Skip to content

about comparison of KURTOSIS in package: moments and fBasics

2 messages · Baize, Harold, Spencer Graves

Baize, Harold
# 
Thanks to Spencer Graves for providing links to explain
the various types of kurtosis reported by R packages. 

Spencer Graves>>(http://mathworld.wolfram.com/k-Statistic.html).

Spencer also said:

SG>>	  However, these are little used, as the estimates are known to be so 
SG>> highly variable.  It is generally preferred to transform to normality or 
SG>> to use some other distribution and then use maximum likelihood.

This advice is good if your interest is comparing models, but what 
if variation in kurtosis itself is your interest? I am wondering if 
someone could provide some direction for answering questions about 
differences between samples in kurtosis. There are tests of 
significance for means and variance. How would one test hypotheses 
of difference in kurtosis between samples? 

Thanks in advance. 

Harold Baize
Youth Services Evaluator
Butte County Department of Behavioral Health
3 days later
Spencer Graves
# 
There are doubtless tests for kurtosos by itself, though I'm not 
familiar with any.  When I'm conderned about kurtosis (which is often), 
I routinely make normal probability plots of observations and residuals 
from model fits.  If I see roughly a straight line, I conclude that I 
won't likely be too misled by assuming normality.  If I see a smooth "S" 
shape with long tails, I would be inclined to try to fit a Student's t 
model.  If I see relatively sharp breaks, that's the fingerprint of a 
mixture.

	  For mixture distributions, I like Titterington, Smith, and Makov 
(1986) Statistical Analysis of Finite Mixture Distributions (Wiley). 
There are more recent books available, but I haven't seen a better 
discussion of normal plots of mixtures.  There are R packages for 
estimating mixtures of various kinds.  To find them, I suggest you try 
"RSiteSearch" with a variety of different key phrases.  For Student's t 
models, I suggest you try "fitdistr" in library(MASS) or "RSiteSearch" 
(or Google).

	  If you'd like further help from this listserve, I suggest you PLEASE 
do read the posting guide! "www.R-project.org/posting-guide.html". 
Anecdotal evidence suggests that post closer to the style recommended in 
this guide tend to get more useful replies quicker.

	  hope this helps.
	  spencer graves
Baize, Harold wrote: