Nate Silver needs this election to be close. (Or have Trump win).
So yesterday I tweeted out that Nate Silver was going to look pretty bad (whether that is fair or not is a different question) if this election is a blow out because that basically means that he got two-thirds of the big predictions (Dem Primary, Rep Primary (which he botched badly), and General Election) of this election. And this would go along with badly missing the British election in 2015 (the final result wasn’t even in their prediction intervals). So if he misses this election badly, it’s just one more high profile miss after basically getting everything right in 2008 and 2012.
And then someone responded to my tweet asking what’s with all the “Silver hate”? First off, I wasn’t hating on Silver in that tweet. I was merely trying to point out that his public image has a lot riding on this election. What I mean is most people only think about Nate Silver every four years (he’s like and Olympian; he’s doing stuff all the time, but most people only care every 4 years) during the general election, and if he get’s that right the general public believe that he is a genius because that is the story that the media will tell about him. Now, the average American is probably not following Silver’s other predictions very closely, but they look to him only during the general election. And if he get’s this wrong, the media will start looking back at his other predictions more closely and see that he has a pretty bad recent track record. In summary, no matter how bad Silver has been of late, if he get’s the general election correct the public will largely ignore the recent poor forecasts.
But there is apparently a bunch of Silver hate out there. Some pundits and politicians have always disliked Silver. This year’s the most high profile hate is coming from Ryan Grim, the Washington Bureau Chief for the Huffington Post. In his article, Grim contends that Silver in “unskewing the polls — all of them — in Trump’s direction”. Silver responded in a series of tweets. He did not seem happy about this criticism.
So why is there so much hate for Silver? My theory is that it’s in part jealousy. It’s objectively true that Silver is incredibly successful. When he was with the NY Times he was single handedly drawing huge amounts of traffic to the site. Most journalists aren’t doing this. So I think there are a lot of people out there who are rooting for Silver to fail and recently they’ve had a lot of stuff to work with in terms of not so great forecasts. But this General Election is really all that most people care about and Silver’s numbers are so much different than everyone else. So it’s easy for reporters who already aren’t huge Silver fans to start to pile on. I also think that journalists may dislike him (he didn’t fit in at the NY Times) because he essentially invaded their territory and almost immediately became more popular and drove more traffic than they did. So I think a lot of this kind of hate is jealousy.
And just as I think some journalist are annoyed that Silver came into their territory and took it over with ease, I think there is some annoyance (I wouldn’t say hate) among statisticians. My claim here among other statisticians in anecdotal, though I will admit that I myself have been annoyed by Silver. But in reflecting on that position, I now think that is stupid. I’m not really annoyed with Nate Silver; I’m annoyed with the general public’s REACTION to Nate Silver. Look Silver did a great job in 2008 getting 49 of 50 states correct and then in 2012 he got 50 out of 50. This was labeled a “Triumph of the Nerds” and he’s been called a “Math Wizard” and “Polling Data Guru“. What annoys me about this is that these elections weren’t that difficult to predict. Plenty of people got this election correct and while Silver was very good in 2012, he wasn’t the best (Note: I will be scoring the 6 prediction methods listed on the Upshot live on election night using log-loss and Brier score). But no one seemed to be talking about Drew Linzer and Wang and Ferguson. So I guess what really annoys me is that Silver is treated in the media like some singular talent that sees and predicts things that no one else on earth can predict, when in fact that’s just not true. I don’t blame Silver for this. In fact, I don’t know who to blame, but it annoys me. And some other people seem to feel similarly:
- Obama’s big win does not mean Nate Silver is a towering electoral genius.
- Nate Silver’s genius isn’t math. It’s journalism.
So all of this brings us to today: November 7, 2016. The day before election day. The general consensus is that Hillary Clinton is a big favorite to win this election tomorrow. Over at The Upshot, the New York Times, Silver’s former employer, is tracking a whole bunch of different prediction sources including their own predictions, FiveThirtyEight, Huffington Post, PredictWise, Princeton Election Consortium, Daily Kos, and a few others. Of the six sites that give a fully quantitative probability of a Clinton win, 5 of the 6 are over 80% and 2 of the 6 are over 90%. The only exception to this is FiveThirtyEight which currently has a Clinton win probability of 64%.
