Definition of Kurtosis value in Statistics and GroupBy Nodes

[johnprime]

Hi All

I’m trying to find out whether the Kurtosis value calculated in the Statistics and GroupBy Nodes is the standard Kurtosis value, excess Kurtosis (kurtosis - 3) or an alternative form.
The description info for the Statistics node has no guidance on this at all, while the description tab in GroupBy node’s configuration window contains:-
Kurtosis
“Calculates the kurtosis per group. Attention: calculation is bias-corrected and at least four values per group are required. If the latter does not hold, a missing cell is returned.”

It’s not clear what type of bias-correction is used and if this correction is on standard or excess kurtosis or is in fact the calculation for the latter.

From my testing I don’t believe the value can be standard kurtosis as I can get negative values (standard kurtosis should always be positive) so the output could be excess kurtosis or calculated by an alternative method.

Can anybody provide a definitive answer on this?

In general it would be good to capture this level of detail in node descriptions so users fully understand how statistics/values are calculated by a particular node.
As much as I love KNIME, currently I find it difficult to go out there and advocate for it as effectively as I would like when I can’t defend it from accusations of it being a black box solution because of these kinds of issues. I would be happy to contribute to the updating of node descriptions to help improve this situation.

Many thanks for your help

John

[izaychik63]

You can look at Alternative Definition of Kurtosis here
itl.nist.gov/div898/handbook/eda/section3/eda35b.htm

[Mark_Ortmann]

@johnprime,
this is how it is calculated.
https://commons.apache.org/proper/commons-math/javadocs/api-3.3/org/apache/commons/math3/stat/descriptive/moment/Kurtosis.html

Hope this helps?!?

Mark

[johnprime]

Thank you @Mark_Ortmann and @izaychik63 for the info and links

It took a bit of checking and testing to reach a conclusion on this.
The value returned is excess kurtosis
The java calculation used in the Statistics and GroupBy Nodes exactly matches the 2nd kurtosis equation in this article - https://www.spcforexcel.com/knowledge/basic-statistics/are-skewness-and-kurtosis-useful-statistics

Here it not only subtracts 3 from the kurtosis value to give excess kurtosis (with a normal distribution = 0); it also takes into account the sample size.

Hope folks find that useful

Thanks

John