[Refute] Placebo Test

[elakhatibi]

Hi,

I have tried to understand the source code of "placebo Test" ("https://github.com/py-why/dowhy/blob/master/dowhy/causal_refuters/placebo_treatment_refuter.py)"

I was wondering (especially in hereunder line of code) whether the algorithm tries to assign one of the available features as a "new_treatment" variable and shut down backdoor paths, then calculate the NEW causal-effect estimation?

[new_treatment = self._data[self._treatment_name].iloc[permuted_idx].values]

Meanwhile, could you please share with me some docs/materials about validation parts?

Looking forward to hearing from you.

Thanks a lot for your precious time and your kind consideration.

Best, Ela

[amit-sharma]

The algorithm shuts down the backdoor path by permuting the treatment variable data. Since the treatment is randomized, it cannot be affected by the confounders. This paper on negative controls describes the core idea. https://pubmed.ncbi.nlm.nih.gov/20335814/

We are also preparing an updated chapter on this topic, will post it here next week.