Public political polling has been dying for the last 10 years. The end of the landline telephone era and the rise of the Internet have combined to turn it from a reasonably accurate public service, into a farce existing solely to draw clicks.
Polling averages served to take the polls, which were increasingly inaccurate, and piece together some sort of coherent message from them. But now the polls are so wrong, even the polling averages have outlived their usefulness.
New York last night was a great example of this. Trump’s 59.8% of the statewide vote (number accurate as of this writing, via Decision Desk) last night didn’t just differ from all of the polls taken in the last week. It was outside of the margin of error of nearly all of the them. Meaning we had a series of improbable statistical anomalies, or they were just garbage polls.
Note that the Margin of Error is an estimate of the element of randomness in the sample. Half the time the actual result should be above poll, and half the time it should be below, with the result within the Margin of Error 19 times out of 20, on a bell curve near the actual result. So how did the pollsters do, mathematically?
- Emerson College? Far below, and barely within the margin of error at 55, with Margin of Error 5.1.
- CBS News/YouGov? Also far below, and just at the margin of error with Trump 54, MoE 5.9.
- 0ptimus? Wrong. Not even close, at Trump 49, MoE 1.
- NBC 4/WSJ/Marist? Wrong. Trump 54, MoE 4.5 doesn’t make the cut.
- Public Policy Polling? Wrong. Trump 51, MoE 4.5.
- Siena? Wrong. Trump 50, MoE 5.
- Quinnipiac? Wrong. Trump 55, MoE 4.2.
- Gravis got close, at Trump 57, MoE 4.5. That’s the only reasonable result of the eight last polls.
Remember what I said above? Mathematically, we expect half of the polls to be above the actual result, and half of them to be below. That’s what the Margin of Error predicts. So the fact that we had three of eight polls getting the result within the margin of error (by less than half a point in two cases), doesn’t actually exhonerate the polls. At least one of the polls should have had Trump over 60, and preferably over 62. None of them did, therefore, as a group, all of the polling failed. That’s not opinion. That’s math.
Some may say that primaries are hard to poll, but these failures have gone just beyond the New York primary. In the UK general election last year, every single poll undershot the Conservative victory. At 36.8%, with a 6.3 point lead over Labour, half the polls should have ranged from 37 to 43. None of them even made it to 37, and none of them got close to that 6.3 point lead. Of the 11 final polls, 7 even got the final result wrong, picking a Labour win or a tie, and every single poll-based forecast projected a hung parliament instead of the decisive Conservative win.
Then it went wrong again in Canada. The Liberals hit 39.5% nationally, and only one poll managed to get there at 40. Every other poll was below, instead of half the polls being above, the way the math says they should.
Public polling is dead. It’s wrong every time now, when it counts. There are systemic errors, failing to predict the results, in some cases even failing to predict the winners, as in the UK or Iowa. And they’re all wrong in the same direction, suggesting herding and faked polls. They say it’s because of ‘late breaking shifts’, but do we really think it’s plausible that happens every single time? Does anyone think Donald Trump had a late breaking shift his way?
Public polling is dead. It’s time to accept that.