Calculate ATE when T is continuous in double machine learning about econml HOT 4 CLOSED

LinearDML (and most of our other estimators) assume that the effect is linear in (a featurization of) the treatment T.

effect(X, T0, T1) will give you the effect (change in output) given covariates X of moving treatment assignment from T0 to T1.
If you don't featurize the treatment and have a single treatment, then effect(X, T0, T1) will be linear in T1-T0, so just using T0=0 and T1=1 is sufficient to compute the result for any other pair of treatments (since effect(X, T0, T1) == (T1-T0) * effect(X, 0, 1)).

If you want to calculate a non-linear treatment effect, then the easiest way to do that is by using a treatment featurizer, which will expand the treatment you provide into multiple treatment columns - see this notebook for examples.

from econml.

My treatment is continuous and ranged from 0 to 80000, is that mean I need to set T0 = 0 and T1=80000 to get the ATE?

from econml.

In the case of continuous treatments there isn't a single ATE, there is (in general) an average treatment effect for each pair of source and target treatments, saying how the output would change on average if instead of treating everyone at level T0 you treat them at level T1. For linear models, this is just some factor times T1-T0, but for non-linear effects that's not necessarily true.

So in your case if you want to know how the average effect of moving from one extreme to the other, you could use effect(X=None, T0=0, T1=80000), but it really depends on exactly what you're trying to do (and if you use a treatment featurizer to model non-linear treatment effects, then it may not be the case that maximal treatment effect is achieved at the maximum treatment level, for example).

from econml.

Thanks!

from econml.