Wednesday, November 2, 2016

Nate Silver: Master of statistical guitar face

As promised, here is a reassessment of Nate Silver in light of the tightening presidential race.  As I had been saying, when one candidate is way ahead at the national level, you don't play state-by-state games based on the faulty assumption that popular vote-electoral college splits are plausible when the race is a blowout at the national level.  That's what Silver was doing, and why he overestimated Trump's chances all those many... days ago.  What a year, huh?!  Does Silver get credit for giving Trump more of a chance than I did since Comey has now given Trump a chance?  No, because that wasn't how he arrived at his numbers.  My criticism then was over how he handled, or failed to handle the relationship between state-level and national-level trends.  That criticism still stands.

Anyway, the race is tightening.  Clinton's lead in the RCP average is down to 1.7 points.  Time to look at the state-by-state polls!  Here's the RealClearPolitics map, being cautious.  Clinton is still ahead, but Trump has cut into her lead.  She stands at 246 electoral votes, and if she adds Virginia, Colorado and New Hampshire back into her column, she wins.  Will she?  Well, it's just those Remington polls that put Virginia and Colorado into the toss-up columns, and I talked about Remington yesterday.  So, will Clinton win VA, CO and NH?

Isn't this what Silver does?  Doesn't Silver call states correctly?  After all, he called every state correctly in 2012, right?

And this is where we get into guitar face.  I'm kind of a music fan.  Did'ya notice?  And I've been into guitar for a long time.  My first real inspiration to play was my mother's record collection, and specifically Crosby, Stills, Nash & sometimes Young.  The first tune I taught myself, of course, was Stairway to Heaven, because of course it was (Jimmy Page is a hack, or I couldn't have done that), but then I turned to CSNY.  Some guitar parts were easy to play.  Others-- no way.  The parts I could never play?  Those were the Stephen Stills parts.  That dude is a SERIOUS guitarist.  If you don't play, you don't understand.  The easy parts?  Neil Young.  Guy's a hack.  But he makes GREAT facial expressions when he plays.  Observe.

And it isn't just the face-- it's everything.  This tune is an extended "jam" during which Neil's face contorts with the lip curls and expressions of pain, he does full body spasms and throws himself around like he's doing something impressive.  As a guitarist?  He isn't.  All he's doing is playing the same note over and over again, and bending the strings with a bit of distortion.  If you like the song, then fine.  It sounds kind of cool, doesn't it?  I can't really listen to it, and Stephen Stills is really more my kind of musician, but my point is about Neil as a performer.  He does something that is actually really simple, but through gimmicky motions and behavior, makes it look more difficult than it really is, so people act like he can actually play guitar.  Actual guitarists generally know better.

Neil isn't a good guitarist.  He does easy things and makes them look more impressive than they are.  I promise, I'll connect this to Nate Silver.

Now, have you ever heard of Django Reinhardt?  Django was a Belgian gypsy whose left hand was severely burned in a caravan fire.  Nerve damage essentially paralyzed his left ring finger and pinky, leaving him unable to play chords, and at the time-- early 20th century-- the guitar was nothing but a strummed rhythm instrument.  Django, though, wanted to play.  So, he had to figure out how to do it with two fingers.  He invented a way to do it.  He figured out how to play single note lead lines on guitar.  With TWO FINGERS!  And he was better at it than any mortal guitarist with four.  And watch his facial expression as he played (the camera isn't always on it).

OK, who impresses you more?  Who makes the more pained expressions?

Now, let's get back to Nate Silver.  In 2012, Nate Silver correctly called every state.  You know who else did?  EVERYONE who looked at the polling averages and predicted that the candidate ahead in the polling would win the state.  What made Silver different?  In addition to saying that Romney would win Missouri, he put a probability on it.  Was his probability assessment right?  Um, we can't actually check that.  But Silver made really cool facial expressions and full statistical body spasms when he put those probabilities on each state, right?

Here's my point.  Predicting each state the day before the election is incredibly easy.  The probability assessments you put on each one are difficult to assess, particularly since we can generally be pretty confident about each individual state the day before the election.

You know what's hard?  Predicting the election six months before.  That's some Django Reinhardt shit right there.  In every election before this one, political science has done pretty well.  See, for example, Alan Abramowitz and his "Time for a Change" model.  Of course, that one predicted a Trump victory because of a tepidly growing economy and the fact that Democrats have already won two presidential elections in a row.  Of course, if Trump does win, Abramowitz's model will look like some serious Django Reinhardt playing.

