Ryan Faulk and I, who own this site, made a number of predictions about the 2016 presidential election:
- Trump would win
- Trump would out-perform polls, especially in the Rust Belt
- Trump would drive new voters to the polls and this would bias pollsters against Trump
- The existence of this “monster vote” would be predictable based on primary turnout.
- Pollsters were using samples which were biased against Trump in terms of party ID by oversampling democrats and undersampling republicans/independents
Let’s look at how these general points fared and then at the more specific electoral predictions we made. Obviously, Trump won, so we got that right.
Trump over-preformed polls generally (1), but he especially did so in the rust belt:
So, we were right about that too. We also said that the polls would underpredict Trump, in part, because they were not picking up on new voters who would vote Trump. We said that the existence of this “monster vote” would be predictable based on the changes in primary turnout for the republican and democrat party.
This would predict that Trump would outperform polls to an especially large degree in the rustbelt:
(Primary advantage is defined as the percent change in republican primary turnout relative to 2008 minus the percent change in democrat turnout relative to 2008 (2)).
Of course, Michigan is not in line with this theory. Michigan’s data is misleading. There was a huge increase in democratic primary turnout there relative to 2008 However:
- Obama was not on the ballot in the 2008 Michigan primary and this massively depressed 2008 primary turnout
- Sanders won Michigan, out performing polls by a huge margin. These were not Clinton supporters.
Given this, our hypothesis is largely confirmed by these numbers so long as you interpret them with a little common sense.
More generally, the bigger the turnout change advantage for the republicans was in a state the less accurate the polls were in that state:
The republican turnout advantage also correlated with Trump’s overall margin of victory (r=0.37). So, we were right about the importance of primary turnout aswell.
As we expected, there was indeed an increase in new voters this year relative to past elections. New voters basically always vote democrat because they are typically very young people. Given this, the relevant question when looking at new voters is not whether or not they voted democrat but, rather, by what margin. In 2008, 69% of new voters voted for Obama. This year, only 56% of new voters voted for Clinton. That is a huge drop-off and suggests that Trump did indeed bring in new voters. (I could not find data on 2012).
Thus, the monster vote was real and was predictable based on primary turnout, just as we said it would be.
We also said that pollsters were giving overly pro-Clinton results in national polls because they did not properly weight by party ID. We were right. The Real Clear Politics average for national polls had Clinton winning by 3.3 points on the night of the election. She actually won by 0.2 points. Thus, they were off by 3.1 points.
Five Thirty Eight was even less accurate. They predicted that Clinton would win the popular vote by 3.6 points and so were off by 3.4 points.
I applied the same weighting formula we advocated months ago to the Bloomberg, IBD, Economist/YouGov, ABC/Washington Post, Monmouth, Ipsos/Reuters, and CBS polls in the final Real Clear Politics average for the election. (The other polls lacked the necessary data.) This changed their prediction from Clinton winning by 3.3 points to Trump winning by 1.8 points. Thus, a prediction based on unskewed numbers would have been off by 2 points, which is about one-third more accurate than the unadjusted pollsters were.
Of course, these polls would have still been wrong by more than we would like them to be. We never said that unskewing would make polls perfect, just more accurate. And we were right about that too.
So, in terms of the general picture, we got this election right. That being said, we did not get every detail correct. This is what I predicted would happen in an article uploaded the day before the election:
I got three states wrong: New Hampshire, Michigan, and Wisconsin.
New Hampshire, I got wrong, but barely. I never said that it would be a landslide, and Clinton won by 0.2%. So I don’t feel too bad about that.
Based on what I have said thus far about our general theory of this election, it may be surprising that I was wrong about Michigan and Wisconsin. I had considered what our theory said about both of them. However, the polls there were really bad for Trump, and I assumed that the “experts” couldn’t be that wrong and so trusted them over my own data and hypothesis. This was clearly a mistake and, moreover, the only major mistake I feel I made when analyzing this election.
- Polling averages for each state were taken from FiveThirtyEight
- Data on primary change was taken from our main article on the subject which can be found here.