Question: How/where does this compare to EconML about dowhy HOT 3 CLOSED

thanks @tszumowski for starting this discussion. Yes, you are right---DoWhy and Econml overlap only in the estimate section. DoWhy implements the full process of causal reasoning including model, identify, estimate and refute. In comparison, Econml implements only the estimate step.

In designing DoWhy, we kept a focus on the "ideal" process of doing a causal analysis, which includes identification and more importantly, refutation so that modeling assumptions can be tested. The estimators in DoWhy currently are the standard estimators for causal inference. As you rightly point out, EconML has much more advanced estimators for estimating the conditional average treatment effect (CATE). This is why we are implementing an interface in DoWhy so that you can call Econml methods directly from dowhy's estimate function. Here's an experimental Jupyter notebook to see it in action.

For your second question, there are actually two considerations when designing a causal analysis. One is about the modeling framework, and the other is about the target estimand.

Potential outcomes and structural equations are ways to construct a causal model. Another way is to use structural causal model which is based on a graphical model. The differences between these frameworks are often a matter of detail (and of big academic debate). But in practice, both econml and dowhy are compatible with these different ways of expressing a causal model. In DoWhy specifically, we use the structural causal model framework in the identify step, and rely heavily on methods derived from the potential outcomes framework in the estimate step.

The other consideration is the target estimand: do you want an effect for the full population (average treatment effect, ATE) or for a specific population, e.g., conditioned on "Gender=Female", (conditional average treatment effect, CATE). EconML methods are designed to estimate CATE, which is a subject of active research. Most of DoWhy's methods focus on estimating the ATE so far, although we are extending some of the methods to also estimate CATE.

Hope this helps.

from dowhy.

Wow. @amit-sharma thank you for the fantastic summary. I'm only recently trying to ramp up on causal analysis, coming from a stats/ML/bayesian background. I am noticing what you mean by "big academic debate" as I was getting a bit lost in the various different methods. This clarified a lot for me.

Regarding the EconML integration, I ran into an issue when trying to run that notebook. I'll create a separate issue for that. Otherwise, this ticket can be closed. Thank you.

from dowhy.

we use the structural causal model framework in the identify step

What does it mean by this? Is the identify step used as input for the estimate step? How are we integrating the prior into the estimation?

from dowhy.