Bryan recently put up two interesting posts, one on personality, the other on econometrics.

On personality, he points to some analysis saying that the correlation is higher between certain personality traits and life outcomes than it is between socioeconomic status and life outcomes. On econometrics, he joins a discussion with fellow George Mason economists Russ Roberts and Tyler Cowen on whether sophisticated technique, such as instrumental variables estimation, is ever persuasive.

On the latter point, I agree with Russ. The way I look at it, instrumental variables are an attempt to deal with problems, such as possible reverse causality. However, the sophisticated techniques rarely solve the problem completely, and they often introduce worse problems. The use of instrumental variables is to me a signal that the data are not going to prove decisive in answering the question at hand.

Basic facts and simple correlations can be convincing. Instrumental variables estimates are tools for sowing doubts.

Unfortunately, I have not been able to locate the paper to which Bryan referred in his post on personality. The results are interesting, but we should beware of confirmation bias (contrary to what Bryan says, the results tend to confirm my own biases).

However, one of my biases is that there are strong interaction effects and nonlinearities that determine individual outcomes. Those interaction effects are obscured by simple correlations. For example, my guess is that the interaction between IQ and the various personality traits matters. Show me someone who is very disagreeable and neurotic with a low IQ, and I will show you a homeless person. Show me someone who is somewhat disagreeable and neurotic with a very high IQ, and I will show you a college professor.

I would expect extraversion to have a positive effect on outcomes when it coincides with high IQ and high conscientiousness, but it might have a negative effect otherwise. There probably is a similar interaction among openness, IQ, and conscientiousness.

It would be really interesting to get hold of one of these large datasets and try some nonlinear specifications that include interaction effects. However, as with instrumental variables, introducing nonlinear specifications could sow doubt and confusion rather than provide persuasive evidence.