[FEA] Support Generalized Adjustment Criterion for Estimation plus Add Example Notebook

[nparent1]

Adds support for the generalized adjustment estimands at the estimation stage. The PR primarily does two things:

• Abstracts the get_backdoor() method of the IdentifiedEstimand class to instead be a get_adjustment_set() method. Then if the identifier_method is specified as 'general_adjustment', the get_adjustment_set() method will return the general adjustment set (rather than explicitly a backdoor set). Then throughout the repo where get_backdoor_set() was called for estimation purposes, we instead call get_adjustment_set(). Note that because we have get_adjustment_set() default to the backdoor set, the overall behavior is the same as before, unless a user explicitly sets general_adjustment as the identifier_method.
• Adds various tests and test utils. In particular, adds support for specifying explicit graphs in the estimation tests, for more specific tests of the general_adjustment criterion.

An example notebook is also added, dowhy_generalized_covariate_adjustment_estimation_example.ipynb

This is a follow-on PR to #1292

Issue Ref: #402

[nparent1]

This will need to be rebased once #1292 merges

[nparent1]

@amit-sharma - this is a follow on to #1292 to add support for general adjustment for estimation. It is ready for review

[nparent1]

Tagging in @bloebp as another possible reviewer!

Also - it seems like I don't have permission to actually request a reviewer in the top right

[amit-sharma]

thanks for the reminder, @nparent1 Was travelling and could not get to this sooner.

The notebook example is very clear and well-written. Thanks for adding.

To clarify, even if we miss changing get_backdoor_variables to get_adjustment_set in some places, the code should still work if it was assuming backdoor, right?

I'm just trying to understand how wide these changes would apply and if there is a risk of breaking existing functionality.

Currently there is an instrumental variable estimator test failing---can you take a look at that in the build_docs run? Adding other comments inline

[nparent1]

Thanks for the review @amit-sharma, regarding your questions

To clarify, even if we miss changing get_backdoor_variables to get_adjustment_set in some places, the code should still work if it was assuming backdoor, right?

I'm just trying to understand how wide these changes would apply and if there is a risk of breaking existing functionality.

Correct - anywhere get_backdoor_variables() is being called should still work exactly as before, since we are still storing the backdoor variables exactly as before on the identified estimand. This PR allows users to set the identifier method to use the general adjustment criterion instead, but unless they do so everything should behave exactly as before

Currently there is an instrumental variable estimator test failing---can you take a look at that in the build_docs run? Adding other comments inline

Thanks for calling out - that should be working now. Fixed in commit ff6c08a

That said - there is still a build failing but it looks like potentially a fluke? It says "ModuleNotFoundError: No module named 'torch._dynamo'", an error I've seen in past PRs (like #1292)

I think this is ready for review again!

[amit-sharma]

Thanks for these changes, @nparent1 Looks good to me.