Re: [R] Analysing ordinal/nominal data

From: Piotr Majdak <>
Date: Fri 17 Jun 2005 - 20:54:15 EST

Prof Brian Ripley wrote:

> You have still not defined v' and w', nor the scores (are they estimated > or not). But the model I suggested is such a model with scores 1,2,...

Sorry for that, here it is:
scores v and w: integer scores, reflecting the ordering of columns/rows. Agresti suggests to use 1,2,3,... as you did. v' and w': mean of v and w, respectively

> As I suggested before, a binomial logistic model is appropriate here,
> not a Poisson log-linear one. (They are equivalent, but the binomial
> version is easier to interpret and less wasteful to fit.)

After your suggestions I've read the logit-Chapter in Agresti (1984) and tried to fit a logit model to my data:

log(m_ij2/m_ij1) = intercept + beta^PR_i(u_i-u') +


mijk: expected frequencies for cell with indicies i,j,k u_i: scores of PR with u={0,1,2,3}
v_j: scores of ENV with v={0,1,2}
beta^PR: association parameter for PR with responses beta^ENV: association parameter for ENV with responses

I wrote in R:



fit=glm(countM ~ as.integer(PR)+as.integer(ENV),


R gives me:


                 Estimate Std. Error z value Pr(>|z|)
(Intercept)      0.18435    0.07737   2.383 0.017179 *
as.integer(PR) -0.10538 0.03085 -3.416 0.000635 *** as.integer(ENV) -0.07075 0.04748 -1.490 0.136191

Looking at my data, I know that for PR=0 I have a very weak dependence on ENV and for PR=3 very strong dependence on ENV. How can I get this interpretation from the summary above? I'm new in R: is fit$linearpredictors what I'm looking for?

Piotr Majdak
Institut für Schallforschung
Österreichische Akademie der Wissenschaften
Reichsratsstr. 17
A-1010 Wien
Tel.: +43-1-4277-29511
Fax: +43-1-4277-9296

______________________________________________ mailing list
PLEASE do read the posting guide!
Received on Fri Jun 17 21:26:16 2005

This archive was generated by hypermail 2.1.8 : Fri 03 Mar 2006 - 03:32:47 EST