[Repo Assist] fix: pass encoded effect modifiers to EconML when categorical variables present

[github-actions]

🤖 This is an automated PR from Repo Assist, an AI assistant.

Closes #820.

Root Cause

When effect modifiers include categorical columns (dtype "category"), _set_effect_modifiers() one-hot encodes them via _encode() / OneHotEncoder. The resulting self._effect_modifiers DataFrame has new column names such as x7_0 / x7_1, but self._effect_modifier_names still held the original name 'x7'.

Four call sites then did df[self._effect_modifier_names] on the already-encoded DataFrame or on the full dataset and then expected to pass the result to the fitted EconML model:

Method	Problem
effect()	df is already the encoded EM frame from estimate_effect(); column selection with original names raises KeyError
effect_interval()	same
effect_inference()	same
effect_tt()	selects raw EM cols from full dataset but does not encode before passing to fitted EconML model
shap_values()	same as effect_tt()
estimate_effect() (target_units=DataFrame path)	raw user DataFrame not encoded before being passed as X_test

Fix

• effect / effect_interval / effect_inference: df entering these methods is already the encoded effect-modifier sub-frame assembled by estimate_effect(). Remove the redundant (and broken) column selection; pass df directly to apply_multitreatment.
• effect_tt / shap_values: df is the full raw dataset. Select the effect-modifier columns and then encode them with the stored encoder so the EconML model sees data in the same feature space it was trained on.
• estimate_effect (target_units=DataFrame path): encode the incoming DataFrame via the stored encoder before using it as X_test.

Changes

• dowhy/causal_estimators/econml.py — six targeted call-site fixes
• tests/causal_estimators/test_econml_estimator.py — regression test test_categorical_effect_modifiers using a categorical effect modifier with backdoor.econml.dml.DML

Test Status

• ✅ black --check passes (reformatted one file)
• ✅ isort --check passes
• ✅ All flake8 errors are pre-existing long-line docstrings and unused variables in existing tests; no new errors introduced by this PR
• ℹ️ Full test suite with EconML installed could not be run in this CI environment; however the logic change is a straightforward removal of redundant column selection and addition of the same _encode() call already used in fit()

Note

🔒 Integrity filter blocked 90 items

The following items were blocked because they don't meet the GitHub integrity level.

• Issue with distance_metric_params for dowhy.causal_estimators.distance_matching_estimator #1390 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• data subset refuter bug when dataframe has categorical columns #1372 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Error in the tutorial of "Finding optimal adjustment sets" #1360 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Error in do() operator for interventions outside original treatment distribution #1344 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• PC: background_knowledge not strictly enforced (tiers & forbidden_edges); results vary by uc_rule / uc_priority #1343 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Issue with Logistic Regression in Mediation Analysis #1335 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Interventional samples gives inconsistent results. #1307 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• falsify_graph uses np.math which does not exist #1268 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Counterfactual Samples giving invalid values of effect in dowhy gcm #1241 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Error with Do operator.  #1240 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• model.estimate_effect and model.refute_astimate throws 'A column-vector y was passed ...' error #1212 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• No Backdoor Path Available #1188 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• numpy.dual is dropped but it still occurs in dowhy #1181 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Inconsistency in the placebo_treatment_refuter when using estimate_effect of IV #1180 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Causal Graph not provided. DoWhy will construct a graph based on data inputs. #1125 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• numpy has no attribute 'long' #1052 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• ... and 74 more items

To allow these resources, lower min-integrity in your GitHub frontmatter:

tools:
  github:
    min-integrity: approved  # merged | approved | unapproved | none

Generated by Repo Assist · ◷

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@b897c2f3e43bde9ff7923c8fa9211055b26e27cc