Fix 40420: Interpret NaN in clip() as no bound.

doc/source/whatsnew/v1.3.0.rst

@@ -855,6 +855,7 @@ Other
 - Bug in :meth:`DataFrame.equals`, :meth:`Series.equals`, :meth:`Index.equals` with object-dtype containing ``np.datetime64("NaT")`` or ``np.timedelta64("NaT")`` (:issue:`39650`)
 - Bug in :func:`pandas.util.show_versions` where console JSON output was not proper JSON (:issue:`39701`)
 - Bug in :meth:`DataFrame.convert_dtypes` incorrectly raised ValueError when called on an empty DataFrame (:issue:`40393`)
+- Bug in :meth:`DataFrame.clip` not interpreting missing values as no threshold (:issue:`40420`)
 
 .. ---------------------------------------------------------------------------
 

pandas/core/generic.py

@@ -7352,6 +7352,14 @@ def _clip_with_one_bound(self, threshold, method, axis, inplace):
                 threshold = self._constructor(threshold, index=self.index)
             else:
                 threshold = align_method_FRAME(self, threshold, axis, flex=None)[1]
+
+        # GH 40420
+        # In order to ignore nan values in the threshold, set the values in
+        # subset that correspond to these na values to True. This indicates to the
+        # final where() to not clip.
+        if is_list_like(threshold) and threshold.isna().any(axis=None):
+            subset_kwargs = {"axis": axis} if threshold.ndim != subset.ndim else {}
+            subset = subset.where(threshold.notna(), True, **subset_kwargs)
         return self.where(subset, threshold, axis=axis, inplace=inplace)
 
     @overload
@@ -7483,10 +7491,12 @@ def clip(
         ----------
         lower : float or array_like, default None
             Minimum threshold value. All values below this
-            threshold will be set to it.
+            threshold will be set to it. A missing
+            threshold (e.g `NA`) will not clip the value.
         upper : float or array_like, default None
             Maximum threshold value. All values above this
-            threshold will be set to it.
+            threshold will be set to it. A missing
+            threshold (e.g `NA`) will not clip the value.
         axis : int or str axis name, optional
             Align object with lower and upper along the given axis.
         inplace : bool, default False
@@ -7547,6 +7557,25 @@ def clip(
         2      0      3
         3      6      8
         4      5      3
+
+        Clips using specific lower threshold per column element, with missing values:
+
+        >>> t = pd.Series([2, -4, np.NaN, 6, 3])
+        >>> t
+        0    2.0
+        1   -4.0
+        2    NaN
+        3    6.0
+        4    3.0
+        dtype: float64
+
+        >>> df.clip(t, axis=0)
+        col_0  col_1
+        0      9      2
+        1     -3     -4
+        2      0      6
+        3      6      8
+        4      5      3
         """
         inplace = validate_bool_kwarg(inplace, "inplace")
 
@@ -7559,9 +7588,17 @@ def clip(
         # so ignore
         # GH 19992
         # numpy doesn't drop a list-like bound containing NaN
-        if not is_list_like(lower) and np.any(isna(lower)):
+        isna_lower = isna(lower)
+        if not is_list_like(lower):
+            if np.any(isna_lower):
+                lower = None
+        elif np.all(isna_lower):
             lower = None
-        if not is_list_like(upper) and np.any(isna(upper)):
+        isna_upper = isna(upper)
+        if not is_list_like(upper):
+            if np.any(isna_upper):
+                upper = None
+        elif np.all(isna_upper):
             upper = None
 
         # GH 2747 (arguments were reversed)

pandas/tests/frame/methods/test_clip.py

@@ -144,17 +144,25 @@ def test_clip_with_na_args(self, float_frame):
         tm.assert_frame_equal(float_frame.clip(np.nan), float_frame)
         tm.assert_frame_equal(float_frame.clip(upper=np.nan, lower=np.nan), float_frame)
 
-        # GH#19992
+        # GH#19992 and adjusted in GH#40420
         df = DataFrame({"col_0": [1, 2, 3], "col_1": [4, 5, 6], "col_2": [7, 8, 9]})
 
         result = df.clip(lower=[4, 5, np.nan], axis=0)
         expected = DataFrame(
-            {"col_0": [4, 5, np.nan], "col_1": [4, 5, np.nan], "col_2": [7, 8, np.nan]}
+            {"col_0": [4, 5, 3], "col_1": [4, 5, 6], "col_2": [7, 8, 9]}
         )
         tm.assert_frame_equal(result, expected)
 
         result = df.clip(lower=[4, 5, np.nan], axis=1)
         expected = DataFrame(
-            {"col_0": [4, 4, 4], "col_1": [5, 5, 6], "col_2": [np.nan, np.nan, np.nan]}
+            {"col_0": [4, 4, 4], "col_1": [5, 5, 6], "col_2": [7, 8, 9]}
         )
         tm.assert_frame_equal(result, expected)
+
+        # GH#40420
+        data = {"col_0": [9, -3, 0, -1, 5], "col_1": [-2, -7, 6, 8, -5]}
+        df = DataFrame(data)
+        t = Series([2, -4, np.NaN, 6, 3])
+        result = df.clip(lower=t, axis=0)
+        expected = DataFrame({"col_0": [9, -3, 0, 6, 5], "col_1": [2, -4, 6, 8, 3]})
+        tm.assert_frame_equal(result, expected)

pandas/tests/series/methods/test_clip.py

@@ -49,8 +49,13 @@ def test_clip_with_na_args(self):
         tm.assert_series_equal(s.clip(upper=np.nan, lower=np.nan), Series([1, 2, 3]))
 
         # GH#19992
-        tm.assert_series_equal(s.clip(lower=[0, 4, np.nan]), Series([1, 4, np.nan]))
-        tm.assert_series_equal(s.clip(upper=[1, np.nan, 1]), Series([1, np.nan, 1]))
+        tm.assert_series_equal(s.clip(lower=[0, 4, np.nan]), Series([1, 4, 3]))
+        tm.assert_series_equal(s.clip(upper=[1, np.nan, 1]), Series([1, 2, 1]))
+
+        # GH#40420
+        s = Series([1, 2, 3])
+        result = s.clip(0, [np.nan, np.nan, np.nan])
+        tm.assert_series_equal(s, result)
 
     def test_clip_against_series(self):
         # GH#6966