BUG: Can't add a missing value to an int64 or Int64 column without it being upcast inconsistently

[Xnot]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

import numpy as np
import pandas as pd

# gets upcast to object
df = pd.DataFrame({"a": [1, 2, 3]}, dtype="int64")
df.loc[4] = pd.NA

# gets upcast to float64
df = pd.DataFrame({"a": [1, 2, 3]}, dtype="int64")
df.loc[4] = np.NaN

# gets upcast to object
df = pd.DataFrame({"a": [1, 2, 3]}, dtype="Int64")
df.loc[4] = pd.NA

# gets upcast to Float64
df = pd.DataFrame({"a": [1, 2, 3]}, dtype="Int64")
df.loc[4] = np.NaN

# can hold a missing value when initialized with it and remain Int64
df = pd.DataFrame({"a": [1, 2, 3, pd.NA]}, dtype="Int64")
# then gets upcast anyway when you add a second missing value
df.loc[4] = pd.NA

Issue Description

Int64 can hold missing values, however when adding a missing values to an int64 or Int64 column, it gets upcast to float64, Float64, or object.

Expected Behavior

int64 should be upcast to Int64 and Int64 should not be upcast at all.

Installed Versions

1.4.2

[rhshadrach]

Thanks for the report! Certainly 1, 3, and 5 look like definitive bugs to me. For the other two:

• 2: np.NaN is a float, so this looks correct to me. We do not want to upcast to Int64 while it's marked as experimental
• 4: This is discussed in ROADMAP: Consistent missing value handling with new NA scalar #28095 (comment) and still open

[rhshadrach]

For 1, my thinking here is that if the user encounters pd.NA, then they are working the nullable dtypes and so upcasting to them is okay.

[phofl]

1 is not a but currently, we treat pd.NA as object in numpy dtypes by design as far as I am aware.

3 and 5 are bugs but have the same cause. We seem to cast to object when enlarging the DataFrame. This works as expected when overwriting an existing value

[simonjayhawkins]

3 and 5 are bugs but have the same cause.

duplicate of #32346?

(There is already some discussion there on the conversion to object dtype. if 3 and 5 are covered by that discussion, 1 and 2 are not bugs and 4 is already covered also, should probably close this issue to help keep discussion in one place)

[phofl]

Yes you are correct. I‘ll try to look into this.

can close here but maybe copy the NA case over to add tests later