[R] factors in multinom function (nnet)

Bill.Venables@csiro.au Bill.Venables at csiro.au
Wed Apr 13 01:57:56 CEST 2005


Alexandre,

I have a couple of remarks to make, not all of which you might find
immediately helpful, I regret to say.

* The choice between using predictors linearly or in factor versions is
a modelling choice that is in no way specific to multinom.  It is a
general aspect of modelling that has to be faced in a whole variety of
situations.  Indeed the full spectrum of choices is much wider than
this: linear, polynomials, splines, different sorts of splines, harmonic
terms, factors, ...  In fact the idea behind gam's was really to allow
some of this extensive field of choices to be model driven, but I
digress.  Point 1: you need to learn about modelling first and then
apply it to multinom.

* It is curious to me that someone could be interested in multinomial
models per se.  Usually people have a context where multinomial models
might be one approach to describing the situation in a statistically
useful way.  Another could be something like classification trees.  The
context is really what decides what modelling choices of this kind might
be sensible.

* There is an obvious suggestion for one reference, a certain notorious
blue and yellow book for which multinom is part of the support software.
I believe they discuss some of the alternatives as well, like
classification trees, and some of the principles of modelling, but it's
been a while since I read it...

* Frank Harrell recently issued an excellent article on this list on
brain surgery in a hurry to which you may usefully refer.  I believe it
was on April 1.

Bill Venables.

-----Original Message-----
From: r-help-bounces at stat.math.ethz.ch
[mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of Alexandre Brito
Sent: Wednesday, 13 April 2005 8:20 AM
To: r-help at stat.math.ethz.ch
Subject: [R] factors in multinom function (nnet)


Dear All:

I am interested in multinomial logit models (function multinon, library
nnet) but I'm having troubles in choose whether to define the predictors
as factors or not.

I had posted earlier this example (thanks for the reply ronggui):

worms <- data.frame(year = rep(2000:2004, c(3,3,3,3,3)),
				age = rep(1:3, 5), 
	mud = c(2,5,0,8,7,7,5,9,14,12,8,7,5,13,11),
	sand = c(4,7,13,4,14,13,20,17,15,23,20,9,35,27,18), 
	rocks = c(2,6,7,9,3,2,2,10,5,19,13,17,11,20,29))

k <- as.matrix(worms[,3:5])

(mud, sand and rocks are factors;  age and year are predictors)

Now there are several possibilities:

m1 <- multinom(k ~ year+age, data = worms)
m2 <- multinom(k ~ factor(year)+age, data = worms)
m3 <- multinom(k ~ year+factor(age), data = worms)
m4 <- multinom(k ~ factor(year)+factor(age), data = worms)
m5 <- multinom(k ~ year:age, data = worms)
m6 <- multinom(k ~ year*age, data = worms)
m7 <- multinom(k ~ factor(year):age, data=worms)
m8 <- multinom(k ~ year:factor(age), data=worms) 

and so on.

I am far from an expert on this, and I would like to learn more about
the utilization of multinom function in R and the kind of doubts I
described above. So I hope that someone can recommend me some references
in this matter (internet, books...) if any is available. 

Thanks in advance, best wishes 

Alexandre
______________________________________________
R-help at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html




More information about the R-help mailing list