BUG/ENH: should idxmax/idxmin work on object types?

[jseabold]

• numeric_only=True flag to min,max,idxmax,idxmin to select out numeric only
• fix idxmax on object dtypes (related to BUG: min on non-numeric data with nans #4147)

related is #6287

Just happened to come across this. Related to #2982

df = pd.DataFrame([["zimm", 2.5],["biff", 1.],["bid", 12.]])

This is fine (I was surprised it worked for objects)

df.max()

This falls down

df.idxmax()

I'd be fine if there's a strong performance reason not to support this, though it would be nice to not return an error at least. Maybe a stop-gap solution is to only return a solution for numeric types similar to the quantiles method.

[jreback]

yep...related is #4147, should prob work

could prob support a numeric=True keyword to only work on numeric (and dates?) types, otherwise return for all types? (you can of course do this with df._get_numeric_data(), but maybe easier this way)

what do you think?

[jseabold]

Yeah, I found _get_numeric_data looking at quantiles. Didn't want to point people to the private method but would make an easy PR.

[jreback]

yep...have the numeric_only flag on quantile....I think that's the best way to go here (add to min,max,idx,idxmax...

[jseabold]

Might be nice to also support this for objects at some point. I'm pretty sure there's a deterministic ordering for objects in Python 2/3, maybe using __repr__? Not sure. I remember discussing this for patsy. Can't find the discussion right now though.

[jreback]

yep...see issue #4147, this is actually currently a numpy bug (works for max but not for min, when having nan present)....pretty easy fix....should default then be numeric_only=False? as max...et al do make sense generally here