Date: Thu, 4 Aug 2005 10:44:48 -0300
Reply-To: Hector Maletta <hmaletta@fibertel.com.ar>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Hector Maletta <hmaletta@fibertel.com.ar>
Subject: Re: Reference Category for Categorical Vars in NOMREG
In-Reply-To: <ba.7735aab3.3023748c@aol.com>
Content-Type: text/plain; charset="US-ASCII"
Very good idea, Marija, for the case of dichotomous variables (reversing the
coding from 0 to 1 and vice viceversa would also do).
In the case of variables with more than two categories, the solution would
be to choose an intercept category (a.k.a. reference category) and code it
with the highest number or letter so it will be the last one.
Hector
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU]
> On Behalf Of Marija Norusis
> Sent: Thursday, August 04, 2005 10:39 AM
> To: SPSSX-L@LISTSERV.UGA.EDU
> Subject: Re: Reference Category for Categorical Vars in NOMREG
>
> If you declare the dichotomous variable to be a covariate
> instead of a factor, the original coding is maintained.
>
> Marija Norusis
> _www.norusis.com_ (http://www.norusis.com)
>
> In a message dated 8/4/2005 8:28:33 A.M. Central Standard
> Time, mscoppettone@IAGR.NET writes:
>
> The lesson on the semantics of "reference category" aside,
> my point is the
> following:
>
> For a binary variable coded as 0 or 1, NOMREG is estimating
> the effect of taking the value of 0 while setting the
> coefficient for level 1 equal to 0. In general, a "0" is
> used to indicate that the characteristic is not present, and
> vice versa. Thus is would be helpful is one could get NOMREG
> to estimate the effect of the variable taking the value 1
> for convenience purposes.
>
> Clearly, a little simple arithmetic can be applied to the
> output to get this, but it would be helpful if there were an
> option for this.
>
>
> On Thu, 4 Aug 2005 00:29:49 -0300, Hector Maletta
> <hmaletta@fibertel.com.ar> wrote:
>
> >Marija,
> >thanks for the clarification, but on the whole I found it slightly
> confusing
> >anyway. However, my point was chiefly to make a distinction between
> >this
> use
> >of "reference category" and the use of the same name in other
> >procedures such as Logistic or Cox Regression, when the
> conversion of
> >categorical variables into dummies is concerned. In NOMREG
> the category
> >in question is not omitted: it is the one for which the estimated
> >effect is equivalent to the intercept.
> >
> >On the other hand, leaving to the user the option of designating
> >another category for that role may be extremely nice, and it is
> >indeed
> regrettable,
> >though harmless, that SPSS programmers had chosen not to leave that
> option.
> >
> >Besides, even if that option is not implemented, it should be more
> >clearly stated that alternative models are possible where the
> >"intercept" category is one of the other categories in the factor,
> >other than the last one, and provide the formulas to
> convert one model results into another.
> >
> >Hector
> >
> > _____
> >
> >From: Marijaspss@aol.com [mailto:Marijaspss@aol.com]
> >Sent: Wednesday, August 03, 2005 11:41 PM
> >To: hmaletta@fibertel.com.ar; SPSSX-L@LISTSERV.UGA.EDU
> >Subject: Re: Reference Category for Categorical Vars in NOMREG
> >
> >
> >
> >
> >On page 46 of the Advanced Statistical Procedures Companion it
> >states,
> >
> >"The SPSS Multinomial Logistic Regression procedure treats the last
> category
> >of a categorical or factor variable as the reference
> category and sets
> >its coefficient to 0. The variable male is coded 0 for no and 1 for
> >yes, so males are the reference category."
> >
> >Marija Norusis
> >www.norusis.com
> >
> >
> >
> >In a message dated 8/3/2005 5:30:22 P.M. Central Standard Time,
> >hmaletta@fibertel.com.ar writes:
> >
> >By the way, in Marija Norusis' book (SPSS 13.0 Advanced Statistical
> >Procedures Companion) I find the explanation of this matter rather
> confusing
> >(p.47). First, it is not clear in her example which gender is 0 and
> >which
> is
> >1: in the text it seems that the variable (confusingly
> called MALE) is
> zero
> >for males, since she says "For males the estimated value is
> just the
> >intercept, since the coefficient for the predictor variable is 0"
> >whereas
> in
> >the accompanying table the category with B=0 is MALE=1.
> Secondly, the
> table
> >has a footnote regarding that zero, indicating that "The
> parameter is
> >set
> to
> >zero because it is redundant", when in fact it is not
> exactly so: the
> >true explanation IMHO is that the parameter is zero by
> construction:
> >the model being tested establishes that the effect of being male is
> >just the intercept, and the effect of being female is the intercept
> >plus the effect of gender.
> >
> >
> >
> >
> >
> >__________ Informacisn de NOD32 1.1185 (20050801) __________
> >
> >Este mensaje ha sido analizado con NOD32 Antivirus System
> >http://www.nod32.com
>
> __________ Informacisn de NOD32 1.1185 (20050801) __________
>
> Este mensaje ha sido analizado con NOD32 Antivirus System
> http://www.nod32.com
>
>
| 
