According to a new model, if
the U.S. presidential election were to have taken place on Oct. 28, former Vice
President Joe Biden would have an 88.3 percent chance of winning. That’s the
finding of a group of U.S. university researchers based on new research
published in SIAM Review .
This finding assumes that
Americans vote the way that they say they will in publicly available polling
data and that voters not accounted for in existing polling data will turn out
equally for both candidates.
How did the researchers –
from Northwestern University, UCLA, Augusta University and The Ohio State
University – arrive at their conclusion? By applying a modeling framework like
ones experts use to forecast the spread of infectious diseases (such as
COVID-19) to the high-stakes challenge of forecasting election outcomes.
“What we assume is
that, similar to how an infected person can cause – or influence – a
susceptible person to become infected with a virus, a Republican or Democratic
voter can influence an undecided voter,” said lead researcher Alexandria
Volkening, an NSF-Simons Fellow at Northwestern University, who co-authored the
study with UCLA Mathematics professor Mason Porter, Augusta University
Biostatistics and Data Science professor Daniel Linder, and Ohio State
University Biostatistics and Mathematics professor Grzegorz Rempala. Their 2020
forecasts are also done in collaboration with Volkening’s students Samuel
Chian, William He and Christopher Lee.
“I think we were all
initially surprised that a disease-transmission model could produce meaningful
forecasts of elections, but one of the benefits of mathematical modeling is
that you can apply similar methods to shed light on many different problems,”
she added.
The group’s election
forecasting model – which is based on “compartmental modeling” – was shown to
have a similar success rate to popular forecasters FiveThirtyEight and Sabato’s
Crystal Ball.
Researchers treated
Democratic and Republican voting inclinations as two possible kinds of “infections”
that can spread between states. Undecided, independent or minor-party voters
were considered “susceptible” individuals and infection was interpreted as adopting
Democratic or Republican opinions. ‘Recovery’ represented the turnover of
committed voters to undecided ones.
Unlike election
forecasts that combine polling data with other data, such as historical voting,
the economy and approval ratings, the researchers’ model uses only publicly
available polling data and treats all polls on equal footing. Transmission is
interpreted as opinion persuasion, influenced by campaigning, media coverage
and debates, and opinions spread both within and between states. Despite its
simplicity, the model performs surprisingly well, Volkening explained. For
example, it was as effective as popular analysts were at predicting (known as
“calling”) the 2012 and 2016 races for governors, senators and presidents in
the U.S. using historical polling data, she said.
"One important
limitation is that we assume all undecided individuals who are left at the end
of our simulated elections vote for minor-party candidates or turn out equally
for the Democratic and Republican candidates,” Volkening said. “If undecided
voters all vote in one direction or voter turnout is heavily partisan, it is
very possible for a trailing candidate to win.”
Though the paper is
being published in the midst of a global pandemic, UCLA’s Porter is quick to
point out that the idea to use a disease transmission model was made long
before the COVID-19 pandemic surfaced.
“When we first
discussed using this approach, it was on the heels of the 2016 election when
pollsters were predicting a Clinton win and of course that’s not what
happened,” said Porter, noting that the researchers speculated that something
was wrong with the forecasting models that were being applied, the polling data
itself, or the interpretations of forecast uncertainty.
“There are many tools
already available for compartmental modeling because people have studied
infectious diseases for quite some time with great success, so it made sense to
try a similar approach to study election forecasting,” he said.
The group’s model and U.S.
election forecasts are publicly available at https://modelingelectiondynamics.gitlab.io/2020-forecasts/index.html and
the researchers strongly encourage readers to try out their modeling framework
and build on it further.
To read the entire study, visit SIAM Review.