8th January 2013 at 5:34 pm #1850
I am trying to build a regression model with single level predictors that are theoretically hierarchical (proximal individual traits to distal societal factors)and I am using GENLIN in SPSS (my outcome is dichotomous)
Now instead of putting all the predictors in the model at once, I am trying to enter them as blocks of variable (to reflect the theoretical hierarchy) but I was wondering about how to carry one step to the other…If I start with all the sociodemographic variables for example (sex, age, SES) and I note that only sex is significant do I remove the other two variables before entering the next block or leave them? and then when I enter the next block of variables, say attitude variables, Do I ignore any changes in the coefficients and significance of the previous block variables and just look at the new ones?
Basically how do I use the theoretical hierarchy in model interpretations.
Hope I get some answers!
Nora13th January 2013 at 6:58 pm #1859
It sounds like what you want to do is a Type III (simultaneous) SS GLM. If you do that and then remove n.s. variables, you risk Type II errors and decreasing your explained variance.
I would do a Type I (sequential) SS GLM. Why not just do this:
- Dummy code the DV as 0/1 (“pretending” that it is continuous, a la a discriminant analysis)
- Prioritize all of your data for ‘most likely a cause’ to ‘least likely a cause’ and run a sequential SS regression (or mixed GLM)
DV = A B C D E F
Run this model:
DV = A
then this one:
DV = A B
You can test for ‘significant contribution of B’ by using the equation in the attached file. This is a handout that I gave out for the graduate stats course this year (for internal use only, if you would like to reproduce any parts, please request written permission).
This will allow you to prioritize your model by blocks and theoretically prioritize your predictors within the blocks as well. Additionally, using the Type I SS will allow you to eat up the variance according to your theory.
If you insist on using Type III SS, I’d suggest running your first block, assessing the critical variables and then running a Type I SS model with the critical variables. Run the Block 2 analysis on the residuals of that Type I SS model. Repeat for all subsequent blocks.
Hope that all makes sense.
Raf15th January 2013 at 5:28 pm #1858
thank you Rafael! I must admit I am not familiar with this typology, what does SS stand for?. I will try to study this further.
What I have noticed is that GENLIN on SPSS does not give the option of block analysis or step wise option as in logistic regression. What I attempted to do is to manually enter each block of data and I did not remove the insignificant variables from each block but rather observed if any of the changed as I added the next block. I looked also at the liklihood ratio chi-square in each of the three models to compare the fit.
Anyway, as I said I will read further about your explanation… very kind of you.15th January 2013 at 5:40 pm #1857
SS stands for Sum of Squares.
Why not use GLM in SPSS instead of GENLIN?15th January 2013 at 6:07 pm #1856
you mean treat my outcome as a continuous?15th January 2013 at 8:07 pm #1855
Yes. What is your outcome?16th January 2013 at 3:24 pm #1854
meeting recommendations for fruit intake, 0=does not meet (<2 per day) ,1=meets recommendations ( 2 or more per day)16th January 2013 at 4:02 pm #1853
If this were my study (and I couldn’t break the intake into the number of servings), I’d treat the DV continuously. Though if you want to use GENLIN and model it categorically, I’m sure you could modify the suggestions I posted and adapt them to GENLIN (I’m not very familiar with SPSS).
Raf16th January 2013 at 6:28 pm #1852
Working on these suggestions as we speak. Thank again 🙂
Nora16th January 2013 at 6:29 pm #1851
Feel free to message me directly if you have any other questions.
- The forum ‘Default Forum’ is closed to new topics and replies.