David McCoy

Scenario: Driving home late last night.   Listening to NPR when I could find it.  Heard a commentator say something to the effect of, “No Democratic president has ever won office without carrying the state of Missouri.”  Sudden urge to rip the radio from the dash and never listen to any newscast again.

I just searched for “without winning Missouri” on Google.com.  You don’t have to be smart to search nowadays – all you have to do is enter the key snippet.  Lots of results, most saying effectively the same thing: “Wow! How did Obama win without Missouri?”  Apparently, a Democrat being elected president has always coincided with Missouri’s going Democratic (but not the other way around – Missouri has predicted a Democrat and been wrong, as several sites bemoan).  Whoopee!  A way to predict the winner!   If a Democrat won then Missouri must have gone Democrat.  But this time, it failed.

Failed indeed. This is a fine example of pure statistical hype.  Good stats classes teach ways to avoid this kind of math mistake.  Missouri was never a validated predictor of the Democratic wins, and the so-called “American opinion.”  “What?  It aligns so perfectly.  How can you say that?”  Even if the next 10 Democrats carry Missouri and win the presidency, it is still an invalid predictor in my book.  I can’t even muster a reasonable case for real covariance. This was just the same bad math that has sportscasters saying:

“Well Tom… following a line-drive single, this batter has never hit the second pitch from a left-handed pitcher on a Tuesday night in the third inning”

Richard Feynman would have loved this one.  It’s the same mentality that makes people see faces in cookies and pancakes.  The brain wants to spot patterns – it is desperate for pattern detection.  When you take a big ol’ bunch of data – like all the state-by-state election results since who-knows-when – and you start to look for patterns, you know what?  You are going to find a few patterns.  Some of those patterns may hide a great predictive model – this is what well-cross-validated discriminant analysis is all about.  However, just digging in the data to find a pattern – any pattern – is a fool’s errand.  You will always find some alignment that looks like it’s predictive, but in reality, it’s just the randomness of life at work.  What real predictive capability takes is a lot of work on defining the precise criterion you are measuring, a lot of construct validity checking, and a lot of work on cross-validation against the criterion, etc.  Oh, and a little solid theory on the relationship between the independent and dependent variables wouldn’t hurt either.  In other words, sometimes what you see is just a nice reflection in the mirror, but there is nothing on the other side.

Missouri is a cool place.  And Missourians may be so close to the pulse of America that their collective voice echoes the blended wishes of the other 49 states.  But I don’t see the proof.  I don’t see the validity that I would want to see, and, in my opinion, Missouri is not a valid oracle for the American pulse, even if it names the right presidential party for as long as I live.   And that pancake that you keep locked in that safe is not a spitting image of Barney Fife.  Sorry, it’s just not real.  Come back to me with a rock-solid, statistically sound case on why Missouri is “the bellwether” – so called – against the loose criterion “How America Will Decide” and we can talk.  Do the hard work, show me how you cross-validated your predictive, discriminant model, and prove your point, and I will back down.  My stats are rusty – you may be able to take me.  But, don’t come to me with some interesting presidential election patterns and say that proves Missouri is the guide to all-things-American. Until then, just enjoy the pretty pictures randomness can draw.

P.S. Don’t get me started on “historically the stock market has always outperformed other investments” statements being used to predict a rosy future either… not that you would anytime soon.