[Repo Assist] fix(cit): correct conditional_MI entropy computation for discrete variables

[github-actions]

🤖 This is an automated PR from Repo Assist.

Problem

The conditional_MI function in dowhy/utils/cit.py was producing always-incorrect values for the conditional mutual information of discrete variables.

Root cause: Iterating over a pandas DataFrame yields column labels (strings), not row values. For a single-column DataFrame data[['Foo']], zip(X_DataFrame, Z_tuples) produced exactly one pair — ('Foo', first_z_tuple) — regardless of sample size.

This made:

• H(X,Z) = H(Y,Z) = H(X,Y,Z) = 0 (entropy of a 1-element sequence)
• CMI = 0 + 0 - 0 - H(Z) = -H(Z) (always negative)

Since the GraphRefuter uses the threshold cmi_val <= 0.05, a CMI of -H(Z) always passes, meaning the conditional independence test for discrete variables silently did nothing — it always declared CI to hold, even when variables were strongly dependent.

The same bug also caused a KeyError when column names had more than one character (e.g., 'Foo'): list('Foo') → ['F', 'o', 'o'], which then fails the DataFrame index lookup. This was reported in #949.

Changes

dowhy/utils/cit.py:

• Wrap string x/y arguments in a list ([x]) to avoid character iteration
• Call .squeeze(axis=1) on single-column DataFrames to get a Series — iterating a Series yields row values correctly
• Remove two redundant intermediate variables for Z
• Rename lambda parameter from shadowing outer x to v for clarity

tests/utils/test_cit.py (new file):

• test_multi_char_column_names — regression for Error handling column names in CIT as used by CausalModel.graph_refute method #949 (no KeyError on 'Foo', 'Bar', 'Baz')
• test_single_char_column_names — ensures single-char names still work
• test_independent_variables_low_cmi — CMI < 0.05 for truly independent discrete variables
• test_dependent_variables_high_cmi — CMI > 0.5 for fully dependent discrete variables

Before vs After

Scenario	Before (broken)	After (fixed)
Independent X, Y | Z	-H(Z) ≈ -1.0 bit	~0.001 bits ✓
Fully dependent X = Y | Z	-H(Z) ≈ -1.0 bit	~1.0 bit ✓
Column name 'Foo'	KeyError	Works ✓

Test Status

The four new unit tests pass locally (verified via direct Python execution with numpy, pandas, scipy). The full CI suite requires Poetry with all optional extras.

Closes #949

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