Donald Trump’s widely unexpected victory in the 2016 U.S. presidential election has raised questions about the accuracy of public opinion polling, the aggregation of polling into probabilistic election forecasts and the interpretation of election polling by data analysts, journalists and the general public. While national-level polls on average proved as accurate as in past elections in predicting the popular vote (with an average error on the margin of about 2 points), there were substantial polling errors at the state level, particularly in Midwestern swing states (Eten 2016; Silver 2017; Cohn, Katz and Quealy 2016).
These significant misses, amplified by a proliferation of overconfident probabilistic forecasts, have (fairly or not) placed a cloud over the polling industry, provided ammunition to critics of survey research and led the leading industry association to study how the 2016 misses can be avoided in the future. This paper demonstrates the advantages of a statistical approach developed over the past two decades, multilevel regression with poststratification (MRP), in improving survey estimates in general and, specifically, in producing superior forecasts of election outcomes. CONT. – pdf
Chad P. Kiewiet de Jonge & Gary Langer, Langer Research Assoc.