Thread by @guilhermejd1

A lot of pathbreaking work (see recent Nobel Prize!) has focused on estimating the local average treatment effect (LATE). Two underappreciated points: LATE might not be the quantity you want; other causal quantities might be easier to describe when key assumptions fail. 1/N

Angrist, Imbens & Rubin (1996) show LATE describes a very particular type of causal effect among “compliers,” those who accept/reject treatment, X, as instructed by encouragement, Z. This narrow focus is what makes the estimand “local.” 2/N

In some cases, researchers are interested in the LATE. In others, they resort to targeting it because unobserved confounding prevents point identification of the average treatment effect (ATE), as in the DAG below. 3/N

But a point often missed is that even when such confounding consists, you can often bound the ATE. And when the strong assumptions needed to estimate the LATE fail, we show it can be easier to describe the range of possible ATEs than it is to say anything about LATE. 4/N

For instance, see

https://twitter.com/jonmummolo/status/1447541370856656896

5/N

The reason LATE is so hard to recover is that it is, in the parlance of @yudapearl’s causal hierarchy, a “Layer 3” or “counterfactual” quantity. In contrast, the ATE is a “Layer 2” or “interventional” quantity. There is a subtle but important difference. 6/N

With Layer 2 (interventional) quantities, we ask how units respond to treatment, e.g. “What happens if I take aspirin?” With Layer 3 (counterfactual), we ask hypothetical questions *contrary to fact*, e.g. “What would have happened if I took aspirin *even though I did not*?” 7/N

See Elias Bareinboim et al for the PCH theorem. causalai.net/r60.pdf 8/N

Because LATE considers “compliers,” it is asking for the treatment effect among those who were encouraged to take treatment and accepted it *but would not have if they were assigned control* and vice versa. This makes LATE a Layer 3 quantity. 9/N

Intuitively, it is much harder to answer a question about counterfactuals than interventions, as we cannot simply run an experiment to identify it. In practice, this means that the assumptions needed to identify the LATE are strong. 9/N

And when they fail, it is hard to even informatively bound the quantity. 10/N

In contrast, even when some LATE assumptions fail, we have found it is sometimes possible to say something informative about the ATE (i.e. sign it), which may be more interesting anyway, since researchers often resort to the LATE when the ATE is not identified. 11/N

Causal inference forces us to ask “what is your estimand?” This is one of the greatest contributions of the credibility revolutionaries rewarded by the Nobel this week. As techniques for causal inference w/ observational data mature, this question will remain front and center.N