MORE LIKE NATE SLIVER AMIRITE

I don't even know what that title is supposed to mean.

Of interest this election season has been fake pundit Carl "The Dig" Diggler, the creation of two comedy writers, who not incidentally has predicted correctly the outcome of twice as many primaries as Almighty Beltway Knowledge God Nate Silver. Their prediction "method" involves little more than "gut feelings" and comedic stereotypes of the residents of different states, and so to even call it a method is unjustified. But that's precisely the point.

Nate Silver is, on the whole, a force for good. Attempts to provide analysis that relies on empirical data are, and always will be, an unqualified positive. His (at this point it is hard to separate him from his FiveThiryEight colossus, which of course involves other analysts and writers) reputation has taken a blow in 2016, though, and frankly I'm not sorry to see it happen. His analysis has always been terribly basic – on the order of something a good undergraduate statistics course would cover – and the reputation he has built as some sort of data god is a bit much. He has become, intentionally or otherwise, a liberal Bill Kristol; it doesn't matter if he's always wrong, he's still brilliant.

Two aspects of Silver's predictions deserve serious criticism, one of which Mr. Diggler emphasizes. He has a really annoying tendency to hide behind probability – "I didn't say Clinton would win, I merely said there was a 99% chance she would win!" Empirically, this makes perfect sense. Probabilistic analysis is never 100% accurate and does not claim to be. In the face of a large number of incorrect predictions, though, someone treated as an idol should have a better defense of his supposedly brilliant methods than The Simpsons' classic "Well, when you're right 52% of the time, you're wrong 48% of the time" gag. ("OK Jimmy, you're off the hook!")

The part that always has bothered me – and yes, of course I'm jealous – is that Silver became A Genius by predicting the outcomes accurately of two very, very easy to predict elections – 2008 and 2012 – in which a simple average, even an unweighted one, of barely-scientific polls by state was sufficient to see that the Electoral Vote would be lopsided. Those were not especially close elections, and it is not difficult to predict the outcomes of elections that are not especially close. As for his correct predictions of many other statewide races such as Senate and gubernatorial races, his model amounts to little more than averaging poll results obtained by other organizations and which are publicly available. In 2016, so far we see that the same magical techniques that told us McCain was going to get blown out (duh) are of minimal use in predicting an outcome that isn't totally obvious.

Yes, primaries are much more volatile and difficult to predict due to a number of factors like low voter turnout and a large, shifting field of candidates. In that sense we would expect predictions to be less accurate. But that's exactly the problem for Silver; it's becoming very easy to say "Well if you can't predict a race correctly unless all the conditions for making a correct prediction exist, what are you really doing? What good is this?" That has been the thorn in my paw with Silver all along. It's like saying that you can hit a lot of home runs, provided the pitcher throws the ball exactly where you want it, how you want it, with the wind blowing out at 50 mph. The limits of the Big Data approach and worship thereof are becoming very obvious. Like public opinion polling (on which much of 538/Silver is based), it is a useful tool when the gap between or among options is greater than the margin of error. When it isn't, the data don't tell you much at all. At that point you're effectively guessing. And Silver puts a number on his guesses, which gives them the imprimatur of scientific authority ("Clinton has a 63% chance to win!" – so precise!). But in reality he's telling you that Clinton is slightly more likely than Sanders to win a given race; the odds are about 3 in 5. That's an improvement on a coin flip, but it isn't much of an improvement.

It is far better to hear someone talk about data than to listen to some empty suit talk about his hunches or his conversations with various cab drivers. I'll take a Nate Silver column over 99% of what's out there for consumption. But people really need to stop chanting his name like it's a magical talisman that all but guarantees victory. What he's doing is not that complicated and, more importantly, not that useful unless the outcome of a given race is not in doubt. If the new working definition of genius is the ability to avoid being wrong about outcomes that are obvious, then I wish someone had sent out a memo in the mid-00s that the position was being filled. I, or anyone else with a basic understanding of political data, could just as easily have filled it.

26 thoughts on “MORE LIKE NATE SLIVER AMIRITE”

  • I'll give Silver this much credit–when Michigan blew up in his face, he didn't pretend that it hadn't. He started off with a complete admission of being not just wrong, but very, very wrong. Then he went back and showed his math–yes, there was a certain amount of ass-covering of the "Primaries Are Always Less Predictable" variety–and showed why he'd been wrong, and in general handled it about as well as an adult person could. (Purists might have required him to set himself on fire and/or quit the business, but come on. Real world.)

