To Come

In the aftermath of the 2016 general election polling miss (mostly at the state level) there are more questions than ever about polling accuracy, quality, and methodology. In this talk I’ll use the 2016 pre-election polls from our HuffPost Pollster database to investigate how polling in the aggregate has been affected by poll proliferation, and in particular, how accuracy of our estimates was affected by different types of polls. I will also look at whether choices about excluding polls from the aggregate affected results. HuffPost Pollster uses model-based approaches to aggregating polls, and for our general election polls we use a Bayesian time series model. This was the baseline for our forecast model. I will discuss the advantages and disadvantages of using this type of model, as well as the parameters that can be adjusted to make it more or less responsive to individual polls. I will also address the decisions we made about state correlations, national polls' influence, and whether our criteria for including polls affected the model.