Elias Bareinboim (Columbia) – Causal Data Science - 2019
Details
Title : Elias Bareinboim (Columbia) – Causal Data Science Author(s): FODSI Link(s) : https://www.youtube.com/watch?v=wPYFuIgad_4
Rough Notes
One of the current challenges is that all data are not created equal, there are:\
- Differing experimental conditions.
- Different underlying populations.
- Sampling procedures that are not random.
- Treatment assignments that are not random.
- Many variables that are not measured.
Good news for us - these problems are now formalized as causal nference problems - there are conditions and algorithms to decide about what is entailed from certain data collections.
Some common inference tasks:
- Statistics - Descriptive : Going from observational samples to the observational distributions.
- Statistics - Experimental : Going from interventional samples to interventional distributions.
- Causal inference from observational studies - Going from observational samples to interventional distributions.
- Experimental inference (generalized instrumental variables).
- Sampling selection bias.
- Causal Transportability/External validity.
(Elias's opinion - causal inference from ~1974-2011 was concerned with confounding bias, which has been solved by Rubin, Robins, Dawid, Pearl).
Let's focus on transportability (extrapolation and robustness of causal claims). The research question here is: "Is it possible to compute the effect of \(X\) on \(Y\) in a target environment \(\Pi^*\) (where experiments are not feasible), using experimental finding from a different environment \(\Pi\). The most common example of this is when moving from the lab to the real world. Another important point is that the interesting things happen between the extreme cases of the real world being exactly the same as the lab environment and the real world having everything (i.e. all causal mechanisms) being different from the lab.
But in the real world, we have some mechanisms being the same across different environments - e.g. biology remains the same in Los Angeles and New York. The speaker gives a concrete example, where the causal effect in \(\Pi^*\) can be computed from experimental data in \(\Pi\), via a formula called the transport formula/recalibration.
A causal quantity \(Q\) is transportable from \(\Pi\) to \(\Pi^*\) if and only if there exists a do-calculus reduction of \(Q(\Pi^*)\) to an estimand that is a function of the observed distributions.