method to analyse ordinal and nominal variables

[bgreen]

I am hoping for some advice on a question that is more a statistical than
an SPSS question.

I have been asked to look at data that consists of people from 3 groups. I
wanted to examine whether there are group differences (e.g age, marital &
employment status, previous treatment history, previous drug use). With the
exception of age I originally had thought of using the chi-square, to
determine whether than was any difference in the percentage of each group
with certain characteristics. However, in addition to an initial concern
about interpreting the chi-square when there are 3 groups there is the
issue that the groups are in effect also ordinal variables -  as group
membership is based on crime committed, which vary in severity.


My questions are:

1. Is it problematic treating group membership both as a dependent and an
independent variable? If so, what alternative analysis in SPSS should be used?

2. If not

(a) is the chi-square the preferred method of analysis to test for sample
differences for the categorical variables such as employment status & other
variables which are not to be used to model offence severity?

(b) is an ordinal logistic regression using the symptom variables (present
or absent) as predictors and group as the IV a reasonable option?


Any assistance is appreciated,

regards

Bob Green

[Keith McCormick]

Hi Bob,

If you are content to do a series of bi-variate analyses (I can't tell
from your post), there is no harm in doing chi-square of group versus
all but age.  (With age you can simply do an F-test.)  With chi-square
you aren't declaring a DV, so re: Q1 you are all set.  As for the fact
that group is ordinal, that is not a problem.  Chi-square is run on
ordinals all the time. In addition to chi-square you could request an
"association measure" found in the same menu at Chi-Square within
crosstabs.  If you have two ordinal variables, you could request a
Gamma, or one of the others.  You can also chose one for nominal
variables.  These will help you determine which of the variables has
the strongest association with group.

I am not sure why you are worried about group having three categories,
but I have a hunch.  Chi-square doesn't require that you have only
two, but you might be worried that it won't tell you that group 1 is
different from group 2, and 2 from 3 etc.  That is true, but I
wouldn't worry about that since you will want to establish overall
differences first.

I am not thinking PLUM as I read your post (ordinal regression).  When
you mention it, I think you mean DV, when you post IV.  If so, I guess
you could run PLUM or loglinear ... Although, you imply that you want
to keep it pretty basic.

If you want something that would be straightforward to explain to
others, I think I would go with CHAID, if you have access to it.
(Watch out, however, as this would be using a Data Mining technique -
I don't know who you are performing the analysis for).  CHAID
(classification trees) will allow you to make Group the DV, and put
all of the others, including age, as IV.  The advice might not work
for you since it is not true multivariate stats, but I mention it
because you were considering doing a series of crosstabs. All CHAID is
is a kind of "stepwise chi-square". It would allow you to look at
basic unhypothesized interactions, and it can handle nominal, ordinal,
and scale IVs.

There are a lot of related issues, but I will let you (or others) post
again, if you want to follow-up.

My main concern is whether the current analysis has to be definitive,
in which case using data mining might be the wrong choice.  Do you
have specific hypotheses that you are testing?  Or are you exploring?

best, Keith

www.keithmccormick.com

On 9/2/06, Bob Green <[hidden email]> wrote:

> I am hoping for some advice on a question that is more a statistical than
> an SPSS question.
>
> I have been asked to look at data that consists of people from 3 groups. I
> wanted to examine whether there are group differences (e.g age, marital &
> employment status, previous treatment history, previous drug use). With the
> exception of age I originally had thought of using the chi-square, to
> determine whether than was any difference in the percentage of each group
> with certain characteristics. However, in addition to an initial concern
> about interpreting the chi-square when there are 3 groups there is the
> issue that the groups are in effect also ordinal variables -  as group
> membership is based on crime committed, which vary in severity.
>
>
> My questions are:
>
> 1. Is it problematic treating group membership both as a dependent and an
> independent variable? If so, what alternative analysis in SPSS should be used?
>
> 2. If not
>
> (a) is the chi-square the preferred method of analysis to test for sample
> differences for the categorical variables such as employment status & other
> variables which are not to be used to model offence severity?
>
> (b) is an ordinal logistic regression using the symptom variables (present
> or absent) as predictors and group as the IV a reasonable option?
>
>
> Any assistance is appreciated,
>
> regards
>
> Bob Green
>