# Bias and Excess Variance in Election Polling: A Not-So-Hidden Markov Model

**Authors:** Graham Tierney, Alexander Volfovsky  
**arXiv:** [2206.14570](https://arxiv.org/abs/2206.14570)  
**Date:** 2022-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.14570v2/

## Key points

- The study explores the measurement of polling errors in US Presidential and Senate elections.
- The authors highlight the increased interest in measuring polling errors due to historic misses in the 2016 and 2020 US Presidential elections.
- The most common method for measuring directional errors and non-sampling excess variability is by comparing poll results with election results conducted within a few days of the election.
- Typical models are highly sensitive to the time between the poll and the election, making analysis challenging.
- Tierney and Volfovsky propose leveraging hidden Markov models traditionally used for election forecasting to address this issue.
- Their approach allows for flexible capturing of time-varying preferences and treats the election result as a peak in the hidden Markovian process.
- Results using data from polls conducted during US Presidential elections from 2004 to 2020 and US Senate elections from 1992 to 2020 show that previous estimates of bias were too extreme by 10% in Presidential elections and by 25% in Senatorial elections.
- Previous estimates of excess variability were also found to be too large.
- Utilizing hidden Markov models provides less sensitivity to the choice of time window, avoids conflating shifting preferences with polling error, and offers more interpretable results despite using a highly flexible model.
- This research contributes to improving our understanding of polling errors in elections.

## Summary

The study titled "Bias and Excess Variance in Election Polling: A Not-So-Hidden Markov Model" by Graham Tierney and Alexander Volfovsky explores the measurement of polling errors in US Presidential and Senate elections. The authors highlight the increased interest in measuring polling errors due to historic misses in the 2016 and 2020 US Presidential elections. They note that the most common method for measuring directional errors and non-sampling excess variability is by comparing poll results with election results conducted within a few days of the election. However, analyzing such polling error data has proven to be challenging, as typical models are highly sensitive to the time between the poll and the election. To address this issue, Tierney and Volfovsky propose leveraging hidden Markov models traditionally used for election forecasting. These models allow for flexible capturing of time-varying preferences and treat the election result as a peak in the hidden Markovian process. The authors demonstrate their approach using data from polls conducted during US Presidential elections from 2004 to 2020 and US Senate elections from 1992 to 2020. Their results show that previous estimates of bias in Presidential elections were too extreme by 10%, while estimates of bias in Senatorial elections were too extreme by 25%. Additionally, they find that previous estimates of excess variability were also too large. By utilizing hidden Markov models, Tierney and Volfovsky's approach provides less sensitivity to the choice of time window, avoids conflating shifting preferences with polling error, and offers more interpretable results despite using a highly flexible model. This research contributes to improving our understanding of polling errors in elections and provides valuable insights for future analysis.