[Scipy-tickets] [SciPy] #1425: kurtosistest requires at least 4 observations
SciPy Trac
scipy-tickets@scipy....
Sun Sep 11 19:54:54 CDT 2011
#1425: kurtosistest requires at least 4 observations
-------------------------+--------------------------------------------------
Reporter: bsouthey | Owner: somebody
Type: defect | Status: needs_work
Priority: normal | Milestone: Unscheduled
Component: scipy.stats | Version: devel
Keywords: |
-------------------------+--------------------------------------------------
Changes (by josefpktd):
* priority: lowest => normal
Comment:
I was trying to look at this. I think that the stats.stats version should
raise a ValueError for n<5, similar to the change that Warren made. I
don't think returning (0, 1) is a good answer if the sample is to small.
R raises an exception
{{{
> x2 = c(0.31190987, 0.5403142 , -1.51626563, -0.64241117)
> anscombe.test(x2)
Error in if (pval > 1) pval <- 2 - pval :
missing value where TRUE/FALSE needed
In addition: Warning message:
In sqrt(1 + 4/m3) : NaNs produced
}}}
The values differ a little bit compared to R, there is a mistake in the
formula for A
{{{
A = 6.0 + 8.0/sqrtbeta1 *(2.0/sqrtbeta1 + np.sqrt(1+4.0/(sqrtbeta1**2)))
}}}
should be instead without the final `**2`, i.e.
{{{
A = 6.0 + 8.0/sqrtbeta1 *(2.0/sqrtbeta1 + np.sqrt(1+4.0/(sqrtbeta1)))
}}}
equation 3 in the reference (I don't have CRC available, but it wouldn't
be the first imprecision or typo in there).
correct reference
Distribution of the Kurtosis Statistic b2 for Normal Samples
F. J. Anscombe and William J. Glynn
Biometrika
Vol. 70, No. 1 (Apr., 1983), pp. 227-234
Published by: Biometrika Trust
Article Stable URL: http://www.jstor.org/stable/2335960
I'm trying to prepare a pull request in the next few days, but the **2 bug
should be removed, and some tests against R added, for 0.10.
--
Ticket URL: <http://projects.scipy.org/scipy/ticket/1425#comment:3>
SciPy <http://www.scipy.org>
SciPy is open-source software for mathematics, science, and engineering.
More information about the Scipy-tickets
mailing list