,
10 tweets,
2 min read
Read on Twitter
This post gets to the heart of my disagreement with @yudapearl. I think this accurately describes the difference between epi and econ. But in this case, I think econ is right! The identification problem is NOT principally about determining the logical consequences of assumptions.
DAGs are a useful framework for doing that, but this is not a recipe for clean identification. Good identification comes from searching institutional and contextual knowledge to find a "ready-made" DAG, such as a variable which we know satisfies the assumptions of a good IV.
Economists are trained to search for these "ready-made" DAGs. This is is what leads to clean and transparent identification strategies. An identification strategy should come first, then, if necessary, the DAG should appear.
Contrast with an alternative approach: simply state our assumptions and let DAGs do the work. If we don't have identification, make more assumptions. This is the road to identification hell (population: nutritional epidemiology).
Take Elias's seemingly innocent question: if you can't solve a 4 variable DAG, how can you expect to solve a 100 variable DAG that arises in practice?
I would (and did) counter: no one should ever write a 100 variable DAG! It is hard enough to defend *one* exclusion restriction. With a 100 variable model, you are making thousands of guesses and hoping they happen to work out.
Far better to write models that rely on one or a small number of exclusion restrictions that can be interrogated using contextual and institutional knowledge and whose robustness can be graphically assessed by plotting the raw data appropriately.
Now, it might be that in addition to having a mental model of IV in their heads when they search for a "clean DAG", economists should also have a mental model of the "front-door criterion".
But before we get to that stage, we will need many real-world examples where the relevant assumptions seem supportable.
To me, the strongest case for DAGs so far is in helping think through what to control for once we have an ID strategy. But even then, there is work to do: find real examples where we should omit controls due to M-bias. I suspect they exist and economists should pay attention.
