No evidence of crisis in election polling, study finds
When polling was first introduced during the 1940s and 1950s, the mean error was 2.1%.
A new study has found that claims of a crisis in the accuracy of election polling are false and that there is no evidence to suggest that poll errors have increased over time.
The research, led by Professor Will Jennings at the University of Southampton and Professor Christopher Wlezien from the University of Texas at Austin, examined more than 30,000 national polls from 351 general elections in 45 countries during the period between 1942 and 2017.
The paper, titled Election polling errors across time and space, found that the average absolute error of pre-election polls has fluctuated over the years, but has not increased.
Our study shows the importance of testing polling accuracy over the long term and in cross-national perspective Prof Will Jennings
When polling was first introduced during the 1940s and 1950s, the mean error was 2.1%. This average was also consistent during the 1960s and 1970s and has been 2% since 2000.
Also, in countries where there is regular pre-election polling over a period of almost 40 years, the report says that polls have become more accurate, not less.
Prof Jennings said: “It is understandable that people tend to focus on high-profile polling misses, and shock election results, but the evidence dispels the claim that polling is any more inaccurate than at any point in history.
“Of course polling is a tricky business and has to adapt to changes in politics and wider society, and will inevitably go wrong from time to time. But it is important not to rush to judgment and generalise from individual cases.
“Our study shows the importance of testing polling accuracy over the long term and in cross-national perspective.”
The report, published in the journal Nature Human Behaviour, also claims that there is no significant trend of increasing polling inaccuracy in recent times. The researchers suggest that declining response rates and the growing variation in survey modes have had little effect on the performance of pre-election polls.
A University of Southampton spokesman said: “The research found that while all polls contain error, it tends to be greatest for parties or candidates receiving a larger share of the vote, as was the case in the 2015 and 2017 UK general elections and the 2016 US presidential election.
“Characteristics of political systems also appear to influence the accuracy of polling. Errors tend to be lower in proportional representation systems as well as presidential elections, particularly in the USA and France.”