BUG: groupby nunique not respecting observed

doc/source/whatsnew/v1.4.0.rst

@@ -943,6 +943,7 @@ Groupby/resample/rolling
 - Bug in :meth:`GroupBy.nth` failing on ``axis=1`` (:issue:`43926`)
 - Fixed bug in :meth:`Series.rolling` and :meth:`DataFrame.rolling` not respecting right bound on centered datetime-like windows, if the index contain duplicates (:issue:`3944`)
 - Bug in :meth:`Series.rolling` and :meth:`DataFrame.rolling` when using a :class:`pandas.api.indexers.BaseIndexer` subclass that returned unequal start and end arrays would segfault instead of raising a ``ValueError`` (:issue:`44470`)
+- Bug in :meth:`Groupby.nunique` not respecting ``observed=True`` for Categorical grouping columns (:issue:`45128`)
 - Bug in :meth:`GroupBy.head` and :meth:`GroupBy.tail` not dropping groups with ``NaN`` when ``dropna=True`` (:issue:`45089`)
 - Fixed bug in :meth:`GroupBy.__iter__` after selecting a subset of columns in a :class:`GroupBy` object, which returned all columns instead of the chosen subset (:issue:`#44821`)
 - Bug in :meth:`Groupby.rolling` when non-monotonic data passed, fails to correctly raise ``ValueError`` (:issue:`43909`)

pandas/core/groupby/generic.py

@@ -1449,6 +1449,7 @@ def _iterate_column_groupbys(self, obj: DataFrame | Series):
                 selection=colname,
                 grouper=self.grouper,
                 exclusions=self.exclusions,
+                observed=self.observed,
             )
 
     def _apply_to_column_groupbys(self, func, obj: DataFrame | Series) -> DataFrame:

pandas/tests/groupby/test_categorical.py

@@ -1779,3 +1779,18 @@ def test_groupby_last_first_preserve_categoricaldtype(func):
         Categorical([1, 2, 3]), name="b", index=Index([1, 2, 3], name="a")
     )
     tm.assert_series_equal(expected, result)
+
+
+def test_groupby_categorical_observed_nunique():
+    # GH#45128
+    df = DataFrame({"a": [1, 2], "b": [1, 2], "c": [10, 11]})
+    df = df.astype(dtype={"a": "category", "b": "category"})
+    result = df.groupby(["a", "b"], observed=True).nunique()["c"]
+    expected = Series(
+        [1, 1],
+        index=MultiIndex.from_arrays(
+            [CategoricalIndex([1, 2], name="a"), CategoricalIndex([1, 2], name="b")]
+        ),
+        name="c",
+    )
+    tm.assert_series_equal(result, expected)