[VenishPaneliya]

sensitivity_summary and sensitivity_interval take c_y and c_t, but all six overloads document the second one as c_d, which they do not accept:

file	methods	documented	in signature
dml/causal_forest.py	both	c_d	c_t
dml/dml.py	both	c_d	c_t
dr/_drlearner.py	both	c_d	c_t

The existing description, "The level of confounding in the treatment. Ranges from 0 to 1.", already matches c_t, so only the name changes.

Two things I deliberately left alone rather than guess at:

• rho is undocumented in all six. It has no docstring anywhere in the package — the implementation these forward to, validate/sensitivity_analysis.py, has none either — so I did not invent a description.
• sensitivity_summary in dr/_drlearner.py does not document its T argument, though sensitivity_interval next to it does.

Happy to add those if you tell me how rho should be described.

Docs only, no behaviour change. ruff check reports the same before and after.

[kbattocchi]

Thanks for the contribution! rho is the worst case correlation between the estimation errors in Y and T, so should be in the range [-1, 1], with ±1 being the most conservative choices. It would be good to document this and also document the T argument to sensitivity_summary.

[VenishPaneliya]

Thanks, that's much clearer than what I would have written. I've used your wording for rho on all six overloads and documented T on sensitivity_summary in _drlearner.py so it matches the sensitivity_interval next to it.

One thing I wasn't sure about: rho defaults to 1., i.e. the most conservative end of the range. Is that intentional as a worst-case-by-default, or would a user typically be expected to pass an estimate from their own domain knowledge?

[kbattocchi]

Thanks, that's much clearer than what I would have written. I've used your wording for rho on all six overloads and documented T on sensitivity_summary in _drlearner.py so it matches the sensitivity_interval next to it.

One thing I wasn't sure about: rho defaults to 1., i.e. the most conservative end of the range. Is that intentional as a worst-case-by-default, or would a user typically be expected to pass an estimate from their own domain knowledge?

Yes, it's intentionally the worst case by default and we don't generally expect users to modify it; the point of the sensitivity analysis is to understand how susceptible our estimates are to unmeasured confounders, so typically worst-case confounding is what we expect people to care about.

[kbattocchi]

Looks good, thanks!