Mario Rodriguez wrote:
>
>>> It seems that MAXIMA is well suited for doing many common types of least
>>> squares analysis (e.g., regression, analysis of variance). Much of the
>>> necessary code could be adapted from IML, SAS, etc. Is anyone working on
>>> this?...
>
> The stats package contains only a few inferential procedures. I have not
> worked on it last year for two main reasons:
>
> a) I needed a flexible drawing tool for plotting statistical graphics;
> so I began writing the draw package (now that nobody is hearing me, I
> can confess!). But this package is taken me more time than expected.
>
> b) Some users asked me why should I write new procedures in Maxima/Lisp,
> when there are so good and tested programs such as R (written in C).
>
> At this point I am not sure what to do with stats. Should we continue
> with the development of stats? or should we wait for a foreign functions
> call system in Maxima to call R routines? (I have no idea how hard will
> it be to implement this second alternative.)
Calling the R routines would be great, especially because statisticians
are constantly working on R's code. It's where the action is in
statistical computing.
On the other hand, I've found some useful things that maxima can do with
respect to bayesian methods that can be easier than in R. For example, I
was doing some bayesian inference by numerical normalization of the
posterior distribution, and using quadpack, rationals, and bigfloats can
be extremely helpful. Naive integration of likelihood functions is prone
to underflow (since they're the product of a lot of probabilities which
are often very small) and having maxima's numerical types allows for a
relatively simple and therefore understandable implementation of
posterior inference. I'd like to work some more on this type of thing
over the next few months.
As far as I can see, Bayesian methods are the future of statistics. if
maxima can become strong in this one area due to its useful symbolic,
and numerical combination, that would be great.