Evan Cohen is a Principal at The Brattle Group where he leads case teams and supports government, private clients and expert witnesses in complex business litigation. He has an MBA from MIT and a BA in economics from Cornell University.
One of his interests is in economic and statistical modeling of elections. On Tuesday, October 25th, he addressed the school on how statistical methods are used in predicting elections. He began by talking about the constitutional history of American elections beginning with the 12th amendment which changed the Vice President and President to be elected as a ticket. He then explained the mechanics of the electoral college and how the size of the electoral college is calculated.
He then shifted to the focus of his talk - on how statisticians use various methods to construct polls and evaluate polls to overcome the margin of error implicit in any poll. He described what Nate Silver had done in 2008 in building a model methodology that predicted the national results correctly in 49 of 50 states. He described how there are a number of sites similar to Silver’s approach that were essentially weighted polls of polls. These various sites all differed in terms of the fine points of their methodology but utilized similar forecasting techniques.
The regional and state poll results are weighted through historical and other correlations to describe likely voting patterns. As an example, he mentioned that if Trump were to win the state of Michigan, that based on past voting patterns, he would have a 99% chance of winning in Ohio as well. After constructing these correlations, the models are then subjected to a large number of Monte Carlo simulations, essentially “what if” scenarios based on likely statistical chance. These results are then tabulated to consider the likelihood of a particular candidate winning. So a result that reports that Candidate X has a 70% chance of winning the election means that 70% of the Monte Carlo simulations favored Candidate X.
He said based on the poll data now available Hillary Clinton had somewhere from a 85-99% chance of winning the election. The difference between those two numbers depended entirely upon the model employed and the difference that different modelers saw in the underlying historical correlations in data and accuracy of polls.
Perhaps surprisingly, he said that this election was the most stable in the last 65 years. He pointed to a heavily divided electorate as the primary reason for this phenomenon. He said that since Trump had no path to victory without Ohio and Florida (traditional swing states), that the interesting states to watch would be Pennsylvania, Colorado and NH. He closed by mentioning the importance this year of NH in both the Presidential and Senatorial races.