US election polls could show Hillary Clinton win over Donald Trump because of ‘herding’, not because they are right

Pollsters have a tendency to disavow polls that sit outside of the mainstream consensus – but sometimes they can be right

Andrew Griffin
New York
Monday 07 November 2016 16:03 GMT
Trump v Clinton: US Election forecast - November 7

Most election polls are showing Hillary Clinton ahead. And that might be a worry for her campaign.

The clear consensus that has emerged around a win for Hillary Clinton, with polls predicting that she will win the popular vote by three to four percentage points. That consensus tends to be held up as evidence of the fact that such a result is likely – but it might not be.

Almost every major poll is showing Ms Clinton in the lead, usually by around two to four percentage points. They have appeared to come together in what appears to be a consensus in recent weeks.

But some notable polls are major outliers, in particular the LA Times poll that showed Trump winning by five percentage points. That has been partially disavowed by people including the paper itself, but is also a highly respected poll.

The similar results might not be a consequence of the fact that different data points are showing the same result. Instead, pollsters might be skewing those data points to take cues from other polls – a phenomenon called “herding”.

The idea of herding is that because pollsters want to make sure that their findings are correct, and so tend to skew their results so that they reflect the prevailing mood. That can mean that they will avoid producing unusual results, potentially making their findings less accurate but also allowing them to blend in with the other results.

Nate Silver of 538 has suggested that the effect is probably at play in this election, where almost all polls are showing very similar results. It would make particular sense this year, since many previous assumptions have already been proven incorrect and trust in polling data is low.

That has led to wrong consensus in the past. In an Iowa election in 2014, polls consistently showed the winner taking the lead by only about four points – but they ended up with a lead of twice that, with some outlying polls that suggested that result being ignored because they didn’t fit with the herd.

A study from Vanderbilt University in 2012 that looked at polls taken in advance of that year’s New Hampshire primary showed that polls tended to “take cues” from one another because of the huge stakes involved in getting their predictions right.

“If so, reported polls should not be assumed to be independent of one another and so-called ‘poll-of-polls” will therefore be misleadingly precise,” authors Joshua D Clinton and Steven Rogers wrote in their paper.

Join our commenting forum

Join thought-provoking conversations, follow other Independent readers and see their replies


Thank you for registering

Please refresh the page or navigate to another page on the site to be automatically logged inPlease refresh your browser to be logged in