Skip to content

scores from multinomial logistic regression

4 messages · Jacqueline Hall, Brian Ripley

#
What do you mean by the scores?

What multinom does is to fit probabilities (which you can extract by 
fitted()): the response is a discrete probability distribution.
There is an underlying linear predictor but

(a) it is K-dimensional and
(b) there is a degree of ambiguity, usually resolved by setting the 
predictor for one category to zero (but not in this code).

That linear predictor is only generated in the underlying C code.
On Fri, 7 May 2004, Jacqueline Hall wrote:

            

  
    
#
Hi,

Sorry for not making this clear,

By "score" I meant the score from the score test, for assessing the addition
of a new variable to the model. (first derivative of the log likeihood/
information matrix, the ratio (score)having a chi squared distribution of
appropriate df)

I'm looking for something similar/appropraite for logistic regression, my
outcome (response) variable has 4 categories (hence the interest in
multinom), the covariates are continuous.

Thanks again,

Jacqui

-----Original Message-----
From: Prof Brian Ripley [mailto:ripley at stats.ox.ac.uk] 
Sent: 07 May 2004 15:07
To: Jacqueline Hall
Cc: r-help at stat.math.ethz.ch
Subject: Re: [R] scores from multinomial logistic regression


What do you mean by the scores?

What multinom does is to fit probabilities (which you can extract by 
fitted()): the response is a discrete probability distribution. There is an
underlying linear predictor but

(a) it is K-dimensional and
(b) there is a degree of ambiguity, usually resolved by setting the 
predictor for one category to zero (but not in this code).

That linear predictor is only generated in the underlying C code.
On Fri, 7 May 2004, Jacqueline Hall wrote:

            

  
    
#
Note the coxph docs call these `efficient scores', which helps. However,
what the `score' part of a coxph object gives is not the score for the
addition of a new variable, but for the addition of all the variables, as
I understand it. (You could do repeated fits using init=, but that's not
what you said.)

I am not aware of a good way to do that even for a glm. (Smart people
could make add1.lm do it, I guess.) I would just do likelihood ratio tests
since computation time is unlikely to be an issue these days.
On Fri, 7 May 2004, Jacqueline Hall wrote: