gcm.counterfactual_samples for multiple and related variables

[bernddude]

Hello i am trying to perform a counterfactual analysis in which i try to evaluate the changes of two related variables .
For example i have two variables (stress factors) in the model and i like to understand their role .
Lets say there is a total stress of 1 and it could be broken down in a = .8 and b = .2 or it could be a = .2 and b = .8
Now in my case the total stress varies for each item in the model, so i can't just hard code items.
Is there a way to do this via the current dictionary approach ? I didn't see a solution ?

any advise would be very welcome

[bloebp]

Hey,

I am not sure I fully understand how your data looks like. Do you always have "a" and "b"? And, do you know their values for each item?

[bernddude]

Thank you for the response and sorry if my explanation isn't clear.
What i am speaking of is a case in which two or more variables of the model might be compared.
Say for example you have a machine and could either route a process through option a or b and i like to compare which of these options has different side effects on a target value under investigation.

So i might set up a counterfactual model which takes the load of path a m sets it to zero and adds the original value of a to the path b.
However, a and b are not fixed value , they are different for each entry in the data set.

I did actually find a way to make this work, however it isn't the most elegant way.
I simply created a dictionary like mapA2B = {a:b for a,b in zip(valuesA,valuesB} and
used that in the lambda function of the counterfactual routine .
I made all values unique by ensuring a small random offset

as i said not elegant, however it did the job for the study i was interested in

[bloebp]

You can simply do something like:

{'a': lambda a: a + offset_a, 'b': lambda b: b + offset_b}

by referencing the a (and b respectively), you can use the actual values and add an offset or take a fraction of it etc.