Dubbers Asks: The Living Bejeebers

Is foreign intervention euvoluntary? Not according to Chris Coyne. He's found that even using the barest criteria for success forwarded by State Department elites, foreign adventures fail to meet the standards proposed. One cynical way of looking at it is that Coyne felled a lot of trees and spilled a lot of ink proving something everyone already knows: politicians lie. This is hardly a revelation.

What is a revelation is to whom they're lying. And it's a further revelation who it is that endeavors to believe those lies.

On Twitter, to Thy Dungeon Master, Dubbers asks a simple question: "how does war enthusiasm relate to income? My prior: negative correlation." The freight on this question is stamped "GSS" and it so happens that I know how to skipper that vessel, folks. I offered to do the leg work, and I also admit that my priors were the same as Carden's: high income means more aversion to war.

I was disappointed to discover that the GSS doesn't have that specific question (that I could find anyway; there are more than 5000 variables in that bad boy), but it does have a few plausible proxies. I went with USINTL, "take active part in world affairs" with a binary output: [take active part | stay out].

More below the fold. Lots more. Warning: what you'll find below may look pretty technical, but I promise I'll make it as easy as possible for the average reader to understand.

Here are the results of a plain-jane linear regression. Most of the terms should be self-explanatory. The DV is a pseudo-dummy for "should the US take an active part in world affairs?", with 0 being "stay out" and 1 being "take active part". I switched the order of the responses in my constructed variable for clarity. That's why the name you see in the table is "iusintla" instead of just "usintl". The only independent variable that should raise eyebrows is WORDSUM, which is a scaled vocabulary test, sort of a proxy for IQ. Get 10 words right and you're all erudite or something. 0 words right means you're either a yahoo or a troll. Not that there's anything wrong with that.

As for the columns in the table, the coefficient represents the linear relationship between the row variable and the outcome (whether or not the person supports more foreign intervention). Take Age, for example. A coefficient of 0.013 means that for for each year of age, the respondent is 1.3% more likely* to be of the opinion that the US should interfere around the world. The "Robust Std. Error" column helps give error bands for the coefficient (and yes, I know there are better methods for using survey data, but this is just a blog post, please calm yourself). The t-score lets us know how many standard deviations the estimate is from the null hypothesis that the coefficient is zero; a big absolute value means more statistical significance. The "P>t" column translates the t-score into a probability. The important one for our purposes is the beta estimate. This gives us a nice standardized way to test the relative magnitudes of the effects. You see, one of the problems with comparing job status to political party affiliation or to age is that it's totally apples to oranges. Beta estimates take a step to making everything a batch of pluots. It finds how far from the mean the intensity of the effect is. Age has a beta estimate of .366, meaning that it's a pretty strong effect, particularly compared to most of the other covariates. Note also the beta estimate for Age squared, which means that there are big ol' non-linearities in there too. You see, statistical significance can only tell you if the effect is indistinguishable from zilch. The beta estimate tells you if the effect is strong. Careful, thoughtful analysis tells you whether or not your results are meaningful. If you've not spent much time refereeing papers, you'll find that far too many authors make the cardinal error of conflating these things. My advice? Read more McCloskey.

*this isn't exactly the right interpretation, since OLS doesn't return probabilities, but what OLS does do is give beta estimates. There's a tradeoff here, and I apologize to anyone whose econometric sensibilities I have offended.