Anyway, yes, we are six days from the election, and the election is close at the national level.  That means we need to focus on state-by-state polls.  That means Silver's method is not as completely wrong in the current environment as it was when the polls were further apart.  I'm still going to focus more on RealClearPolitics and their polling averages than Silver's guitar face probability estimates.

I'll end with some anti-guitar face.  Embedding has been disabled for this clip, but here's a link to a great Roy Buchanan performance.  You have probably never heard of him.  However, you probably have heard of Jeff Beck.  Roy was Jeff's favorite guitarist.  Roy was a virtuoso blues-rock guitarist at least on par with Hendrix.  At least.  Yes, that great.  Sometimes, it is hard for non-guitarists to understand just how amazing he was because he played around with weird techniques like artificial harmonics, and unlike Hendrix, he relied on a Telecaster without an effects box rather than some of the more wild and obvious tricks that Hendrix used, but Roy was amazing.  (I love Hendrix:  I'm just trying to convey how inhumanly great Buchanan was).  At about 5:45, Roy wanders over to the keyboard, mid-solo, drinks a beer, yawns, and picks his nose, while playing, to demonstrate how easy it was for him to play.  Sort of the anti-Neil Young.

Alan Abramowitz's Time for a Change model may get it wrong this year, but his record is far more impressive than Nate Silver's.  He just doesn't do guitar face as well.  Now, go listen to some Roy Buchanan, and some Django Reinhardt, and remember that any kid can play Neil Young.  And watch the state-by-state polls.


  1. Ah, but can you defend RCP vs HuffPollster vs Silver on their collections of state polls?

    RCP and HuffPollster differ rather significantly because of which polls they include (and for how long).

    Also, where's my credit for calling every state in 2004 December 2000? Am I Svengali? Or, have our elections been ridiculously easy to predict for the last 3 cycles?

    Hint: it's both.

    And THAT is where Silver is wrong. His model assumes that the inter-state correlations from the last few elections (I think he uses 5) still hold. That's a LOT of correlations doing some work in his models...correlations based on 5 data points. AKA: interesting, and actually capturing some reality, but also kinda bullshit and not reliable.

    Just look at North Carolina. LOOK AT IT! WTF is North Carolina doing, all blue and shit, while South Carolina says "fuck you, we're still pissed about the whole 'Fort Sumter incident'?" What it's doing is showing that creating a place that draws smart college graduates (let's call it: Science Rhombus) is a good way to begin to crawl out from under 150 years of stupid.

    Silver's model is wrong because those 5-election correlations have trouble picking up on that stuff. 1996 is holding NC back in his model.

    1. I don't remember you calling each state in December 2000. (Honestly, I don't). Where's the email? Putin, are you reading this? Assange? Get on this!

      Anyway, the differences across aggregators even out after you get away from the big events, like Comey. The problem is when an election is too close to the big event. Without an a priori way to assess how much to weight prior polls, there is no way to pick one method over the other. So, no, I can't defend any one over the other. However, I expect that we'll have enough polls over the next couple of days that the differences should even out anyway, at least enough to color each state the same way for each aggregator. We'll see, though, won't we?

      As for the number of cycles issue, um, yeah. The NC/SC issue pretty much makes your methodological point, although Silver currently has NC colored just slightly blue, and I wouldn't stake much money on it. Incidentally, remember Adam? History Department-- University of *South* Carolina, so, ya' know...

    2. Yep. Incumbency effect + polarization = 2004 map.

      (Of course, I would have been wrong had the economy crashed in the meantime, but it didn't, so I wasn't)

      Yeah, but he probably just REPLACED a different liberal professor. Research Triangle has brought a huge number of smart people to NC. Similar to (and underappreciated cause of) New Mexico going blue.

    3. New Mexico is rather different. You've got the Latino factor, combined with Trump. That makes it solidly blue this year for a reason that has nothing to do with the trends in NC. And you do know that the relationship between education and partisanship isn't really that clean, right? You must have run those numbers at some point. Shall I put them up? Just cross-tabulate education and PID with NES data. C'mon. I know you have done that at some point.

    4. No doubt. But education is a much bigger factor in the vote in 2016 than it was before. (Hey! Campaign effect!)

      White, no college: -27.5
      White, college: +12.2

      That's a 40 point effect for college. That's fucking insane, and the old NES data don't have that because it didn't exist.

    5. Source? Elaboration? Is this a Trump vote difference?

    6. Source:
      (It's an average of polls that have reported their various marginals)

      And my 2004 prediction was NOT an example of this, and you just have to take my word for it:

    7. Potentially interesting, but I'll wait for NES numbers. As for the scam, I've seen the stock version, but trust you? I know you.