It's pretty lame to paste an image of the NY Time's 82% chance of Clinton winning the election in an article that calls out Nate by name. Nate gave Clinton like a 70% chance.
Can model building predict human behavior? It's fuzzier than the hard sciences, but not that much fuzzier. And if you do believe it's impossible to model human behavior, you have to throw out vast swaths of knowledge and areas of study: there goes sociology, anthropology, economics, psychology, epidemiology...
To be honest, this reads as quite the Luddite article.
Not to mention the focus on Simulmatics is akin to using the 19th century use of bleeding as the basis for a criticism of modern medicine.
Also, the modern example of Cambridge Analytica obviously has essentially nothing to do with what 538 actually does...
Towards the end, the criticism of if a predictive model is falsifiable is an accurate one given that its output is probabilistic, and given the small number, and relative infrequency of, American elections (especially presidential ones) - which impacts the model's calibration. But, Nate actually brings this up quite often (which I assume the author would attribute to him "covering his bases").
The article is way too long, and doesn't really talk about Nate Silver much.
However, one interesting item is the idea that there is no way to verify these predictions. Because they are forecasting one time events, there is no way to validate if their models are in the least bit accurate. If you have a X% chance of event happening and it doesn't, were you wrong? No because it wasn't 100%, and even if you are right, it's not terribly meaningful. It would be interesting to see if a site like five thirty-eight modeled all elections results, and then see how accurate they were across all of their predictions. Maybe you would have to look at all predictions with > 70% chance of happening and see if they got 70% of those correct.
However, in the end those predictions seem like they could be done by throwing darts at a dart board and we wouldn't really know.
I don't think this article provides strong evidence that they are well calibrated on the presidential election specifically (sample size N=3), or that they are correctly accounting for rare black swan events, but it does seem to imply that the criticisms about "538 claims victory no matter what because they always have non-zero probabilities" are oversimplified.
2016 wasn't a black swan event. It was a polling error, which do happen, if rarely. It was not unforseeable, and 538 included a probability of that happening which is why they gave Trump higher chances than most others did.
And 538 does do backtesting on elections back to 1972. That's not particularly trustworthy since it invites over-fitting, but internally they do have a little bit more than N=3 to work from.
(I have fairly minor quibbles with some of Nate's modeling ideas, but I broadly mean to be defending him).
I don't mean to imply 2016 was a black swan event- I agree that ~30% was probably as accurate a take as could be achieved (most evidence that seems reasonable to use indicated a lead for Clinton, but that it wouldn't be that surprising for that lead to be overcome). I just mean that the model assumes a fairly normal election environment, without like a huge attack on Election day or something on election day.
The N=3 comment was meant specifically for evaluating their calibration, not the data they use for their model.
It's not as dire as you suggest. To simplify a bit, you can take the set of ~75% predictions created by a model (which can be an actual model, or it can be just the meta model of whatever model Nate Silver wants us to trust right now), and then see what percentage of them were right. If that comes to 75%, then the model is well-calibrated, and it's a solid enough basis to assign a 75% probability to future 75% predictions, which is often as good as you can get when trying to make decisions and plans.
> those predictions seem like they could be done by throwing darts at a dart board
They are. They're made by throwing 40,000 darts at a dart board. 40,000 simulations of the election based on the polls, historical trends, demographics, and more.
The real threat, the real reason why the 2016 election was mostly mis-called, is when the polls themselves make mistakes in a common way. In 2016 it was education that wasn't factored strongly enough.
The models can be run on past elections and polling + data before those elections. That's how the develop the models.
I haven't finished this article, but what I've read so far just reads like anti-empiricism fluff, so I don't plan on finishing unless someone here suggests otherwise.
Nate Silver and 538 gave Trump a 28% chance of winning on Election night in 2016. Nobody else gave him even close to that kind of probability. Sam Wang, considered another very good polling expert said he would eat a bug if Trump won, giving him less than 1% of a chance (and he DID[0]).
And here is the New Atlantic saying not to trust Nate Silver? Give me a break.
Statistical models give probabilities, not certainties.
I think Nate is pretty sharp -- but I don't trust the people who answer polls. Who has time to pick up the phone? Most of the time it's some spammer. And if you do answer, why tell the truth about your vote?
I think that the point is that both the pollsters and Nate are looking at how various polling methodologies correlate to actual past election results.
The raw numbers from a poll are no good though, because you have a biased sample. The people who answer the survey are not representative of the people who show up to vote. They use the actual election results to calibrate the correction for this.
As long as these relationships stay stable, polls should be pretty accurate. In my understanding, the biggest problem in 2016 was that the polls were undercorrecting for the difference between college educated voters vs. not college educated voters. Historically, this made less of a difference than other factors in voting behavior, so it had less impact in the weighting model. In 2016, it had a lot of impact, and college educated voters were more likely to answer the survey, so their tendency to vote Hillary was overrepresented in the polls.
That said, the polls were off by less than 2% nationally. Perhaps that's not a great result, but it's not as if they were wildly wrong, it was just a close election.
Unfortunately this election is also somewhat close this year, and voting behavior is going to be different with the pandemic, mail in voting, etc. I think it's very hard to reliably call Pennsylvania.
538 does rank pollsters based on "historical accuracy and methodology"[1], which are then weighted accordingly in their prediction model.
And Nate and Elliot Morris (from the economist) have talked about how polling error tends to be normally distributed, so using multiple pollsters with different biases could correct for polling error. Pollsters are picking representative samples and weighting it to reflect the population, so it's not like they're just counting on who picks up their phone (and phone calls aren't the only way to poll anymore anyway).
As I recall, in 2016 the pollsters did oversample college-educated white voters, which is what gave Clinton her misleading lead. This apparently has been fixed.
There's also the fact that any demographic under or oversampling should be accounted for in the uncertainty range produced by the prediction. I haven't been able to find decent write-ups of this particular part of model forecasting, but from what I've been able to piece together : I believe they quantify the uncertainty of their presidential predictions by building logistic regressions of the output based on highly-correlated predictor data (i.e demographic information, GDP) and then simulating thousands of possible scenarios where demographics trend differently. I don't exactly know how these simulations are set up (I would love to know though). And presumably this is updated in a bayesian matter everytime they get more input data, so the prior probabilities will have an impact on the posterior.
If anyone has a different or more detailed idea of how this is done, let me know. In fact, here's the actual code for Elliot Morris' forecasting model: https://github.com/TheEconomist/us-potus-model if anyone can figure it out better then me.
That's also true of every election in recent history, and yet polls still historically have had predictive power.
Even shy Tory stories are perennial fixtures in the poll criticism genre; perhaps it could exist as a stronger effect this year, but as yet there's no evidence that it is.
This what scares me about the polls - that they don't account for Trump supporters being less truthful than the rest of us. Just look at Trump - he hasn't said anything true for years.
Can model building predict human behavior? It's fuzzier than the hard sciences, but not that much fuzzier. And if you do believe it's impossible to model human behavior, you have to throw out vast swaths of knowledge and areas of study: there goes sociology, anthropology, economics, psychology, epidemiology...