Traditional polling models are dead

In recent times we have seen a considerable amount of shock results. Brexit, the election of Donald Trump and now a hung parliament after a predicted landslide to the Conservatives to name a few.

It appears something is amiss here – our democratic votes keep defying traditional polling. But one poll did predict the election outcome correctly, and that was an experimental poll conducted by YouGov.

Usually, pollsters gather a sample size of 2000 or so people who fairly represent the demographics of the country, they then work out changes in opinions and ways in which they plan to vote and calibrate this against previous election results and the likelihood of that demographic to turn up on voting day. This traditional method predicted a significant swing towards Conservatives in the recent election, and while the Conservatives did get an increase in their overall vote share, it failed to identify some interesting trends and changes in voting behaviour.

So what did YouGov’s experimental poll do differently?

Well, the short answer is they applied modern statistical big data techniques used by today’s best data analysts and data scientists to the electorate. It came out with the result that closely aligned with what happened when voters went to the polls – predicting a hung parliament.

The long answer is they interviewed over 50,000 people without worrying about how fairly they represented each demographic and used multilevel regression and post-stratification procedures to adjust the results.

“Our survey may have only included 70 to 100 people in Canterbury, but it still had many other similar people who are not in Canterbury. The trick was to identify the similarities,” says Andy Morris, the chief innovation officer at YouGov. By identifying clusters, they could fill in gaps in constituencies by comparing similarities from elsewhere.

This led the poll to not only cause a stir by rejecting the accepted hypothesis from traditional polling but proving them wrong as well. YouGov’s poll predicted the Conservatives would win 302 seats vs the 318 they won which was within their margin of error.

Let’s hope in the future advanced statistical methods are taken more seriously and prove to make up for the inaccuracies in our current polling methods during these changing times.

We would love to know your thoughts on this. What do you think has caused traditional polling to get results so wrong recently? Is the method used by YouGov here the best way to go about polling the electorate?