BUG: to_numeric converting StringArray to object or float64

pandas/_libs/lib.pyx

@@ -2325,10 +2325,14 @@ def maybe_convert_numeric(
                 if not seen.coerce_numeric:
                     raise type(err)(f"{err} at position {i}")
 
-                seen.saw_null()
-                floats[i] = NaN
                 mask[i] = 1
 
+                if allow_null_in_int:
+                    seen.null_ = True
+                else:
+                    seen.saw_null()
+                    floats[i] = NaN
+
     if seen.check_uint64_conflict():
         return (values, None)
 

pandas/core/tools/numeric.py

@@ -19,8 +19,8 @@
     is_integer_dtype,
     is_number,
     is_numeric_dtype,
-    is_object_dtype,
     is_scalar,
+    is_string_dtype,
     needs_i8_conversion,
 )
 from pandas.core.dtypes.generic import (
@@ -30,6 +30,7 @@
 
 from pandas.core.arrays import BaseMaskedArray
 from pandas.core.arrays.arrow import ArrowDtype
+from pandas.core.arrays.string_ import StringDtype
 
 if TYPE_CHECKING:
     from pandas._typing import (
@@ -196,6 +197,8 @@ def to_numeric(
     else:
         values = arg
 
+    orig_values = values
+
     # GH33013: for IntegerArray & FloatingArray extract non-null values for casting
     # save mask to reconstruct the full array after casting
     mask: npt.NDArray[np.bool_] | None = None
@@ -220,17 +223,23 @@ def to_numeric(
                 values,
                 set(),
                 coerce_numeric=coerce_numeric,
-                convert_to_masked_nullable=dtype_backend is not lib.no_default,
+                convert_to_masked_nullable=dtype_backend is not lib.no_default
+                or isinstance(values_dtype, StringDtype),
             )
         except (ValueError, TypeError):
             if errors == "raise":
                 raise
+            values = orig_values
 
     if new_mask is not None:
         # Remove unnecessary values, is expected later anyway and enables
         # downcasting
         values = values[~new_mask]
-    elif dtype_backend is not lib.no_default and new_mask is None:
+    elif (
+        dtype_backend is not lib.no_default
+        and new_mask is None
+        or isinstance(values_dtype, StringDtype)
+    ):
         new_mask = np.zeros(values.shape, dtype=np.bool_)
 
     # attempt downcast only if the data has been successfully converted
@@ -265,7 +274,7 @@ def to_numeric(
 
     # GH33013: for IntegerArray, BooleanArray & FloatingArray need to reconstruct
     # masked array
-    if (mask is not None or new_mask is not None) and not is_object_dtype(values.dtype):
+    if (mask is not None or new_mask is not None) and not is_string_dtype(values.dtype):
         if mask is None:
             mask = new_mask
         else:

pandas/tests/tools/test_to_numeric.py

@@ -723,12 +723,12 @@ def test_precision_float_conversion(strrep):
 @pytest.mark.parametrize(
     "values, expected",
     [
-        (["1", "2", None], Series([1, 2, np.nan])),
-        (["1", "2", "3"], Series([1, 2, 3])),
-        (["1", "2", 3], Series([1, 2, 3])),
-        (["1", "2", 3.5], Series([1, 2, 3.5])),
-        (["1", None, 3.5], Series([1, np.nan, 3.5])),
-        (["1", "2", "3.5"], Series([1, 2, 3.5])),
+        (["1", "2", None], Series([1, 2, np.nan], dtype="Int64")),
+        (["1", "2", "3"], Series([1, 2, 3], dtype="Int64")),
+        (["1", "2", 3], Series([1, 2, 3], dtype="Int64")),
+        (["1", "2", 3.5], Series([1, 2, 3.5], dtype="Float64")),
+        (["1", None, 3.5], Series([1, np.nan, 3.5], dtype="Float64")),
+        (["1", "2", "3.5"], Series([1, 2, 3.5], dtype="Float64")),
     ],
 )
 def test_to_numeric_from_nullable_string(values, nullable_string_dtype, expected):
@@ -738,6 +738,24 @@ def test_to_numeric_from_nullable_string(values, nullable_string_dtype, expected
     tm.assert_series_equal(result, expected)
 
 
+def test_to_numeric_from_nullable_string_coerce(nullable_string_dtype):
+    # GH#52146
+    values = ["a", "1"]
+    ser = Series(values, dtype=nullable_string_dtype)
+    result = to_numeric(ser, errors="coerce")
+    expected = Series([pd.NA, 1], dtype="Int64")
+    tm.assert_series_equal(result, expected)
+
+
+def test_to_numeric_from_nullable_string_ignore(nullable_string_dtype):
+    # GH#52146
+    values = ["a", "1"]
+    ser = Series(values, dtype=nullable_string_dtype)
+    expected = ser.copy()
+    result = to_numeric(ser, errors="ignore")
+    tm.assert_series_equal(result, expected)
+
+
 @pytest.mark.parametrize(
     "data, input_dtype, downcast, expected_dtype",
     (