How ATE is calculated when using generalized_linear_model method with binomial/logistic regression

[kiroah]

Hi, I'm wondering if someone can shed some light on my question regarding ATE (average treatment effect).

So I believe the textbook answer of ATE when using logistic regression is the coefficient of the treatment variable (in log-odds). I am comparing the ATE output from the estimate summary with the underlying model's (estimate.estimator.model) coefficient and it doesn't match, even if I convert the log-odds to odds.

My guess is that DoWhy is calculating the ATE by getting the difference of outcome when the treatment value is fixed as 1 VS 0 (both the outcome and treatment is binary in my case) for each row and getting the average, which is essentially the same as what S-learner does.

I want to confirm this is the case. I did some testing myself by creating a logistic regression model manually and doing the same, but I have tiny differences so just wanted to get another opinion.

Thanks in advance!

[kiroah]

Nevermind, I think I got the textbook answer wrong :)

[abhilasha-workday]

Hi @kiroah,
If you dont mind can you share what is the correct way to interpret ATE for a logistic regression model based on your findings?