Online Program

A Poll-Based Bayesian Hierarchical Model for American Presidential Election (309991)

Keywords: Bayesian modelling, political polling

In this presentation, a Bayesian Hierarchical polling aggregation model for American Presidential Elections is presented. It considers the support for the Democratic and Republican candidates relative to each other. It has multiple variations that control the uncertainty, the weight of the polling data, and the number of polls to include. It is not a forecast and attempts to aggregate the polls at various points in election cycle. At the end of an election cycle it can be used to predict the outcome. It uses previous elections results and poll data from other states to create partially pooled estimates of support for a candidate even when no polling data exists in a state. This model is computationally efficient and can be run on a laptop in minutes. This model was used to predict the 2020 election, and was tested on data from the 2008, 2012, and 2016 elections. This method outperforms a polling average and provides reasonable estimates of uncertainty in each state. It also produces highly similar estimates to the FiveThirtyEight and Economist model while not taking hours to run. This method could also be adapted to predict Senate or House races.