S3 vs S4 classes

I have developed a package, bayesm, which uses existing classes of 
objects.  I would like to add a new class corresponding to objects from 
this package.

I have been reading about classes and all sources tell me that I should 
use so-called "new" or S4 classes.

For new projects, perhaps: the issue is far less clear cut for existing 
ones.  In particular, the structure of model-fitting classes inherited 
from the White Book (lm, glm and so on) is S3 classes, and so related 
model-fitting functions (e.g. my polr) fit most naturally into the S3 
paradigm.

However, a major purpose of defining a class for my package would be to 
add methods to the existing generic functions:  print, plot, and 
summary.  My understanding is that these functions work with the "old" 
or S3 classes.

S4 classes should use show() rather than print(), according to the 
established practice (and I think the Green Book, but I do not have it to 
hand).  Certainly auto-printing uses show() for S4 objects in preference 
to print().

If I want to use S4 classes, do I need to write new generic functions or 
is there as way to "extend" the existing generics to work with S4 
classes and objects.  I may be using the wrong word here "extend" but I 
hope that everyone understands the point. If I have to write new generic 
functions, it is a simply matter of registering the pre-existing methods 
based on S3 objects?  Or do you basicaly have to start from scratch -- 
i.e. the two types of classes, methods and generic functions are 
distinct?

If you define an S4 method for any function, e.g. plot, it becomes S4 
generic and after looking for S4 methods will use the previous function 
(and if that is S3 generic, it will still look for S3 methods).

To my mind the advantage of S4 classes is that they are what JMC sometimes 
calls 'formal' classes: they are rigidly defined.  That suits them to new 
projects where you can think about the classes you want and define their 
structure and relationships in advance.  On the other hand, for 
development and experimenting that rigid definition can be a 
straightjacket.  I have moved various model-fitting functions to S4, and 
moved them back to S3 in frustration at the lack of fit to existing tools.

Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595