EgorKraevTransferwise
commented
2 years ago To elaborate, the set A below could be represented via an outcome_modifiers init arg and class member in the CausalModel, and adjusting the DML classes to take an additional optional argument that would allow to use different feature sets for y_model and t_model.
Outcome modifiers could be processed in `CausalModel.init() like this:
if model_from_graph_string:
if outcome_modifiers is not None:
logger.warn("Model supplied via graph, ignoring the outcome_modifiers argument")
outcome_modifiers = [n for n in incoming_edges(outcome) if n not in common_causes]
if outcome_modifiers is not None:
if effect_modifiers is None:
effect_modifiers = outcome_modifiers
else:
assert all([e in common_causes + outcome_modifiers for e in effect_modifiers])
else:
if effect_modifiers is not None:
outcome_modifiers = [e for e in effect_modifiers if e not in common_causes]
amit-sharma
emrekiciman
First of all, let me say how awesome I find the integration between econml and dowhy - I very much prefer expressing my problem in the graph language, and then choosing the right EconML machinery to solve it.
But it seems to me the conflation of variables (let's call them A) that in the causal graph are connected to the outcome only (rather than both treatment and outcome) with the X variable in double machine learning is actually incorrect. It is fine to think of them as 'effect modifiers', so have the treatment effect on outcome be conditioned on them.
However, it contradicts a causal graph that only connects A to the outcome to have the t_model consume it as one of its outputs, X, along with W. One could say that if there is no actual dependence, the fitted t_model will have zero coefficients on X anyway, but apart from the theoretical incorrectness, would it not create unneeded variance/noise?
Would it not be cleaner to supply A as one of the inputs to the y_model but not to the t_model, and use another way to distinguish between common causes that are and are not used as effect modifiers in DML?
To make it more precise: let's suppose we have a causal graph with common causes C, treatment T, outcome O, and some factors A that have no effect on the outcome. The graph is then C -> T, C-> O, T -> O, A->O.
The DML equations imply that both X and W have an effect on both T and O, so in the DML equations as presented in EconML C is the union of X and W, and conversely X is a subset of C.
If we want to add additional factors A that only affect the outcome, then X would be a subset of C + A instead of subset of C; so the clean way of translating a causal graph into a DML model would be to separately specify X (list of variable names) as one of the model_params, check that it's a subset of C+A, and then have t_model consume C and y_model consume C+A, with the treatment effect still being f(X)*T?
The integration of causal graphs with EconML is awesome, but would it not be even better if it'd be theoretically correct? :) Or am I missing something?
Happy to try to spell out in code what that might look like, if that PR has a chance of being accepted :)