The example provided is not the best example. Anyway, let's assume that X
has some effect and I want to test whether this effect is different for
CASES A (vs CONTROLS) and for CASES B (vs CONTROL). How do I test this?
Gianluca
-----Original Message-----
From: Prof Brian Ripley [mailto:ripley@stats.ox.ac.uk]
Sent: Wednesday, 18 August 2004 5:15 PM
To: Gianluca Severi
Cc: 's-news@lists.biostat.wustl.edu'
Subject: Re: [S] constrained multinom()
I think you misunderstand the model. You have case A vs control and case
B vs control. Neither has a significant X coefficient, and so you almost
certainly have nothing significant.
On Wed, 18 Aug 2004, Gianluca Severi wrote:
> Dear S-plus users,
>
> I have fitted a polytomous logistic regression model with the function
> multinom in the nnet library. I have a response Y with 3 categories and a
> predictor with 2 categories. Now I'm interested in testing whether the
> effect of X is different according to the type of response. Any idea on
how
> to do this?
>
> > my.table <- expand.grid(Y=0:2,X= 0:1)
> > temp <- c(612,477,209,107,83,42)
> > my.table <- apply(my.table,2,function(x){rep(x,temp)})
> > my.table <- data.frame( Y =
> factor(my.table[,"Y"],labels=c("CONTROL","CASE A","CASE B")),
> + X =
> factor(my.table[,"X"],labels=c("NO","YES"))
> + )
> > library(nnet)
> > library(MASS)
> > options(contrasts=c("contr.treatment","contr.poly"))
> > fit <- multinom(Y ~ X, data = my.table)
> # weights: 9 (4 variable)
> initial value 1680.876802
> final value 1559.227816
> converged
> >
> >
> >
> > fit
> Call:
> multinom(formula = Y ~ X, data = my.table)
>
> Coefficients:
> (Intercept) X
> CASE A -0.2492187 -0.004772402
> CASE B -1.0744047 0.139230847
>
> Residual Deviance: 3118.456
> AIC: 3126.456
>
> > summary(fit)$standard.errors
>
> Re-fitting to get Hessian
>
> (Intercept) X
> CASE A 0.06107717 0.1585067
> CASE B 0.08011675 0.1989328
>
>
> The effect of X is clearly not different in CASE A and CASE B how can I
> formally test this hypothesis? How can I fit the same model constraining
the
> two coefficients for X to be identical?
--
Brian D. Ripley, ripley@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
|