Heckman on Causality, Do-Calculus and Econometrics

To give a more neutral context than the paper:

Causality, identification and counterfactual analysis is an increasingly important topic in "all things statistics". For example, it is crucial in AB testing, "lifting" approaches and so on. Furthermore, however, it likely offers the correct theoretical background for understanding which Machine Learning models "generalize", and why.

Given its importance, the spotlight has fallen on researchers engaged with this question. Pearl is a computer scientist, and outspoken "enfant terrible", who developed the DAG approach and do calculus on it. It is a reasonable general approach (see, however, this paper) which however requires "large models" in terms of distributional assumptions. It has proven very useful to understand causality mechanisms, for instance (e.g. front-door and back-door criterion), but I don't know many instances of someone using a full on DAG estimate in, say, an academic paper.

Statistics, of course, also treats this topic. Here, much lingo comes from experiments, because experiments offer simple restrictions on the the DGP that allow us to identify causality. But experiments are restrictive. For instance, AB testing in websites is rarely a good pure experiment. Statistics has therefore developed further approaches - for instance in Biology, Medicine and also Marketing - to treat such cases. As champion I am going to nominate Andrew Gelman, who writes prominently and well from the Stats perspective. Gelman has debated with Pearl and his sphere several times.

Finally, there is a field that has been - since at least the 1940s - concerned with counterfactuals and causality, and never really had experimental data to begin with: Economics. Here, Heckman (Nobel price - econ - winner) is an equally abrasive personality as is Pearl. Econometrics has - perhaps secretly to other fields - developed a lot of the tools we use today. It's impact in causal analysis, structural modeling, time series, natural experiments etc. is clear.

On the backdrop of these streams of the literature, Heckman writes this rather uncompromising paper. On the surface, it is adressed to economists, but we an be sure both Pearl and Gelman will fashion a response.

To see why, let me quote the article in question: "Recent work in computer science has begun to reinvent the logic of econometric forecasting using its own colorful private language but without any fresh insights or acknowledgement of a large body of econometric thought"