The Monkey Cage

But there’s no proof of causality or lack thereof in what’s shown here. We’ve got heaps of omitted variables and these are just OLS regressions. Causality shouldn’t even be mentioned, not by Crook or Galston or Sides.

For more on causation & correlation see my post (just one of many I’ve done on the subject): http://theincidentaleconomist.com/causation-without-correlation-is-possible/

Austin: Sure, fine. In this case, as with many questions in the social sciences, there can be no randomized controlled trial. So we’re left thinking about the significance of correlations, or lack thereof. I think the evidence I’ve accumulated in this post is useful on that score.

A randomized controlled trial isn’t necessary. Look into instrumental variables. In fact, read my blog. I’ve written about them before and next week will have some tutorial posts on the subject. The following is a start, but it’s advanced. Check back on my blog on Monday (two relevant posts are scheduled, one in the AM, one in the PM).

http://theincidentaleconomist.com/what-took-con-econometrics/

Thanks for the incredibly valuable advice on instrumental variables. I think it would be fair to say that neither John nor any other statistically inclined political scientist has ever heard of them. This is really, really exciting news. Perhaps we can write “Freakopolitics?” Or maybe, not so much.

Ah, the scatterplot takedown. I saw a conference presentation about this sort of blog post just last week.

Nice piece, John. Kudos.

I’m not a statistician, but aren’t you comparing two populations? The mistrust data seems drawn from the general public and the polarization data seems to be taken from Congressional members.

Austin - political science isn’t that backwards. I’m sure John knows about IVs - they’re part of the standard econometrics sequence in most better Ph.D. programs in polisci.

They’re also, in many ways, very 1990s… not in the sense that they are never useful (they clearly are) but in that they’re not a panacea to identification, as it’s incredibly hard - and very, very rare - to find an IV that actually fulfills 2SLS assumptions. And using bad instruments is worse than not using them at all. Both economics and polisci shared a period of wild enthusiasm about structural equation models as the solution to identification problems in general, an enthusiasm which has been replaced by a much more cautious (and less widespread) use. (and I think Angrist and Pischke would agree with that - that’s why they’re emphasizing research design - something that seems hardly possible here - or can you think of a “natural experiment” with trust???

In addition, take the limited number of possible data points, the time-series nature of the data - and it’s pretty clear that you can’t get clear identification out of this question.

So we go to step two: We ask ourselves: “What type of story fits with the data and the type of associations that we see?”
And the associations clearly support one story (econ growth) and throw serious doubts on another one (polarization). Sure, it’s still possible that it’s the other way around - but very, very unlikely.

Finally, remember that polisci (at least of the type that John does) lives in a (roughly speaking) Popperian Universe, in which we’re concerned with falsification and particularly with avoiding type II errors.
So if someone wants to make a claim about the relationship between polarization and trust - s/he needs to go ahead and present convincing evidence - but as John demonstrates, the evidence isn’t even suggestive of such a relationship.

Sebastian - Very well put. I think we got our signals crossed though (internet—grrr…). My first comment was a standard caution against overuse of causality type language, here and in general.

My second was not meant to imply that one can always find good instruments. It was merely a caution not to give up so easily on questions for which no randomized trial can be conducted. To the extent one can find instruments, one should put them to use.

I did not mean to suggest instruments exist in this instance. But, I have given some thought as to whether any might exist here, and I agree it is a challenge to think of some obvious good ones. Perhaps if polarization is measured narrowly (something to do with party behavior) while trust is measured broadly (something to do with the populace), one might imagine a shock to the former that doesn’t immediately affect the latter (like a profound but subtle change in parliamentary procedure that only the inside players understand). Or, thinking about state or local government, one might hypothesize that trust is more highly correlated across political boundaries but some appropriately narrow definitions of polarization are not (is the polarization of Ohio’s legislature highly correlated with that of Utah’s?). If that’s a reasonable hypothesis then one might use trust in neighboring political regions as an instrument for local trust, one that is perhaps arguably uncorrelated with local polarization.

You can tell I’m straining, though this idea of using neighboring region values is a common trick in industrial organization and can sometimes be justified and tested. Polysci isn’t my field, so I do not want to make any claims about it or my ability to say anything sensible (I’m a consumer of it, not a practitioner).

Meanwhile, on technical grounds alone, if one is interested in a gentle introduction (or refresher) of IV technique, it is something I’ll be covering Monday and beyond.

I failed to add that there have been some recent developments that move IV beyond 2SLS and into nonlinear models, so satisfying 2SLS requirements is not necessary. They’re surprisingly simple to implement. Again, they’ll be covered in my posts.

Research with which I have been involved suggests that polarization is likely to be more associated with trust in government.

When analyzing the relationship between the value of clean lakes and rivers (through government action to clean them or keep them clean) with political affiliation, Democrats had higher values than Republicans, but both had higher values than Green party members or members of other parties.

It is likely that partisans whose parties have some hand in controlling government tend to trust the government more, not less, than non-partisans or unaffiliated citizens.