Hi Lucas,
This is how I read it. They've defined the model with only
the intercept in as the final model ~ there's a note at the bottom of the
table.
Then each variable is added. I read this as not one
variable after another being added, but as in each case only that variable being
added. You'll see why in a moment.
AIC and BIC are explained elsewhere, but basically the
lower the better.
Now we get to the key parts of the table. You need to look
at the Chi-square value as the difference between the -2LL value for the case in
point and the -2LL value for "the final model" (defined as the one with only the
intercept).
Because the Chi-Square value in each case equals the -2LL
of that case minus that of the final model, I read this as looking at a number
of models each of which contains only one variable (plus the intercept). I'm
more used to tables like this where the chi-square value comes from the
difference between the -2LL of the previous model and that where the extra
variable has been added in.
The degrees of freedom (df) depend on how the variables
have been defined. I'd assume 3 race categories, 5 marital categories, 2 cappun
categories (yes/no), and that age is continuous.
Using the chi-square value and the df value from tables,
or however you do it, you can see whether omitting that one variable is
significant or not (p<0.05).
I would urge caution here, however, because even though
this univariate approach might identify apparently significant variables, they
may not remain significant once other variables are entered into the model. The
text goes onto this. The approach is very much the same, except that the number
of combinations and permutations of which variables to include rises quickly as
the number of possible variables rises. Some people argue that using the
univariate approach first allows for better selection at the multivariate stage.
Others disagree.
I hope this helps,
Martin Holt
----- Original Message -----
Sent: Saturday, April 04, 2009 10:26
AM
Subject: model selection logistic
regression with interaction terms
Hi everybody,
at the moment Im doing a logistic
regression with additional interaction effects. Now I want to do some kind of
model selection to get a parsimonious model.
I found at http://faculty.chass.ncsu.edu/garson/PA765/logistic.htm
(search for Likelihood Ratio test, point 4.) that you can do a LR-test for
the individual model parameters. Can somebody help me how I get to on the
above mentioned site illustrated table of the LR-Test? (By the way, Im using
SPSS 16.0).
Or is there another possibility to select
easily the most parsimonious model with the interaction
terms?
Thanks for help in
advance.
Lucas
Locked
