BUG: DataFrame[Int64].mean().dtype is object, should be Float64

[jbrockmendel]

This breaks the last corner case I want to test for #33036.

arr = np.random.randn(4, 3).astype("int64")
df = pd.DataFrame(arr).astype("Int64")
df.iloc[:, 1] = pd.NA  # <-- incorrectly casts to object, lets cast back and ignore that for now
df = df.astype("Int64")

res = df.mean()

>>> res
0     0.0
1    <NA>
2   -0.25
dtype: object

Trivial to fix with 2D EAs, not sure how/if to fix it without.

[Demetrio92]

df.iloc[:, 1] = pd.NA  # <-- incorrectly casts to object, lets cast back and ignore that for now

this here is a whole separate issue.

Reported here #44199

[Demetrio92]

Also, I now know how to fix your example.

Fixed:

arr = np.random.randn(4, 3).astype("int64")
df = pd.DataFrame(arr).astype(pd.Int64Dtype())
df.iloc[:, 1] = pd.NA  # no more casting issues
# df = df.astype("Int64")  # not needed anymore

res = df.mean()
>>> res
0   -0.50
1     NaN
2    0.25
dtype: float64

[jbrockmendel]

Also, I now know how to fix your example.

Not quite. The point of the example is that we get unwanted behavior when we have an all-NA Int64 column

[Demetrio92]

@jbrockmendel pd.NA is experimental. You can use np.nan instead, and it again will fix your issue.

This is really not about .mean or an all-NA column. Anything in pandas that will touch pd.NA will be "broken" if using current default types.

[jreback]

@Demetrio92 maybe it's not clear
@jbrockmendel opened this issue to have a reference in order to work in it - how else do you think things get fixed?

[Demetrio92]

@jreback
As I see it @jbrockmendel is trying to test whether .mean is correctly working on nullable integers. I can show that it does, and the problem is in .astype("Int64"), which seems not to do what it is intended.

So, for testing .mean, you can use .astype(pd.Int64Dtype()), and open a separate issue on why .astype(pd.Int64Dtype()) does not do the same as astype("Int64"). Which deserves a way more thorough investigation and a separate issue.

Issues get fixed, when devs see what is the problem. We can create 1000 tickets all stating the same ".dtype is object, should be Float64" by using .mean, .var, .median. It will create a lot of noise and get nothing fixed.

[jbrockmendel]

you can use .astype(pd.Int64Dtype()), and open a separate issue on why .astype(pd.Int64Dtype()) does not do the same as astype("Int64").

They look the same to me. If you think there is something different, please open a separate issue for that to avoid disrailing this one.

test whether .mean is correctly working on nullable integers. I can show that it does

Incorrect. In both your example and the OP, after setting df.iloc[:, 1] = pd.NA, we end up with df.dtypes[1] == object. Without re-casting (the need for which you correctly identify as a separate issue) the .mean call is no longer testing the dtype of interest.

[Demetrio92]

I was digging into the related issue. And figured out that despite that casting problem .mean() actually works as OP expects it if the second casting is removed.

arr = np.random.randn(4, 3).astype("int64")
df = pd.DataFrame(arr).astype("Int64")
df.iloc[:, 1] = pd.NA  # <-- incorrectly casts to object, lets cast back and ignore that for now
# df = df.astype("Int64")

res = df.mean()

Not sure why, but the above code works as expected

0    0.00
1     NaN
2   -0.25
dtype: float64

Tested using pandas 1.3.4