BUG: incorrect broadcasting that could casuse dtype coercion in a groupby-transform

closes #14457

Author: jreback

doc/source/whatsnew/v0.19.1.txt

@@ -43,9 +43,13 @@ Bug Fixes
 - Bug in string indexing against data with ``object`` ``Index`` may raise ``AttributeError`` (:issue:`14424`)
 - Corrrecly raise ``ValueError`` on empty input to ``pd.eval()`` and ``df.query()`` (:issue:`13139`)
 
+
 - Bug in ``RangeIndex.intersection`` when result is a empty set (:issue:`14364`).
 - Bug in union of differences from a ``DatetimeIndex`; this is a regression in 0.19.0 from 0.18.1 (:issue:`14323`)
 
+- Bug in groupby-transform broadcasting that could cause incorrect dtype coercion (:issue:`14457`)
+
+
 - Bug in ``Series.__setitem__`` which allowed mutating read-only arrays (:issue:`14359`).
 
 

pandas/core/groupby.py

@@ -3454,7 +3454,6 @@ def _transform_general(self, func, *args, **kwargs):
         from pandas.tools.merge import concat
 
         applied = []
-
         obj = self._obj_with_exclusions
         gen = self.grouper.get_iterator(obj, axis=self.axis)
         fast_path, slow_path = self._define_paths(func, *args, **kwargs)
@@ -3475,14 +3474,24 @@ def _transform_general(self, func, *args, **kwargs):
             else:
                 res = path(group)
 
-            # broadcasting
             if isinstance(res, Series):
-                if res.index.is_(obj.index):
-                    group.T.values[:] = res
+
+                # we need to broadcast across the
+                # other dimension; this will preserve dtypes
+                # GH14457
+                if not np.prod(group.shape):
+                    continue
+                elif res.index.is_(obj.index):
+                    r = concat([res] * len(group.columns), axis=1)
+                    r.columns = group.columns
+                    r.index = group.index
                 else:
-                    group.values[:] = res
+                    r = DataFrame(
+                        np.concatenate([res.values] * len(group.index)
+                                       ).reshape(group.shape),
+                        columns=group.columns, index=group.index)
 
-                applied.append(group)
+                applied.append(r)
             else:
                 applied.append(res)
 

pandas/tests/test_groupby.py

@@ -1336,6 +1336,18 @@ def nsum(x):
         for result in results:
             assert_series_equal(result, expected, check_names=False)
 
+    def test_transform_coercion(self):
+
+        # 14457
+        # when we are transforming be sure to not coerce
+        # via assignment
+        df = pd.DataFrame(dict(A=['a', 'a'], B=[0, 1]))
+        g = df.groupby('A')
+
+        expected = g.transform(np.mean)
+        result = g.transform(lambda x: np.mean(x))
+        assert_frame_equal(result, expected)
+
     def test_with_na(self):
         index = Index(np.arange(10))