Controlling for a combination of continuous, dichotomous and ordinal variables
Locked
Controlling for a combination of continuous, dichotomous and ordinal variables
 
|
Dear all,
I'd like to compare the means of one dependent variable between 2 independent samples in order to isolate the effect of a manipulation. Unfortunately, it wasn't possible (due to political reasons) to randomize groups and they differ on 9 background variables which are expected to have some causal impact on the dependent variable. 4 of these are dichotomous, 3 are continuous and 2 are (rather strictly) ordinal. My sample sizes (unweighted) are 729 and 468 respondents. I've a rough idea about how to proceed (some sort of GLM, entering all factors together) but I'm rather uncomfortable about the details. Which procedure is to be preferred? Can I just add these variables together with 'manipulation' into some sort of GLM? And how can I obtain the adjusted means for the dependent variable? Isn't it a problem that the control variables have different measurement levels? Should I dummy code the ordinal ones? Am I not 'pushing the model over its natural limits' with so many control variables and a limited number of observations? Any help or literature references (hopefully not too technical!) are highly appreciated! Actually, if someone can point out just which procedure to use (and perhaps some major pitfalls), that would already be a good first step for me. Have a good weekend! Ruben Express yourself instantly with MSN Messenger! MSN Messenger |
Re: Controlling for a combination of continuous, dichotomous and ordinal variables
|
Dear
Ruben,
Lots of questions,
isn't it?
1. GLM indeed is a simplest possibility here.
a) The
easiest way to handle your data is to enter (in
GLM):
- your controlled factor
as well as dichotomous & ordinal variables as factors
- and continuous variables as
covariates.
b) You shouldn't stop here. Different top-down or
bottom-up strategies for fitting the best model can be implemented, it depends
on your analysis' purposes (goal of the research, meaning of
data...). c) I would expect some problems with
collinearity. Variables selection or data reduction (PCA or - because of mixed
measurement levels - CATPCA) could be helpful here. In case of experimental
factor x background covariate correlation, using SEM (here: it would be a path
model) has been suggested as an alternative.
d) Ordinal character of variables is lost in GLM
(but it's not necessary a problem). You might consider using Helmert or
repeated contrasts to test if the ordinal variable effect is
"lineary-ordinal".
e) To make model interpretation easier, it's
good to center continuous variables.
2. What do you mean by adjusted means? Model
predictions? You can save predictions from model to data if you would like to
process them, define profile plots if you want to visualise
results, use EMMEANS (in GLM window: "Display means for" from "Options"
submenu) if you want your results in
tables....
3. GLM gives broader capabilities of modelling data
structure then dummy coding.
4. "Pushing model over limits...": It's impossible to
say without looking at data... It depends on effects you are interested in
(e.g. you need more data for stable estimates of interaction effects), on data
structure, etc.
Two classical, easily digestible books with
reference to SPSS:
(with 90% certainty: GLM is in 1st
part)
(GLM with catchy 'sex, drugs & rock'n'roll'
examples)
Good luck!
Best regards,
Mariusz Trejtowicz
|
Re: Controlling for a combination of continuous, dichotomous and ordinal variables
|
|
In reply to this post by Ruben Geert van den Berg
I would possibly think about propensity scores. Use your
covariates to predict the group differences (ignoring the outcome variable),
and then output the predicted values and use them as a single covariate to test
the effect of the intervention on the outcome. Dr. Paul R. Swank, Professor and Director of Research Children's Learning Institute University of Texas Health Science Center-Houston From: SPSSX(r) Discussion
[mailto:[hidden email]] On Behalf Of Ruben van den Berg Dear all, Express
yourself instantly with MSN Messenger! MSN
Messenger |
| Free forum by Nabble | Edit this page |