Social science, done honestly

William Dickens has spent the last several years thinking hard about racial differences in IQ. The stylized facts didn't make much sense to him, taken together:

• Large differences in average intelligence by race (about 1 standard deviation)
• High estimates of heritability of intelligence and little discernible lasting effect of environment
• Rising intelligence over time (the Flynn Effect).

If variance in intelligence is largely explained by genetics, not by environment, how can we explain the Flynn Effect, which happens far more quickly than could be explained by any evolutionary story?

Dickens worked with Flynn to come up with a model that allows for complex interactions between genes and the environment that would fit the three stylized facts. The model works as follows, and is nicely described here. Kids with small initial differences in genetic endowments sort into more or less cognitively demanding environments. The ones in the more cognitively demanding environments reinforce their initial advantages. Because genes cause the initial environmental sortition, all then loads onto genetics in variance decomposition. But the observed differences are still a function of both genes and environment. And, if one group systematically has poorer access to cognitively demanding environments, then large group differences can be observed despite there being small underlying genetic differences. If the average environmental cognitive complexity increases for all groups over the period, we get the Flynn Effect.

Cute model and seems plausible. Bill then took the model to the data. Or, rather, took one implication of the model to the data: a full test would require some rather expensive experimentation. If the model holds, then a decomposition of variance into transient environmental effects, permanent environmental effects, and measurement error should show that at least some differences across siblings come down transient environmental effects. Instead, he found most of the environmental variance is permanent with the rest being measurement error (see slide 43 at the above link).

BUMMER! (Did I mention that one of the “advantages advantages” of tightly linking your theory to your empirical work by modeling is that you can be proved decisively wrong?)

Moreover, differences across identical twins in childhood are almost entirely measurement error. So if one twin randomly is assigned to a more cognitively demanding environment in childhood, we can't see any effect of it in the data.

And so Bill heads back to the drawing board, working on a new model where folks switch environments to match the one that meets their current level of cognitive ability, but where there's noise in the process and fewer opportunities for change as time goes on.

That's how social science, or any science, is supposed to be done. Hypothesis leads to test leads to starting over if the hypothesis is rejected. DeLong and Lang worry that all economic hypotheses might well be false because of publication bias; Ed Leamer worries that folks choose the method that gives them the results they want. Nice to see folks who work against both. And, entirely consistent with other observations on Bill, who visited here as an Erskine Fellow a couple of years back and who derived results on a scratch-pad while role-playing a low-intelligence half-ogre named Grissumpf in Bryan Caplan's excellent all-economist D&D campaign (I was a sage-assassin who threatened his way into sage school, charisma of 4). Good times.