Problems with quantitative polling, and answers from qualitative data

The results of the US elections this week show a surprising trend: modern quantitative polling keeps failing to predict the outcome of major elections. In the UK this is nothing new, in both the 2015 general election and the EU referendum polling failed to predict the outcome

Problems with quantitative polling, and answers from qualitative data

The results of the US elections this week show a surprising trend: modern quantitative polling keeps failing to predict the outcome of major elections.

In the UK this is nothing new, in both the 2015 general election and the EU referendum polling failed to predict the outcome. In 2015 the polls suggested very close levels of support for Labour and the Conservative party but on the night the Conservatives won a significant majority. Secondly, the polls for the Referendum on leaving the EU indicated there was a slight preference for Remain, when voters actually voted to Leave by a narrow margin. We now have a similar situation in the States, where despite polling ahead of Donald Trump, Hillary Clinton lost the Electoral College system (while winning a slight majority in the popular vote). There are also recent examples of polling errors in Israel, Greece and the Scottish Independence Referendum.

Now, it’s fair to say that most of these polls were within the margin of error, (typically 3%) and so you would expect these inaccurate outcomes to happen periodically. However, there seems to be a systematic bias here, in each time underestimating the support for more conservative attitudes. There is much hand-wrangling about this in the press, see for example this declaration of failure from the New York Times. The suggestion that journalists and traditional media outlets are out of touch with most of the population may be true, but does not explain the  polling discrepancies.

There are many methodological problems: numbers of people responding to telephone surveys is falling, perhaps not surprising considering the proliferation of nuisance calls in the UK. But this remains for most pollsters a vital way to get access to the largest group of voters: older people. In contrast, previous attempts to predict elections through social media and big data approaches have been fairly inaccurate, and likely will remain that way if social media continues to be dominated by the young.

However, I think there is another problem here: pollsters are not asking the right questions. Look how terribly worded the exit poll questions are, they try to get people to put themselves in a box as quickly as possible: demographically, religiously, and politically. Then they ask a series of binary questions like “Should illegal immigrants working in the U.S. should be offered legal status or deported to their home country?” giving no opportunity for nuance. The aim is clear – just get to a neat quantifiable output that matches a candidate or hot topic.

There’s another question which I think in all it’s many iterations is poorly worded: Who are you going to vote for? People might change whether they would support a different politician at any moment in time (including in a polling booth), but are unlikely to suddenly decide their family is not important to them. It’s often been shown that support for a candidate is not a reliable metric: people give you answers influenced by the media, the resdearcher and of course they can change their mind. But when you ask people questions about their beliefs, not a specific candidate, they tend to be much more accurate. It also does not always correlate that a person will believe a candidate is good, and vote for them. As we saw in Brexit, and possibly with the last US election, many people want to register a protest vote – they are not being heard or represented well, and people aren’t usually asked if this is one of the reasons they vote. It’s also very important to consider that people are often strategic voters, and are themselves influenced by the polls which are splashed everywhere. The polls have become a constant daily race of who’s ahead, possibly increasing voter fatigue and leading to complacency for supporters of who ever is ahead the day of the finishing line. These made future predictions much more difficult.


In contrast, here’s two examples of qualitative focus group data on the US election. The first is a very heavily moderated CBS group, which got very aggressive. Here, although there is a strong attempt to ask for one word answers on candidates, what comes out is a general distrust of the whole political system. This is also reflected in the Lord Ashcroft focus groups in different American states, which also include interviews with local journalists and party leaders. When people are not asked specific policy or candidate based questions, there is surprising  agreement: everyone is sick of the political system and the election process.


This qualitative data is really no more subjective than polls based on who answers a phone on a particular day, but provides a level of nuance lacking in the quantitative polls and mainstream debate, which helps explain why people are voting different ways – something many are still baffled by. There are problems with this type of data as well, it is difficult to accurately summarise and report on, and rarely are complete transcripts available for scrutiny. But if you want to better gauge the mood of a nation, discussion around the water-cooler or down the pub can be a lot more illuminating, especially when as a researcher or ethnographer you are able to get out of the way and listen (as you should when collecting qualitative data in focus groups).

Political data doesn’t have to be focus group driven either – these group discussions are done because they are cheap, but qualitative semi-structured interviews can really let you understand key individuals that might help explain larger trends. We did this before the 2015 general election, and the results clearly predicted and explained the collapse in support for the Labour party in Scotland.

There has been a rush in the polling to add more and more numbers to the surveys, with many reaching tens or even hundreds of thousands of respondents. But these give a very limited view of voter opinions, and as we’ve seen above can be very skewed by question and sampling method. It feels to me that deep qualitative conversations with a much smaller number of people from across the country would be a better way of gauging the social and political climate. And it’s important to make sure that participants have the power to set the agenda, because pollsters don’t always know what issues matter most to people. And for qualitative researchers and pollsters alike: if the right questions don’t get asked, you won’t get the right answers!

Don't forget to try Quirkos, the simplest and most visual way to analyse your qualitative text and mixed method data. We work for you, with a free trial and training materials, licences that don't expire and expert researcher support. Download Quirkos and try for your self!