    So I think it's unfair to yoke his name to that of Bill Kristol, who A. is Never Wrong Just Ask Him, and B. is always wrong seriously just read the fucking autopsy. True that they're both invoked by assholes as oracles, but Silver's an oracle who uses numbers and expresses his opinions in conditionals, and Bill just guts a black goat on his back patio every morning and lets the voices in his head read the entrails.

  • Thank you, thank you, thank you!

    Nate Silver actually deserves more criticism this election season than he's getting because he went full-on pundit. His mantra has been to "ignore the pundits and look at the data," but he completely went against that this election cycle. As early as August, all of the polls had Trump up and some states like South Carolina he was up by insurmountable margins. This was also after he openly mocked John McCain for being a POW and told everyone Megyn Kelly isn't fit to moderate a debate because she was on her period. If you looked at the polls of the early primary states, it became obvious that, at the very least, Trump was going to do some serious damage.

    Nate Silver ignored ALL of the poll numbers and instead went with his gut that The Party would sort all this out. Nate got blind-sided by the same hubris he was mocking pundits in 2008 and 2012 for – the numbers proved to be more accurate than what his gut was telling him.

  • I'm not sure if they play baseball in strong gale force winds…unless you meant to choose an unlikely condition…ah, layers upon layers, that's why I keep coming back.

  • @J. Dryden

    I will give Silver credit for admitting mistakes after Michigan. I really wish we would see more of this.

    My biggest beef with him this cycle is that he has written so authoritatively that Trump had no chance and at one time was espousing Christie (!) as the possible nominee. Instead, if he had said "This is what my model shows me, but there are a lot of factors that make this primary different than ones in the past such as an unusually high number of candidates…" I would have respected him a lot more. Instead he took the role as Mega Math Wizard Man and would definitively say Trump had no chance when all the polls were saying otherwise.

  • Assistant Professor says:

    Sam Wang's always seemed to have a better handle on this. Silver's got his own proprietary "special sauce" that he uses mainly to keep things in doubt and generate suspense. This was especially clear in the 2012 election when liberals were worried that maybe the supreme, blithe, reality-denying obliviousness of Republicans might have something to it and so they would reflexively hit Refresh on the Time's Five Thirty-Eight. In order to avoid basically being "Calm down, folks, Barry's got this one" and then not getting page views, he would sometimes make the election closer based on special sauce that would keep the page views rolling in.

  • Of the people I know who know of him, liberals tend to hold him in higher -but not top- esteem, whereas conservatives see him as the dark witch king of aggamar who used his evil academic powers to put that anti-muslim-christ lord Obama on the throne. I guess they need someone to focus blame on.

  • lord karnage says:

    agree with pretty much everything in the article except, but i don't think this is entirely fair…

    > His analysis has always been terribly basic – on the order of
    > something a good undergraduate statistics course would cover –

    While it's probably useful to conceptualize his approach as a simple average of polls, that's not what he's doing. if that was the instruction you gave to a typical undergraduate, I don't think they would be able to implement results noticeably better than average or pundits.

    i am not seeking to argue a point to exasperation, just to point out that the methods involved are not really that basic. his role as a popularizer of analysis should be separated from his role doing that analysis. he did a good job at the analysis, and the methods are not trivial.

    there may be some undergraduates who can conceptualize and implement repeated-draw sampling as in monte carlo simulations, but even that doesn't quite get at the underlying complexity of things like baysian model averaging, where the model terms themselves come from repeated distributions of data and variables. Even undergraduate data scientists (at least the ones we are interviewing) have trouble describing the computational mechanics behind this. I laugh out loud that many of them will say that it's what gelman does, or what silver does, but their level of sophistication doesn't seem to go beyond knowing who uses it, and that there's a library somewhere with a function for that.

  • @Tom

    I agree that Silver refused to believe some of his own data, but seriously, some of his own data was self-contradictory when it came to assessing Trump's chances. If doing well in polls strongly has historically correlated with getting the nomination, and getting lots of endorsements has also historically correlated with getting the nomination, and Trump is doing very well in one metric but terribly in another, then what does that mean?

    Trump was so unlike any modern candidate that his candidacy was very difficult to prognosticate. Silver probably got out ahead of himself last summer/fall in saying Trump wouldn't be the nominee, but at the same time, it isn't as simple as saying Silver just ignored the polls.

