He difference between a “fixed”
factor and a “non fixed”
or random ones arises only in random experiments, which is not your case.
Apparently you haveonly two predictors: the prior diagnosis and the
intervention measure. Thus use the two in a logistic regression and that’s
it. The resulting odds or probabilities of the DV will tell you by how much the
odds or probability of meeting the criteria are changed by the intervention
measures, for people with a given prior diagnosis status. The results should be
treated with caution due to the non randomized nature of the groups and the non
pairing of subjects. To correct for this you may want to introduce also some
other relevant control variables (such as age, sex, perhaps some medical
condition relevant to the diagnosis, say blood pressure or whatever, etc.),
just to discard a few possible confounding factors.
Hector
From: SPSSX(r)
Discussion [mailto:[hidden email]]
On Behalf Of Tuerk, Elena
Sent: 15 June 2009 12:48
To: [hidden email]
Subject: basic logistic regression
help
I would appreciate clarity on a basic LR
question:
I want to determine whether an
intervention effect exists with a categorical dependent variable (coded 0/1 for
No/Yes): “Meets criteria for X diagnosis.” The study groups
(control/intervention) are also coded 0/1 and the design is a non-randomized
trial. The file is set up horizontally (one row per case), but can be reverted
to a vertical setup if necessary.
I want to control for pre-test status on
the DV.
I would like confirmation on the
variable entries in SPSS:
Dependent: diagnosis.1 (post)
Covariate 1: diagnosis.0 (pre) –
identified as a categorical covariate with the change contrast listed as
“indicator” and reference as “last?”
Covariate 2: group?? (also listed as a
categorical variable with the same options as covariate 1). I am confused
because in the analyses I typically conduct, the IVs are usually clearly
identified as such (e.g., fixed factor, independent).
I would greatly appreciate any
assistance on this. Also, if a different method is preferable given the design
(e.g., generalized estimating equations), I would appreciate any advice on that
as well.
Thanks,
Elena
Locked
