[Repo Assist] fix(dummy_outcome_refuter): correct preprocess_data_by_treatment bugs

[github-actions]

🤖 This PR was created by Repo Assist, an automated AI assistant.

Related to #1665 (pandas 3.x + scikit-learn compatibility audit).

Problem

preprocess_data_by_treatment() in dowhy/causal_refuters/dummy_outcome_refuter.py contained five bugs, some of which had been present since the function's introduction.

Bug 1 — assert instead of ValueError

assert len(treatment_name) == 1, "At present, DoWhy supports a simgle treatment variable"

Python's assert can be disabled with -O (optimised mode). Production code should use an explicit ValueError.

Bug 2 — num_bins computed from a per-column Series, not a scalar

num_bins = (data.max() - data.min()) / (bucket_size_scale_factor * std_dev)

data.max() - data.min() returns a Series (the range of all columns), not the scalar range of the treatment column. Passing a Series to pd.cut(data[t], num_bins) works coincidentally only if the column ranges happen to be monotonically increasing; otherwise pd.cut raises ValueError: bins must increase monotonically.

Fix: compute the scalar range from the treatment column only:

t_range = data[treatment_variable_name].max() - data[treatment_variable_name].min()
num_bins = max(1, int(t_range / (bucket_size_scale_factor * std_dev)))

Bug 3 — Input DataFrame mutated

The function called data["bins"] = pd.cut(...) and data.drop("bins", ..., inplace=True) on the caller's DataFrame. This is surprising behaviour and is incompatible with pandas 3 Copy-on-Write semantics, which returns a read-only view for sliced DataFrames.

Fix: call data = data.copy() at the start of the function.

Bug 4 — Dead no-op line

data = data  # no-op

Removed.

Bug 5 — Unreachable categorical branch + immediate overwrite

elif "categorical" in variable_type.name:   # BUG: "categorical" not in "category"
    groups = data.groupby(treatment_variable_name)
    groups = data.groupby("bins")            # BUG: overwrites; "bins" doesn't exist
    return groups

pandas names its categorical dtype "category", not "categorical". Since "categorical" in "category" is False, this entire branch was dead code and could never execute.

The second line also immediately overwrote the first, and the "bins" column was never created at that point.

Fix: change the guard to "categor" in variable_type.name (a substring of both "category" and "categorical"), remove the overwrite.

Additional — observed=True missing on groupby calls

All three groupby calls now include observed=True to suppress the pandas FutureWarning about the default changing.

---

Changes

Only two files changed:

File	Change
dowhy/causal_refuters/dummy_outcome_refuter.py	Fix all five bugs in preprocess_data_by_treatment
tests/causal_refuters/test_dummy_outcome_refuter.py	Add 3 regression tests

---

Test Status

3 new regression tests — all pass:

Test	Verifies
test_preprocess_data_by_treatment_multiple_treatments_raises	Multiple treatment names raises ValueError (not AssertionError)
test_preprocess_data_by_treatment_continuous_treatment	Continuous treatment bins correctly; input DataFrame is not mutated
test_preprocess_data_by_treatment_categorical_treatment	Categorical treatment (pd.Categorical) now reachable and returns correct groups

black --check, flake8 --select=E9,F63,F7,F82: no errors.

Generated by 🤖 Repo Assist, see workflow run.

Generated by 🌈 Repo Assist, see workflow run. Learn more.

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@11c9a2c442e519ff2b427bf58679f5a525353f76