we find any statistical differences correlated with accessibility technologies or with different voting technology vendors.
The tests uncovered nothing suspicious. That supports a conclusion that voting machines themselves were not hacked.
[interstitial_link url=”https://www.washingtonpost.com/news/monkey-cage/wp/2017/06/06/does-trump-have-the-power-to-block-comey-from-testifying-probably-not/?utm_term=.44ae54817441″]Does Trump have the power to block Comey from testifying? Probably not.[/interstitial_link]
We did run into a problem with the Wisconsin votes: We cannot be sure which technology was used to produce the record of each vote. We often do not know which votes are cast by hand on paper and which votes are cast using Direct Recording Electronic (DRE) technology. That makes it impossible to trace each vote from when and how it was cast to how it was counted — and therefore those votes are impossible to analyze precisely.
We ran into a different problem in Michigan: only a subset of precincts were recounted. Serious problems have been documented in Detroit that potentially could have changed the outcome in the state. Votes in Detroit were not recounted and are not part of our data.
For reasons we detail in the working paper, it’s not feasible to say anything reliable about the associations between voting technologies and the exact numbers of votes the candidates received.
Nonetheless, our analysis offers evidence that voting technology did not distort the votes in Wisconsin or Michigan. How a vote was treated appears not to have depended on which candidate the vote was for. If there was a hack, it appears not to have changed the results.
Walter R. Mebane, Jr. is a research associate at the Center for Political Studies, professor of political science and professor of statistics at the University of Michigan.
Matthew Bernhard is a PhD student in computer science at the University of Michigan.