ENH: support downcasting of nullable EAs in pd.to_numeric

pandas/core/tools/numeric.py

@@ -7,6 +7,9 @@
  ensure_object,
  is_datetime_or_timedelta_dtype,
  is_decimal,
+ is_extension_array_dtype,
+ is_float_dtype,
+ is_integer_dtype,
  is_number,
  is_numeric_dtype,
  is_scalar,
@@ -15,6 +18,8 @@
 from pandas.core.dtypes.generic import ABCIndex, ABCSeries
 
 import pandas as pd
+from pandas.core.arrays.numeric import NumericArray
+from pandas.core.construction import extract_array
 
 
 def to_numeric(arg, errors="raise", downcast=None):
@@ -118,10 +123,14 @@ def to_numeric(arg, errors="raise", downcast=None):
  is_series = False
  is_index = False
  is_scalars = False
+ is_numeric_extension_dtype = False
 
  if isinstance(arg, ABCSeries):
  is_series = True
  values = arg.values
+ if is_extension_array_dtype(arg) and isinstance(values, NumericArray):
+ is_numeric_extension_dtype = True
+ values = extract_array(arg)
  elif isinstance(arg, ABCIndex):
  is_index = True
  if needs_i8_conversion(arg.dtype):
@@ -142,6 +151,12 @@ def to_numeric(arg, errors="raise", downcast=None):
  else:
  values = arg
 
+ if is_numeric_extension_dtype or (
+ is_extension_array_dtype(arg) and isinstance(values, NumericArray)
+ ):
+ is_numeric_extension_dtype = True
+ mask, values = values._mask, values._data
+
  values_dtype = getattr(values, "dtype", None)
  if is_numeric_dtype(values_dtype):
  pass
@@ -188,6 +203,16 @@ def to_numeric(arg, errors="raise", downcast=None):
  if values.dtype == dtype:
  break
 
+ if is_numeric_extension_dtype:
+ if is_integer_dtype(values):
+ from pandas.core.arrays import IntegerArray
+
+ values = IntegerArray(values, mask)
+ elif is_float_dtype(values):
+ from pandas.core.arrays import FloatingArray
+
+ values = FloatingArray(values, mask)
+
  if is_series:
  return arg._constructor(values, index=arg.index, name=arg.name)
  elif is_index:

pandas/tests/tools/test_to_numeric.py

@@ -725,3 +725,19 @@ def test_to_numeric_from_nullable_string(values, expected):
  s = Series(values, dtype="string")
  result = to_numeric(s)
  tm.assert_series_equal(result, expected)
+
+
+@pytest.mark.parametrize(
+ "data, input_dtype, downcast, expected_dtype",
+ (
+ ([1, 1], "Int64", "integer", "Int8"),
+ ([450, 300], "Int64", "integer", "Int16"),
+ ([1, 1], "Float64", "float", "Float32"),
+ ([1, 1], "Float64", "integer", "Int8"),
+ ),
+)
+def test_downcast_nullable_numeric(data, input_dtype, downcast, expected_dtype):
+ arr = pd.array(data, dtype=input_dtype)
+ result = pd.to_numeric(arr, downcast=downcast)
+ expected = pd.array(data, dtype=expected_dtype)
+ tm.assert_extension_array_equal(result, expected)