Piketty Makes “Typos,” Bryan Caplan Cites “Noise”
This is odd. Bryan Caplan apparently wrote in his book on irrational voters:
There are countless issues that people care about, from gun control and abortion to government spending and the environment… If you know a person’s position on one, you can predict his view on the rest to a surprising degree. In formal statistical terms, political opinions look one-dimensional. They boil down to roughly one big opinion, plus noise.
So Weeden and Kurzban decided to test that out, and claimed Caplan was wrong.
Now the odd part. Here’s how Caplan responded: “Just one problem: I’m well-aware that the data aren’t tidy. Accurately predicting individual opinions is hard. I deliberately included the words ‘plus noise’ to ensure that readers knew I was not claiming great predictive powers” (emphasis original). Caplan concluded his post by saying, “If Weeden and Kurzban really wanted to dispute the one-dimensionality of political opinion, they should have been correlating specific issue views with ideology, not specific issue views with each other.”
I was going to expound on just how crazy this was, but in the comments Weeden himself chimes in:
This is just getting weird. You explicitly said of individual issues that “If you know a person’s position on one, you can predict his view on the rest to a surprising degree.” And we ran a test of that proposition, showing that it can be pretty weak stuff, depending on the issue pair. (I’ve got a blog post showing similar points here:http://www.pleeps.org/2014/06/30/if-being-routinely-liberal-or-conservative-is-a-human-universal-why-is-it-true-only-of-recent-college-educated-whites/)
But now you’re saying we should have known we were testing wrong thing! We shouldn’t have tested the thing we quoted you as having said, but, instead (obviously!) we should have tested something else…
It almost sounds like Caplan considers “ideology” the “one big belief” to which he refers. In that case, he’d effectively be saying, so it seems, that self-identified political affiliation predicts individual policy positions. It’s pretty trivial, but presumably accurate for a majority policy advocacies. In that sense, it might just be two trains passing in the night: it would make the case for Weeden and Kurzban reasonably misinterpreting a piece of disembodied text, but it would also trivialize a seemingly large part of Caplan’s thesis. Hard to say what’s really going on.
Bottom line: when Bryan reaches into a drawer for a t-shirt, he just doesn’t care what he grabs!
Weeden’s piece identifies a significant correlation only among white bachelors degree 1994-2012. The Pew research article identifies a doubling of the proportion of Americans who consistently express liberal or conservative opinions over the last two decades. This seems consistent.
The Pew piece uses 10 political values, whereas Weedon used only two, going back 35 years to a period where opinions were less polarised. Given the overlap apparent in the Pew graphs, it is not too surprising that there is little correlation. The Pew piece is a much better way to represent the polarisation.
Caplan did not say if you knew one opinion you could predict one other opinion – he said “the rest”, suggesting to me some aggregating of the opinions. However, Weeden is right to point out that this does not mean every opinion correlates with every other individual opinion, and Caplan has probably phrased it in such a way that suggests a stronger correlation than is actually the case. However, Weeden is attacking a straw man when he questions whether being liberal or conservative is a “human universal.” He says he will “cut to the chase” when he selects abortion and income redistribution as the sole representative opinion of liberalism and conservatism, but that is not cutting to the chase but over-simplifying at best, cherry picking at worst. This is cutting to the chase of proving what nobody said was true anyway. The Pew research clearly shows large overlap, although diminishing. The other research he cites says in the abstract “In this article, we argue that one organizing element of the many differences between liberals and conservatives is the nature of their physiological and psychological responses to features of the environment that are negative.” – notice that one element of many – hardly a human universal.
Aggregating of opinions?
Oh I get it. If you tell me that you are in favor of abortion, then I can predict to a high degree of accuracy, notwithstanding some noise, that you don’t like to eat 40,000 different varieties of poop.
I’ll be at least 40,000 for 40,010 wouldn’t I?
Not quite sure what you mean. As far as I am aware, the only poop that is consumed with great delight and considerable expense is Kopi Luwak – or civet cat coffee, although I have not tried it myself.
If I tell you I am in favor of progressive taxes, you may be able to predict some other preferences of mine. Maybe not all, and you may be wrong about all of them. But if you ask 100 people with that preference, then grouping you predictions according to “liberal” and “conservative” will give you better than chance over the range of questions. Caplan does say your success will be “to a surprising degree”, I suppose technically that would include a surprising lack of correlation too.
If those political opinions didn’t look one dimensional before being reduced to a simple aggregate figure; they very likely will do afterwards.
If strict parenting and rigid social conditioning can encourage a whole generation of people to react against it (e.g. the beatniks and then hippies) then surely strict and rigid mathematical equations that tell people what they will do, can do the same, thus ruining the predictability of the equation.
I had interpreted Caplans postion as something like, “There exists a position that people hold that allows you to know most position that they hold, with a large amount of certainty”. In the language of machine learning, it would be ” The largest vector in the Priciple Component Analysis breakdown contains most of the information”.
However Caplans reply didn’t clarify anything for me as to what he means.
Caplan says:
“Specific issue views, like individual items on the SAT, are very noisy. Ideology, like overall SAT score, is not so noisy. When you correlate two noisy things with each other, you get a really tiny correlation. When you correlate one noisy thing with a not-so-noisy thing, you get a moderate correlation. If Weeden and Kurzban really wanted to dispute the one-dimensionality of political opinion, they should have been correlating specific issue views with ideology, not specific issue views with each other.”
So when he says “If you know a person’s position on one, you can predict his view on the rest “, then if one can accept that by “view[s] on the rest ” he means someones “ideology” , rather than his “specific issue views”, he is being consistent (if a bit unclear) here.
OK thanks, you guys have convinced me that Caplan wasn’t speaking nonsense. But, I think his defensible claim is pretty obvious and hardly “surprising” as he said. I can understand why Weeden honestly thought Caplan was claiming something else.