  • What?

    "He has a really annoying tendency to hide behind probability"

    That's like saying Coca Cola is hiding behind soda. What do you even mean by this? Probability is his product. He runs some statistical analyses on polls and provides his results in terms of the probability of various outcomes. If you want a different product go to a different aisle on the internet. Don't go to the soda pop display and complain that you can't find the noodles.

    I think Nate Silver's best contribution isn't in predicting things but in educating people a little about statistics. He is bringing facts to a mostly fact free discussion.

    I do find him personally annoying though, so this was fun to read. But liberal Bill Kristol??? I don't think he is particularly liberal at all, except in the "facts have a liberal bias" sense. Plus, he isn't a drooling grinning moron.

    And to be fair, those elections were only that easy to predict in retrospect.

  • cackalacka says:

    What is the liberal equivalent of pushing for the nomination of Sarah Palin or bombing a 3rd world populace?

  • Emerson Dameron says:

    A lot of Obaptimists loved Silver because he accurately predicted that the Good Guy would win. He may have gotten in over his head after that.

  • Leading Edge Boomer says:

    When Mr. Silver quotes, e.g., "63%", he is giving you the result of his computer simulations. He has said more than once that these computer-generated numbers must be taken with some grains of salt because they are, sometimes more and sometimes less, sensitive to the inputs. 63% cannot be thought of as different from 57% or 70%.

  • You would think smart guys like Nate Silver would know the difference between statistics and polling, alas not so much.

    Statistics is a legitimate branch of mathematics and is as proven as these things can get. Polling humans is a whole different matter.

    First, statistical sampling fundamentally relies on the idea of randomness. That is to say that no member of your selected sample has the option of refusing to be sampled. If I take a truly random sample from a machine that make ball bearings and measure their diameters then my sample remains random. This doesn't work so well polling humans. Every sample of humans is by definition self selecting so your attempt at a random sample is mitigated by whoever chooses by their own volition to participate.

    Second, ball bearings can not lie, forget or get distracted. Whatever measurement of their diameter you get is a fact. Yes, granted your measuring device might be inaccurate but mirabuli dictus your stats analysis would still be correct largely.

    Not to say that polling is entirely useless but it always gets my rhubarb up when people use probabilistic terms and pretend that polling is "science".

  • HoosierPoli says:

    That's not actually how Silver made his name though. Under the handle Poblano, he came to prominence on the hated Daily Kos, where he outperformed everyone in predicting the outcomes of 2008 Democratic primaries, ignoring the conventional wisdom that Hillary would waltz through to the nomination. I mention this because I read every post and yet this seems to have been deleted from his "official biography", probably because it involves a seriously left-wing outlet being right about something.

    The way he pulled it off, as I recall, was detailed demographic breakdowns of each congressional district of each state. The current 538 model seems an awful lot less sophisticated than what he did in 2008, and some of the other models (mainly demographic) that have performed much better seem to indicate that he's leaving out important data that he used to understand the value of. The interesting question is why.

  • HoosierPoli says:

    Just one more thing to add:

    "hen it isn't, the data don't tell you much at all. At that point you're effectively guessing. And Silver puts a number on his guesses, which gives them the imprimatur of scientific authority ("Clinton has a 63% chance to win!" – so precise!). But in reality he's telling you that Clinton is slightly more likely than Sanders to win a given race; the odds are about 3 in 5. That's an improvement on a coin flip, but it isn't much of an improvement."

    The benefit of putting numbers to your predictions is not to give it "the imprimatur of scientific authority". The point is to be PRECISE about what your prediction IS. It's a long-standing problem, which political science has studied in the context of decision-making processes, that a phrase like " a good chance" SOUNDS definite but is actually totally subjective – one person might thing it means 80/20, another might think it means 20/80. This is a way that the pundit class weasles out of bad predictions, or revises their wrong predictions into right ones.

    So when 538's model say a 99 percent chance and the result is hugely unexpected, the fact that we can lambaste him for it is a feature, not a bug.

  • I wonder if part of Nate's problem is the quest for novelty. Now that he's an Internet Brand, he needs to keep the punters coming back for more. It's not enough to say, "I fed the polls into my model, and Trump's going to win" or worse yet, "I fed the polls into my model, and I have no goddamned idea who's going to win".

    So instead Silver writes about handwaving bullshit like the "endorsement primary". He concocts increasingly baroque models, with fewer and fewer data points, which as any data scientist can tell you is a recipe for junk results.

