Re: Analysis with Categorical Independent and Dependent Variables

Posted by Art Kendall on Nov 14, 2012; 8:44pm
URL: http://spssx-discussion.165.s1.nabble.com/Analysis-with-Categorical-Independent-and-Dependent-Variables-tp5716193p5716202.html

I do not have a great deal of experience  with these procedures.
If I understand correctly, plum and logistic are special cases for GenLin and or NomReg.

CATREG has built-in ways to test the degree to which level of measurement makes much of a difference in fit.  Depending how small N's are, treating education and age as not too discrepant from interval level of measurement means that fewer cases are needed. (I.e., 1 predictor per variable vs # of categories minus 1 per variable).

If this were a problem I were working on I would use it as an opportunity to compare the detailed results and the substantive conclusions from CATREG, Genlin, and NomReg. If I had some time I would see how correspondence analysis

it would be interesting to hear from the Leiden people who designed the CATEGORIES module, or from people who have experience with a pair of more of these procedures,  about comparing CATREG correspondence analysis, GenLin and NomREG

Without a more detailed understanding of the context, the present problem looks like a discriminant function type of question except that some of the predictors are mixed levels of measurement nominal (religion), ordinal or interval (education,age) and interval (gender).
In other words, the question might be "what distinguishes people who voted Obama/Romney/other.

 It makes intuitive sense that just as interaction terms can be used in regression, there is most likely a way to use interaction terms in CATREG.

From practical point of view 1) I would be surprised if the there were enough "other" votes to make  very fine distinctions. 2) It would be an unusual context where there were cases in each value of religion to do much.  Of course a lot depends on how large the total pop is  in the legislative district 

I would be very hesitant about generalizing from a district with such an unusual representation of religious subgroups.
On 11/14/2012 2:55 PM, Maguin, Eugene wrote:

Art,

 

If you’ve used the CatReg proc, what is your experience with the concordance of coefficients and standard errors between using CatReg and GenLin, NomReg, Plum, or Logistic to analyze the same datasets? It seems as if there are different assumptions in the underlying model between CatReg and the other procedures and, certainly in the estimation method. Given a categorical DV, when would you choose CatReg over the other categorical methods?

 

Thanks, Gene Maguin

 

From: SPSSX(r) Discussion [[hidden email]] On Behalf Of Art Kendall
Sent: Wednesday, November 14, 2012 2:44 PM
To: [hidden email]
Subject: Re: Analysis with Categorical Independent and Dependent Variables

 

See the CATREG procedure,  you can mix nominal, ordinal,  and interval variables.  Are you stuck with age as a grouped variable? (for example the cell with age 18-25 would have a small n for advanced degrees.


It would be a good idea to first check cell sizes etc via crosstabs. Treating age and education as ordinal may offset potential problems of empty or small cells.  you might or might not have to ignore higher way interaction.

You might also consider clustering some of the categorical variables 

Art Kendall

Social Research Consultants

On 11/14/2012 10:33 AM, Dates, Brian wrote:

I’ve had a question from a former student who has been asked, as a legislative aide, to analyze data from a legislative district. All variables are categorical.  For example, he would like to know what the effect of Religion (Roman Catholic, Maronite, Chaldean, Melkite, Other), Gender (Male, Female), Education (Less than High School, High School, Some College, Completed College, Advanced Degree), and Age Group (18-25, 26-54, 55 or older) had on voting behavior (Obama, Romney, Other). Factorial logistic regression would be in order if the dependent variable was dichotomous, or at least that’s my understanding.  But it’s not, and there will be other analyses in which the dependent will have more than three categories. Any ideas would be welcome.  Thanks.

 

 

Brian G. Dates

Director of Evaluation and Research

Southwest Counseling Solutions

1700 Waterman

Detroit, MI     48209

313-841-7442

[hidden email]

 

Leading the Way in Building a Healthy Community

 

 

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD