What Do We Know at 7 PM on Election Night? Read About Nate Silver and Andrew Gelman's Election Forecast Model
This article was published in the October 2010 issue of Mathematics Magazine.
On the evening of November 8, 1988, I (Andrew Gelman) was working with my colleague Gary King in his Harvard office. It was election night, and Massachusetts Governor Michael Dukakis was a candidate for President against Vice President George H. W. Bush. Gary somehow had gotten his hands on a pair of tickets to Dukakis’s victory party in Boston, and we were trying to decide whether to go. Dukakis was expected to lose, but—who could say, right? We had the TV on, and the first state to report, at 7 PM, was Kentucky, which Bush won by over 10 points. Gary informed me that the election was over: Kentucky, at the time, was near the political center of America, and there was no way that Dukakis would do much better nationally than he did in Kentucky. So we saved ourselves a subway ride and kept on working.
What about the most recent election night—November 4, 2008, when the candidates were Senators Barack Obama and John McCain? Was it possible for a viewer to play along at home with the election and decide at 7 PM, or 8 PM, or 9 PM what the outcome would be? On the night before the election we (AG and NS) analyzed a probabilistic election forecast to make some guesses at what might be known at different times during the evening. This article is a report of that analysis. Most of the results presented were obtained on November 3, the night before the election, and were intended as a guide for interpreting what we would hear the next night, as the votes were counted.
We performed one set of calculations using statewide vote margins; that is, reports like we remembered from Kentucky in 1988; and another set of calculations using only the tally of states won or lost, without the margins of victory.
In the next sections we describe the model we used, and then present our results from November 3. We then discuss what actually happened and ways in which election-night reporting could be improved in the future.
Our model on the night before the election
If one wants to make probabilistic forecasts, one needs a model. We used the model developed by one of us (Nate Silver) to make election forecasts at the website fivethirtyeight.blogs.nytimes.com. It is described at that site [6] but it is too rich a model for us to describe in detail here. We will describe it in general terms, and discuss the tools one might use to develop a similar model of one’s own.For subscribers, read the full article online (Requires MAA Membership)
Not a member? Join the MAA today!