ENH: More permissable conversion to StringDtype

[topper-123]

Currently, converting directly to StringDtype is not really safe unless you know you've got a str type:

>>> pd.Series([1,2, np.nan], dtype=str)[0]
'1'  # string, works
>>> pd.Series([1,2, np.nan], dtype="string")
ValueError  # does not work
>>> pd.Series([1,2, np.nan], dtype=str).astype("string")
# works
>>> pd.Series([1,2, np.nan]).astype("string")
ValueError  # does not work
>>> pd.Series([1,2, np.nan]).astype(str).astype("string")
# works

The net result is that we have to wrap conversions, so in general will have to do x.astype(str).astype("string"), which seems unnecessary.

This extends to other conversions to StringDtype, e.g. for read_excel we have to do:

>>> df = pd.read_excel(..., dtype={"col_x": str}).astype({"col_x": "string"})

where directly setting dtype={"col_x": "string"} would seem more natural.

I think we should be able to convert directly to StringDtype by internally converting first to str, if needed. I might be missing some nuance why this is not allowed, though.

Any comment on loosening up here?

[jorisvandenbossche]

This might be a duplicate of #31204, at least the astype("string") is discussed there.

And I suppose that if astype("string") works, doing pd.Series(.., dtype="string") should work similarly?

[topper-123]

Yeah, it’s similar. I’ll close this.