ENH: make Series.ptp() handle missing values

[ajcr]

Currently (in master), Series.ptp() is just implemented using np.ptp() and so the method will return nan for any Series that has one or more missing values:

>>> s = pd.Series([5, 0, np.nan, -3, 2])
>>> s.ptp()
nan

It is simple to write s.max() - s.min() instead, but the ptp() result is surprising as most pandas methods are designed to handle missing data gracefully. I think most users would expect the ptp() method to ignore NaN.

If there is any agreement as to whether ptp() should be changed, I would like to work on a pull request!

---

Extending the idea, it might be useful to have both DataFrame.ptp() and groupby.ptp() methods.

For this example DataFrame...

df = pd.DataFrame({'a': [1, 2, 2, 1, 1],
                   'b': [3, 11, 72, 46, 32],
                   'c': [1.2, 6.7, 13.9, np.nan, -7.7],
                   'd': ['v', 'w', 'x', 'y', 'z']})

...I would expect the following behaviour:

>>> df.ptp()
a      1
b     69
c   12.7
dtype: float64

>>> df.ptp(axis=1)
0     2.0
1     9.0
2    70.0
3    45.0
4    39.7
dtype: float64

>>> df.groupby('a').ptp()
    b    c
a         
1  43  8.9
2  61  7.2

Again, if there is any consensus from the community on whether these additional methods should be added, I'd be happy to work on the pull request.

[jreback]

absolutely. this should do .align, then .max()-.min(). Its here prob just as a convience.

want to do a pull-requests?

[ajcr]

Sure, I can work on this. It looks pretty straightforward, although I guess non-numeric columns will have to skipped over as 'str' - 'str' will raise an error, for example.

[jorisvandenbossche]

Certainly OK to fix the NaN issue in Series.ptp! (we could consider it a bug)

But, I am a bit more hesitant on the second part, adding it to DataFrame. Some reasons: 1) we already have many methods, and personally I don't think ptp has enough added value to justify addition, 2) it is really easy to do this yourself and 3) I also don't find ptp a very clear name.

[ajcr]

I agree ptp is very easily implemented by the user if they need ever it, so maybe it's not worth adding it as a new method to DataFrame and groupby for now.

In that case, I'll just change Series.ptp() so that it's written as max() - min() and so is NaN-aware.