Mon, Jun 12, 2017 - Page 7 News List

The outlier pollster that got the British election right

An LSE-developed approach formulates a prediction by employing big data analytics to evaluate an opinion survey, taking into account a constituency’s demographics and past voting patterns

By Leonid Bershidsky  /  Bloomberg View

Illustration: Yusha

Once again, well-informed people, including British politicians, have been surprised by an election result because they put too much trust in traditional polls. It is time to get smarter about assessing voter preferences. The tools are there for those who want them.

On the eve of Thursday’s election, British newspapers and many outside observers expected a landslide victory for the Conservative Party. Even a 100-seat majority in the 650-seat parliament was still being discussed. That is because both polls and prediction markets signaled such an outcome.

The Financial Times’ polling average showed that British Prime Minister Theresa May’s Conservatives would win 44 percent of the vote to 36 percent for Jeremy Corbyn’s opposition Labour Party. On Wednesday, Betfair had the Conservatives at odds of 2-to-9 to win a majority and no majority at 5-to-1. On the surface, there was every reason to expect success for May’s plan to get a strong mandate for her party.

That is unless you followed the YouGov election model developed by Benjamin Lauderdale of the London School of Economics (LSE). An outlier compared to traditional polls, it consistently predicted the actual outcome — a hung parliament, in which no party would have a majority.

The model is a hybrid of an opinion survey and big-data exercise. It uses a small number of responses to survey questions to figure out how voting is related to individual characteristics of people living in each constituency, and then projects the result based on that constituency’s demographic breakdown and past voting record.

This approach — the scientific name is multi-level regression and post-stratification — is a relatively cheap way of matching the survey methodology to the UK’s electoral system, in which politicians contest individual constituencies. The big data part of the model allows pollsters to avoid the cost of surveying large samples in each constituency.

Most of the other UK polls assessed the national vote breakdown — the share of the vote each of the parties was expected to get — by using nationwide samples. They would have been relevant — assuming the samples were representative — in, say, a Dutch parliamentary election, where parties compete in a single nationwide constituency.

National polling can also work in Germany, where every citizen gets two votes — one for a specific candidate in an electoral district and one in a Dutch-style nationwide proportional system; the two votes are usually well correlated. It can also work in a French presidential election, where the nation is a single constituency, too. That is why respectable Dutch, French and German polls using nationwide samples are worth following.

In the US, where the successful presidential candidate needs to win the most electoral votes, outcomes in individual states matter more than the results of the nationwide popular vote. Just as polls based on national samples predicted, failed Democratic US presidential candidate Hillary Rodham Clinton won the popular vote last year.

The controversial University of Southern California and Los Angeles Times poll, which used a sample weighted differently from most others, predicted then-Repulican US presidential candidate Donald Trump would win — but it was wrong, too. Yet Trump won the presidency.

The US needs more complex models than these nationwide polls so it can better gauge local voting patterns. However, these are rare so far. The polling firm Morning Consult applied a YouGov-style approach in April last year and found that Clinton was the strongest candidate at that time.

Comments will be moderated. Keep comments relevant to the article. Remarks containing abusive and obscene language, personal attacks of any kind or promotion will be removed and the user banned. Final decision will be at the discretion of the Taipei Times.

TOP top