Robert Kern wrote:
On 2009-09-02 14:15 PM, Raymond Hettinger wrote:
On Sep 2, 6:51 am, Thomas Philips<tkp...@gmail.com> wrote:
While the random module allows one to generate randome numbers with a
variety of distributions, some useful distributions are omitted - the
Student's t being among them.
I'm curious to hear what your use cases are.
My understanding is that t-distribution is an estimation tool
used with small samples of a population where the variance or
standard deviation is unknown.
So, when do you ever need to generate random variables with
this distribution? ISTM that this is akin to wanting
a generator for a Kolmogorov distribution -- usually the
distribution is used to test empirical data, not to generate it.
In more complicated models, estimates of one parameter need to be
propagated through the model, particularly if you are looking at
sensitivity to parameters. Student's t describes the variation of an
estimate of a mean of a sample from a Gaussian distribution. If I were
modeling a processing wherein someone makes an estimate of a mean and
then acts on that estimate, I would want to generate random t variates
to feed that model.
I think most of the existing generators were chosen because they
are useful in simulation programs. AFAICT, the Student's t-
distribution
doesn't fall into that category (usually, you know the population
standard deviation when you're the one generating data).
Student's t distribution is also used as a sort of generic fat-tailed
distribution in some models and is not tied to the "estimate of a mean"
description.
Yes. One can run the simulation with various df's to see the effect of
such data and possibly how and when a model breaks.
ISTM, there ought to be a statistics module that can calculate
cumulative distribution functions for a variety of distributions.
This would be far more helpful than creating more generators.
Yes, scipy.stats.
Is that stable enough so that all or part could be added to stdlib?
tjr
--
http://mail.python.org/mailman/listinfo/python-list