Some people think that Silver’s prediction is way off:
Nate Silver is unskewing polls — all of them. And he’s panicking the world. https://t.co/Ig471ixSK6
— Ryan Grim (@ryangrim) November 5, 2016
Silver was not happy (#understatement) about this article :
This article is so fucking idiotic and irresponsible. https://t.co/VNp02CvxlI
— Nate Silver (@NateSilver538) November 5, 2016
The article by Grim is short on technical details of WHY he thinks what Silver is doing is wrong. I’m sure there are technical criticisms that could be made about Silver’s methods (or anyone’s method), but Grim’s argument is simply that Silver is adjusting the polling results. I don’t have a problem with that in principle as long as it’s done well, and I have no real basis to judge whether it is being done well or poorly.
But it does seem that Silver is hedging his bets here. As Drew Magary says (emphasis added):
I tweeted about this yesterday, but I’m gonna put the take here because maybe you’re already freaking out about Tuesday and starting to drink heavily. Here’s the deal: Donald Trump is going to get his ass kicked. Anyone who says otherwise is either a) afraid of jinxing it and/or making Hillary Clinton voters complacent (understandable); b) afraid of being wrong (Nate Silver); c) supporting Trump; or d) interested in making this a “horse race” for the sake of maintaining public interest (most of the television media, along with grotesque shitbags like Mark Halperin).
So what would be good for Silver? Let’s take a look.
I ran a simulation based on the state by state predictions posted on The Upshot using my correlation estimates where I simulated the election 50,000 times. Based on these simulations with my correlation estimate gives mostly similar overall win percentages as the Upshot has posted. I have:
- NY Times: 86.07%
- 538: 63.16%
- Huffington Post: 99.98%
- Predictwise: 88.5%
- Princeton Election Consortium (PEC): 88.54%
- Daily Kos: 99.01%
A few of these are notably different from what is posted. For instance, PEC is basically predicting that Clinton has a >99% chance of winning the election. My simulation only has them at 88.54%. This would be consistent with PEC either not accounting for state by state correlation or estimating less correlation between state than I am. Daily Kos is also notable in that they only have the probability of a Clinton win at 87%, but based on their state by state probabilities, the implied win probability is over 99%. I can’t explain this. The other four overall predictions essentially match with their implied state by state predicted win probabilities.
Next let’s take a look at the distributions of outcomes that are being implied by these state by state probabilities. Below are the density estimates of these 6 prediction. I’ve also added a rather naive projection for reference. The gray line is a model where I gave 15 states/districts to Clinton with certainty and 21 to Trump. The remaining 20 states/districts I made a coinflip.
The notable aspect of this plot that jumps out is the left tail of the 538 prediction. It’s way above the other 5 forecasts. Also, the New York times is allowing for, more so than other forecasts, for a very large Clinton victory (well over 400 EV!!!!!). So what is good for Silver? The only outcome that makes him look really good to the public at large is if Trump actually wins this election. If Clinton wins in a close race he’s also probably ok in the court of public opinion. But, if Clinton beats Trump and gets more than say 320 electoral votes basically everyone got that except Silver.
I’ve also added boxplots of the 6 major forecasts. Look how tight Huffington Post is!!! The IQR is 324 to 351 (Is that reasonable?) At the other end of the spectrum you have 538 whose IQR is 244 to 351, which is a rather large range. Silver’s median prediction is also much lower than the other 5 at 308. The other five forecasts have median predictions of 323, 325, 326, 336, and 342. S0 basically anything in the 320+ range and people are going to start asking why everyone else got it and Silver, who wouldn’t have “missed” if the Clinton total is 320, was off. So the ideal outcome for Nate Silver, who’s a Democrat, is a very close Clinton victory. That way he gets the candidate he prefers, and he looks good. (And America will still exist…..)
So what do I think will happen? I really like this map that @bennyjet34 posted to me on Twitter a few hours ago. This is essentially the consensus map based on all 6 forecasts.
But what fun is the consensus? Here is my final prediction:
I think Clinton get’s Nevada based on the early turnout there and Trump gets New Hampshire (because NH is essentially the deep south….). I’m also going to predict that Clinton wins Ohio. I have no numeric argument for this, but I’ve got a feeling.
Finally, I leave you with this from xkcd: