Date:         Wed, 28 Sep 2005 10:34:24 -0400
Reply-To:     "S. E. Hayman-Abello" <hayman2@uwindsor.ca>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         "S. E. Hayman-Abello" <hayman2@uwindsor.ca>
Subject:      Re: Interpretation of logistic regression ' results
In-Reply-To:  <S32875AbVI1Ngd/20050928133636Z+5948@avas-mr13.fibertel.com.ar>
Content-type: text/plain; charset="US-ASCII"

Hector and all,

Continuing with logistic regression. How do you interpret the same finding
as Jan reports (i.e., significant predictor at model level but parameter
estimate not significant) with continuous variables?

I have 3 outcome (DV) categories (Worse/No Change/Better) and two of my
continuous variables show significance at the model level but not within
either the 'Worse' or the 'No Change' parameter estimates. My reference
category is 'Better'. I wonder now, based on Hector's advice to Jan, what
would happen if I changed the reference category to 'Worse'. Would I do that
in SPSS NOMREG by assigning the Worse category the highest rather than the
lowest value?

There are the current findings:

Model
ANARTIQ                        (chi) 9.826      2       .007


 (Par. Est.)    B       S.E.        Wald          df      Sig     Exp(B)

'Worse'
ANARTIQ     -.089   .049    3.310         1       .069     .915

'No Change'
ANARTIQ     .041    .054        .578        1       .447     1.042



Thank you for any help,
Sue

PS I also have a significant dichotomous categorical predictor (MEDPROB:
yes/no) that has a negative parameter estimate (B = -1.730) and a very low
odds ratio (Exp(B) = .177). I'm having a brain cramp about interpreting this
finding. My take is that the odds of a subject being in reference (Better)
group (compared to the No Change group) are lower (bc. of the negative B?)
for subjects with Yes on MEDPROB. If that's right, what do I say more
specifically about the  .177 odds ratio? I'm used to thinking in terms of
odds ratios >1.



Hector Maletta said ...

> The coefficients of dummies measure the significance of the difference
> between one category and the reference category. In your case, the reference
> category seems to be "in the middle", since two of the other categories have
> a negative coefficient and one category (represented by the DIAG(3) dummy)
> has a positive coefficient. I suspect that if you use the category
> represented by DIAG(3) as your reference category, the other two shown in
> the table would show a more significant (i.e. lower) probability, because
> they have a larger difference with that category than they have with your
> current reference category.
> However, I think what is more important is the coefficients of the
> categories, and not so much the overall effect of the categorical variable
> as such. If DIAG alludes to diagnostic, I suppose it is more important what
> the diagnostic is (i.e. which category), than the fact that some diagnostic
> existed.
>
> Hector