[R] R routines vs. MATLAB/SPSS Routines

[Douglas Bates]

On 10/26/07, Frank Thomas <news.ftr at free.fr> wrote:
...
> BTW: Contrary to some ideas both R  & SPSS can be programmed and the
> algorithms for both have been published. So, no matter whether open
> source or private property you know what you do (if you want).

This is off the point of Matt's original question and I apologize for
hijacking the thread but where have the algorithms that underly the
SPSS procedures been published?

My particular interest is in the methods for the linear mixed models
implemented in MIXED in SPSS (and also PROC MIXED in SAS).  A person
who was quite enthusiastic about the MIXED procedure in SPSS sent me a
PDF file about MIXED that I suppose could be considered a description
of the algorithms as long as you didn't read it too closely.  However,
the descriptions are far too vague to use them as a basis for writing
code and furthermore they jump back and forth between two or three
different representations of the model without tying the different
threads together.  There is no indication of what representation forms
the basis of the code and how the calculations are implemented.  Even
more alarming, parts of it are flat-out wrong.  Even the mixed-model
equations as given in this document are wrong, as one would quickly
find out if one tried to implement them.  The organization is
disjointed and generally the language and grammar indicate that it has
not been copy edited carefully.  I would not give it a good grade if
it were submitted as a project report in my statistical computing
course.

I have been unable to trace the source of this document.  It is
definitely a discussion of the computational algorithms in MIXED but I
haven't been able to track its original source.  In a way I hope it
was a preliminary draft or something like that.  If SPSS released this
version as an official publication it is a sign that they have fallen
on hard times.

If one simply needs to have a reference to cite regarding the
calculations done n your analysis then a document like this, suitably
corrected, might do.  I don't know if this is what you meant by saying
that the algorithms have been published but if it is then this is not
sufficient for the purposes of investigating the computational
methods.  A few years ago Brian Ripley mentioned in a presentation, I
believe to the Royal Statistical Society, the need for a reference
implementation of a statistical computing technique.  That is, one
should provide code that is accessible to, usable by and modifyable by
other researchers in the field.  (Brian: If you know the presentation
to which I am referring, can you provide a reference or URL please?)

Perhaps not for students who simply need to apply a few statistical
techniques but certainly for researchers in the statistical techniques
and computational methods there is a world of difference between a
closed-source implementation with a supplementary document describing
the alleged computation method and an open-source implementation.