This week’s edition of ‘Metrics Monday will be a slight departure from the usual post in that I won’t be making any specific point about applied work. Rather, this will be more of a meta-post on econometrics focusing on how econometrics is taught.
The way I see it, econometrics has two general objectives:
- Causal inference, and
- Forecasting or properly modeling the data-generating process (DGP).
According to a new working paper by Angrist and Pischke (2017), the way econometrics is taught needs to be rethought, because although many of the problems economists are currently studying involve causal inference, a lot of the tools and the language that is used to teach econometrics to undergraduates (some of whom will go on to learn nothing else in econometrics after that one introductory course) is a holdover from the days when econometrics was all about forecasting or properly modeling the data-generating process.
Among other things, here is what Angrist and Pischke recommend:
In other words,
I have touched upon a number of these things in the past in this series of posts (see here and here, for instance), but I am often struck by the methodological rift I sometimes witness between younger and older applied economists at seminars and conferences–a rift which mirrors the causal inference vs. proper modeling of the DGP distinction. Likewise, I still too often read manuscripts wherein the authors are clearly interested in the causal effect of one variable but also spend time talking about the IV they use for a control variable, or wherein the reader is never explicitly told: Look, this is our variable of interest, and these are our controls. (As a matter of fact, I reviewed one such paper this morning!)
This isn’t to say that control variables aren’t important–regressions are ecosystems, and at the end of the day, any identification strategy makes a selection-on-observables argument–but they are not on an equal footing with the variable of interest.
Ultimately, Angrist and Pischke argue for an overhaul of the way econometrics is taught at introductory levels, and in the second half of their paper, they look at a number of widely adopted textbooks to make their case that causal inference ought to be taught first.
The way I understand their paper, they are not saying that forecasting or proper modeling of the DGP should not be taught–just that it should come after teaching students the relatively simpler, less arcane, and no-less-useful ideas that underlie causal inference and issues of research design.
The way I see it, many undergraduates will never take another econometrics class ever again. Since our job as teachers of undergraduates is to form responsible citizens, and not researchers, I feel like our students are better served when we teach them not to get hoodwinked by causal claims made on the basis of mere correlations. For me, that is a much more important component of critical thinking than knowing which specification test to apply to a regression.