FIX: silent data loss in MultiIndex __setitem__ with object-dtype level

[Bahtya]

Problem

df[key] = df[key] / x silently drops the assignment (no error, no warning) when:

1. DataFrame has a column MultiIndex
2. Level 1 has object dtype due to mixed types (e.g. string + int)
3. The top-level label has exactly one sub-column

cols = pd.MultiIndex.from_tuples(
    [("info", "M"), ("info", 0), ("earnings", 1), ("earnings", 2), ("prices", 0)]
)
df = pd.DataFrame(np.arange(20, dtype=float).reshape(4, 5), columns=cols)
df["prices"] = df["prices"] / 100  # silent no-op! values unchanged

This is a regression from 2.3.x → 3.0.x and causes silent data loss.

Root Cause

In _set_item_frame_value, the guard added for GH#62518/GH#61841 checks:

is_string_dtype(cols_droplevel.dtype) and not cols_droplevel.any()

When maybe_droplevels produces Index([0], dtype="object"):

• is_string_dtype(object) → True
• Index([0]).any() → False (because 0 is falsy)

This causes the early return to trigger incorrectly, silently discarding the assignment.

Fix

Replace not cols_droplevel.any() with (cols_droplevel == "").all() to explicitly check for empty strings:

and len(cols_droplevel) > 0
and (cols_droplevel == "").all()

This correctly identifies actual empty-string columns without being fooled by falsy integer values in object-dtype Indexes.

Testing

Verified locally that the fix resolves the issue from the bug report:

# Before fix: prices values unchanged [4. 9. 14. 19.]
# After fix: prices values correctly divided [0.04 0.09 0.14 0.19]

Also confirmed the original GH#62518/GH#61841 cases are still protected since their columns genuinely contain only empty strings.

Fixes #65118

[jbrockmendel]

Pls add test

[Bahtya]

Thanks for the review! I'll add a test for the silent data loss case in MultiIndex __setitem__ with object-dtype levels.

[Bahtya]

Hi @jbrockmendel, thanks for the review! I'll add tests for this fix. Working on it now.

[Bahtya]

Hi @jbrockmendel, I've added three test cases for this fix:

1. test_multiindex_setitem_object_dtype_level_no_silent_drop — exact reproduction from the bug report (mixed string/int level values, df["prices"] = df["prices"] / 100)
2. test_multiindex_setitem_object_dtype_level_single_subcolumn — minimal case with integer 0 as level value
3. test_multiindex_setitem_object_dtype_level_falsy_values — ensures columns with empty string "" alongside falsy integer 0 are still writable

All tests pass locally. Ready for re-review!