This past weekend, I launched my 2016 presidential forecast. It’s a statistical model that attempts to predict who will become the next President of the United States – Democratic nominee Hillary Clinton or Republican nominee Donald Trump.
The model makes predictions based on an analysis and weighting of hundreds of state and national-level polls that are adjusted along several dimensions. For example, poll results are adjusted according to whether each poll was conducted using likely voters, registered voters, or all adults. Polls are also adjusted according to an estimate of each polling firm’s house effect, which is defined as the tendency for a specific polling firm to show bias toward one or another candidate relative to other polls in the same state. And although polling data constitutes the bulk of the model’s basis for making predictions, other information is also incorporated, such as religious and racial/ethnic demographics in each state and the prior political leaning of each state.
If you’re a fan and frequent reader of FiveThirtyEight – the data-driven website run by Nate Silver – then this sort of thing should be pretty familiar to you. In fact, my model is broadly similar to FiveThirtyEight’s “polls-only” model, with a few relatively minor exceptions.
Not surprisingly then, my model’s predictions roughly track those of FiveThirtyEight and other statistical models, such as the one operated by the New York Times blog, The Upshot.
For instance, as of today (September 28th) my model estimates that Hillary Clinton has a 64% chance of becoming the 45th President of the United States. That’s a bit more generous than the current FiveThirtyEight estimate (56%) but a bit more conservative than the current New York Times estimate (70%).
Nonetheless, all the models agree that Clinton’s chances of winning in November have declined markedly over the past month or so. She’s down in my estimates from a high of around 85% in early August. She’s also down from a high of 89% in FiveThirtyEight’s model and from 90% in the New York Times’s model.
Clinton’s decline throughout September is largely due to several weeks-worth of bad poll numbers following her campaign’s botched handling of her recent pneumonia diagnosis.
You can see this in the the figure below, which shows movement in the polls for both Clinton and Trump.
Movement is defined here as the difference in poll results between successive editions of the same poll within a given state. So, for instance, if Public Policy Polling had Clinton at 43% in North Carolina in August and at only 40% today, then that would constitute a movement of -3 points for Clinton from August to present.
Even though the data are quite noisy, you can see that, on average, Clinton was moving up in the polls until around 87 days before the election, which was August 13. After that point, the rate at which she was moving up in the polls started to decrease until finally she started losing ground around the time of September 11th, which not coincidentally is the day she was filmed nearly collapsing after abruptly leaving a ceremony in New York commemorating the 15th anniversary of the 9/11 attacks.
But with the first presidential debate now behind us, the question of course is where things will go from here. Clinton is widely regarded as having won the first debate, but did she do enough to put the breaks on her falling poll numbers? Could the first debate possibly even turn things around? Then again, are her falling poll numbers even all that meaningful? Perhaps her recent drop in the polls merely reflects a sort of reversion to the mean.
Only time will tell. But we should begin getting our first hint of the answer to these questions soon, as high quality post-debate poll results begin coming in toward the end of this week and this weekend.
Brian Kurilla is a psychological scientist with a Ph.D. in cognitive psychology. You can follow Brian on Twitter @briankurilla