    Also, I find it amusing when Silver is described as using Big Data. Dude, please. Ten thousand human genomes at 3 billion base pairs each is 30,000,000,000,000 data points. That's Big Data. Silver uses anywhere from a few hundred to less than 10 (!) data points. That's not even Medium Data. At best, it's Slightly More Than Tiny Data.

  • I’d like to say “who cares about Nate Silver and his predictions?” but alas, I know that folks care quite a lot. I’m not among their number. What little I have learned about him became suspect once he joined the pundit class and branded himself with a soapbox website. So I tune him out, knowing that he’s as fallible as the next.

    Regarding polling, I really have to wonder about its usefulness anymore. The fun of gaming a poll or vote seems to have overridden the outcome (Boaty McBoatface, anyone?). Just today I read an admonition that everyone should skew their answers and online opinion oversharing as an act of rebellion against being constantly tracked and commoditized. For that reason, I’m definitely going to vote for Trump and run up my debts to the max.

  • ClockworkSteve says:

    Thanks for saying this. Silver's choice of headlines for the last few months, for example, all the "Bernie can't win" posts, makes me question his objectivity. He's also getting very defensive recently now that his precitions are being questioned.

  • I believe the reason Nate Silver (of whom I am a great fan) stands out in this day and age is that the 'horse race' aspect of political analysis has taken over most media…everyone wants to pretend that each race is between roughly equal opponents and that there is suspense in the results. Relying on data had become/is still unfashionable, and everyone wants to think there is some other X factor which will swing each race in a dramatic fashion.

    Relying on data, of itself, is valuable. As opposed to listening to a random 'man on the street' interview or your gut feelings and asserting those have as much sway a taking a crack at interpreting polls with a statistical background.

    I believe Nate himself would strongly agree with most of your commentary, and he often course-corrects his own analyses and comments upon non-objective factors found within his own work.

  • fledermaus says:

    "Regarding polling, I really have to wonder about its usefulness anymore."

    I think this is the case too, but for much more banal reasons. I think that is just getting harder and harder to get an accurate sampling of people who want to spend 10-20 minutes answering a bunch of questions. In an age of cable cutting, cell phones, caller ID and general disgust at the political and media establishment, people who still respond to polling are going to be a self selecting group. And it's not just contained to the US, the accuracy of political polls in Europe is also declining IIRC

  • I always thought Mr. Silver made his statistical bones on the general election, not the primaries. his methodology, though seemingly simplistic, has the advantage of pretty much eliminating the outlier's impact (either way), thus giving a more accurate picture of an election, at a given point in time. kind of like a balance sheet, it's good for one moment only, and completely changes in the next 5 minutes.

    primaries and caucuses are a whole other kettle of fish, making pre event polling a heck of a lot more complex than the general, which has one set of rules for everyone. to this day, I still don't completely understand exactly how a caucus works, and why any state with brains would even have one. part of the problem we're all suffering from, is the "Trump Effect": how a bozo who had no business running at all, is totally lacking in self-awareness, and seemingly has no idea of how our form of national government actually works, has not only been very successful in his first bid for any public office, but is going to be the Republican nominee for president, because the carny barker in him instinctively knew how to appeal to the knuckle dragging republican base, way better than the "Pro's from Dover" did. who the hell could have predicted that?

  • Trump's effect on the usual process of election prediction is like the Mule's effect on Hari Selden's psychohistory –

  • alwayscurious says:

    Nate references the Pareto Principle of Prediction in his book: 20% effort will put you ahead of 80% of the crowd. I've followed the predictions on his site (largely ignoring the commentary)–they seem to be pretty well in line with what's happened. And 538 is a reasonable catalog of what the polls ARE saying before the primary. Remember, the big problem in 2008 & 2012 is that FEW in the Republican party publicly admitted to what was happening in polls. So ~45% of the population was blown away that some geek correctly predicted the result…using NUMBERS. Nate certainly isn't the best statistician in the world–as he says, he only has to do 20% more than the average Joe to be ahead of 80% of the population.

    Likewise, his book points out that he's trying to raise the level of discourse. If nobody cares about the result, the predictions will be garbage. The only way force better predictions is to get people to care about the outcome (motivating them to also collect better data). He's moved into predicting primaries now BECAUSE the general election data is so much better quality & easier to use for predictions.

Comments are closed.