LSA: Are there things Americans can do to support election security?
Walter Mebane: There’s the usual don’t believe everything you hear. One of the biggest challenges in the current election climate in the United States is misinformation. A lot of this is sourced by bad foreign actors and from so-called echo-chambers that are specifically trying to undermine people’s confidence or willingness to participate in the election. It’s well-established that if people think the election is going to be problematic then voter turnout declines. Why bother to participate if you know your vote won’t count?
Most people can’t do much about election security, but they can check their own registration and report if they discover some kind of problem. You can also look for ways to minimize crowding in polling places and vote early. Try not to get irrationally excited. Be proactive and well-informed.
In addition to being better informed and careful about where you get info, check your ballot. If you’re voting with a pen, double- and triple-check your ballot to make sure you’re not making a mistake. From personal experience, it’s pretty easy to do.
LSA: Your research focuses on election forensics, which uses statistical techniques to identify patterns and the likelihood that such patterns occur by chance. For those of us who are not statisticians, can you explain how you do this?
WM: Since last year, I’ve been working on a model to estimate the number of fraudulent votes in elections. In Russia’s elections or Turkey’s elections or in other countries where most people agree there were big electoral problems, the model can show millions of fraudulent votes. The patterns that you find from the model’s different indicators match stories about those elections.
But the problem is that the method also comes up with a lot of false positives. For example, in the United States, we can lose votes because of voter suppression or because people decide not to vote in a congressional election that’s not going to be close. These kinds of things frequently occur in politics. They are issues, but they’re not usually fraudulent.
There are a variety of models where people expect to see a particular pattern in the data. If they see a different pattern, they read it as a signal that something went wrong. There are many methods that find all kinds of anomalies in the pattern of voter turnout, or the vote count for a candidate or parties. The problem is interpreting why the anomalies occur.
We don’t have gold standards for recognizing patterns, so it’s a challenging measurement task. One method we’re using is simulating elections. When we simulate an election, we know the frauds, the strategic behavior, the election rules, and we can generate the administrative failures. We can then use machine learning to compare the simulated elections to real elections and evaluate all the anomalies and find patterns within them to parse out what is actual fraud and what is an administrative issue or voter suppression or strategic voting.
LSA: Despite our best efforts, elections sometimes have errors. How can you distinguish honest mistakes from intentional electoral malfeasance?
WM: This is the frontier in statistical analysis. It can sometimes be hard to separate fraud from accidents or administrative mistakes because someone wanting to commit fraud can make it look like an accident. Recently in New York, there was a big mechanical problem that sent out ballots with a major error. Administrative problems are a particular risk because the scale of absentee voting this election could overwhelm places that already had shaky procedures.
There are a variety of modeling techniques that allow you to have a bunch of simulated potential voters that each make decisions that can be written as code. They can vote based on lots of information or misinformation, they can be intimidated, they can vote confidently, they can wait in a long line. You can get simulated votes, simulated precincts, and then you can apply all these models to the simulated data. I’ve based our simulations on things I’ve seen in elections around the world. If we generate a million different election configurations, it’s not as vast as the possibilities in the world but I think it is enough to learn from. I bet we’ll know whether there are or aren’t distinctive patterns that one can see across the diverse methods people have deployed to discriminate intentional electoral malfeasance from honest mistakes. See me in a year or two about whether that worked or not.