I want to talk for a moment about polling.
Political polling is terrible. It's been terrible this whole election cycle, as a brief look at
Real Clear Politics can tell you, but the fascinating part is
why it's terrible.
I can't claim to have a fancy degree in political science or anything -though I would like one- but what I have is actual experience on the "front lines" of polling. I was an employee for a polling outfit that did contract work for bigger polling houses during the 2010 Midterm election. Basically we did push polls for one party in one state, flipped over and did one for the opposite party in the other state.
("Let me ask you a question: Did you know that Candidate A is the greatest thing to happen to American democracy since Thomas Jefferson? And did you know that Candidate B has carnal knowledge of a horse? Now, if the election were held today, who would you vote for?" Repeat ad nausem.)
The actual mechanics of working there were simple: get two complete polls an hour, or go home. Since I had big plans to pay the light bill, I stayed on the phones and called people who had the bad luck to have been previously registered voters in an area where someone wants a poll. The problem is that no one really wants to stay on the phone for political polls. I can't tell you how many times I called someone on Sunday, only to
finally have someone pick up the phone, then answers "Excuse me sir, were you aware that this is
the Lord's Day?", followed by a click. A lot of people who were comfortable answering the political questions hared out when I got to basic demographics like age, color, religion and so on that accompanied every poll, and
then hung up. This got so frequent, I kept a bag of dice from my D&D game on my desk and just rolled dice to just finish out the survey when they hung up.
I needed two polls an hour. I
got two polls an hour.
Thankfully, unlike a lot of people at that place, I didn't just make up polls. I did the calls, I made a sincere effort. But it's hard to keep people on the line, and harder still to get people to actually answer a landline phone around dinner time. When that happened, the first question was usually "On a scale of one to ten, how likely are you to vote in the upcoming election?" The answer you're supposed to get is "Seven or above". Frequently, the only person who picked up the phone in the past hour answered "Four". I marked it down as a "Seven" and continued on.
You have not been disheartened until you have to explain to a twenty-three year old over the phone what the difference between a Democrat and a Republican is, and realizing it's the tenth time this month you've had to do it.
Statistics is a lie we tell ourselves that having some of the data is really as good as having all of the data. To make this work, you have to make assumptions. To that end, when I handed polls in, my bosses would then crunch the numbers based on the assumptions their employers made, and later on I'd see those numbers on Fox News or CNN.
The reason that polling was so bad over this election cycle is a one-two punch of bad assumptions and just the sheer unreliability of traditional random sampling polls. After that job, I know I haven't trusted a single one since.
At the risk of inflating their egos,
the LA Times / Dornsife Poll was the most accurate, and did indeed predict the outcome of the election: it was a fistfight in the margin of error with Donald Trump in a slight lead. That's exactly what we got, and the more interesting part to me is how they managed it.
They just got volunteers who were willing to be asked about the election every three days to tell them how they felt about voting for the candidates and who they thought would win. Make a couple of assumptions based on voter enthusiasm and turnout, and voila. The calculation itself assumed Trump would win the
popular vote, but it's colossally more accurate than traditional polling because it completely eliminates the biggest problem with traditional polls: having to talk to people who know nothing and don't care, but logging their data in with everyone else, because some poor pollster's job depends on his employer's unrealistic expectations.
With half the error eliminated, it's no small wonder that it did the best.
