🚀 go-pugleaf

Dear Stat-l members,

I have recently joined this list and thought I would submit this message
as much to check that I can do it as for any other reason.

I have been interested in the Minitab v Glim/Genstat discussion because I
am familiar with all three packages. I thought that I would share my
impressions with the group because I suspect I have encountered them in
a different order from many.

I have used Genstat off and on since the late 70s when I encountered it
while working as a government-employed statistical consultant [NZ Ministry
of Agriculture and Fisheries]. It's great strength is in the analysis of
designed experiments which is original reason we first purchased it.
Glim I met at about the same time but I never really used it much as its
capabilities were duplicated within Genstat. It is a much smaller program,
though, so it is more suited to microcomuters than Genstat.

Genstat is not well-suited to interactive use, however and I found myself
using Minitab more and more for data exploration as my consulting problems
grew messyer and messyer. When I became an academic a few years ago I
adopted Minitab because of is simplicity, and because I wanted to stress
the importance of examination of the data before doing statistical
analyses [and after].

My current research involves generalizations of GLMs [ G^{2}LMs ? ] and
I find myself writing collections of Minitab macros to fit these models
using IRLS. I do this because the Minitab environment is like being
inside a friendly debugger all the time. You can get a lot of insight
into how an algorithm is progressing by doing things like plotting
iterative weights against predictors at different iteration steps.
I still use Genstat, though, to check that my Minitab macros are giving
the same answers as all the Genstat output structures.


Allan Reese writes

                      " ....., I suggest that the Glim
approach and notation IS a useful and usable approach to analysing
binary data. In particular, the specification of models in both Glim and
Genstat, in terms of an error distribution and a set of predictive
variates and factors - with notation for interactions - seems to me
far easier to explain that the Byzantine ramblings of the SPSS Manova
command ... and I say that as an SPSS user of 20 years sitting."


The success of Genstat and Glim is based on the marriage of the
Wilkinson/Rogers (Applied Statistics, 1973) model description syntax to
algorithms of Wilkinson, Nelder and others. A model specification
language based on this is described in Chapter 12 of "Computation for the
Analysis of Designed Experiments" by Richard Heiberger (Wiley, 1989) who
also mentions other implimentations such as in P-STAT. A similar syntax
is also implimented in S : see Chapter 2 of "Statistical Models in S"
by John Chambers and Trevor Hastie, Wadsworth&Brooks/Cole 1992.


One final remark : I am getting tired of all the dodges that I have to get
up to to compensate for Minitab's lack of even rudimentary language
control structures like IF ... THEN ... ELSE . I am thinking of shifting
to XLISP-STAT or S+ for my research computing.

Murray Jorgensen       Department of Mathematics and Statistics
maj@waikato.ac.nz      University of Waikato, Hamilton, NZ