BUG: Unwanted conversion from timedelta to float (#18493)

doc/source/whatsnew/v0.22.0.txt

@@ -151,7 +151,7 @@ Indexing
 - Bug in :class:`IntervalIndex` where empty and purely NA data was constructed inconsistently depending on the construction method (:issue:`18421`)
 - Bug in ``IntervalIndex.symmetric_difference()`` where the symmetric difference with a non-``IntervalIndex`` did not raise (:issue:`18475`)
 - Bug in indexing a datetimelike ``Index`` that raised ``ValueError`` instead of ``IndexError`` (:issue:`18386`).
-
+- Bug in ``TimeDeltaBlock._can_hold_element()`` where masked assignment of a timedelta series converts the series values to float64 (:issue:`18493`)
 
 I/O
 ^^^

pandas/core/internals.py

@@ -1956,7 +1956,8 @@ def _can_hold_element(self, element):
         tipo = maybe_infer_dtype_type(element)
         if tipo is not None:
             return issubclass(tipo.type, np.timedelta64)
-        return isinstance(element, (timedelta, np.timedelta64))
+        return is_integer(element) or isinstance(
+            element, (timedelta, np.timedelta64))
 
     def fillna(self, value, **kwargs):
 

pandas/tests/indexing/test_timedelta.py

@@ -2,6 +2,7 @@
 
 import pandas as pd
 from pandas.util import testing as tm
+from numpy import nan
 
 
 class TestTimedeltaIndexing(object):
@@ -47,3 +48,14 @@ def test_string_indexing(self):
         expected = df.iloc[0]
         sliced = df.loc['0 days']
         tm.assert_series_equal(sliced, expected)
+
+    @pytest.mark.parametrize(
+        "value, expected",
+        [(None, [pd.NaT, 1, 2]),
+         (pd.NaT, [pd.NaT, 1, 2]),
+         (nan, [pd.NaT, 1, 2])])
+    def test_nan(self, value, expected):
+        series = pd.Series([0, 1, 2], dtype='timedelta64[ns]')
+        series[series == series[0]] = value
+        expected = pd.Series([pd.NaT, 1, 2], dtype='timedelta64[ns]')
+        tm.assert_series_equal(series, expected)