|
Abstract:
|
National tracking polls and state-specific polls have long been a popular method of forecasting national voting intention for elections in the United States. The 2016 US Presidential election introduced the widespread use of high-volume national tracking polls, which attempted to infer voting intention estimates at the state level. Such tracking polls relied on rim or post-stratification and achieved poor results. Thus indicating that corrective methods, which pollsters typically rely on, would not alone yield accurate estimates. This paper presents multilevel regression with poststratification (MRP) as a tool for reducing such error. We examine data from the Reuters and Google mega-polls to compare different techniques for adjusting survey data. Using the Stan language we are able to implement a fully Bayesian regression models and compare the results to the estimates produced by traditional methods. As expected, we find that MRP produces significant accuracy gains at the state level over traditional methods. Additionally, we explore ways of integrating external information into the prior to enhance MRP; something that is not possible through empirical Bayes MRP.
|
