Polling bias and undecided voter allocations: US Presidential elections, 2004 - 2016

TL;DR

This paper investigates how different methods of allocating undecided voters in polls affect the accuracy of US presidential election predictions, highlighting the limitations of static approaches and proposing probabilistic alternatives.

Contribution

It demonstrates that static allocation rules are inadequate for election predictions and advocates for probabilistic methods, also analyzing polling biases across multiple elections.

Findings

High number of undecided voters in 2016 affected predictions

Static allocation methods can bias election forecasts

Probabilistic allocations may improve prediction accuracy

Abstract

Accounting for undecided and uncertain voters is a challenging issue for predicting election results from public opinion polls. Undecided voters typify the uncertainty of swing voters in polls but are often ignored or allocated to each candidate in a simple, deterministic manner. Historically this may have been adequate because the undecided were comparatively small enough to assume that they do not affect the relative proportions of the decided voters. However, in the presence of high numbers of undecided voters, these static rules may in fact bias election predictions from election poll authors and meta-poll analysts. In this paper, we examine the effect of undecided voters in the 2016 US presidential election to the previous three presidential elections. We show there were a relatively high number of undecided voters over the campaign and on election day, and that the allocation of undecided voters in this election was not consistent with two-party proportional (or even) allocations. We find evidence that static allocation regimes are inadequate for election prediction models and that probabilistic allocations may be superior. We also estimate the bias attributable to polling agencies, often referred to as "house effects".