Linear regression				Number of obs	= 4591
				F( 39, 4551)	= 15.24
				Prob > F	= 0.0000
				R-squared	= 0.1090
				Root MSE	= .41999
iusintla	Coef.	Robust Std. Err.	t	P>t	Beta
Age	.0129988	.0032206	4.04	0.000	.3658341
Age Squared	-.0001309	.0000365	-3.59	0.000	-.3205856
(log) Real Income	.0309943	.009696	3.20	0.001	.0537577
Female	-.0495057	.013237	-3.74	0.000	-.0558453
Race (white)
black	-.1165278	.0226521	-5.14	0.000	-.0864812
other	.0279967	.0388172	0.72	0.471	.0103651
Work Status (full time)
part time	-.0067998	.0181958	-0.37	0.709	-.0057201
other	-.0274666	.0341774	-0.80	0.422	-.0111287
Degree (less than hs)
high school	.1134515	.023549	4.82	0.000	.1271605
junior college	.1454613	.0334162	4.35	0.000	.0794924
bachelor	.193755	.0270526	7.16	0.000	.1639202
graduate	.1756396	.0303617	5.78	0.000	.1095078
Wordsum (see discussion)
1	.0653227	.1040948	0.63	0.530	.0158982
2	.0152426	.092295	0.17	0.869	.005215
3	.0791207	.0863169	0.92	0.359	.0390428
4	.1709022	.0834355	2.05	0.041	.1128714
5	.194587	.0821057	2.37	0.018	.1600027
6	.2362381	.0815767	2.90	0.004	.2215347
7	.2830815	.0817545	3.46	0.001	.2413197
8	.2652059	.0824434	3.22	0.001	.1841871
9	.2756902	.0825786	3.34	0.001	.1774822
10	.3082281	.0827772	3.72	0.000	.1743316
Likely to lose job? (very likely)
fairly likely	-.0284648	.0406777	-0.70	0.484	-.0144701
not too likely	-.0001526	.0320099	-0.00	0.996	-.0001481
not likely	-.0026678	.0304863	-0.09	0.930	-.0028712
Political Party (strong democrat)
not str democrat	-.0287365	.0224446	-1.28	0.200	-.0271258
ind, near dem	-.0228068	.0251095	-0.91	0.364	-.0174511
independent	-.0437312	.0269507	-1.62	0.105	-.0312856
ind, near rep	-.0428694	.0275389	-1.56	0.120	-.0299344
not str rep	.0039907	.0238844	0.17	0.867	.0035133
str republican	.0470049	.0272599	1.72	0.085	.0306234
other	.0289268	.0690152	0.42	0.675	.0054327
Political Views (extremely liberal)
liberal	.0107359	.0433978	0.25	0.805	.0079434
slightly liberal	.0339694	.0424106	0.80	0.423	.0278215
moderate	-.0001004	.0415717	-0.00	0.998	-.0001092
slightly conservative	.0341171	.0426605	0.80	0.424	.0292143
conservative	.024637	.044173	0.56	0.577	.0188384
extremely conservative	-.0778241	.0604246	-1.29	0.198	-.0261403
Dummy variable for (D) President	-.0055567	.0131418	-0.42	0.672	-.0059967
Constant	-.1845453	.1430876	-1.29	0.197	.

Don't sweat the stuff at the top. I left it in in case curious readers are interested, but if you don't know what an F-statistic is, it won't hurt. At any rate, let's get to it! First off, note that Carden's tweet is crushed. Just annihilated. The sign on real income is positive and the beta estimate is pretty darn strong (relative to the other terms, of course). Mo' money, mo' intervention! A surprise upset! Now, I do have to admit that my priors were the same as the WMOE's: I expected higher income folks to be more wary of foreign intervention, but okay, I can accept that reversal. The one that really shocks me is WORDSUM. Great gobbling googly-moogly! Smarter people are more in favor of foreign intervention. I really did a double-take of these results. The caricature I admit to having was the snooty Eastern elite scoffing at the idea of nation-building coupled with the hick Southern yokel waving a flag from the back of a comically oversized pickup truck. Boy was I wrong. Check out the coefficients on WORDSUM. Positive, the lot of 'em, and increasing the further up you get. Here's a plot of the margins I took from a probit regression with a similar specification:

Fig 1: Wordsum

How about that. Even correcting for a typical suite of controls, smarter people are more likely to favor foreign intervention. This is a complete reversal of my priors. Complete.

Ditto for income. Here are the marginsplots for the uncontrolled and the controlled margins on real income:

Fig 2: Uncontrolled Income

Fig 3: Controlled Income

Yes, the error bands are wider on the left side of the controlled regression, but this is consistent with many of Caplan's findings. Particularly on WORDSUM, we find that low information voters tend towards inconsistency. This is just more evidence of what we already know.

The last thing I looked at (for now) was a partisan hypothesis: maybe folks' support for intervention had something to do with who was in office at the time. We know that the anti-war left became eerily silent after Obama took office, so is there something systematic about partisan bias with regard to the interaction of Oval Office holder and respondents' party identification or ideology? Take a look:

Fig 4: Ideology # President

It's tough to see here in the post editor, but the blue line is the margins on responses with a Republican in the White House (starting with Tricky Dick in '72) and the red line is for a Democrat in office. In retrospect, I should have managed the colors better, but I didn't, and now you all must bear the price. Double apologies.

I got basically the same marginsplot for party affiliation. No need to double up unless I end up turning this into a proper article. 

So what can we make of this for EE folks? What's the nature of the exchange when governments interfere abroad? We might think that business elites could extract some special rents obtained at gunpoint, we might make a Bootleggers and Baptists point about the nature of aid dollars, we might even tell a story of neo-colonialism. But whatever the story, what these results suggest is that all else equal, it's the brightest and the richest among us that most strongly support those policies that have been shown over the course of centuries to be the most dismal failures. People, smart people, appear to believe the kayfabe that is politics. Wow.

Rational irrationality, amplified.

Non-euvoluntarity hot on the heels of intelligence.

This is the bullshit we need to tax. This is the bullshit that causes untold misery around the world. Would betting markets bend that curve downwards? My priors say "yes", but I'm willing to bet I won't find out any time soon.