By around 10 p.m. Eastern time on Nov. 8 it was clear that the presidential election results weren’t playing out as expected. That first became apparent when the states most poll watchers expected to be easy wins by Democratic candidate Hillary Clinton had suddenly become squeakers.
Margins everywhere became very narrow and it wasn’t long before states regarded as solid for Clinton were being called for Republican Donald J. Trump.
Even though a few western states had yet to close their polls, it already was clear that the chances of a Clinton victory were steadily fading. By the time I gave up and went to bed about 1 a.m. Nov. 9, I knew it was only a matter of time before Trump was declared the winner in enough states to get the 270 electoral votes required to become the next U.S. president.
By that time of night, television news reporters and their pollsters were openly kicking themselves for having missed the now obvious strength of Trump voters. The question had already become how they could have missed all of that.
The answer is that the opinion-polling community was still following the data-gathering and -analysis practices that it had for decades. But the world had changed to the point that those practices no longer were completely relevant. When I wrote my Oct. 31 column about how social media analysis was showing a much closer race than predicted by the most recent polls, it was clear that traditional polling methods were not tracking public opinion correctly.
But until Election Day, it wasn’t clear just how far off the polling was. Now that it’s clear, it’s important to understand the pitfalls of traditional polling and find ways to correct the means we use to measure public sentiment. But I also need to mention that social media analysis is not a magic bullet, either. It has its own set of flaws.
In the case of election polling, there’s always been a challenge in finding the right sample and making sure the sample is the right size. A sample that’s not chosen properly will yield incorrect results: A sample that’s too small has a margin of error that’s too large to be useful, and a sample that’s too large can take too long to process or is too expensive to field—or both.
There are other factors that affect polling results once the results have been gathered. In the case of political pollsters, results are weighted to reflect how likely respondents are to actually vote, for example. One such assumption frequently used is that voters in rural areas are less likely to go to the polls, so their responses receive less weight.
Likewise, people in certain demographic categories are assumed to be more or less likely to vote, depending on the specific demographic. Finally, respondents are assumed to be telling the truth about personal decisions such as voting intent.
Conspiring against the previous responsiveness of voters being polled is the dramatic increase in unwanted marketing calls, fraudulent calls attempting to scam people, calls for donations and political calls. People are getting wary of talking on the phone, and have repeatedly asked the Federal Communications Commission and the Federal Trade Commission for relief.