diff --git a/pandas/core/common.py b/pandas/core/common.py
index f8f5928ca7d51..171ce9462452f 100644
--- a/pandas/core/common.py
+++ b/pandas/core/common.py
@@ -1271,14 +1271,23 @@ def _possibly_downcast_to_dtype(result, dtype):
         dtype = np.dtype(dtype)
 
     try:
-
         # don't allow upcasts here (except if empty)
+        print dtype.kind, result.dtype.kind
         if dtype.kind == result.dtype.kind:
             if result.dtype.itemsize <= dtype.itemsize and np.prod(result.shape):
                 return result
 
         if issubclass(dtype.type, np.floating):
             return result.astype(dtype)
+
+        # a datetimelike
+        elif ((dtype.kind == 'M' and result.dtype.kind == 'i') or
+              dtype.kind == 'm'):
+            try:
+                result = result.astype(dtype)
+            except:
+                pass
+
         elif dtype == np.bool_ or issubclass(dtype.type, np.integer):
 
             # if we don't have any elements, just astype it
@@ -1309,13 +1318,6 @@ def _possibly_downcast_to_dtype(result, dtype):
                     if (new_result == result).all():
                         return new_result
 
-        # a datetimelike
-        elif dtype.kind in ['M','m'] and result.dtype.kind in ['i']:
-            try:
-                result = result.astype(dtype)
-            except:
-                pass
-
     except:
         pass
 
diff --git a/pandas/core/groupby.py b/pandas/core/groupby.py
index cb5dedc887bca..4b55b8cced559 100644
--- a/pandas/core/groupby.py
+++ b/pandas/core/groupby.py
@@ -1083,12 +1083,24 @@ def _try_cast(self, result, obj):
     def _cython_agg_general(self, how, numeric_only=True):
         output = {}
         for name, obj in self._iterate_slices():
-            is_numeric = is_numeric_dtype(obj.dtype)
+            if is_numeric_dtype(obj.dtype):
+                obj = com.ensure_float(obj)
+                is_numeric = True
+                out_dtype = 'f%d' % obj.dtype.itemsize
+                values = obj.values
+            else:
+                is_numeric = issubclass(obj.dtype.type, (np.datetime64,
+                                                         np.timedelta64))
+                if is_numeric:
+                    values = obj.view('int64')
+                else:
+                    values = obj.astype(object)
+
             if numeric_only and not is_numeric:
                 continue
 
             try:
-                result, names = self.grouper.aggregate(obj.values, how)
+                result, names = self.grouper.aggregate(values, how)
             except AssertionError as e:
                 raise GroupByError(str(e))
             output[name] = self._try_cast(result, obj)
@@ -2567,9 +2579,9 @@ def _cython_agg_blocks(self, how, numeric_only=True):
             data = data.get_numeric_data(copy=False)
 
         for block in data.blocks:
-
             values = block._try_operate(block.values)
 
+            # TODO DAN
             if block.is_numeric:
                 values = _algos.ensure_float64(values)
 
diff --git a/pandas/tests/test_groupby.py b/pandas/tests/test_groupby.py
index 4077f468d8b1f..734287baaa50d 100644
--- a/pandas/tests/test_groupby.py
+++ b/pandas/tests/test_groupby.py
@@ -41,6 +41,7 @@ def _skip_if_mpl_not_installed():
     except ImportError:
         raise nose.SkipTest("matplotlib not installed")
 
+
 def commonSetUp(self):
     self.dateRange = bdate_range('1/1/2005', periods=250)
     self.stringIndex = Index([rands(8).upper() for x in range(250)])
@@ -603,6 +604,28 @@ def f(grp):
         e.name = None
         assert_series_equal(result,e)
 
+        # ...and with timedeltas
+        df1 = df.copy()
+        df1['D'] = pd.to_timedelta(['00:00:01', '00:00:02', '00:00:03',
+                                    '00:00:04', '00:00:05', '00:00:06',
+                                    '00:00:07'])
+        result = df1.groupby('A').apply(f)[['D']]
+        e = df1.groupby('A').first()[['D']]
+        e.loc['Pony'] = np.nan
+        print(type(result))
+        print(type(e))
+        assert_frame_equal(result, e)
+
+        def f(grp):
+            if grp.name == 'Pony':
+                return None
+            return grp.iloc[0].loc['D']
+        result = df1.groupby('A').apply(f)['D']
+        e = df1.groupby('A').first()['D'].copy()
+        e.loc['Pony'] = np.nan
+        e.name = None
+        assert_series_equal(result, e)
+
     def test_agg_api(self):
 
         # GH 6337
@@ -4365,6 +4388,19 @@ def test_index_label_overlaps_location(self):
         expected = ser.take([1, 3, 4])
         assert_series_equal(actual, expected)
 
+    def test_groupby_methods_on_timedelta64(self):
+        df = self.df.copy().iloc[:4]
+        df['E'] = pd.to_timedelta(['00:00:01', '00:00:02', '00:00:03', '00:00:04'])
+        # DataFrameGroupBy
+        actual = df.groupby('A').mean()['E']
+        expected = pd.to_timedelta(Series(['00:00:03', '00:00:02'], index=['bar', 'foo'], name='E'))
+        assert_series_equal(actual, expected)
+
+        ser = df['E']
+        # SeriesGroupBy
+        actual = ser.groupby(df['A']).mean()
+        assert_series_equal(actual, expected)
+
     def test_groupby_selection_with_methods(self):
         # some methods which require DatetimeIndex
         rng = pd.date_range('2014', periods=len(self.df))