BUG: MultiIndex.get_level_values() replaces NA by another value (#5074)

doc/source/release.rst

@@ -569,6 +569,7 @@ Bug Fixes
   - Fixed a bug where default options were being overwritten in the option
     parser cleaning (:issue:`5121`).
   - Treat a list/ndarray identically for ``iloc`` indexing with list-like (:issue:`5006`)
+  - Fix ``MultiIndex.get_level_values()`` with missing values (:issue:`5074`)
 
 pandas 0.12.0
 -------------

pandas/core/index.py

@@ -393,6 +393,9 @@ def values(self):
     def get_values(self):
         return self.values
 
+    _na_value = np.nan
+    """The expected NA value to use with this index."""
+
     @property
     def is_monotonic(self):
         return self._engine.is_monotonic
@@ -2256,7 +2259,8 @@ def get_level_values(self, level):
         num = self._get_level_number(level)
         unique_vals = self.levels[num]  # .values
         labels = self.labels[num]
-        values = unique_vals.take(labels)
+        values = Index(com.take_1d(unique_vals.values, labels,
+                                   fill_value=unique_vals._na_value))
         values.name = self.names[num]
         return values
 

pandas/tests/test_index.py

@@ -1445,6 +1445,39 @@ def test_get_level_values(self):
         expected = self.index.get_level_values(0)
         self.assert_(np.array_equal(result, expected))
 
+    def test_get_level_values_na(self):
+        arrays = [['a', 'b', 'b'], [1, np.nan, 2]]
+        index = pd.MultiIndex.from_arrays(arrays)
+        values = index.get_level_values(1)
+        expected = [1, np.nan, 2]
+        assert_array_equal(values.values.astype(float), expected)
+
+        arrays = [['a', 'b', 'b'], [np.nan, np.nan, 2]]
+        index = pd.MultiIndex.from_arrays(arrays)
+        values = index.get_level_values(1)
+        expected = [np.nan, np.nan, 2]
+        assert_array_equal(values.values.astype(float), expected)
+
+        arrays = [[np.nan, np.nan, np.nan], ['a', np.nan, 1]]
+        index = pd.MultiIndex.from_arrays(arrays)
+        values = index.get_level_values(0)
+        expected = [np.nan, np.nan, np.nan]
+        assert_array_equal(values.values.astype(float), expected)
+        values = index.get_level_values(1)
+        expected = ['a', np.nan, 1]
+        assert_array_equal(values.values, expected)
+
+        arrays = [['a', 'b', 'b'], pd.DatetimeIndex([0, 1, pd.NaT])]
+        index = pd.MultiIndex.from_arrays(arrays)
+        values = index.get_level_values(1)
+        expected = pd.DatetimeIndex([0, 1, pd.NaT])
+        assert_array_equal(values.values, expected.values)
+
+        arrays = [[], []]
+        index = pd.MultiIndex.from_arrays(arrays)
+        values = index.get_level_values(0)
+        self.assertEqual(values.shape, (0,))
+
     def test_reorder_levels(self):
         # this blows up
         assertRaisesRegexp(IndexError, '^Too many levels',

pandas/tseries/index.py

@@ -495,6 +495,9 @@ def _mpl_repr(self):
         # how to represent ourselves to matplotlib
         return tslib.ints_to_pydatetime(self.asi8, self.tz)
 
+    _na_value = tslib.NaT
+    """The expected NA value to use with this index."""
+
     def __unicode__(self):
         from pandas.core.format import _format_datetime64
         values = self.values