Demystifying the Election Prediction 2016 Contest
September 30, 2016
ASA is sponsoring the Prediction 2016 contest, giving students a chance to take part in this year’s election season by using statistics to predict the next president of the United States. Think you’re up for the challenge? We’ve compiled a few resources to help you get started.
Nate Silver, a well-known American statistician, offers some great tips on how to create good prediction models on his website, FiveThirtyEight. He suggests models be probabilistic and empirical while also responding sensibly to change in inputs and avoiding changes of rules in midstream. There is also a very helpful “User’s Guide” on FiveThirtyEight’s prediction process for the 2016 election that explains the four major steps Silver follows.
So where’s a good first stop on the journey to predicting the next U.S. president? Polls. Data need to be collected, weighted and averaged before any models can be created or any predictions can be made. There are several different places on the web where election polling data can be found, including: Gallup, Quinnipiac and RealClearPolitics, among others. It’s also important to adjust and combine polls with other data points and account for uncertainty when making predictions.
After polling data have been collected and adjusted, it’s time to choose which statistical methods will be used to predict the election. This is an opportunity to get creative, as there are several different methods that can be used. Biostatistician Bob O’Hara wrote an article speculating how Nate Silver correctly predicted the outcome of every state in the 2012 U.S. presidential election, suggesting Silver used Bayes’ Theorem, graph theory and hierarchical modeling to make his predictions.
Jo Hardin, professor of mathematics at Pomona College, offers a different possible approach in her post on R-statistics blog. Hardin provides an analysis of data that uses weighted means and a formula for the standard error of a weighted mean. She also suggests the possibility of using a generalized linear model.
Finally, it may be worth taking a look at the Time for Change forecasting model for some additional inspiration. This model estimates the weights of just three predictors and has correctly predicted every U.S. presidential election since 1988.
For additional reading on statistical approaches for election prediction, see the following:
Forsberg OJ, and Payton ME. (2015) “Analysis of Battleground State Presidential Polling Performances, 2004–2012,” Statistics and Public Policy, 2, 1-10. (Source)
Christensen WF, and Florence LW. (2008) “Predicting presidential and other multistage election outcomes using state-level pre-election polls,” The American Statistician, 62, 1-10. (Source)
Park, D. K., Gelman, A., and Bafumi, J. (2004), “Bayesian Multilevel Estimation with Poststratification: State-Level Estimates from National Polls,” Political Analysis, 12, 375–385. (Source)
Congratulations to all our 2017 data science and statistics graduates! Whether you have reached the end of your academics at the high school, college, or graduate level, this is truly something to celebrate. And when you study statistics, there is even more to celebrate! Your skills in this important emerging field will help you no…
We’re hitting graduation season! Best wishes to all the graduating students out there, whether it be high school or college, as you move on to your next big thing. As you look ahead, might we recommend pursuing an education (or career) in data science? As you already know if you’ve visited This is Statistics before,…