Social Science

Polling is getting more accurate, despite new challenges 6-17-17

Source: "Democracy's whipping boys"  

Three major polling failures in the last three years: first, British pollsters preducted a "remain" victory in June of 2016. Then, in November of 2016, some American pollsters showed Clinton with a 99% chance of victory. Finally, in June of 2018, the Conservatives were predicted to maintain control of Parliament. What happened? If we look at long-term results (see chart), polls have been getting better and better each year. 

Polls ImprovingPoll aggregation or averaging allowed Nate Silver to give Trump a 1-in-3 chance of winning in 2016. And it was a close election - change 78,000 votes in PA, MI and WI and Clinton wins the Electoral College. Her 2.1% win in the popular vote was within 1% of the outcome predicted by most polls. For the 2017 Parliament result, a witch of just 75 voters from Labour to Conservative in the districts with the narrowest Labour margins would have allowed the Conservatives to maintain their majority. As for Brexit, there were polls showing a toss-up but they were discounted by the media.

Current challenges: first, no one answers their phones. 72% of those called by phone in 1980 agreed to take part in a poll versus 8% today. Next, it's hard to get a representative sample. Some demographic groups are far more reluctant to give their opinion. It takes about 20 calls to find an elderly white woman who will participatre versus almost 350 calls to find a young Latino male. Online pollsters like YouGov assemble large, stable panels of each major demographic and ten weight the results based on how large or small that group is in the overall electorate. Weighting is also used by telephone pollsters to reduce non-response bias (that is, those groups with high non-response bias are given higher weights than those who readily participate). But compicated weighing schemes allow pollsters to adjust results to reduce extreme results. This can result in "herding" behavior. Finally, polls have a hard time accounting for voter turnout. Groups with traditionally ow turnout like uneducated whites in the US or the youth in the UK can have their preferences doscounted because of the low turnout. When their turnout is actually high, the polls can miss the mark.