testing on octave 8.0.1 with dataframe 1.2.0:

>> x = dataframe(rand(5,1))
warning: genvarname is obsolete; use matlab.lang.makeValidName or matlab.lang
.makeUniqueStrings instead
x = dataframe with 5 rows and 1 columns
_1      X
Nr double
 1 0.9377
 2 0.8110
 3 0.9436
 4 0.9106
 5 0.1947

>> skewness (x.data)
ans = -1.4022

>> skewness (x)
ans = dataframe with 1 rows and 1 columns
_1       X
Nr  double
 1 -1.4022

>> kurtosis (x.data)
ans = 3.1050

>> kurtosis (x)
ans = dataframe with 1 rows and 1 columns
_1      X
Nr double
 1 3.1050

(all still good there for first report problem)

   嬂᪁>> x = dataframe (zeros (5, 1))
x = dataframe with 5 rows and 1 columns
_1      X
Nr double
 1      0
 2      0
 3      0
 4      0
 5      0
>> skewness (x.data)
ans = NaN
>> skewness (x)
error: x(1): out of bound 0 (dimensions are 0x0)
error: called from
    df_mapper>@<anonymous> at line 28 column 30
    df_mapper at line 28 column 15
    power at line 22 column 8
    skewness at line 122 column 10
>> kurtosis (x.data)
ans = NaN
>> kurtosis (x)
error: x(1): out of bound 0 (dimensions are 0x0)
error: called from
    df_mapper>@<anonymous> at line 28 column 30
    df_mapper at line 28 column 15
    power at line 22 column 8
    kurtosis at line 125 column 10

so status unchanged from comment #9.  retitling to focus on one remaining issue, which was suggested in comment #2 should be handled better by overloading the functions in the dataframe class.

Hello,
I will be away up to June 20th. I will look about this issue after this date.

Regards

Pascal

@jbect can this bug report be closed?

Ok, my question was indeed about the "closed" status of this bug report ;-)

Where is the new bug report ?

Agree with Julien that fixing the problem within the dataframe package makes more sense in this case. Core functions can't be expected to know about arbitrary data types.

In the file statistics/base/skewness.m, a test is performed to detect variables whose std is 0. As the skewness implies taking the inverse of the std, the skewness for those values is simply set to 0, to avoid division by 0. This is accomplished at lines 79-83. retval is initialised to be zeros of the class of the input. For a user-defined class, this fails with
"zeros: invalid data type specified"
Moreover, overloaded funcs for a dataframe object takes care about propagating meta-information. This way, the line
retval(ind) = x(ind) ./ (c * s(ind) .^ 3)
implies a lot of work, as we're dealing with 3 variables. I thus propose the enclosed patch. Ideas:

ind = (s > 0); <= keep tracks of thing to process some way/some other way
retval = sum (x .^ 3, dim); <= retval receive all its meta-inf from x

if !isobject(x),
    retval(~ind) = zeros (1, class (x)); <= fast way
else
    retval(~ind) = 0; %# let the class subsasgn handle the cast
endif   

retval(ind) = retval(ind) ./ (c * s(ind) .^ 3);  

Regards

Pascal