Thread by @amt_shrma

Great to see the discussion on using DAGs versus potential outcomes for applied causal inference work with @yudapearl, @Jabaluck @autoregress @analisereal @PHuenermund

My take: DAGs and PO are compatible, and the best analyses benefit from using both. 1/7

In a KDD tutorial with @emrek, we outline how you can use DAGs and potential outcomes together for causal analysis and discuss empirical examples. 2/7 causalinference.gitlab.io/kdd-tutorial/

In fact, they are not only compatible, we can always represent a proof in one framework in another. For an identification strategy in a recent paper on recommendation systems, I was able to write the same proof using DAGs and without DAGs.
projecteuclid.org/euclid.aoas/15…

@analisereal and @autoregress provide another example of equivalent proofs using DAGs and PO. 4/7

https://twitter.com/analisereal/status/1113225543023861760

While DAGs & PO are compatible, they do have their strengths. DAGs are most useful for causal identification, offering a transparent way of representing & checking assumptions. Potential Outcomes are most useful for estimation, as most statistical methods are built around it.5/7

In my own work, I have benefited immensely from learning both DAGs & potential outcome frameworks. Knowing the tradeoffs, I can't imagine starting with the PO framework for identifying a causal estimand, just as much as I wouldn't recommend starting with DAGs for estimation. 6/7

If you want to see how DAGs and PO can work together in practice, @emrek and I have created a Python library based on this. Key idea is to separate identification and estimation steps. DAGs for identification, PO methods for estimation of causal effect. github.com/Microsoft/dowhy