BUG: Groupby median on timedelta column with NaT returns odd value (#…

doc/source/whatsnew/v3.0.0.rst

@@ -297,6 +297,7 @@ Performance improvements
 Bug fixes
 ~~~~~~~~~
 - Fixed bug in :class:`SparseDtype` for equal comparison with na fill value. (:issue:`54770`)
+- Fixed bug in :meth:`.DataFrameGroupBy.median` where nat values gave an incorrect result. (:issue:`57926`)
 - Fixed bug in :meth:`DataFrame.join` inconsistently setting result index name (:issue:`55815`)
 - Fixed bug in :meth:`DataFrame.to_string` that raised ``StopIteration`` with nested DataFrames. (:issue:`16098`)
 - Fixed bug in :meth:`DataFrame.update` bool dtype being converted to object (:issue:`55509`)

pandas/_libs/groupby.pyi

@@ -12,6 +12,7 @@ def group_median_float64(
     min_count: int = ...,  # Py_ssize_t
     mask: np.ndarray | None = ...,
     result_mask: np.ndarray | None = ...,
+    is_datetimelike: bool = ...,  # bint
 ) -> None: ...
 def group_cumprod(
     out: np.ndarray,  # float64_t[:, ::1]

pandas/_libs/groupby.pyx

@@ -101,7 +101,11 @@ cdef float64_t median_linear_mask(float64_t* a, int n, uint8_t* mask) noexcept n
     return result
 
 
-cdef float64_t median_linear(float64_t* a, int n) noexcept nogil:
+cdef float64_t median_linear(
+    float64_t* a,
+    int n,
+    bint is_datetimelike=False
+) noexcept nogil:
     cdef:
         int i, j, na_count = 0
         float64_t* tmp
@@ -111,9 +115,14 @@ cdef float64_t median_linear(float64_t* a, int n) noexcept nogil:
         return NaN
 
     # count NAs
-    for i in range(n):
-        if a[i] != a[i]:
-            na_count += 1
+    if is_datetimelike:
+        for i in range(n):
+            if a[i] == NPY_NAT:
+                na_count += 1
+    else:
+        for i in range(n):
+            if a[i] != a[i]:
+                na_count += 1
 
     if na_count:
         if na_count == n:
@@ -124,10 +133,16 @@ cdef float64_t median_linear(float64_t* a, int n) noexcept nogil:
             raise MemoryError()
 
         j = 0
-        for i in range(n):
-            if a[i] == a[i]:
-                tmp[j] = a[i]
-                j += 1
+        if is_datetimelike:
+            for i in range(n):
+                if a[i] != NPY_NAT:
+                    tmp[j] = a[i]
+                    j += 1
+        else:
+            for i in range(n):
+                if a[i] == a[i]:
+                    tmp[j] = a[i]
+                    j += 1
 
         a = tmp
         n -= na_count
@@ -170,6 +185,7 @@ def group_median_float64(
     Py_ssize_t min_count=-1,
     const uint8_t[:, :] mask=None,
     uint8_t[:, ::1] result_mask=None,
+    bint is_datetimelike=False,
 ) -> None:
     """
     Only aggregates on axis=0
@@ -228,7 +244,7 @@ def group_median_float64(
                 ptr += _counts[0]
                 for j in range(ngroups):
                     size = _counts[j + 1]
-                    out[j, i] = median_linear(ptr, size)
+                    out[j, i] = median_linear(ptr, size, is_datetimelike)
                     ptr += size
 
 

pandas/core/groupby/ops.py

@@ -415,6 +415,7 @@ def _call_cython_op(
                 "last",
                 "first",
                 "sum",
+                "median",
             ]:
                 func(
                     out=result,
@@ -427,7 +428,7 @@ def _call_cython_op(
                     is_datetimelike=is_datetimelike,
                     **kwargs,
                 )
-            elif self.how in ["sem", "std", "var", "ohlc", "prod", "median"]:
+            elif self.how in ["sem", "std", "var", "ohlc", "prod"]:
                 if self.how in ["std", "sem"]:
                     kwargs["is_datetimelike"] = is_datetimelike
                 func(

pandas/tests/groupby/test_groupby.py

@@ -145,6 +145,15 @@ def test_len_nan_group():
     assert len(df.groupby(["a", "b"])) == 0
 
 
+def test_groupby_timedelta_median():
+    # issue 57926
+    expected = Series(data=Timedelta("1d"), index=["foo"])
+    df = DataFrame({"label": ["foo", "foo"], "timedelta": [pd.NaT, Timedelta("1d")]})
+    gb = df.groupby("label")["timedelta"]
+    actual = gb.median()
+    tm.assert_series_equal(actual, expected, check_names=False)
+
+
 @pytest.mark.parametrize("keys", [["a"], ["a", "b"]])
 def test_len_categorical(dropna, observed, keys):
     # GH#57595