diff --git a/pandas/_libs/groupby.pyx b/pandas/_libs/groupby.pyx
index 12fe78a0f8a18..ff5258e37e352 100644
--- a/pandas/_libs/groupby.pyx
+++ b/pandas/_libs/groupby.pyx
@@ -1238,6 +1238,7 @@ def group_nth(
                     if nobs[i, j] < min_count:
                         if uses_mask:
                             result_mask[i, j] = True
+                            out[i, j] = 0
                         elif iu_64_floating_obj_t is int64_t:
                             # TODO: only if datetimelike?
                             out[i, j] = NPY_NAT
diff --git a/pandas/core/groupby/ops.py b/pandas/core/groupby/ops.py
index 6e6ef14a25941..eadba375c66a3 100644
--- a/pandas/core/groupby/ops.py
+++ b/pandas/core/groupby/ops.py
@@ -367,6 +367,7 @@ def _reconstruct_ea_result(
         """
         Construct an ExtensionArray result from an ndarray result.
         """
+        dtype: BaseMaskedDtype | StringDtype
 
         if isinstance(values.dtype, StringDtype):
             dtype = values.dtype
@@ -375,19 +376,17 @@ def _reconstruct_ea_result(
 
         elif isinstance(values.dtype, BaseMaskedDtype):
             new_dtype = self._get_result_dtype(values.dtype.numpy_dtype)
+            dtype = BaseMaskedDtype.from_numpy_dtype(new_dtype)
             # error: Incompatible types in assignment (expression has type
-            # "BaseMaskedDtype", variable has type "StringDtype")
-            dtype = BaseMaskedDtype.from_numpy_dtype(  # type: ignore[assignment]
-                new_dtype
-            )
-            cls = dtype.construct_array_type()
+            # "Type[BaseMaskedArray]", variable has type "Type[BaseStringArray]")
+            cls = dtype.construct_array_type()  # type: ignore[assignment]
             return cls._from_sequence(res_values, dtype=dtype)
 
-        elif needs_i8_conversion(values.dtype):
-            assert res_values.dtype.kind != "f"  # just to be on the safe side
-            i8values = res_values.view("i8")
-            # error: Too many arguments for "ExtensionArray"
-            return type(values)(i8values, dtype=values.dtype)  # type: ignore[call-arg]
+        elif isinstance(values, (DatetimeArray, TimedeltaArray, PeriodArray)):
+            # In to_cython_values we took a view as M8[ns]
+            assert res_values.dtype == "M8[ns]"
+            res_values = res_values.view(values._ndarray.dtype)
+            return values._from_backing_data(res_values)
 
         raise NotImplementedError
 
@@ -425,12 +424,8 @@ def _masked_ea_wrap_cython_operation(
             **kwargs,
         )
 
-        new_dtype = self._get_result_dtype(orig_values.dtype.numpy_dtype)
-        dtype = BaseMaskedDtype.from_numpy_dtype(new_dtype)
-        # TODO: avoid cast as res_values *should* already have the right
-        #  dtype; last attempt ran into trouble on 32bit linux build
-        res_values = res_values.astype(dtype.type, copy=False)
-
+        # res_values should already have the correct dtype, we just need to
+        #  wrap in a MaskedArray
         return orig_values._maybe_mask_result(res_values, result_mask)
 
     @final